# NANOTECHNOLOGIES IN NEUROSCIENCE AND NEUROENGINEERING

EDITED BY : Ioan Opris, Mikhail Lebedev, Ruxandra Vidu, Victor Manuel Pulgar, Marius Enachescu and Manuel Fernando Casanova PUBLISHED IN : Frontiers in Neuroscience

#### Frontiers eBook Copyright Statement

The copyright in the text of individual articles in this eBook is the property of their respective authors or their respective institutions or funders. The copyright in graphics and images within each article may be subject to copyright of other parties. In both cases this is subject to a license granted to Frontiers. The compilation of articles constituting this eBook is the property of Frontiers.

Each article within this eBook, and the eBook itself, are published under the most recent version of the Creative Commons CC-BY licence. The version current at the date of publication of this eBook is CC-BY 4.0. If the CC-BY licence is updated, the licence granted by Frontiers is automatically updated to the new version.

When exercising any right under the CC-BY licence, Frontiers must be attributed as the original publisher of the article or eBook, as applicable.

Authors have the responsibility of ensuring that any graphics or other materials which are the property of others may be included in the CC-BY licence, but this should be checked before relying on the CC-BY licence to reproduce those materials. Any copyright notices relating to those materials must be complied with.

Copyright and source acknowledgement notices may not be removed and must be displayed in any copy, derivative work or partial copy which includes the elements in question.

All copyright, and all rights therein, are protected by national and international copyright laws. The above represents a summary only. For further information please read Frontiers' Conditions for Website Use and Copyright Statement, and the applicable CC-BY licence.

ISSN 1664-8714 ISBN 978-2-88963-675-4 DOI 10.3389/978-2-88963-675-4

#### About Frontiers

Frontiers is more than just an open-access publisher of scholarly articles: it is a pioneering approach to the world of academia, radically improving the way scholarly research is managed. The grand vision of Frontiers is a world where all people have an equal opportunity to seek, share and generate knowledge. Frontiers provides immediate and permanent online open access to all its publications, but this alone is not enough to realize our grand goals.

#### Frontiers Journal Series

The Frontiers Journal Series is a multi-tier and interdisciplinary set of open-access, online journals, promising a paradigm shift from the current review, selection and dissemination processes in academic publishing. All Frontiers journals are driven by researchers for researchers; therefore, they constitute a service to the scholarly community. At the same time, the Frontiers Journal Series operates on a revolutionary invention, the tiered publishing system, initially addressing specific communities of scholars, and gradually climbing up to broader public understanding, thus serving the interests of the lay society, too.

#### Dedication to Quality

Each Frontiers article is a landmark of the highest quality, thanks to genuinely collaborative interactions between authors and review editors, who include some of the world's best academicians. Research must be certified by peers before entering a stream of knowledge that may eventually reach the public - and shape society; therefore, Frontiers only applies the most rigorous and unbiased reviews.

Frontiers revolutionizes research publishing by freely delivering the most outstanding research, evaluated with no bias from both the academic and social point of view. By applying the most advanced information technologies, Frontiers is catapulting scholarly publishing into a new generation.

#### What are Frontiers Research Topics?

Frontiers Research Topics are very popular trademarks of the Frontiers Journals Series: they are collections of at least ten articles, all centered on a particular subject. With their unique mix of varied contributions from Original Research to Review Articles, Frontiers Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area! Find out more on how to host your own Frontiers Research Topic or contribute to one as an author by contacting the Frontiers Editorial Office: researchtopics@frontiersin.org

# NANOTECHNOLOGIES IN NEUROSCIENCE AND NEUROENGINEERING

Topic Editors:

Ioan Opris, University of Miami, United States Mikhail Lebedev, Duke University, United States Ruxandra Vidu, University of California, Davis, United States Victor Manuel Pulgar, Wake Forest School of Medicine, United States Marius Enachescu, Politehnica University of Bucharest, Romania Manuel Fernando Casanova, University of South Carolina, United States

Citation: Opris, I., Lebedev, M., Vidu, R., Pulgar, V. M., Enachescu, M., Casanova, M. F., eds. (2020). Nanotechnologies in Neuroscience and Neuroengineering. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88963-675-4

# Table of Contents


Josephine Pinkernelle, Vittoria Raffa, Maria P. Calatayud, Gerado F. Goya, Cristina Riggio and Gerburg Keilhoff

*23 Initial Photophysical Characterization of the Proteorhodopsin Optical Proton Sensor (PROPS)*

Jay L. Nadeau


Trevor J. Gahl and Anja Kunze

*153 Recent Advances in the Therapeutic and Diagnostic Use of Liposomes and Carbon Nanomaterials in Ischemic Stroke*

Lorena F. Fernandes, Gisele E. Bruch, André R. Massensini and Frédéric Frézard

*168 Nano-Architectural Approaches for Improved Intracortical Interface Technologies*

Youjoung Kim, Seth M. Meade, Keying Chen, He Feng, Jacob Rayyan, Allison Hess-Dunning and Evon S. Ereifej

*188 Nanoparticle-Based Systems for Delivery of Protein Therapeutics to the Spinal Cord*

Juan C. Infante


Marta d'Amora and Silvia Giordani

*285 Transcytosis to Cross the Blood Brain Barrier, New Advancements and Challenges*

Victor M. Pulgar

*294 Multi-Modal Nano Particle Labeling of Neurons*

Lilac Amirav, Shai Berlin, Shunit Olszakier, Sandip K. Pahari and Itamar Kahn

### *303 Human Brain/Cloud Interface*

Nuno R. B. Martins, Amara Angelica, Krishnan Chakravarthy, Yuriy Svidinenko, Frank J. Boehm, Ioan Opris, Mikhail A. Lebedev, Melanie Swan, Steven A. Garan, Jeffrey V. Rosenfeld, Tad Hogg and Robert A. Freitas Jr.

# Editorial: Nanotechnologies in Neuroscience and Neuroengineering

Ioan Opris <sup>1</sup> \*, Mikhail A. Lebedev 2,3,4, Victor Manuel Pulgar 5,6, Ruxandra Vidu<sup>7</sup> , Marius Enachescu<sup>8</sup> and Manuel F. Casanova9,10

*<sup>1</sup> Department of Biomedical Engineering, University of Miami, Coral Gables, FL, United States, <sup>2</sup> Department of Neurobiology, Duke University, Durham, NC, United States, <sup>3</sup> Center for Bioelectric Interfaces of the Institute for Cognitive Neuroscience, National Research University Higher School of Economics, Moscow, Russia, <sup>4</sup> Department of Information and Internet Technologies of Digital Health Institute, I.M. Sechenov First Moscow State Medical University, Moscow, Russia, <sup>5</sup> Department of Pharmaceutical Sciences, Campbell University, Buies Creek, NC, United States, <sup>6</sup> Department of Obstetrics and Gynecology, Wake Forest School of Medicine, Winston-Salem, NC, United States, <sup>7</sup> Department of Chemical Engineering and Materials Science, University of California, Davis, Davis, CA, United States, <sup>8</sup> Center of Surface Science and Nanotechnology, Politehnica University of Bucharest, Bucharest, Romania, <sup>9</sup> Department of Biomedical Sciences, University of South Carolina School of Medicine at Greenville, Greenville, SC, United States, <sup>10</sup> Department of Pediatrics, Greenville Health System, Greenville, SC, United States*

Keywords: nanotechnology, neuroengineering, nanoparticles, magnetic tunneling junctions, biosensors, multielectrode arrays, memistor, brain machine interfaces

**Editorial on the Research Topic**

#### **Nanotechnologies in Neuroscience and Neuroengineering**

#### Edited and reviewed by:

*Michele Giugliano, International School for Advanced Studies (SISSA), Italy*

> \*Correspondence: *Ioan Opris ioanopris.phd@gmail.com*

#### Specialty section:

*This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience*

Received: *16 December 2019* Accepted: *13 January 2020* Published: *12 February 2020*

#### Citation:

*Opris I, Lebedev MA, Pulgar VM, Vidu R, Enachescu M and Casanova MF (2020) Editorial: Nanotechnologies in Neuroscience and Neuroengineering. Front. Neurosci. 14:33. doi: 10.3389/fnins.2020.00033* Neuroscience and Neuroengineering operate at the cellular level and their association with Nanotechnology is bringing unexpected strides. During the last decade, we have witnessed an unprecedented increase in the successful application of Nanotechnology to both basic Neuroscience and to Clinical Practice. Novel nanotechnologies are expected to bring important insights on brain mechanisms and medical care to patients. The topic theme, "Nanotechnologies in Neuroscience and Neuroengineeringy" details attempts from different fields to improve brain performance in both healthy people and to patients suffering from neurological disabilities. Twenty-seven articles contributed by 122 authors composed this Research Topic on various aspects of Nanotechnologies applied in Neuroscience and Neuroengineering.

### NANOTECHNOLOGIES IN NEUROSCIENCE

Among the nanotechnologies that emerged recently in neuroscience, we cover the nanoparticles (including magnetic nanoparticles) and their involvement in therapy, the blood-brain-barrier, the nano-electrical and chemical stimulation, as well as recent insights into neuro-engineering, involving the characterization of biophysical features of neural cells and the function of neural microcircuits. Finally, we briefly discuss aspects of neural interfacing that have already been confirmed as feasible for brain machine interfaces and sensors.

#### Nanoparticles

Nanoparticles are ultrafine units in the microscopic field of few to hundreds of nanometers, but less than a micron in size. Nanomagnetic and charged particles are endowed with valuable interactive abilities for neuronal cells. Pinkernelle et al., assessed the "bio-functionality" of growth factors using appropriate biological models. Thus, successful "functionalization" of magnetic nanoparticles with growth factors seems dependent on their "binding" chemistry. These magnetic nanoparticles support regeneration within the nervous system. Amirav et al.,

**5**

reviewed some recent magneto-fluorescent markers and highlighted key differences between them, in terms of durability and relevant approaches. They focused on the intracellular labeling potential and basic functional sensing MRI, with assays that enable the imaging of cells at microscopic and mesoscopic scales. Also, the limitations of available imaging markers have been reviewed/discussed while keeping in mind the possibility of in vivo neural imaging and large-scale brain mapping. Infante, proposed to examine the affinity and nanoparticle-based strategies for the delivery of neurotrophic factors to the spinal cord in an adequate, tunable, and safe therapeutic manner.

## Magnetic Nanoparticles and Magnetic Tunneling Junctions

Cellular processes such as the deformation of membrane, the transport of organelles or the migration of cells are sensitive to mechanical forces, operating through the "chaperoning" forceinducing nanoparticles in electrical/magnetic field gradients, with spatial precision in the range of sub-micrometers. Gahl and Kunze used force-mediating magnetic nanoparticles to generate neuronal cell function. Moretti et al., produced the first biomagnetic chip using a novel technology based on magnetic tunnel junction (MTJ) for cell culture, and demonstrated how these sensors are biocompatible. Such advancements of nanomagnetic field in cellular organization/communication/signaling and intracellular trafficking can be used in the next generation of neurotherapeutic devices.

#### Therapeutic Approaches

A series of therapeutically novel approaches emerged and were discussed in our Research Topic due to their potential applications. We mention here the ones involving: (i) carbon nanomaterials, such as nanotubes, graphene, nano-onions, or fullerenes for therapy, (ii) biosensing and imaging approaches that have antioxidant action, (iii) intrinsic photoluminescence, (iv) their ability to cross the BBB, carry oligonucleotides and cells, and (v) to induce cell differentiation. Fernandes et al., used liposomes and carbon nanomaterials in recent diagnosis and therapies in acute ischemic stroke. Liposomes represent a biomimetic system, with composition, structural organization, and properties very similar to biological membranes. Carbon nanomaterials, not being naturally parts of the human body, reveal new modes of interaction and integration with biological molecules and systems, resulting in completely unique pharmacological properties.

Novel genetic neuroprotective cell therapeutics are bringing promising approaches for the regenerative functions of the eye. Nafissi and Foldvari discussed these genetic nanotechnology neuroprotective therapies in glaucoma. The development of highly specific gene delivery methods which are safe and noninvasive are of crucial importance in ophthalmology. Nadeau reported the initial photophysical characterization of a new genetically encoded voltage sensor (based upon the fluorescence of rhodopsins), namely the "proteorhodopsin optical proton sensor" (PROPS). This is the first sensor capable of indicating the changes in membrane voltage by means of changes in fluorescence. Nadeau reported in two strains of Escherichia coli, a nanosecond time-resolved emission of this protein, before and after membrane depolarization.

d'Amora and Giordani have shown that zebrafish is a good animal model for "high-throughput" screening of chemicals, because of their small size, low-price, and transparency. Zebrafish has emerged as a powerful tool for screening developmental neurotoxicity. Convertino et al. elaborate on the graphene's potential for nerve tissue regeneration hinting to novel approaches of active nerve conduits for peripheral neuron survival and outgrowth. Moldovan et al. carried out experiments in adult transgenic mice with fluorescent tagged liposomes that provided insight into the local anesthetic effect of nanomedicines in post-operative pain. The effect of local anesthetic nanomedicines has important implications for humans.

## Blood Brain Barrier

Restorative strategies of brain function after stroke are centered on the repairing of cerebral endothelial and parenchymal cells. Communication between the cells and signaling within the neurovascular unit, including the multicellular brain-vesselblood interface, with its highly selective blood-brain barrier (BBB), are crucial to the homeostasis of the central nervous system. Zagrean et al.'s work highlights the important role of exosomes in mediating the crosstalk of the cells within the neurovascular unit. It further reveals the restorative therapeutic potential of exosomes in ischemic stroke, a frequent neurologic condition still in need of an effective therapy. Pulgar then discussed transcytosis across the BBB. Pulgar draws our attention to the physiological operation of "receptor-mediated transcytosis" (RMT) to carry load across the brain endothelial cells toward brain parenchyma, exemplifying critical advances in RMT-mediated brain drug delivery.

### Nano-Electrical and Chemical Stimulation

A generation of minimally invasive or non-invasive neural stimulation techniques is being developed, supported by nanotechnology to reach high spatial resolution. In these approaches of neural stimulation, as pointed out by Wang and Guo, a nanomaterial transforms a faraway transmitted primary stimulus (like a magnetic or ultrasonic signal), into a localized secondary stimulus, such as, an electric field in order to stimulate neurons. Stimulating neural systems with "applied" electric field (EF) are a common tool for testing network responses. Tang-Schomer et al. used a "gold wire-embedded silk protein film-based interface" culture to examine the effects of "applied" EFs on neuronal networks in in vitro cultures. Cortical cultures displayed large-scale oscillations, synchronized by EF at specified frequencies. These effects of EF on random neuronal networks have significant implications for studies of brain function and neuromodulation. Furthermore, Goldental et al. mimicking the collective firing patterns of connected neurons, which proved the emergence of cooperative phenomena like synchronous oscillations, the coexistence of fast γ and slow δ oscillations, and other dynamical phenomena within large-scale neuronal networks.

Novel nanofluidic mechanisms like hydrophobic gating, suggested by Jones and Stelzle, may support the control of chemical release appropriate for mimicking neurotransmission. Nanofluidic chemical release facilitates fast, high resolution neurotransmitter-based neurostimulation, that could bring improvements over electrical neurostimulation.

### NEUROENGINEERING

Nanotechnology is a fast-developing field, that provides simple and efficient tools to study the brain in health and disease. Of particular importance are biosensors, multi-electrode arrays, memory resistive devices, and brain machine interfaces.

#### Biosensors

Hossain et al. reported the design and implementation of a "GABA microarray probe." The probe consists of two distinct micro-biosensors, one for glutamate (Glu) and the other for GABA detection, modified with Glu oxidase and GABASE enzymes, respectively. The neurotransmitters GABA and Glu may be detected in real time, simultaneously/continuously, both: in vitro and ex vivo. The detection of GABA by such probe is based upon the "in-situ generation of α-ketoglutarate" from the oxidation that occurs at the Glu micro-biosensor. The GABA probe has been successfully tested in a slice preparation from a rat brain. These results show that the developed GABA probe represents a novel and valuable neuroscientific tool that could be utilized in studies of brain disorders involving the combined role of GABA and Glu signaling.

Many challenges of sensor development, including the bioengineered probes and sensors, arise when the physiological and pathological biomarkers are tested in neural cells (Maysinger et al.). The nanoparticle-based sensors have the ability to detect properties (biochemical and physiological) of neurons and glia, and to generate signals proportional to the changes (physical, chemical, or electrical) in these cells (Maysinger et al.). Among the most used nanostructures are the carbon-based structures (such as C-dots, graphene, and nano-diamonds), the quantum dots (QDs), and the gold nanoparticles. They are capable to detect/measure activity of proteases (metalloproteinases, caspases), ions, and other biomolecules under physiological or pathological conditions in neuronal cells. Such genetically manipulated probes and sensors are useful to reveal the changes in protease activities or calcium ion concentrations.

Moretti et al. demonstrated the biocompatibility of a magnetic sensor array for the detection of neuronal signals in the in vitro culture.

#### Multielectrode Arrays (MEA)

MEA has been developed and used extensively in basic and applied research in neuronal- and cardiomyocyte-networks, both in vivo and in vitro (Spira et al.). The MEA platforms consisting of thousands of sensors (with high-density, small diameter, and low impedance), use vertical nanowires that pass through the cultured cell's membrane and record the action potentials in a similar manner to that of a sharp intracellular microelectrode. Spira's team developed a bioinspired approach in-which cell's energetic resources are utilized with extracellular gold microelectrodes to record attenuated synapticand action-potentials with characteristic features resembling those of intracellular recordings. Moreover, the approach allowed to record intracellular potentials by an array of extracellular electrodes.

Intracortical microelectrodes (IME) have been extensively used to study various functions of the nervous system. Recent strategies to enhance interfacing with the brain's systems have been suggested by methods that mimic the biological tissue. Kim et al. review focusses on nano-architecture, a concept that considers the surface of the implant. Different nanoarchitectural approaches have been discussed to enhance the "biocompatibility" of IMEs, increase the recording quality, and augment the longevity of the implant.

Microelectrode material together with cell culture medium play important roles in the health of a cell as derived from in vitro electrophysiological studies. Ryynänen et al. reported an "ion beam assisted e-beam deposition" (IBAD) based process as being an alternative to the titanium nitride (TiN) method of deposition for "sputtering" in the fabrication of "TiN microelectrode arrays" (MEAs). The developed IBAD TiN process enables the MEA manufacturers with more choices as to which method to use in order to deposit TiN electrodes. The medium evaluation results remind that in addition to electrode material the insulator layer and cell culturing medium keep a crucial role in successful long-term MEA measurements.

## Resistive Memory Devices

Resistive memory devices are a pioneering technology inspired by the brain mechanisms. Resistive random-access memory (RRAM) arrays use little energy and hold a potential for enormous densities. An interesting type of RRAM was demonstrated recently to have alternating (dynamic switching) current rectification properties, like those of CMOS transistors (Berco). Such artificial synaptic devices can be switched between two modes (excitatory and inhibitory) to double the array density and to significantly reduce the peripheral circuit complexity. Gavrilov et al. discusses next the "associative spatial-temporal memories" based on neuromorphic networks with restricted connectivity, termed-"CrossNets." Such networks have the capability to be implemented naturally in nanoelectronic hardware with hybrid memristive circuits (a memistor is a nanoelectric element of circuitry used in parallel computing memory technology). This may allow extremely high energy efficiency, comparable to that of the biological cortical circuits, functioning at a much higher operation speed. Numerical simulations performed by Gavrilov et al. and confirmed with analytical calculations, show that the characteristics depend significantly on the method of information recording into the memory. Most importantly, CrossNet memories provide a capacity higher than that of "Ternary Content-Addressable Memories" with the same number of nonvolatile memory cells (e.g., memristors), and the input noise immunity of the CrossNet memories is lower.

#### Neurons and Networks

Collective firing patterns of thousands of inter-connected neurons have been simulated with sophisticated computational approaches. Their monitoring requires simultaneous measurements of connectivity, synaptic strengths, and delays (Goldental et al.). Such a computational tool allows the study of recurrent neural networks that are capable of "dictating" network's connectivity and synaptic strengths. The method proposed by Goldental et al. is based on the response of neurons and depends exclusively on their recent history of stimulation. It uses a sequential chart for stimulation and recording of single neurons, in order to "mimic" a recurrent neural network with simultaneous measurements of neurons' activity. Utilization of this technique provides evidence for the emergence of spontaneous synchronous oscillations and the network's synchrony (Tang-Schomer et al.). In particular, the cooperative phenomena that include coexistence of fast γ and slow δ oscillations opens the possibility for the experimental study of large-scale networks (Goldental et al.).

#### Brain Machine Interfaces

A brain machine interface (BMI) is a direct communication line between the brain and an external device. Silva reviewed the recent technological capabilities for machine learning and artificial intelligence (AI) to implement "smart" nanobrain machine interfaces (nBMI). His view consists of novel technologies that will "communicate" with the brain using approaches that allow contextual learning and adaptation to dynamic functional demands. It applies to both technologies: (i) invasive (e.g., neural prosthesis), and (ii) non-invasive (e.g., electroencephalography, EEG). Advances in computation, hardware, and software (such as algorithms that learn and adapt in a contextually dependent way) will have the ability to leverage the capabilities that nanotechnology provides to the design and functionality of nBMI.

The opportunity to optically connect/interface with the mammalian/human brain in vivo, has favored an unparalleled investigation of functional connectivity of brain's neuronal circuitry. Pisanello et al. reviewed the role of nanotechnology for optical-neuronal interfaces, focusing on the new devices and methods for optogenetic control of neuronal firing, and on the "detection" and "triggering of action potentials" using "optically active colloidal nanoparticles."

Future nanotechnology will allow us to interface the cloud with a human brain. Martins et al. labeled this as a "human brain/cloud interface" ("B/CI"), based on the nano technologies referred to here as "neural nanorobotics." Neural nanorobotics may endow a "B/CI" with "controlled" connectivity between neuronal firing and external storage and processing of data, via the direct "monitoring" of the brain's ∼86 billion neurons and ∼200 trillion synapses. A neural nano-robotically allowed human "B/CI" might serve as a "personalized conduit," enabling subjects to get a direct, instantaneous access to each aspect of human knowledge.

## AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

## FUNDING

ML was supported by the Center for Bioelectric Interfaces NRU HSE, RF Government grant, ag. No. 14.641.31.0003.

## ACKNOWLEDGMENTS

The authors thank the Frontiers team for their professional help with this Research Topic.

**Conflict of Interest:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2020 Opris, Lebedev, Pulgar, Vidu, Enachescu and Casanova. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Growth factor choice is critical for successful functionalization of nanoparticles

Josephine Pinkernelle1, 2 \*, Vittoria Raffa3, 4, Maria P. Calatayud<sup>5</sup> , Gerado F. Goya5, 6 , Cristina Riggio<sup>4</sup> and Gerburg Keilhoff <sup>1</sup>

*<sup>1</sup> Department of Nephrology and Hypertension, Diabetes and Endocrinology, Otto-von-Guericke University of Magdeburg, Magdeburg, Germany, <sup>2</sup> Institute for Biochemistry and Cell Biology, Otto-von-Guericke University of Magdeburg, Magdeburg, Germany, <sup>3</sup> Department of Biology, University of Pisa, Pisa, Italy, <sup>4</sup> Institute of Life Science, Scuola Superiore Sant' Anna, Pisa, Italy, <sup>5</sup> Aragon Institute of Nanosciences, University of Zaragoza, Zaragoza, Spain, <sup>6</sup> Department of Condensed Matter Physics, University of Zaragoza, Spain*

#### Edited by:

*Ioan Opris, Wake Forest University School of Medicine, USA*

#### Reviewed by:

*Jay Nadeau, McGill University, Canada Marius Enachescu, University Politehnica Bucharest, Romania*

#### \*Correspondence:

*Josephine Pinkernelle, Department of Nephrology and Hypertension, Diabetes and Endocrinology, Medical Faculty, Otto-von-Guericke University of Magdeburg, Leipziger Str. 44, Bldg. 1, D-39120 Magdeburg, Germany josephine.pinkernelle@med.ovgu.de*

#### Specialty section:

*This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience*

Received: *12 June 2015* Accepted: *12 August 2015* Published: *02 September 2015*

#### Citation:

*Pinkernelle J, Raffa V, Calatayud MP, Goya GF, Riggio C and Keilhoff G (2015) Growth factor choice is critical for successful functionalization of nanoparticles. Front. Neurosci. 9:305. doi: 10.3389/fnins.2015.00305* Nanoparticles (NPs) show new characteristics compared to the corresponding bulk material. These nanoscale properties make them interesting for various applications in biomedicine and life sciences. One field of application is the use of magnetic NPs to support regeneration in the nervous system. Drug delivery requires a functionalization of NPs with bio-functional molecules. In our study, we functionalized self-made PEI-coated iron oxide NPs with nerve growth factor (NGF) and glial cell-line derived neurotrophic factor (GDNF). Next, we tested the bio-functionality of NGF in a rat pheochromocytoma cell line (PC12) and the bio-functionality of GDNF in an organotypic spinal cord culture. Covalent binding of NGF to PEI-NPs impaired bio-functionality of NGF, but non-covalent approach differentiated PC12 cells reliably. Non-covalent binding of GDNF showed a satisfying bio-functionality of GDNF:PEI-NPs, but turned out to be unstable in conjugation to the PEI-NPs. Taken together, our study showed the importance of assessing bio-functionality and binding stability of functionalized growth factors using proper biological models. It also shows that successful functionalization of magnetic NPs with growth factors is dependent on the used binding chemistry and that it is hardly predictable. For use as therapeutics, functionalization strategies have to be reproducible and future studies are needed.

Keywords: functionalization, nanoparticles, glial cell-line derived neurotrophic factor (GDNF), nerve growth factor (NGF), organotypic spinal cord culture, PC12 cells

#### Introduction

Nanoparticles (NPs) are materials that display different characteristics at the nanoscale than the corresponding bulk material. Usually, the term NP is used for substances with sizes about 1–100 nm, but there is an ongoing discussion about using the term NPs in a more

**Abbreviations:** BSA, bovine serum albumin; BDNF, brain-derived growth factor; CNS, central nervous system; FBS, fetal bovine serum; GFR-α, GDNF family receptor-α; GDNF, glial cell-line derived neurotrophic factor; MRI, magnetic resonance imaging, MAPK, mitogen-activated protein kinase; NPs, nanoparticles; NGF, nerve growth factor; PNS, peripheral nervous system; PBS, phosphate buffered saline; PI3K, phosphatidylinositol 3-kinase; PLCγ, phospholipase Cγ; PEI, polyethyleneimine; PC12, rat pheochromocytoma cell line; TrkA, tropomyosin receptor kinase A.

size-independent way, referring more to the occurrence of new properties or the high ratio of surface area to volume (Kreyling et al., 2010). One feature that changes with nanoscale is magnetism. Iron is a ferromagnetic bulk material, however reducing its size to the nanoscale results in superparamagnetic behavior. Superparamagnetic NPs are non-magnetic until they are exposed to a strong magnetic field. If the magnetic field is removed, they revert to a non-magnetic state again (Mortimer and Müller, 2003; Estelrich et al., 2015). In the case of magnetite Fe3O<sup>4</sup> and maghemite γ-Fe2O3, which are commonly used iron oxide NPs, the superparamagnetic diameter is 25 nm (Fe3O4) and 30 nm (γ-Fe2O3). Both iron oxides are highly used in biomedicine due to their stability, relatively low toxicity, and high magnetization.

Superparamagnetic iron oxide NPs can be used in diagnostic magnetic resonance imaging (MRI), e.g., Resovist <sup>R</sup> underwent successfully clinical trials and is used to enhance the contrast in MRI scans of the liver (Reimer and Balzer, 2003; Weinstein et al., 2010). In addition to tissue imaging, magnetic NPs are also used to track cells with the MRI in vivo (Dunning, 2004; Antonelli et al., 2013).

In cancer treatment, superparamagnetic NPs are used to induce hyperthermia in prostate cancer and glioblastoma (Jordan et al., 2001; Thiesen and Jordan, 2008). NPs are injected directly into tumor tissue and stimulated with an alternating magnetic field. This causes the particles to heat up, thereby destroying the tumor tissue.

To localize functionalized magnetic NPs within a tissue of interest, a static magnetic field can be applied. Chertok et al. (2008) followed intravenously injected NPs in rats with glioblastoma via MRI and showed an accumulation of NPs in the tumors. Cells can also be targeted magnetically. Hamasaki et al. (2007) labeled neural progenitor cells with magnetic NPs and localized them in an organotypic co-culture with the help of an external magnetic field. Corticospinal axon growth was improved by magnetically targeted neural progenitor cells in comparison to non-labeled cells. Labeling cells with magnetic NPs can also help to purify primary cell cultures. Gordon et al. (2011) used magnetic NPs to purify microglia from mixed primary glial cultures.

The magnetofection technique uses magnetic force to deliver nucleic acids or viruses into cells. By applying a strong external magnetic field, cells can be transfected with superparamagnetic NPs, which are functionalized with gene vectors (Scherer et al., 2002). Magnetofection offers a simple method to transfect cells that are normally difficult to transfect, e.g., cells of the central nervous system (CNS). Efficient transfection was achieved in primary neural stem cells (Sapet et al., 2011), oligodendrocyte precursor cells (Jenkins et al., 2011), hippocampal neurons (Buerli et al., 2007), and astrocytes (Pickard and Chari, 2010). Also primary motor neurons could be transfected using this method. Fallini et al. (2010) transfected primary motor neurons with NPs functionalized with GFP-expressing plasmids. Motor neurons showed no signs of cytotoxicity and about 45% of cells could be transfected.

Magnetic NPs have a large surface-to-volume ratio that enables chemical conjugation and changes the surface properties of the NPs. Usually; they are synthesized by wet chemistry approaches, which produce Ferro fluid water dispersions (Vergés et al., 2008). The stability of Ferro fluids depends on the equilibrium between dipole-dipole interactions among particles and particle-solvent interactions. In order to decrease the strength of dipole-dipole interactions and stabilize the water dispersion of NPs as single particles or small clusters, a surface coating is required. The coating increases the hydrodynamic ratio of the particles, which decreases magnetic interactions among particles, stabilizing the dispersion. Additionally, surface properties affect biocompatibility, particle opsonization in biological media (Tenzer et al., 2013), cellular internalization mechanisms, and biological interactions (Gao et al., 2009; Veiseh et al., 2010). For coating, organic polymers (dextran, chitosan, polyethylene glycol), inorganic substances (gold, silica, carbon), and bioactive molecules (liposomes, proteins, ligands) can be used (Shubayev et al., 2009; Estelrich et al., 2015). The functionalization with bioactive groups and proteins allows a wide range of applications, particularly in life sciences (Pankhurst et al., 2003; Gupta and Gupta, 2005). Drug delivery is one application of interest (Arruebo et al., 2007; Estelrich et al., 2015). Huang et al. (2015a) described layer-by-layer caseincoated magnetic NPs, which could be loaded with doxorubicin and indocyanine green. NPs were stable under gastric conditions and drugs were released by degradation of the casein in the intestine. Nazli et al. (2014) showed that doxorubicin-delivery by MMP-sensitive PEG hydrogel-coated magnetic NPs are taken up efficiently into HeLa cells and the drug released within 2 h.

In neurosciences, magnetic NPs have received attention in the fields of CNS and peripheral nervous system (PNS) regeneration. Injury to the nervous system produces high costs for the health systems, partially due to massive lifelong impairments (Noble et al., 1998; Wyndaele and Wyndaele, 2006) and therefore is still a highly interesting field of investigation.

In CNS injury, one distinguishes between primary and secondary injury. The primary injury results from mechanical forces, damaging cells at the injury site. Due to secondary processes the size of the injury is increased and damage is prolonged. Tissue is additionally damaged by ischemia, edema, excitotoxicity, shifting of ion concentrations, production of reactive oxygen species, inflammation, necrosis, and demyelination of axons (Salewski et al., 2013; Siddique and Thakor, 2014).

Regeneration in the CNS is a huge challenge. One reason for this is the complex glial composition of the CNS: astroglia, microglia, and oligodendroglia. The inhibition of oligodendroglia and its apoptosis after loss of axon contact (Beattie et al., 2000; Vargas and Barres, 2007), the activation of microglia, which secrete pro-inflammatory cytokines (Danton and Dietrich, 2003), and the activation of astroglia, which results in the formation of a glia scarring in the injured area, are producing a regeneration-inhibiting environment making regeneration impossible (Fawcett and Asher, 1999; Fitch and Silver, 2008). In addition to surgical therapies, therapy approaches often try to modulate inflammatory actions, e.g., using minocycline or methylprednisolone (McDonald and Sadowsky, 2002; Beattie, 2004; Stahel et al., 2012). Studies trying to influence the glial scar and the regeneration-inhibiting features of residual myelin are done, too (Kwon et al., 2011). Nevertheless, neurons of the CNS have the capacity to regenerate, which was demonstrated by mixing CNS and PNS tissue/cells. In a mixed milieu, CNS neurons can regenerate into a PNS environment (David and Aguayo, 1981; Pearse et al., 2004).

In contrast, the potential for regeneration in the peripheral nervous system is, in principle, good (Navarro et al., 2007; Vargas and Barres, 2007; Svennigsen and Dahlin, 2013). After axon injury in the peripheral nerve several processes of degeneration and regeneration start. The injured neuron undergoes changes in metabolism and protein expression after retrograde signaling and calcium influx at the lesioned axon. Chromatolysis occurs and regeneration-related proteins like growth-associated protein 43 are upregulated. Growth cones are formed and elongate following guidance molecules provided by activated Schwann cells (Dent et al., 2003; Bradke et al., 2012; Patodia and Raivich, 2012). Schwann cells are activated by losing axonal contact and interruption of blood and oxygen supply. Calcium signaling activates intracellular cascades including mitogenactivated protein kinases (MAPK), like the extracellular signalregulated protein kinases (ERK1/2), and c-jun N-terminal protein kinases (JNK1/2/3) (Agthong et al., 2006; Chattopadhyay and Shubayev, 2009; Yamazaki et al., 2009; Arthur-Farraj et al., 2012). These signaling cascades result in a termination of myelin production, dedifferentiation, and then proliferation of Schwann cells at the distal injury site. Schwann cells start phagocytosis of myelin and cell debris before macrophages enter the injury site time-delayed. Schwann cells then start to produce regenerationpromoting factors, like nerve growth factor (NGF), brain-derived growth factor (BDNF), and glial cell-line derived neurotrophic factor (GDNF) (Funakoshi et al., 1993; Xu et al., 2013; Huang et al., 2015b), and align to regular structures, e.g., the Bands of Büngner, which guide regenerating axons to reinnervate their targets. Altogether, Schwann cells help to establish a regeneration-permissive environment for the regenerating axon.

Nevertheless, the regenerative outcome in the PNS is poor. Only 40% of patients regain a normal functionality, 16% show a good functional outcome, whereas 32% of the patients show little improvement, and 12% even show a complete loss of function (Noble et al., 1998). The clinical strategies used to treat patients with peripheral nerve injury have not improved in the last years. Surgery is still the treatment of choice (Scholz et al., 2009). Direct end-to-end repair avoiding tensions shows the best outcome. If tensionless repair is not possible, grafting or conduits are chosen to bridge the gap between the nerve endings. Often sensory nerves from the patient (autograft) or from donors (allograft) are used as grafts. Blood vessels are usually used as biological conduits. Collagen, polyglycolic acid, and caprolactone conduits are also available (Dahlin, 2008; Griffin et al., 2013; Grinsell and Keating, 2014).

One problem, even after successful surgery, is the non-specific reinnervation of target organs due to randomly axonal sprouting and misdirection of regenerating axons. Additionally, projections in the CNS are lost after an injury and new synaptic contacts need a rebuilding of CNS projections and a long-term learning process (Navarro et al., 2007; de Ruiter et al., 2014).

Magnetic NPs now offer a new possibility to promote PNS regeneration. One idea is to facilitate the uptake of magnetic NPs in regenerating neurons and their axons and to pull the axon along an external magnetic field. Therefore, the guidance of sprouting axons is promoted and the specificity of reinnervation is enhanced (Halpern, 2000; Riggio et al., 2012; Calatayud et al., 2013; Goya et al., 2014; Riggio et al., 2014). Previous studies showed initial success. Human neuroblastoma (SH-SY5Y) cells were effectively influenced by an external magnetic field after magnetic NPs uptake. Also primary Schwann cells and olfactory ensheating cells showed satisfying uptake of NPs and could be moved along a magnetic field (Riggio et al., 2012, 2013). Neurite orientation was shown to be directed toward a magnetic force after uptake of magnetic NPs into differentiated PC12 cells (Riggio et al., 2014).

However, peripheral nerve injury is more complicated. To facilitate regeneration of injured axons, NPs have to be taken up into neurons. Primary neuronal cells in mixed cell cultures showed relatively low uptake of iron oxide NPs (Pinkernelle et al., 2012). Petters and Dringen (2015) showed that using serumcontaining media significantly lowers the uptake of magnetic iron oxide NPs in cerebellar granule neurons compared to serumfree media. NPs diluted in serum-containing media develop a protein corona, which seems to impede the efficient uptake of NPs in neurons. There are various strategies to enhance neuronal uptake. Adams et al. (2015) increased magnetite content of NPs using sedimenting forces. Buerli et al. (2007) and Fallini et al. (2010) used magnetofection to promote uptake into neurons and transfected them with DNA. Another possibility is the coupling of growth factors, which utilize receptor-mediated uptake. Ziv-Polat et al. (2014) conjugated iron oxide NPs with different neurotrophic factors. They showed that the stability of neurotrophic factors in media and their functional activity was improved by binding them to the NPs. Altogether coupling of growth factors seems to be the most promising strategy in terms of a future clinical approach.

In our study, we focus on 2 growth factors: NGF and GDNF. Both play a role in PNS regeneration. NGF belongs to the neurotrophin family binding to the tropomyosin receptor kinase A (TrkA) or to p75 receptor (Boyd and Gordon, 2003). Most studies concerning NGF were performed in PC12 cells, which express both TrkA and p75 receptors. They can be differentiated into neuronal-like cells displaying neurite development (Greene and Tischler, 1976; Dichter et al., 1977). Signaling of Trk receptors is mediated via the activation of phosphatidylinositol 3-kinase (PI3K), MAPK, and phospholipase Cγ (PLCγ). These pathways regulate differentiation, survival, and neuritogenesis (Boyd and Gordon, 2003).

GDNF is a neurotrophic factor belonging to the GDNF family. These growth factors are secretory proteins and bind to a receptor complex consisting of a high affinity ligandbinding subunit, the GDNF family receptor-α (GFR-α), and a signal transduction subunit RET receptor tyrosine kinase. GDNF binds to GFR-α1 thereby activating RET, which leads to the activation of the intracellular tyrosine kinase domain. PI3K, MAPK, and PLCγ pathways are activated, regulating cell survival, neurite outgrowth, and synaptic plasticity among other activities

(Airaksinen and Saarma, 2002; Boyd and Gordon, 2003; Sariola and Saarma, 2003).

We used self-made iron oxide NPs coated with polyethyleneimine (PEI) for coupling growth factors (Calatayud et al., 2013). PEI is a polymer, which is also used for the transfection of cells (Vancha et al., 2004). It contains several amine groups allowing covalent binding of proteins via peptide bonds. First, we choose NGF for functionalization. NGF can be internalized with its receptor (Neet and Campenot, 2001; Matusica and Coulson, 2014) and therefore, coupling of NGF to PEI-NPs should result in receptor-mediated uptake, increasing the uptake rate of the NPs.

To check for bio-functionality of NGF after coupling to the NPs, we used PC12 cells and induced the differentiation with the help of our functionalized PEI-NPs. However, these cells cannot be used to model the complex processes of neuronal injury and regeneration. Because of this, we also used a neonatal organotypic spinal cord model to resemble the regeneration of motor neurons after axotomy (Pinkernelle et al., 2012, 2013; Keilhoff et al., 2014). Organotypic cultures keep the cellular organization and the cellcell-contacts. Therefore, they model the in vivo environment more accurately than disperse primary cell cultures or cell lines (Stavridis et al., 2005; Pinkernelle et al., 2013). Proximal axotomy was induced through the preparation of the spinal cord slices and resulted in a loss of motor neurons. The trophic requirements of motor neurons are still under discussion due to the variety of factors needed and the wide range of cells providing them (Oppenheim, 1996; Brunet et al., 2007).

GDNF is one of the most prominent neurotrophic factors playing a role in both the survival of motor neurons and their axonal regeneration (Bohn, 2004; Brunet et al., 2007; Vyas et al., 2010; Pajenda et al., 2014). In organotypic spinal cord cultures, a representative motor neuronal population of about 60% can be found after 1 week of cultivation with GDNF supplement (Rakowicz et al., 2002; Vyas et al., 2010). Therefore, we functionalized our self-made PEI-NPs with GDNF. Like NGF, GDNF may be internalized via endocytosis together with its receptor and moves anterogradely along axons (Neet and Campenot, 2001; von Bartheld et al., 2001). In previous studies, we found no reliable uptake of iron oxide NPs in primary motor neurons and granular cells (Pinkernelle et al., 2012). Therefore, we tried to facilitate the uptake of magnetic PEI-NPs with the help of GDNF functionalization in order to obtain a receptormediated uptake of NPs into the motor neurons of our spinal cord model.

## Materials and Methods

#### Cell Cultures Cultivation

PC12 cells obtained from American Type Culture Collection were cultured in Dulbecco's modified Eagle's media with 10% horse serum, 5% fetal bovine serum (FBS), 100 IU/ml penicillin, 100µg/ml streptomycin, and 2 mM L-glutamine. Cells were cultivated in poly-L-lysine (Sigma, St. Louis, USA) coated dishes and maintained at 37◦C in a saturated humidity atmosphere of 95% air and 5% CO2. For cell differentiation, PC12 cells were incubated in serum-reduced media (2% FBS).

#### PC12 Actin Staining

The cytoskeletal arrangement of PC12 cells was studied by means of actin staining. PC12 were seeded in Petri-dishes (for confocal microscopy) at a concentration of 5 × 10<sup>5</sup> cells/ml and incubated overnight in growth media for cell adhesion. NGF:PEI-NPs were added in reduced serum media at a concentration of 10µg/ml. Cells were cultured for 3 days before staining.

The media was removed and the cells gently washed with phosphate buffered saline (PBS) at 4◦C followed by fixation/permeabilization with a solution of methanol: acetone (1:1 v/v) at −20◦C for 30 min. At the end of incubation, the cells were washed 3x with PBS at 4◦C and incubated overnight at 4◦C with the primary anti-actin antibody (1:200, rabbit polyclonal, Santa Cruz Biotechnologies, Santa Cruz, USA). Then, the sample was washed again 3x with PBS at 4◦C and incubated 1 h at room temperature with the secondary antibody (1:100, goat anti-rabbit IgG-TR, Santa Cruz Biotechnologies). Cells were washed 3x with PBS, dried, and mounted. The images were analyzed by confocal microscopy.

#### Electron Microscopy

Electron imaging was performed with a scanning electron microscope (SEM INSPECT F50.FEI company) and dual-beam (FIB/SEM. Helios 600.FEI company). SEM images were taken at 5 and 30 kV with a FEG column and a combined Ga-based 30 kV (10 pA) ion beam was used to cross-section single cells. The investigations were completed by energy-dispersive X-ray spectroscopy (EDX) for chemical analysis. For this, PC12 cells were seeded on glass coverslips (previously coated with poly-Llysine) at a density of 5 × 10<sup>5</sup> cells/ml. After cell adhesion, the growth media was removed and replaced with the reduced media containing the functionalized NPs (10µg/ml) or media with corresponding NGF concentration. After 72 h of incubation, cells were washed with PBS, fixed, dehydrated, dried, and sputtered with 30 nm of gold for electron imaging.

#### Organotypic Spinal Cord Cultures and Co-cultures Animals

All animal studies were performed in accordance with the guidelines of the German Animal Welfare Act. This study was approved by the Animal Care and Use Committee of Saxony-Anhalt, Germany. A formal approval to conduct the described experiments was obtained from the Animal Subjects Review Board of our institution and can be provided upon request. All efforts were made to minimize the number of animals used and their suffering.

#### Preparation and Cultivation

Organotypic cultures were prepared as described by Vyas et al. (2010) with slight modifications. Neonatal rats (postnatal day 4) were decapitated, and their spinal cords excised. Spinal roots and meninges were removed in dissection buffer (Hank's balanced salt solution, 3.4 mM NaHCO3, 10 mM 4- (2-hydroxyethyl)-1-piperazineethanesulfonic acid, 33.3 mM D-glucose, 5.8 mM MgSO4, 0.03% bovine serum albumin (BSA), 1% penicillin/streptomycin) and the lumbar spinal cord (approximately L1-L6) was cut into 350µm transverse sections using a McIlwain tissue chopper (Mickle Laboratory Engineering, Gomshall, UK). 6–8 slices were used from 1 animal and distributed equal to all experimental groups. The slices were cultured on Millicell membrane inserts (Millipore, Billerica, USA) in 6-well plates. Each well contained 1 ml of media composed of 50% Eagle's minimal essential media, 25% Hank's balanced salt solution, 25% FBS, 35 mM D-glucose, 2 mM L-alanyl-L-glutamine, 1% penicillin/streptomycin. Cultures maintained at 37◦C in a saturated humidity atmosphere of 95% air and 5% CO2.

Culture preparation induces a loss of motor neurons due to the proximal axotomy caused by the slicing. However, a representative population of motor neurons (approximately 60%) can be found after 1 week in culture in the presence of GDNF (R&D systems, Minneapolis, USA) (Rakowicz et al., 2002; Vyas et al., 2010). Thus, control cultures received GDNF treatment to stabilize the motor neuron population.

#### Immunohistochemistry

Cultures were fixed after 1 week of cultivation by replacing the media with 4% paraformaldehyde overnight. The membranes of the Millicell inserts were separated from the carrier and the cultures were stained free-floating. Slices were washed 3x with PBS. Non-specific binding sites were blocked with 10% FBS and 0.3% Triton-X100 in PBS for 1 h. Cultures were incubated overnight with primary anti-pan-neurofilament antibody (1:1000, mouse monoclonal, Sternberger Monoclonals, Baltimore, USA) to visualize neurons including motor neurons and neurites diluted in 10% FBS, 0.3% Triton-X100 in PBS. Next, 3x washing with PBS was followed by secondary antibody incubation for 3 h (1:250, goat anti-mouse Alexa 488, Invitrogen, Carlsbad, USA). The slices were washed again and embedded on glass slides with Immu-Mount (Thermo Scientific, Waltham, USA).

Cultures were imaged with an AxioImager microscope and analyzed with the AxioVision Rel. 4.8 Imaging software by Zeiss (Jena, Germany). The microscopic settings and the exposure time were set on the basis of control slices and kept equal for the corresponding preparation.

Statistical analysis was performed using Graph Pad Prism 4 (GraphPad Software, La Jolla, USA).

#### Nanoparticles

#### Synthesis of PEI-NPs

Magnetic NPs synthesis is based on a modification of the wellestablished oxidative hydrolysis method (i.e., the precipitation of an iron salt in basic media with a mild oxidant) (Sugimoto and Matijevic, 1980 ´ ). In situ polymer coating was achieved by adding PEI (25 kDa) during the reaction, as described previously (Calatayud et al., 2013). The particles exhibit a ferric oxide core (about 25 nm in diameter) and a thin polymer coating of PEI (about 0.7–0.9 nm); the total mass of PEI in the particles was around 14% (w/w) and the amount of the PEI on the particle surface was estimated 10 ± 2µg/mg of NPs (Calatayud et al., 2013).

#### Preparation of NGF-PEI-NPs and NGF:PEI-NPs

NGF-β (Sigma, N1408) was marked with a fluorescent tag by labeling a protein mixture of NGF/BSA (1:6 w/w) with Alexa Fluor 488 (Life Technology, A-10235). TFP ester of the dye efficiently reacted with primary amines of proteins. Purification through a size exclusion resin allowed discarding the unincorporated dye; 90% of the initial protein was labeled with 1.5 moles of dye per mole of protein (n = 3). In order to functionalize the particles with the fluorescent NGF, 2 functionalization approaches were tested: covalent and non-covalent. PEI groups exposed on the surface of PEI-NPs offer primary and secondary amine-groups that have been used to functionalize the particles with the labeled protein.

In the covalent approach, the fluorescent NGF was attached to PEI-NP via EDAC chemistry (Fluka 03450, Sigma, USA). Briefly 0.6 mg EDAC and 1.2 mg NHS were dissolved in 0.2 ml 0.5 M MES-buffer (pH = 6.3) and added to 0.8 ml of protein (NGF concentration 35µg/ml). After a few minutes, 500µg of PEI-NPs were added and mixed for 3 h at 4–8◦C. The unbounded protein was removed by magnetic separation and discharging the supernatant (3 washing steps). The NPs (thereafter labeled as NGF-PEI-NPs) were resuspended in PBS.

The non-covalent approach consisted in adding an equal volume of fluorescent protein (NGF concentration 35µg/ml) to 1 mg/ml of PEI-NPs. The resulting suspension was dispersed at room temperature for 3 h under stirring. Unbound protein was removed by magnetic separation and discharging the supernatant (3 washing steps). The NPs (thereafter labeled as NGF:PEI-NPs) were resuspended in PBS.

For both approaches, the amount of fluorescent NGF bound to the surface of PEI-NPs was calculated by subtraction, i.e., by measuring the absorbance at 280 nm of the supernatant derived from the 3 washing steps. The concentration was calculated by using a calibration curve obtained with known amounts of fluorescent NGF. The composition of the sample was estimated to be 500µg/ml of NPs, 7µg/ml of NGF.

#### Validation of NGF:PEI-NPs respectively NGF-PEI-NPs

For validation of the functionality of the NGF bound to the PEI-NPs, PC12 cells were used as biological indicator. PC12 cells were incubated in serum-reduced media for 72 h either with NGF:PEI-NPs or NGF-PEI-NPs (NGF concentration 140 ng/ml, NP concentration 10µg/ml) to induce differentiation of the cells based on functionalized PEI-NPs.

#### Preparation of GDNF:PEI-NPs

For NP functionalization, a protein mixture of GDNF/BSA (1:4 w/w, GDNF from Sigma Aldrich, USA) was labeled with Alexa Fluor 488. Purification through a size exclusion resin allowed discarding the unincorporated dye; 90% of the initial protein was labeled with 1.5 moles of dye per mole of protein (n = 3). The functionalization of the NPs was carried out by adding the labeled protein (concentration of GDNF 35µg/ml) to PEI-NPs (500µg/ml). The resulting suspension was dispersed at room temperature for 3 h under stirring. Unbound protein was discarded (via magnetic separation) and the functionalized NPs were resuspended in 15% glycerol water solution to stabilize the product. The composition of the sample was estimated to be 500µg/ml of NPs, 2.8µg/ml of GDNF, 11.2µg/ml of BSA, 15% of glycerol.

#### Validation of GDNF:PEI-NPs

To validate the functionality of GDNF bound to PEI-NPs, we analyzed the survival of motor neurons of an organotypic spinal cord culture as biological indicator. Cultures were divided into 4 groups with 2 receiving GDNF supplemented to the media (100 ng/ml or 50 ng/ml GDNF), another without any GDNF supplement and 1 group cultivated with 10µg/ml GDNF:PEI-NPs for the entire cultivation period of 1 week. Spinal cord slices of the GDNF:PEI-NPs group were preincubated for 30 min on ice with 20µg/ml GDNF:PEI-NPs diluted in preparation buffer to increase NP-uptake. Slices of other groups were handled in parallel using preparation buffer without NPs.

To validate the stability of GDNF:PEI-NPs in serum-high media for spinal cord cultures, we pre-incubated 10µg/ml GDNF:PEI-NPs in spinal cord media for 1 h at 4◦C. Afterwards, GDNF:PEI-NPs were separated using a neodym magnet. One group received this media supernatant for cultivation (thereafter supernatant GDNF:PEI-NPs). The separated GDNF:PEI-NPs were resuspended in fresh, GDNF-free spinal cord media. This GDNF:PEI-NPs containing media was also used for cultivation (thereafter resuspended GDNF:PEI-NPs). Controls were fed with GDNF supplemented media (50 or 100 ng/ml) for the whole cultivation time and handled in parallel.

Statistical analysis was performed using a One-Way-ANOVA followed by a Bonferroni post-hoc analysis with p < 0.05 being statistical significant.

#### Results

#### Approaches for NGF Functionalization: Covalent and Non-covalent

The fluorescently labeled NGF and the functionalized NPs were tested on PC12 cells as these possess specific cell surface receptors that bind NGF. In presence of this growth factor, cells undergo a dramatic change in phenotype wherein they acquire a large part of the characteristic properties of sympathetic neurons. PC12 cells treated with 50 ng/ml of fluorescent NGF were found to exhibit a phenotype similar to the cultures treated with 50 ng/ml of NGF. In a previous study Western blot analysis showed an upregulation of phosphorylated TrkA receptor expression by NGF-β, fluorescent labeled NGF, and fluorescently labeled NGF coupled PEI-NPs confirming this result (Riggio et al., 2014).

Experimental results revealed that the fluorescent NGF binds the surface of the PEI-NPs at a ratio of 14 ± 3µg of fluorescent NGF per mg of PEI-NPs. No statistical difference was found between non-covalent NGF:PEI-NPs and covalent NGF-PEI-NPs (data not shown). PC12 cells were incubated with reduced media modified with NGF:PEI-NPs or NGF-PEI-NPs (NGF concentration 140 ng/ml, NP concentration 10µg/ml). In **Figure 1**, PC12 cultures incubated with reduced media (**Figure 1A**), reduced media and NGF (**Figure 1B**), reduced media and NGF-PEI-NPs (**Figure 1C**), and reduced media and NGF:PEI-NPs (**Figure 1D**) are shown. For NGF-PEI-MNPs, a strong reduction of the neurite number and length was observed compared to the control (NGF 140 ng/ml). Otherwise, for NGF:PEI-NPs cell differentiation proceeded similarly to the control cultures treated with NGF. Confocal imaging

FIGURE 1 | PC12 cells incubated for 72 h with reduced media (A), reduced media modified with NGF (B), reduced media modified with NGF-PEI-NP (C), or reduced media modified with NGF:PEI-NP (D) (20X). Confocal microscopy of PC12 cells after 72 h of incubation with NGF:PEI-NPs is shown in (E). Actin staining reveals the presence of NPs (green) in cell body, the growth cone at the tip of the axons of the differentiated cells (40X).

demonstrated that the NGF:PEI-NPs are strongly engulfed by cells, being localized in both the cell body and neurite protrusions (**Figure 1E**). Based on these results, we can conclude that the covalent approach impairs the bio-functionality of the protein. We postulate that the strong interaction between the particles surface and NGF induce the protein wrapping/adsorption onto the particle surface. This could strongly affect the protein mobility, ultimately altering its 3D structure and the interaction with the TrkA receptor. Based on these results, we considered the non-covalent approach to be the most promising. Electron microscopy confirmed that in presence of NGF:PEI-NPs, PC12 cells were properly differentiated. They exhibited long neurites, which were well connected within the network (**Figures 2A,B**). **Figure 2C** shows a representative cell with a cluster of NPs bound to the cell surface, which is highlighted in **Figure 2D**. EDX analysis revealed iron content and therefore confirmed the NGF:PEI-NP–cell interaction (**Figure 2E**). FIB-SEM analysis of milled PC12 cells, a technique that allows sectioning a cell and acquiring information on the constitutive elements inside (Riggio

et al., 2014), illustrated the successful internalization of NGF:PEI-NPs (**Figures 3A,B,D**). EDX analysis, shown in **Figures 3C,E**, confirmed that internalized clusters seen in **Figures 3A,B,D** are from iron and therefore iron oxide PEI-NPs.

#### Validation of Bio-functionality of GNDF:PEI-Nps

First, we tested for bio-functionality of GDNF:PEI-NPs after coupling. For this, we analyzed the number of surviving motor neurons in organotypic spinal cord cultures incubated with these NPs instead of GDNF supplement via media. GDNF:PEI-NPs were able to keep motor neuronal populations alive. There were no significant differences in the total number of surviving neurons or in the number of motor neurons between controls containing 100 or 50 ng/ml GDNF in the media and GDNF:PEI-NPs incubated cultures (**Figures 4A,B**). Parallel handled cultures that were cultivated without GDNF supplement displayed significantly less surviving neurons in the total number and in the number of motor neurons than controls 100 (100 ng/ml GDNF supplement). **Figures 4C–E** show corresponding immunofluorescent stainings for anti-pan-neurofilament. Controls and cultures, incubated with GDNF:PEI-NPs illustrated a prominent motor neuronal population (**Figures 4C,E**). Cultures without GDNF supplement displayed a clear reduction of stained motor neurons (**Figure 4D**). Accordingly, GDNF function was kept during coupling to the PEI-NPs.

The non-covalent approach was used to test the NP functionalization with GDNF. Therefore, we checked the GDNF coupling stability to the PEI-NPs. For this, organotypic cultures were cultivated for 7 days and fed with media containing 100 ng/ml GDNF (control 100), 50 ng/ml GDNF (control 50, reduced GDNF to compare with the lower GDNF concentration, which is delivered by the GDNF:PEI-NPs), supernatant of GDNF:PEI-NPs or resuspended GDNF:PEI-NPs.

GDNF:PEI-NPs were diluted in spinal cord media without supplemented GDNF (NP concentration 10µg/ml, GDNF concentration about 55 ng/ml) and incubated for 1 h at 4◦C. Then, NPs were separated from media with a magnet and the NPs were resuspended in fresh media without GDNF and used for cultivation of the resuspended GDNF:PEI-NPs-group. The supernatant, which was left after the magnetic separation, was also used for cultivation (supernatant GDNF:PEI-NPs). GDNF is required for the survival of the motor neuronal population in the spinal cord cultures. Therefore, we quantified the number of surviving motor neurons in both groups and compared them to controls. If the number of motor neurons of the resuspended GDNF:PEI-NPs-group was similar to the number of motor neurons in the controls, then the GDNF was delivered by the GDNF:PEI-NPs. But if the number of motor neurons of the supernatant group were similar to that of the controls then the GDNF was dissociated from the PEI-NPs and left in the supernatant.

**Figure 5** displays the result. Organotypic cultures incubated with the supernatant of former diluted GDNF:PEI-NPs revealed the same survival rate of the total number of neurons and the motor neurons as was seen in the controls (**Figures 5A,B**). At the same time, cultures that were cultivated with the resuspended GDNF:PEI-NPs showed a significant decrease of the number of

surviving motor neurons compared to control 50. **Figures 5C–E** shows corresponding immunofluorescent stainings for antipan-neurofilament. Controls and cultures incubated with the supernatant of GDNF:PEI-NPs, but not with the particles itself, illustrated a pronounced motor neuronal population (**Figures 5C,E**). Cultures receiving resuspended GDNF:PEI-NPs showed a clear reduction of stained motor neurons (**Figure 5D**). So, GDNF bio-functionality did not correspond to the availability of the GDNF:PEI-NPs. The GDNF coupling was not stable.

### Discussion

demonstrate *p* < 0.05.

#### Functionalization with NGF

Aim of this study was to functionalize magnetic NPs with growth factors to facilitate a receptor-mediated intracellular uptake into neurons. Enhanced neuronal uptake would allow the magnetically targeted delivery of drugs and growth factors to injury sites of the CNS and PNS.

There are only a few studies using growth factors to functionalize NPs. Polak and Shefi (2015) published an overview about growth factors used. Besides NGF, basic fibroblast growth factor (FGF2), BDNF, and GDNF are reported.

To check the bio-functionality of growth factors after binding to NPs a fast, easy, and reliable model is needed. In vitro assays are the methods of choice because they require no or fewer animals, are faster, and also cheaper than in vivo models.

In our study, we first chose NGF for loading the self-made magnetic PEI-NPs. For determining the bio-functionality of the NGF loaded PEI-NPs, we used PC12 cells as indicator model. PC12 cells can be differentiated with NGF into cells displaying functions of sympathetic neurons. NGF-treated cells terminate proliferation, develop neurites, become electrically excitable, and express neuronal proteins like synapsin and growth associated protein 43 (GAP43) (Greene and Tischler, 1976; Fujita et al., 1989; Das et al., 2004). The outgrowth of neurites is used as a differentiation marker because it is easy to quantify. PC12 cells

are used in many studies concerning NPs and functionalization of NPs, e.g., Pisanic et al. (2007) used PC12 cells to analyze nanotoxicity of iron oxide NPs, Roy et al. (2010) used PC12 cells to check for the targeting of NGF-conjugated NPs and evaluated the expression levels of Trk- and p75-receptors, and Mittnacht et al. (2010) functionalized siRNA for RhoA, a kinase playing a role in neurite outgrowth. We started with a covalent approach to bind NGF to the amine groups of PEI-NPs via EDAC chemistry. Covalent binding worked, but obviously changed the structure of NGF, as these NGF-PEI-NPs were not able to differentiate PC12 cells properly. It seems that the protein function was impaired by covalent binding to PEI-NPs. NGF binds to its receptor TrkA via N-terminal residues. The loops L2 and L4 of the protein structure appear to be responsible for bio-functionality of NGF (Wiesmann et al., 1999). The covalent binding of NGF to amine groups of the PEI-NPs occurs randomly. Because we observed an impairment of bio-functionality of NGF, it is possible that binding to N-terminal residues or in the loops L2 and L4 took place. Alternatively, we bound NGF using a non-covalent approach, which is based on electrostatic interactions between molecules. These NGF:PEI-NPs were able to differentiate PC12 cells to neuronal cells. Functionality was maintained during binding to the PEI-NPs. Interestingly, experimental data demonstrated that the NGF:PEI-NPs complex is also stable when mixed to the cell culture media, which indicates that non-specific adsorption of plasma proteins does not alter the stability and the integrity of the conjugate (Riggio et al., 2013).

#### Functionalization with GDNF

Additionally, we used GDNF for functionalization of our PEI-NPs. GDNF was first described in 1993 by Lin et al. (1993) as a growth promoting factor for embryonic midbrain dopaminergic neurons. Since then, it was shown that GDNF also promotes the survival of dopaminergic neurons in rodents and non-human Parkinson models (Gash et al., 1998), that it can prevent the programmed cell death of motor neurons during embryonic development (Oppenheim et al., 1995), and that GDNF increases survival and neurite sprouting in models of spinal cord injury (Li et al., 1995; Pajenda et al., 2014). GDNF also has nonneurotrophic functions: it regulates the ureteric branching during embryogenesis (Vega et al., 1996) and spermatogenesis (Meng et al., 2000). The various functions of GDNF limit the choice of useful in vitro models. One model that allows analyzing the GDNF functionality fast and reliably is the organotypic culture of dorsal root ganglions. The group of Ziv-Polat et al. (2014) measured sprouting and the onset of myelination to evaluate GDNF function. In contrast, we used a neonatal organotypic spinal cord culture to assess the bio-functionality of GDNF. In such cultures, GDNF is known to be a prominent neurotrophic factor for motor neuronal survival, as it is able to keep about 60% of the motor neurons alive during cultivation (Rakowicz et al., 2002; Vyas et al., 2010). Therefore, quantifying the number of surviving motor neurons after a sufficient cultivation time already allows one to judge the successful functionalization of NPs.

Based on our previous results of NGF coupling to PEI-NPs, we decided to use a non-covalent approach to bind GDNF to our PEI-NPs. To check the functionality of the bound GDNF, we cultivated our organotypic spinal cord cultures with GDNF:PEI-NPs only and compared the number of surviving motor neurons and the total number of neurons after 7 days of cultivation. Compared to control cultures, which received GDNF via media, cultures incubated with GDNF:PEI-NPs showed a similar total number of surviving neurons and also motor neurons at the end of the cultivation. Therefore, GDNF function was kept during coupling to PEI-NPs. Because of the 3-dimensional structure of the organotypic culture, visualizing an uptake of GDNF:PEI-NPs into neurons is difficult. Moreover the use of serum-high media for the cultivation of organotypic spinal cord cultures can cause problems because the NPs become surrounded by a protein corona (Lundqvist et al., 2008). Thus, we tried to confirm the stability of the non-covalent binding of GDNF to the PEI-NPs. According to the protocol already used to test the stability of NGF:PEI-NPs in biological media (Riggio et al., 2013), we incubated the GDNF:PEI-NPs in the serum-rich media for these cultures. After separating these NPs with a strong magnet, we used either the supernatant of the NPs for cultivation of the spinal cord cultures or resuspended them again in GDNF-free, fresh media and used this for the cultivation. If now the number of surviving motor neurons in cultures incubated with resuspended GDNF:PEI-NPs was similar to the number of surviving motor neurons in controls (receiving GDNF via media), the GDNF functionality was kept by the GDNF:PEI-NPs. If the number of surviving motor neurons in cultures cultivated with the supernatant (no GDNF:PEI-NPs left) was similar to the number of surviving motor neurons of the controls, the GDNF biofunctionality passed on to the media. In our study, the number of surviving motor neurons was not significantly different between controls and cultures incubated with the supernatant. But it decreased significantly in cultures cultivated with resuspended GDNF:PEI-NPs compared to controls. Therefore, the GDNF bio-functionality was kept in the supernatant and the GDNF separated from the PEI-NPs.

In contrast, non-covalently NGF:PEI-NPs were found to be stable in binding. They were up taken into PC12 cells and induced a differentiation of PC12 cells. One main difference between the media of PC12 cells and organotypic spinal cord cultures is the amount of serum. Spinal cord culture media contains 25% FBS compared to the 2% used in the media for differentiation of PC12 cells. As already mentioned, NPs are coated with a protein corona in serum-containing media (Lundqvist et al., 2008; Calatayud et al., 2014). The protein corona contains a variety of proteins like immunoglobulins, albumin, apolipoproteins, fibrinogen, and many more (Lundqvist et al., 2008). Non-covalent coupling of GDNF to PEI-NPs is not a fixed coupling; it is based on electrostatic interactions between residues of molecules. These interactions are weaker than covalent interactions where atoms share electrons. Due to the high serum content in the spinal cord media, it can be assumed that the protein corona is more pronounced than in the low-serum media of the PC12 cells. Thus, serum proteins and GDNF appear to compete for binding, which results in a loss of GDNF bound to PEI-NPs.

The group of Ziv-Polat et al. (2014) compared the biofunctionality of NGF, FGF2, and GDNF conjugated to iron oxide NPs. They showed that the conjugated neurotrophic factors were more stable and also the bio-functionality of the growth factors increased significantly. In particular, GDNF bound to the NPs enhanced the myelination in their organotypic dorsal root ganglion model. To conjugate all 3 neurotrophic factors, Ziv-Polat et al. (2014) used a covalent approach via activated double bonds of dextran- or gelatin-coated NPs. Also Marcus et al. (2015) used a covalent binding strategy to bind NGF to human serum albumin-coated iron oxide NPs. They successfully induced the differentiation of PC12 cells in a neuronal-like cell type, thereby proving the bio-functionality of NGF-HSA-NPs. In our experiments, covalently binding was shown to impair the biofunctionality of NGF. We used PEI-coated iron oxide NPs, which allow binding covalently proteins via amine groups. Obviously, binding proteins covalently to NPs is dependent on the chosen chemical strategy. Interactions between proteins and NPs are hardly predictable and since the exact binding site in the structure of the protein is somehow random, the outcome regarding the bio-functionality seems to be random as well. Even if the catalytic site is not involved in chemical binding, the overall 3D structure of the protein could be affected by the strong chemical bonding with the particle, resulting in structural deformation, and loss of function. Designed, small peptides could help to solve these issues and prevent unintentional binding effects and loss of functionality. Alternative approaches already tried to develop small peptides mimicking growth factor functions. Smaller peptides have certain advantages in comparison to native proteins. Growth factors address various signal pathways leading to a variation of actions which can include unintentional effects, too (Boyd and Gordon, 2003). GDNF overexpression during embryogenesis, for example, is preventing naturally occurring programmed cell death of motor neurons leading to hyperinnervation of muscles (Zwick et al., 2001). Small peptides which are designed with special functional sites give the chance to generate distinct signaling profiles. Forte et al. (2014) described the binding of 2 growth factor mimicking peptides, NGF (1–14) and BDNF (1–12) on a gold carrier, a chemistry which can be

### References


used to conjugate to, e.g., gold nanoparticles. Previously, the same group reported biological actions of a NGF (1–14) peptide which could activate single intracellular signal cascades in PC12 cells (Travaglia et al., 2015). Bradley et al. (2010) specified another synthetic peptide showing GDNF functions. This peptide DNSP-11from the proGDNF-domain was shown to support neuronal survival and neuritic outgrowth in vitro, but also increasing dopamine levels in an animal model of Parkinson disease.

Taken together, we could show that functionalization of NPs with growth factors can be complex and difficult. Functionalization of PEI-NPs with NGF using covalent and non-covalent binding strategies displayed contrasting results. Covalently bound NGF-PEI-NPs were not able to differentiate PC12 cells, proving a lack of bio-functionality after binding. Non-covalent bound NGF:PEI-NPs differentiated PC12 cells into neuronal-like cells, thereby showing bio-functionality of NGF. Due to the experiences of NGF-functionalization, we used a non-covalent approach for functionalization of PEI-NPs with GDNF. This was shown to keep bio-functionality of GDNF in an organotypic spinal cord model, but was additionally found to be unstable bound to the PEI-NPs. We suspect that the different serum content in the media of PC12 cells and organotypic spinal cord cultures is a problem for using non-covalent approaches. All in all, our study shows the importance of checking the biofunctionality of growth factors bound to NPs with a proper biological model. The final future aim of all studies using NPs in life sciences is to use these particles in humans. For this, it is important to develop reproducible models that allow a prognosis of NPs action in vivo.

### Acknowledgments

This work was supported by DFG (KE488/15-1) and grants from NanoscieE + 2008 for the MARVENE project (magnetic nanoparticles for nerve regeneration). We grateful thank Dr. Jonathan Lindquist for proof-reading and Susanne Bonifatius for excellent technical assistance.

to generate a repair cell essential for regeneration. Neuron 75, 633–647. doi: 10.1016/j.neuron.2012.06.021


in neurooncology and central nervous system inflammatory pathologies, a review. J. Cereb. Blood Flow Metab. 30, 15–35. doi: 10.1038/jcbfm.2009.192


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2015 Pinkernelle, Raffa, Calatayud, Goya, Riggio and Keilhoff. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Initial photophysical characterization of the proteorhodopsin optical proton sensor (PROPS)

#### Jay L. Nadeau\*

Graduate Aerospace Laboratories, California Institute of Technology, Pasadena, CA, USA

Fluorescence is not frequently used as a tool for investigating the photocycles of rhodopsins, largely because of the low quantum yield of the retinal chromophore. However, a new class of genetically encoded voltage sensors is based upon rhodopsins and their fluorescence. The first such sensor reported in the literature was the proteorhodopsin optical proton sensor (PROPS), which is capable of indicating membrane voltage changes in bacteria by means of changes in fluorescence. However, the properties of this fluorescence, such as its lifetime decay components and its origin in the protein photocycle, remain unknown. This paper reports steady-state and nanosecond time-resolved emission of this protein expressed in two strains of Escherichia coli, before and after membrane depolarization. The voltage-dependence of a particularly long lifetime component is established. Additional work to improve quantum yields and improve the general utility of PROPS is suggested.

#### Edited by:

Ioan Opris, Wake Forest University School of Medicine, USA

#### Reviewed by:

Thomas Knöpfel, Imperial College London, UK Adam E. Cohen, Harvard University, USA

#### \*Correspondence:

Jay L. Nadeau, Graduate Aerospace Laboratories, California Institute of Technology, Firestone 010 (M/C 105-50), 1200 E, California Blvd., Pasadena, CA 91125, USA jnadeau@caltech.edu

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 24 June 2015 Accepted: 21 August 2015 Published: 04 September 2015

#### Citation:

Nadeau JL (2015) Initial photophysical characterization of the proteorhodopsin optical proton sensor (PROPS). Front. Neurosci. 9:315. doi: 10.3389/fnins.2015.00315 Keywords: voltage-sensitive dyes, genetically encoded voltage sensor, proteorhodopsin, time-correlated single photon counting (TCSPC)

## Introduction and Background

Electrophysiology is the most sensitive technique available for measuring cell membrane potential, but patch-clamp recordings are labor intensive, can only be performed on a limited number of cells at a time, and are extremely difficult to perform on very small cells. One of the greatest technical challenges in neuroscience is to be able to perform optical recordings of real-time processes in large networks of coupled cells, the so-called "optical patch clamp." To resolve a single action potential, a voltage-sensitive optical probe must have a potential resolution of ∼100 mV or better, and a time resolution of milliseconds. Until recently, the best results were obtained using voltagesensitive dyes, in particular a class of organic dyes called the amino-naphthyl-ethenyl-pyridinium (ANEP) dyes, such as di-4-ANEPPS and di-8-ANEPPS (Fluhler et al., 1985). While some groups have obtained action-potential data using these dyes, the technique is not widespread because of the specialized equipment needed and the low signal to noise in the best data (Tominaga et al., 2000; Tsutsui et al., 2001). Another, more sensitive dye-based approach involves detecting polarization changes in neurons by photo-induced electron transfer through a synthetic molecular wire to a dye (Miller et al., 2012). The speed of the electron transfer process makes this an ideal approach to monitoring fast voltage changes. However, dyes cannot be used in targeted cell populations or in whole animals. Genetically encoded alternatives have been sought for several decades, with significant breakthroughs appearing within the past several years.

#### Approaches to Genetically Encoded Voltage Indicators (GEVIs)

GEVIs have been thoroughly reviewed in several articles (Baker et al., 2008; Frommer et al., 2009; Akemann et al., 2010; Mutoh et al., 2012; Ohba et al., 2013). The general approach to creating a genetically encoded voltage sensor is to fuse a fluorescent reporter, usually from the family of green fluorescent protein (GFP), with a voltage-sensing domain (VSD) in such a way that the conformational changes of the sensor with voltage result in a change in the fluorescence of the reporter. However, because of the robustness of GFP fluorescence, slowness of fluorescent response to perturbations in the molecule, and lack of expression of membrane protein-tagged GFPs (Baker et al., 2007), changing emission substantially in this fashion has proven to be a difficult task.

An entirely new alternative approach emerged in 2011 based upon microbial opsins rather than the GFP family. These proteins transduce light into cellular signals, including changes in membrane potential; the concept behind engineering them into voltage sensors was to reverse this relationship, transducing changes in membrane potential into changes in fluorescence emission. The first demonstration of this principle was made using a proteorhodopsin-based optical proton sensor (PROPS) from green light-absorbing bacteria (**Figure 1A**) (Kralj et al., 2011). The principle of PROPS is that a Schiff base is located on the proteorhodopsin inside the membrane. When Vm < 0, protons move from the base to the cytoplasm, causing the protein to become non-fluorescent. When Vm > 0, protons move from the cytoplasm onto the base, causing an increase in fluorescence. The ratio of protonated to deprotonated Schiff bases depends upon the voltage drop between the membrane protein and the cytoplasm (**Figure 1B**).

This proof of principle has been significantly developed over the past few years. PROPS does not target well to plasma membranes of eukaryotic cells, so the group developed a similar sensor based upon archaerhodopsin-3 (Arch), an optogenetic control tool (Kralj et al., 2012; Maclaurin et al., 2013; Hou et al., 2014; Venkatachalam et al., 2014). Most recently, the opsin principle has been used to develop electrochromic FRETbased voltage sensors (**Figure 1C**) (Gong et al., 2014; Zou et al., 2014). These rely upon FRET between the opsin and attached fluorescent protein, and required significant optimization of the protein choice and length of linker.

#### Bacterial Ion Channels and Neuroscience

Although it cannot be used in mammalian cells, PROPS remains interesting as both a proof of principle and as a bacterial sensor. In order to study electrogenic properties of bacterial membrane proteins, the proteins are usually cloned and expressed in Xenopus oocytes, which removes the downstream effects seen in the native cells (Schmies et al., 2001). The role of membrane potential in prokaryotic cell signaling is well known, but not fully understood (Szmelcman and Adler, 1976; Margolin and Eisenbach, 1984; Ordal, 1985; Tisa et al., 1993). When elucidated, the mechanisms used by bacteria to regulate membrane potential may help shed light on evolution of memory, olfaction, and other complex functions (Eisenbach, 1982; Eisenbach et al., 1983a,b; Goulbourne and Greenberg, 1983; Vladimirov and Sourjik, 2009; Lyon, 2015). Bacterial ion channels are also often good models for the function of mammalian ion channels, and their relationship to membrane potential may perhaps provide new approaches to drug screening. For example, the bacterium

as a result of membrane potential depolarization. (B) Detailed mechanism of PROPS voltage sensing (concept from Kralj et al., 2011). When membrane potential is negative, protons move away from the Schiff base, causing the chromophore to become less fluorescent. As the membrane depolarizes, protons move toward the Schiff base and increase the quantum yield of the chromophore. The ratio of protonated to deprontonated species depends upon the voltage drop V between the Schiff base and the cytoplasm; in general V < Vmem. (C) A new concept for opsin-based GEVIs fuses fluorescent proteins to the opsin and uses differences in FRET efficiency between the fluorescent protein and the retinal for sensing.

Arcobacter butzleri was recently found to have a voltage-gated Na<sup>+</sup> channel, whose selectivity filter is profoundly different from that seen in mammalian cells (Payandeh et al., 2011). The significance of this remains unknown. Despite limited sequence homology, bacterial sodium channels and transporters share binding sites with mammalian homologs, and often respond to the same ligands (Henry et al., 2007; Bagnéris et al., 2014). The development of bacterial-based optical screening techniques for drugs affecting the sodium channel, as well as other channels and transporters, would have immense application in neuroscience (Chakrabarti et al., 2013; Bagnéris et al., 2014).

In this paper we perform steady-state and time-resolved spectroscopy of PROPS expressed in two bacterial strains. The dependence of emission brightness, spectrum, and lifetime were studied as a function of wavelength and power of excitation. Fits to 1–3 Gaussian distributions were necessary to describe the lifetime decays, consistent with the chromophore being embedded within a protein. Depolarization of the cells with CCCP or exposure to violet light led to greater population of the longer-lifetime state, consistent with changes in steady-state intensity observed during microscopy. Voltage dependence of this fluorescent state was observed. Based upon these results, a preliminary model for fluorescence in PROPS is suggested, with ideas for future work.

#### Materials and Methods

#### Strains and Expression

The E. coli strains containing PROPS were a gift of Adam Cohen, Harvard University. They were prepared as reported previously: E. coli was grown to early-log phase (OD600 = 0.3–0.4) in Lysogeny Broth (LB) at 33◦C. Arabinose was added as an inducer along with 5µM all-trans retinal; further growth was conducted in the dark. The cells were harvested 3.5 h after induction and washed with minimal medium (1x M9 salts, 0.4% glucose, pH 7), then resuspended in minimal medium. Cultures were stored at 4 ◦C for up to 1 week before use. Two different strains of E. coli were used: BW25113 (referred to here as "BW") (1(araDaraB)567, 1lacZ4787(::rrnB-3), lambda-, rph-1, 1(rhaDrhaB) 568, hsdR514; and UT5600 (the "UT" strain) (F- ara-14 leuB6 secA6 lacY1 proC14 tsx-67 1(ompT-fepC)266 entA403 trpE38 rfbD1 rpsL109 xyl-5 mtl-1 thi-1). Concentrations of arabinose used for induction were 0.02% w/v (BW strain) or 0.2% (UT strain), and protein expressed was gauged by the color of the pellet.

#### Steady-state Spectroscopy and TCSPC

Steady-state spectra were recorded on a Fluorolog-3 (Jobin Yvon) spectrometer. Measurements were made with and without the addition of 50µg/mL carbonyl cyanide m-chlorophenyl hydrazone (CCCP; Sigma-Aldrich). Photoluminescence decays from bulk samples were obtained by the time-correlated single photon counting (TCSPC) technique. Eight hundred nano meter laser pulses (∼70 fs) out of a Coherent RegA 9050 Ti/sapphire regenerative amplifier operating at 250 kHz repetition rate were used to pump an OPA (Coherent 9450) which produced tunable visible light with an average power of ∼30 mW. The beam was focused into the sample with a focal spot diameter of ∼0.25 mm. The excitation power delivered to the sample was set at 3 mW ("High Power") or 40 µW ("Low Power"). For violet light exposure, full power at 400 nm was used, providing ∼50 W/cm<sup>2</sup> ; exposure was performed before spectroscopy because only one illumination wavelength was possible at a time. The luminescence was collected with a 3.5 cm focal length lens placed perpendicular to the excitation beam and the collimated luminescence focused into a monochromator with a 10 cm focal length lens. The monochromator was a CVI CMSP112 double spectrograph with a 1/8 m total path length in negative dispersive mode with a pair of 600 groove/mm gratings (overall f number 3.9). The slit widths were 2.4 mm and based on a monochromator dispersion of 14 nm/mm, provided 10 nm resolution. A Hamamatsu RU3809 microchannel-plate photomultiplier was mounted on the monochromator exit slit. A Becker and Hickl SPC-630 photon counting board was used to record the time-resolved emission. The reference signal was provided by a portion of the excitation beam sent to a fast photodiode. To ensure good statistics, count rates were held at <1% of the laser repetition rate to avoid pulse pile up. Typical acquisition times were 10 min for a single scan. The instrument response function (IRF) was determined from scatter off a solution of dilute coffee creamer. The full width at half-maximum of the IRF was 37 ps.

#### Curve Fitting

Data analysis was performed using FluoFit 4.0 (PicoQuant, Berlin). Goodness of fit was assessed by χ 2 -values and by examination of residuals; χ 2 -values <1.1 and a random distribution of residuals were required for a fit to be considered accurate. A sum of exponentials, up to 4 terms, was insufficient to describe the decays, as was a stretched exponential (up to 3 terms). The best fit was obtained to a sum of Gaussian distributions, which is appropriate for a collection fluorophores within inhomogeneous environments such as proteins, with the mathematics and physics developed by Prendergast et al. (Alcala et al., 1987a,b,c; Togashi and Ryder, 2006). This model is described by the equations.

$$\begin{split} I(t) &= \int\_{-\infty}^{t} \text{IRF}(t') \int\_{-\infty}^{\infty} \rho(\mathbf{r}) \exp\left(-\frac{t - t'}{\tau}\right) d\tau dt' \\ \rho(\mathbf{r}) &= \sum\_{i=1}^{n} \frac{A\_i}{\sigma\_i \sqrt{2\pi}} \exp\left[-\frac{1}{2} \left(\frac{\mathbf{r} - \mathbf{r}\_i}{\sigma\_i}\right)^2\right] \\ \sigma\_i &= \frac{\Delta\_{FWHM}}{\sqrt{8\ln 2}}, \end{split} \tag{1}$$

where n = 1 − 3 for our samples. Both amplitude-weighted average lifetimes:

$$\langle \tau \rangle = \sum\_{i} \frac{A\_i \tau\_i}{A\_i}$$

and intensity-weighted average lifetimes:

$$\langle \tau \rangle = \sum\_{i} \frac{A\_i \tau\_i^2}{A\_i \tau\_i}$$

were calculated. The intensity average corresponds to the amount of time the fluorophore spends in the excited state. The amplitude average is the lifetime a fluorophore would have if it had the same steady-state fluorescence as the fluorophore with several lifetimes (Sillen and Engelborghs, 1998).

#### Results

All spectroscopy was performed on PROPS expressed in E. coli cells. Steady-state spectroscopy showed a fluorescence emission peak at ∼660 nm in both strains after induction; there was no measurable emission without induction. In the BW strain, an increase in emission with CCCP was seen across the emission peak when samples were excited at 580 nm (**Figure 2A**). In the UT strain, smaller differences were noted at wavelengths both bluer and redder than the emission peak (**Figure 2B**).

Although the pellets of both strains appeared equally pink, comparison of the spectral changes seen with CCCP at different excitation wavelengths revealed significant differences for the BW strain. At excitation wavelengths from 510 to 530 nm, essentially no change was seen. Between 535 and 580 nm, the difference with CCCP grew in a roughly linear fashion, then declined again in approximately a mirror image of the increase. Excitation wavelengths >600 nm were not used so that the entire spectral peak could be captured (**Figures 3A,B**). The UT strain was significantly different. Less dependence upon excitation wavelength was seen in the difference spectra, and rather than reflect a simple enhancement or quenching, the spectral changes

between spectra with and without CCCP with excitation at 580 nm in the UT strain.

showed a negative and a positive peak. Overall changes were significantly smaller than with the BW strain. Although the absolute value of the emission was approximately half as strong in the UT strain (peak ∼7000 counts for BW vs. ∼3500 counts for UT), the differences with CCCP were almost 10-fold lower in this strain than in the BW strain (difference of ∼300 counts with CCCP for UT vs. nearly 3000 for BW) (**Figures 3C,D**).

TCSPC was then performed at 532 and 600 nm excitation with the two strains, beginning with the BW strain at low power (40µW). With excitation at 532 nm and emission at 650 nm, three terms were required in Equation (1) to obtain a good fit. Although there was a longer-lifetime component apparent in the samples after the addition of CCCP, both the intensity-weighted and amplitude-weighted average lifetimes were indistinguishable in the two cases (**Figure 4A**, **Table 1**). Emission at 710 nm could be fit to a single Gaussian-distributed exponential without CCCP. Addition of CCCP again caused the appearance of a longer lifetime component, but without change in mean lifetime (**Figure 4B**, **Table 1**). Excitation at high power (3 mW) caused the complete disappearance of the longer-lifetime component, with a significant reduction in mean lifetime (**Table 1**).

In contrast, at 600 nm excitation voltage dependence of the lifetime decays could be observed. At 650 nm emission, the

FIGURE 3 | Dependence of voltage-sensitivity upon wavelength of excitation. (A) Difference spectra (CCCP-no CCCP) for multiple excitation wavelengths in the BW strain. The different excitation wavelengths are indicated by numbers next to the curves. (B) Difference at 650 nm emission vs. excitation wavelength in the BW strain. (C) Difference spectra (CCCP-no CCCP) for multiple excitation wavelengths in the UT strain. (D) Difference at 600 and 700 nm emission vs. excitation wavelength in the UT strain.

longest-lifetime component shifted from 1.2 to 2.4 ns with the addition of CCCP. In addition, the fractional intensity of the longer component increased. Both the intensity-weighted and amplitude-weighted lifetimes were approximately doubled with CCCP addition at 650 and 710 nm emission (**Figures 4C,D**, **Table 2**). Pre-exposure to violet light (400 nm, 50 W/cm<sup>2</sup> ) had a nearly identical effect. CCCP plus violet light led to a further increase in the longest lifetime and a slight increase in its fractional intensity. High power excitation further increased the magnitude of the long lifetime (**Table 2**). **Figure 5** illustrates the long-lifetime component and mean intensity-weighted lifetime as a function of selected test conditions. It can be readily seen from **Figure 5A** that a lifetime component >2 ns occurred with CCCP or high-power excitation with 600 nm light. While a long component was seen with 532 nm excitation in the presence of CCCP at low excitation power, high excitation power at 532 suppressed this component. From **Figure 5B**, it can be appreciated that the mean lifetimes were voltage-dependent only with 600 nm excitation.

The results for the UT strain were qualitatively similar, though the magnitude of the changes was smaller than with the BW strain, consistent with what was observed in the steady-state spectra. Only emission at 710 nm was recorded because this was a maximum in the difference spectra. Excitation at 600 nm led to a small increase in mean lifetime. High-power excitation reduced lifetime. Interestingly, a small decrease in mean lifetime was seen with CCCP addition at 532 nm excitation (**Table 3**).

#### TABLE 1 | TCSPC fit parameters for the BW strain with excitation at 532 nm.


#### TABLE 2 | TCSPC fit parameters for the BW strain with excitation at 600 nm.


#### Discussion

The results obtained here are consistent with the original work on PROPS, which reported increased red fluorescence in PROPS-expressing E. coli upon depolarization with CCCP or exposure to violet light (Kralj et al., 2011). Our results suggest that there are multiple red-fluorescent species in heterogeneous environments within the protein, but that most of them have much shorter lifetimes than the species responsible for the voltage-sensitive emission, which has a lifetime of ∼2.5–3 ns. Only a fraction of the molecules probed in these experiments showed this slower lifetime, suggesting that improvements in PROPS yield and voltage sensitivity could be obtained through exact identification and

FIGURE 5 | Comparison of fit parameters. (A) Length of the longest lifetime component as a function of sample conditions for selected samples. (B) Value of the intensity-weighted average lifetime as a function of sample conditions. Uncertainties in the fits are <2%.



mutation of this state to yield greater stability or ease of excitation.

The differences seen between the two tested strains might have been due to a two-fold difference in expression levels, with the BW strain expressing more highly. However, it may also reflect differences in membrane potential due to strain variations or differences in the phase of the cell cycle in which the cells were harvested, which can affect membrane potential (Bot and Prodan, 2010). Control for the phase of growth should be performed in future studies of PROPS in E. coli if quantitative comparisons are desired. It is also possible that PROPS is less well trafficked in the UT strain. Poorly trafficked proteins will appear as inclusion bodies in cells, so would be readily identified upon high-resolution optical microscopy if the cells were used for optical recording. A comparison of growth rates and viable cells might also show differences in toxicity of the protein to the different strains.

The association of the emissive state with stages in the proteorhodopsin photocycle cannot be done precisely from these data, but a simplified model may be suggested based upon the observations in analogy with what is known about other proteins. Proteorhodopsin has a photocycle similar to that of bacteriorhodpsin, where a proton is moved across the membrane by means of a series of conformational changes (**Figure 6**). Retinal begins as all-trans in the ground state (G), and the Schiff base is protonated. Photoisomerization of retinal to the cis results from visible excitation and results in the L state. The Schiff base is then deprotonated to the extracellular side (M1), then becomes accessible to the cytoplasmic side (M2). It is reprotonated to form the N state, and the retinal returns to the trans state (O). The O state then returns to ground. Fluorescence can result from on-pathway states or from off-pathway states that are created by the light excitation. Key observations from the current study are: (1) the longer-lifetime state is excited efficiently at 600 nm, but not at 532 nm; (2) high laser power (3 mW) prevents the observation of this state with 532 nm emission, but enhances it with 600 nm emission; (3) depolarization and violet light exposure both enhance the fraction of molecules in this state, and the effects of the two are additive.

The observed CCCP dependence of the fluorescence might be suspected to result from changes in pH of the cells, rather than necessarily from depolarization. The original study deconvolved these effects by co-expressing pHluorin with PROPS. When cells were treated with CCCP, intracellular pH became equal to extracellular pH, leading in most cases to a change in fluorescence of pHluorin. However, PROPS fluorescence increased sharply regardless of medium pH—even when the internal and external pH were identical—suggesting that its fluorescence changes were due to membrane potential rather than pH (Kralj et al., 2011). In the current work, with external pH ∼7, the fluorescence changes in PROPS are due to both membrane potential and pH.

The photophysics of bacteriorhodopsin (Cao et al., 1993; Kamiya et al., 1997, 1999) and of an opsin-based GEVI, Arch (Maclaurin et al., 2013), have been investigated in great detail. The fluorescence of Arch was initially believed to result from the ground state, but instead was found to be the result of a state formed from the N state by exposure to yellow light, called the Q state. A similar Q state has also been observed in bR due to the sequential absorption of 3 photons (Ohtani et al., 1992, 1995). While green light is sufficient to create Q from N, orange light is required for excitation of Q.

It is likely that the fluorescence here results from Q. Increased laser power increases PROPS quantum yield with orange light excitation but not with green. Since Q is a 3-photon process, its formation should increase at higher laser power, and it is excited with orange rather than green light. It is unlikely that the voltage-sensitive PROPS fluorescence arises from the O state or the ground state, since in these states the Schiff base is extracellular, and the voltage-sensitive state it is cytoplasmic. Further studies using simultaneous red and violet light exposure, and perhaps transient absorption, will be needed to elucidate the precise identity of the emitting states. The utility of the present work is in identifying the lifetime of the fluorescent state, which is comparable to that seen with the GFP family (Pepperkok et al., 1999). These results illustrate the utility of nanosecond-scale fluorescence measurements, and suggest experiments to screen for mutants that show better quantum yields than the currently available constructs. Complementing steady-state brightness results, time-resolved measurements

green light and hν2 represents violet light.

distinguish between mutations that increase lifetime and those which increase the fraction of emission from the longer-lifetime state. Such measurements may help identify mutations that create novel states with substantially increased lifetime, as has been done with cyan fluorescent protein (Goedhart et al., 2010).

These results also suggest approaches to the use of GEVIs in fluorescence lifetime imaging microscopy (FLIM). FLIM is a valuable technique for quantitative fluorescence microscopy because it does not depend upon fluorophore concentration. The existing construct has too low of a quantum yield to be seen with commercial FLIM (data not shown), but custom microscopes and/or improved mutants may make this a valuable approach to voltage sensing. The current work shows that the appearance of the 2–3 ns lifetime is a signal for protonation of the Schiff base, an ideal lifetime range for FLIM, as the signal is able to decay between laser pulses. The small magnitude of the voltagedependent change seen in the current studies is almost certainly due to the presence of many inhomogeneous proteins in these bulk samples, so it is likely that changes of single molecules will be resolvable on a system that is set up for imaging PROPS.

#### Conclusion

The fluorescence emission of PROPS results from a variety of states, of which the voltage dependent state has a lifetime of ∼2– 3 ns and probably corresponds to the orange-excited Q state. Fluorescence lifetime measurements can provide insight into the photophysics of GEVIs, which can lead to improved sensors and to the use of such sensors in applications such as FLIM.

#### Acknowledgments

This work was performed in the laboratory of Stephen Bradforth at USC. JN's salary support was provided by the Canada Research Chairs. Thanks to J. Lanyi for useful discussions.

#### References


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2015 Nadeau. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Neuroprotective therapies in glaucoma: II. Genetic nanotechnology tools

Nafiseh Nafissi and Marianna Foldvari\*

*School of Pharmacy and Waterloo Institute of Nanotechnology, University of Waterloo, Waterloo, ON, Canada*

Neurotrophic factor genome engineering could have many potential applications not only in the deeper understanding of neurodegenerative disorders but also in improved therapeutics. The fields of nanomedicine, regenerative medicine, and gene/cell-based therapy have been revolutionized by the development of safer and efficient non-viral technologies for gene delivery and genome editing with modern techniques for insertion of the neurotrophic factors into clinically relevant cells for a more sustained pharmaceutical effect. It has been suggested that the long-term expression of neurotrophic factors is the ultimate approach to prevent and/or treat neurodegenerative disorders such as glaucoma in patients who do not respond to available treatments or are at the progressive stage of the disease. Recent preclinical research suggests that novel neuroprotective gene and cell therapeutics could be promising approaches for both non-invasive neuroprotection and regenerative functions in the eye. Several progenitor and retinal cell types have been investigated as potential candidates for glaucoma neurotrophin therapy either as targets for gene therapy, options for cell replacement therapy, or as vehicles for gene delivery. Therefore, in parallel with deeper understanding of the specific protective effects of different neurotrophic factors and the potential therapeutic cell candidates for glaucoma neuroprotection, the development of non-invasive and highly specific gene delivery methods with safe and effective technologies to modify cell candidates for life-long neuroprotection in the eye is essential before investing in this field.

Keywords: neurodegenerative disease, glaucoma, neurotrophic factor therapy, gene and cell therapy, nanotechnology, minicircle DNA vectors, transposon systems, ZFN- TALEN-CRISPR/Cas

## Introduction

Glaucoma is one of the most common causes of blindness in the world with over 70 million people (79 million by 2020), including 400,000 Canadians, affected by this disease (Foster and Resnikoff, 2005). While preventable with proper diagnosis and continual treatment, a patient's vision cannot be recovered once it has been affected. The underlying mechanisms associated with glaucoma progression are still under investigation, but it is well established that the damage to retinal ganglion cells (RGCs) is mainly the result of mechanical injury resulting from increased intraocular pressure (IOP) caused by disruption of the trabecular meshwork (Margalit and Sadda, 2003).

Local vascular insufficiency at the optic nerve head can also lead to a decrease in neurotrophic factors (NFs) levels (Bessero and Clarke, 2010), which results in RGC death (Fang et al., 2010).

#### Edited by:

*Ioan Opris, Wake Forest University School of Medicine, USA*

#### Reviewed by:

*Hun-Kuk Park, Kyung Hee University, South Korea Hari S. Sharma, Uppsala University, Sweden*

#### \*Correspondence:

*Marianna Foldvari, School of Pharmacy and Waterloo Institute of Nanotechnology, University of Waterloo, 200 University Avenue West, Waterloo, ON N2L 3G1, Canada foldvari@uwaterloo.ca*

#### Specialty section:

*This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience*

Received: *29 June 2015* Accepted: *17 September 2015* Published: *14 October 2015*

#### Citation:

*Nafissi N and Foldvari M (2015) Neuroprotective therapies in glaucoma: II. Genetic nanotechnology tools. Front. Neurosci. 9:355. doi: 10.3389/fnins.2015.00355*

**33**

Currently, glaucoma management relies on pharmacological and invasive surgical treatments mainly by reducing the IOP, the most important risk factor for the progression of the visual field loss.

There is strong evidences from several research groups that repeated administration of neurotrophic factors (NTFs) such as neurotrophin-4 (NT-4), brain-derived neurotrophic factor (BDNF) (Di Polo et al., 1998), ciliary neurotrophic factor (CNTF) (Ji et al., 2004), and glial cell line-derived neurotrophic factor (GDNF) (Jiang et al., 2007) increased the survival of neurons in rodent models. Among the NTFs, BDNF appears to provide the highest level of protection by supporting both protective and regenerative functions (Danesh-Meyer, 2011). BDNF has a direct effect on RGCs and appears to correct problems with bidirectional transport of NTFs and by indirect influence on other retinal cells, helps direct damaged axons. Initial damage to axonal transport has already been attributed to a deficiency in NTFs. For instance, neurotrophin deprivation due to bidirectional axonal transport obstruction within RGCs has been shown to result in axonal damage (Iwabe et al., 2007). Retrograde flow at the optic nerve head prevents proteins made by RGCs from reaching their axonal extensions and perturbed retrograde transport of NTFs produced in the superior colliculus (SC) in the brain to reach the RGCs (Lim et al., 2010). Several studies have now identified the potential role of astrocytes, microglia and Müller cells in RGC survival within the optic nerve head region through their secretion of NTFs (Johnson and Morrison, 2009). This has been suggested by studies involving adenovirus-mediated intravitreal delivery targeted toward Müller cells (Di Polo et al., 1998) and after a single intravitreal injection of BDNF in a cat model (Weber et al., 2008). The targeting of regenerative factors to Müller glial cells can also stimulate their dedifferentiation into multipotent progenitor cells, which may differentiate into new RGCs or photoreceptor cells to replace the ones that became damaged during injury by high IOP. BDNF also appears to support axonal path finding to the brain (Benowitz and Yin, 2008, 2010).

It is now widely recognized that lowering IOP in the treatment of glaucoma is not enough. In addition to neuroprotective and neuroregenerative approaches, poly-therapeutic strategies may be the future. Combination treatments such as IOP-lowering drugs with neurotrophic factors and/or antioxidants and/or antiapoptotic agents may be necessary. The common challenge to all these therapeutic possibilities lies in the delivery and maintenance of NTF levels in the retina for a prolonged period.

Neurotrophic factor-based gene therapy may meet this challenge. It could be performed either by direct transfer of transgenes coding for NTFs into the patient or by using living cells that express NTFs persistently as vehicles to transport the NTF transgenes. In general, gene delivery and prolonged or stable expression of the intended therapeutic transgenes in target cells are the key factors to a successful gene therapy approach.

In part I of this paper we discussed the supportive effects of different NTFs in glaucoma, identified different methods of NTF transgene delivery, and discussed potential cell candidates for cell-mediated therapy. Here, we focus the review on advanced techniques for production of safer and more efficient DNA vectors as well as innovative non-viral approaches for ex vivo gene delivery/gene editing in order to provide stable and long-term expression of therapeutic genes such as NTFs in suitable candidate cells.

## RGC Rescue Therapy in Glaucoma Treatment

Exogenous supplementation of NTFs, apoptosis inhibitors and survival factors as transgenes or their recombinant protein products is a promising approach to stop or decline RGC death in progressive glaucoma (Thumann, 2012).

Interrupting the apoptosis cascade by delivering genes encoding caspase inhibitors or expressing anti-apoptotic genes such as Bcl-2, or interfering with the expression and activity of pro-apoptotic factors by siRNA technology have been successfully tested in preclinical studies of glaucoma treatment in animal models (Liu et al., 2009; Thumann, 2012).

Additionally, increasing NTF levels by gene delivery has been widely investigated in vitro and in vivo. NTFs are small proteins that are secreted by the central and peripheral nervous system and are critical in their own development and maintenance (Lim et al., 2010). NTFs are classified into several groups and among them, the nerve growth factor (NGF) family members such as GDNF, BDNF, NTs, and CNTF have been the subject of more detailed studies for gene therapy in glaucoma(Johnson et al., 2011).

NTF supply is important for RGC survival or regeneration during development, and extensive experimental strategies have been tested to supply exogenous NTF to protect and promote survival of injured RGCs in glaucoma.

Although direct gene therapy by delivering the exogenous transgenes encoding for NTFs using different viral and nonviral carriers are particularly attractive, in certain cases such as glaucoma, life-long neuroprotective support through exogenous NTF therapy is essential. Therefore, cell therapy would be a more sustainable approach via delivering NTFs by living cells and direct replacement of growth factors and NTFs by cells that are genetically modified ex vivo. Application of genetically modified cells as gene delivery vehicles has certain advantages such as relative simplicity of manipulation and evaluation of cells in vitro compared to in vivo gene modification. Furthermore, some of these modified cells continue to divide in vitro under certain culture conditions, which facilitates expansion of these cells for further investigations. Finally, some of these engineered cells show a tendency to localize into particular tissues.

Recent studies showed that several stem and progenitor cells expressing and secreting the NTFs provide neuroprotective support when transplanted into animal models of glaucoma and other retinal diseases (Johnson et al., 2011). In this paper we focus on advanced non-viral nanotechnology tools for genetic modification of candidate cells aiming to achieve long-term expression of NTFs therapeutics.

## New Generation of DNA Therapeutics

The necessity to generate safe and efficient DNA vectors for transgene delivery via a variety of non-viral approaches has spurred many different proposals. Among them bacterial sequence free DNA vectors in two forms such as supercoiled circular covalently closed and linear covalently closed DNA, termed as "minicircle" and "ministring," respectively, are considered the most promising (Darquet et al., 1997, 1999; Chen et al., 2003; Nafissi and Slavcev, 2012; Nafissi et al., 2014; Slavcev et al., 2014; Slavcev and Nafissi, 2014). Replication and largescale production of plasmid DNA vectors is dependent on the prokaryotic backbone and specific selection markers to isolate and propagate plasmid-containing bacterial strains after bacterial transformation. However, these sequences are undesirable in clinical applications because of the following reasons: (A) the bacterial sequences are recognized as invading factors and trigger host innate immune response that leads to systematic removal of the vector (Klinman et al., 1996; Mitsui et al., 2009); (B) the horizontal transfer (importing genes from environment or from other bacteria) of antibiotic resistant genes from plasmid DNA to normal microbial flora is a risk factor for the generation of antibiotic resistant flora (Chen et al., 2008); (C) residual selection markers in the final plasmid product, due to unsuccessful removal, can cause allergic reaction and hypersensitivity in sensitive individuals after gene delivery (Cavagnaro, 2013); and (D) the bacterial sequences are reported as the main cause for heterochromatindependent silencing of the intended transgene (Chen et al., 2003; Mayrhofer et al., 2009). In contrast, the new generation of DNA vectors that are bacterial sequence free offer higher and more persistent expression, generally at levels 100–1000 times greater than their standard plasmid precursor (Kay, 2011). Previously, purification of miniDNA vectors from bacterial extracts was labor-intensive, time-consuming, and a multi-step process that needed digestion of the bacterial backbone by a restriction enzyme (Schakowski et al., 2001, 2007) followed by purification of miniDNA vector and removal of digested sequences by cesium chloride ultracentrifugation (Bigger, 2001; Chen et al., 2005). However, employing prokaryotic-derived site-specific recombination systems, mainly from bacterial viruses (phage) such as λ integrase (Int), P1-derived, Cre, phiC31Int, N15 derived TelN, and PY54-derived Tel, dramatically facilitated the production and purification of miniDNA vectors (Nafissi and Slavcev, 2014). These systems show limitations that have been improved over the time (**Table 1**). In general, the first step in generating bacterial sequence-depleted DNA vectors with phagederived enzymes is to engineer a bacterial cell (mainly E.coli) that express these enzymes, insert the recognition sequence of these enzymes in to the plasmid DNA vector closely upstream and downstream of the therapeutic transgene expression cassette, and transfer the plasmid into engineered E.coli cell (Nafissi and Slavcev, 2012). Consequently, the in vivo intramolecular recombination at the recognition sites results in generation of two well-characterized molecules from the single parent plasmid:(i) the miniDNA vector (circular or linear covalently closed) comprising the therapeutic transgene expression cassette, and (ii) the miniplasmid containing the bacterial backbone elements (**Figure 1**). Combining the endonuclease I-SceI together with its recognition site in the plasmid backbone allows simultaneous digestion of the bacterial backbone into small pieces and production of purified miniDNA vector. For instance, further purification of the minicircles by affinitybased chromatography allows the isolation of highly pure and pharmaceutical-grade minicircles by this technique (Rodríguez, 2004; Thyagarajan et al., 2008). For the first time, Darquet et al. (1997, 1999) showed that "minicircle" DNA confers much higher transgene expression levels in vitro and in vivo, respectively, compared to the parental plasmid precursors or other conventional control plasmids encoding the same transgene. This result was further confirmed, and showed an even more significant increase in the expression of the encoded transgene, when the same amount (weight-to-weight basis) of minicircle DNA and parental plasmids were delivered (Darquet


et al., 1997, 1999; Chen et al., 2003, 2005; Vaysse et al., 2006; Jia et al., 2010). After developing the convenient production systems mentioned above, miniDNA vectors have been extensively studied by different research groups by comparative studies with the parent plasmids (**Table 2**). The following examples reflect the broad applicability of the new generation of DNA vectors from gene therapy to more recently in stem cell research and regenerative medicine considering dynamic aspects from formulations and optimization of miniDNA vectors complexes with synthetic vectors. Neural stem cells (NSC) that are very difficult to transfect were successfully transfected by minicircle DNA vector by microporation and showed higher transgene expression and NSC survival when compared to their plasmid counterpart (Madeira et al., 2013).

Minicircle DNA vectors have been also used as integrating vectors for recombinase-mediated cassette exchange (RMCE) (Jakobsen et al., 2013) or as a sleeping beauty transposition system (Sharma et al., 2013), and minicircles significantly improved the integration efficacy and gene modification capacity, respectively, compared to the standard plasmids. In a different study, same molar ratio and same copy numbers of the minicircle and plasmid DNA vectors were complexed with lipid-based nanoparticles and were transfected into a variety of human cell lines and in animal models. This study confirmed previous results and showed that transgene delivery by minicircle DNA vectors significantly improves transgene expression levels both in vitro and in vivo. This group also concluded that improved minicircle delivery in vivo and higher transgene expression levels are due to the compact size of minicircle-nanoparticles, better cellular uptake and cell entry, higher intercellular minicircle copy numbers in transfected cells, better intracellular trafficking toward nuclei, and better nuclear uptake. These results also showed higher mRNA transcription levels in minicircle-mediated transgene delivery compared to parent plasmid-mediated transgene delivery (Kobelt et al., 2013). They concluded that the euchromatin structure and more accessibility of bacterial-sequence-free DNA vectors to transcription machinery of host cell is a key reason for higher expression level of transgene. Even further modification of miniDNA vectors to combat intracellular barriers such as nuclear membrane would dramatically improve transgene delivery, especially in slow or non-dividing cells. For example it was shown that the addition of nuclear-targeting sequences (DTS), such as the SV40 enhancer, or karyopherins (Miller and Dean, 2009) to the DNA sequence in parallel with removing undesired bacterial sequences improved transfection efficacy and expression levels of the transgene and its protein product (Nafissi et al., 2014).

New generation of DNA vectors represent a promising alternative to conventional plasmids in terms of biosafety, bioand immuno-compatibility, improved gene transfer, potential bioavailability and cytoplasmic diffusion due to their smaller size (Nafissi et al., 2014), and low immunogenicity due to the



\**Bacterial sequence depleted.*

\*\**Depends to the size.*

\*\*\**Depends to the size of therapeutic transgene.*

lack of bacterial sequences and immunogenic motifs (Chen et al., 2003; Vaysse et al., 2006). Although miniDNA vectors provide more sustained expression of the transgene relative to standard plasmid DNA vector, in more rapidly dividing cells and tissues with higher regenerative capacity, most DNA vectors become diluted after each mitosis and eventually disappear. Therefore, permanent introduction of a therapeutic transgene followed by its sustained expression is more desirable where life-long pharmaceutical effect is needed. In particular, in the case of gene or cell mediated therapy of glaucoma and other neurodegenerative disorders that require life-long and stable expression of NTFs, tightly controlled integrating DNA vectors and strategic genetic engineering systems are required (Jandial et al., 2008; Blurton-Jones et al., 2014). Therefore, better and safer techniques of genome editing could open a new avenue to investigate and treat neurodegenerative disorders more effectively.

Permanent expression of therapeutic genes is generally carried out by three different approaches: (i) random integration (illegitimate insertion); (ii) homologous recombination (HR); or (iii) site-specific insertion of the transgene into the chromosome of an anticipated cell. Random integration is mostly carried out by viral vectors to achieve insertion of an exogenous transgene into the host human cell's genome. However, lack of control over the site and position of integration would result in undesirable side effects such as unpredictable expression level or silencing of the integrated transgene, potential mutagenesis of neighboring genes, activating oncogenes/ deactivating tumor suppressor genes, and eventually cancer. Therefore, random integration by viral vectors is not a safe method for permanent and sustained expression of therapeutic transgene in ex vivo cellmediated therapies. HR and site-specific insertion are indeed the methods of choice for the optimal control over the site of transgene integration, the number of copies inserted per target cell, the expression level of therapeutic transgene, and to reduce the risk of oncogenesis. Safe and efficient sitespecific insertion is carried out through bacteriophage integrasemediated and transposon systems, and HR through recent and efficient techniques involving Zinc Finger Nucleases (ZFN), Transcription Activator-Like Effector Nucleases (TALENs), and Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR).

## Site-specific Recombination for Cell Therapy

The interest in precise gene modification is growing dramatically owing to recent advances in targeted genome engineering. Targeted genetic engineering techniques allow very specific modifications such as gene insertion, deletion, and replacement. Permanent modification of cells has been extensively used in regenerative medicine as a new clinical tool.

Some examples are engineering glial progenitor cells to permanently express adhesion molecules to increase homing of these cells into the brain (Gorelik et al., 2012), or enhancing the therapeutic effects of stem cells by permanent insertion of NTF genes (Crigler et al., 2006; Janowski and Date, 2011). Here we briefly review some of the most popular and safe non-viral methods for genetic engineering of clinically relevant cells and insertion of therapeutic transgenes such as NTF encoding genes into anticipated cells by transposon systems and bacteriophagemediated integrase systems.

### Transposon Systems

Non-viral transposon systems have been widely used to generate genetically modified and clinically relevant human cells including but not limited to induced pluripotent stem (iPSC), embryonic stem cells (ESC), and mesenchymal and hematopoietic stem cells (Saha et al., 2015). "Sleeping Beauty transposon system (SBTS)" and "piggyback (PB) transposon system" are the two transposon systems that have been successfully used as nonviral gene delivery carriers for gene modification and generation of clinical grade human cells for gene and cell therapy. These systems have lower cost for design, construction, and production of pharmaceutical grade products on a large scale as well as stimulating low levels of innate immunity and the capacity to co-deliver multiple therapeutic transgenes compared to viral integrating systems.

## Sleeping Beauty Transposon System (SBTS)

"Sleeping Beauty transposon system (SBTS)" is a non-viral plasmid-based integration system that combines the integration benefits of viral vectors and ease of production and manipulation of naked DNA vectors and ease of delivery by nanoparticles. The SBTS consists of (A) a plasmid transposon carrying the therapeutic transgene expression cassette flanked by the terminal Inverted Repeats (IR) that contain binding sites for the transposase enzyme; and (B) a source of transposase enzyme, which is a 360-amino acid DNA-binding protein with a transposon binding domain, a nuclear localization signal (NLS), and a catalytic domain.

The SBTS systems mostly target the TA rich sites of the hosting genome to insert the intended therapeutic transgene expression cargo. The gene encoding for the transposase enzyme can be assembled either on the same plasmid that express the therapeutic transgene (sis) or can be located at a different plasmid (trans). In case of trans expression, the two DNA vectors (one carries the transposase enzyme encoding gene and the other carries the therapeutic transgene expression cassette) could be packed within the same nanoparticle carrier and co-transferred into the intended host cell. Few hours after delivery, the transposon system would reach the nucleus of the host cell, the transposase enzyme would be expressed in the cytoplasm and the NLS that has been added to its sequence would direct import of the transposase into the nucleus, cut the IR sequences, and facilitate relocation and insertion of the therapeutic transgene expression cargo from DNA vector into the host genome (Aronovich et al., 2011) (**Figure 2**). Similar to other technologies, SBTS system needs to be improved before entering into clinical trials. For example, in order to achieve pharmacologically relevant expression level of the therapeutic gene and just sufficient transposase expression to carry out the cut-and-paste reaction in difficult-to-transfect target cells, better delivery techniques are required. In addition, precise understanding of the target cell genome would assure insertion of the therapeutic gene into the TA sites and avoids the side effects of insertion into undesired loci.

Transposase expression level and duration of expression is critical in transposon systems because very high expression level of the transposase would cause insertion of the transgene into many different locations of the hosting genome that are rich in TA. In addition, long-term expression of the transposase is also not desirable because it might cause the excision of the integrated transgene out of genome and reverse the effect. Therefore, a good understanding of the host's genetic background, type of tissue to be targeted, and therapeutic gene vs. transposase expression level is necessary to improve transposition efficacy (Aronovich et al., 2011). The CMV promoter would provide a high but

FIGURE 2 | Schematic representation of the non-viral transposon-mediated site-specific gene therapy. The non-viral transposon is used as a bi-component DNA vector system for delivery and insertion of therapeutic transgenes from a DNA vector into the target cell chromosome. This system is composed of either one plasmid (cis) or two plasmids (trans) contains the therapeutic gene of interest (GOI) (green arrow) and its expression cargo such as promoter (orange box) flanked between the two transposon Inverted Terminal Repeats (ITR) (red circles), and the transposase enzyme (blue arrow) expression cargo. (1) After delivery into the intended target cell and transportation into the nucleus, through a "cut and paste" transposase activity, (2) the transposon is excised from the transposon donor DNA vector, and (3) the therapeutic transgene is integrated at a chromosomal site.

short-term expression of the SB transposase enzyme in many human cells and therefore would inhibit the re-mobilization of the integrated therapeutic gene after a longer time (Hackett et al., 2011). On the other hand, tissue-specific mammalian promoters derive an alternative promoter for the therapeutic transgene for sustained expression of the transgene in the right site (Kachi et al., 2006). The SB100X "plasmid-based integrating" transposon system in combination with electroporation technology (Mátés et al., 2009) was recently applied to permanently knock-in the anti-angiogenic and neuroprotective factor PEDF into the ARPE-19 human retinal pigment epithelial cells under the control of either CMV or cell specific CAGGS promoters. In this study, the PEDF positive cells were successfully transplanted into the retina to treat age-related macular degeneration (AMD) (Johnen et al., 2012).

As the proof of principle for licensed SBTS technology in clinical trials it is worth to point out few examples, even though these trials are not specifically toward the treatment of glaucoma. (1) In a in phase II clinical trial of cancer immunotherapy research, patients-derived T cells were genetically modified by SBTS system to express a chimeric antigen receptor (CAR) and were injected back to patients with B-cell malignancies in order to redirect the specificity of human T cells in these patients (Hackett et al., 2011). (2) SBTS technology provided life-long expression of Fanc-C and Fanc-A genes to treat Fanconi anemia, or Factor VIII gene to treat Hemophilia A (Nienhuis, 2008). (3) SBTS-mediated life-long expression of IL13-HSVTK was achieved to treat brain tumors (Di Matteo et al., 2012).

In the field of stem cell research, plasmid-based SB systems have been used in combination with nanoparticles for the expression of therapeutic genes. For example, four human iPSC lines were successfully generated and characterized from fetal fibroblasts using SB-based DNA vector that were combined and delivered by nanocarriers (Davis et al., 2013). Also, rat mesenchymal stem cells (rMSCs) were efficiently transfected by a nanoparticle called "liposome protamine/DNA lipoplex (LPD)." LPD was electrostatically assembled from cationic liposomes and an anionic complex of protamine, SB transposon system, and NLS targeting peptides in order to enhance cell specific targeting and nuclear uptake. This complex dramatically improved transposon-mediated gene insertion in MSCs (Ma et al., 2013). As such, the non-viral DNA-SB-based therapeutics in combination with engineered cell specific nanoparticles serve as a promising and long-lasting way to insert therapeutic genes into the chromosomes of host cells without using a viral integration systems and their consequent undesirable side effects (Aronovich et al., 2011).

## PiggyBac Transposon System

The PiggyBac (PB) transposon derived originally from Lepidopterans is composed of two identical short inverted terminal repeats (700 bp) and a transposase-encoding sequence (coding a 594 amino-acid transposase). The PB transposase catalyzes the transposition of the therapeutic gene expression cargo that is flanked by the inverted terminal repeats carrying by a plasmid and insertion of therapeutic transgene into TTAA rich sites of the host genome. The short IR are the key elements of the PB transposition system (**Figure 2**). One of the benefits of PBtransposon system over SBTS is the capacity of this system to deliver larger transgenes, complex genes, or multiple therapeutic genes together with their regulatory regions. However, low transfection efficiency of large plasmids is still the major limitation of PBtransposon system (Kim et al., 2011). In terms of safety, one of the concerns about using PB systems in clinical applications in human cells is the distribution of over 2000 PB-like elements in the human genome which raises the risk of genome rearrangement and the potential for remobilization of the integrated transgene by endogenous transposase expression. However, recently the potential undesired PB-mediated genomic rearrangements have been investigated to determine the safety of PB systems in clinical applications. No genomic rearrangement and re-mobilization of the integrated transgene were observed, but it was also suggested that long-term evaluation of the safety of transposase systems in animal models is needed to reach a final conclusion (Saha et al., 2015). Therefore, despite these observations, insertional mutagenesis in unwanted sites or rearrangement of the neighboring genes and gene remobilization after integration of the therapeutic gene into host genome still remains the major limitation of PB transposon system in the precise modification of clinically relevant stem/progenitor cells (Li et al., 2013). The PB system was efficiently applied to reprogram murine and human embryonic fibroblasts and to generate pluripotent stem cells. As such, four transcription factors (c-Myc, Klf4, Oct4, and Sox2) were inserted into the genome of embryonic fibroblast cells to reprogram them and to generate pluripotent stem cells. Furthermore, applying the natural tendency of PB transposase to excise the inserted transposon, transcription factors were removed from well-established iPSC cell lines (Woltjen, 2009) post reprogramming. The PB system was also successfully applied in field of cancer immunotherapy by isolating T cells from patients with malignancies and generating genetically modified tumor-antigen-specific T cells. These cells were further injected to patients and the anti-tumor activity of modified T cells with one or multiple insertions was evaluated (Nakazawa et al., 2009). Several studies were carried out using the PB transposon plasmid DNA vectors combined with polymer- or lipid-based nanoparticles to enhance transgene delivery and insertion of the intended gene through PB transposase activity in mammalian cells (Palavesam et al., 2013; Chakraborty et al., 2014).

The miniPB system was recently generated and applied to insert the GOI into the host genome with the same integration efficiency but using a much smaller plasmid vector with the advantage of enhanced transfection efficiency in vitro and in vivo (Solodushko et al., 2014). The mini PB plasmid is a singleplasmid system (cis) carrying the transposase gene and very short inverted terminal repeats (35 bp) flanking the therapeutic transgene expression cassette (Yusa et al., 2009; Li et al., 2011).

Compared to the PB transposon system, the SB100X shows superior efficacy in most human cells including hematopoietic cells with lower safety risk of random integration and insertional mutagenesis (Aronovich et al., 2011). However, for multiple therapeutic gene insertion, the PB system remains the method of choice (Saha et al., 2015).

#### 8C31 Integrase

Site-specific DNA insertion systems are derived from prokaryotes and unicellular yeasts and a number of them have been exploited to facilitate efficient DNA exchange in human cells (Nafissi and Slavcev, 2014). In fact, all site-specific integration systems typically mediate efficient "cut-and-paste" type of DNA exchange between recognition sites in the range of 30–40 bp or longer. Invading bacterial viruses (phage) use this system to integrate their genome into their bacterial host chromosomes by a reaction catalyzed by integrase enzymes at short sequences, termed as the "phage attachment site (attP)" and "bacterial attachment site (attB)" (Nafissi and Slavcev, 2014). Calos and her colleagues showed that sequences very similar to wild type attP and attB sites are available in the human genome and called them as "pseudo attP" and "pseudo attB," sites because an exogenous gene can be integrated specifically into these sites in presence of a phage integrase (Chalberg et al., 2006). The actinophage 8C31 integrase, discovered in the 1990s, efficiently catalyzes a site-specific genomic integration between two DNA recognition (Groth and Calos, 2004) sequences: attB containing DNA vector and a pseudo attP site within the genome of an anticipated human cell, leading to permanent transgene expression (Chavez et al., 2011). Using this system, the insertion of a foreign DNA such as a therapeutic transgene is characterized by the following events: (1) recombination occurs at a specific site on the interacting DNA molecules: DNA vector and hosting cell genomic DNA; (2) expression and synthesis of the recombinase enzyme in the host cytoplasm using host protein synthesis machinery; (3) re-location of the integrase to the host cell nuclei using nuclear localization signals; (4) strand exchange occurs at small regions of DNA homology within the recognition sites; (5) pairing of the interacting insertion sites followed by strand exchange results in structural intermediates; and (6) resolution of intermediates followed by strand migration (Groth et al., 2000) (**Figure 3**). In the human genome "pseudo attP" sites show 40% similarity to wild-type attP sequence. In comparison with the "TA" and "TTAA" rich sequences that are the target sites for the SB and PB transposon systems, respectively, pseudo attP are mostly located in transcriptionally active, euchoromatin sites, and exons with less frequency of introns. 8C31 integrase provides one-copy integration of the exogenous transgene per cell without disrupting the endogenous genes. Therefore, integrase-mediated site-specific gene insertion of transcription factors offers an alternate method of producing iPS cells without interfering with endogenous gene functions (Lan et al., 2012). 8C31 integrase-mediated insertion of exogenous gene occurs at 10-fold higher rates than random integration and often

FIGURE 3 | Schematic representation of the non-viral integrase-mediated site-specific gene therapy. This diagram shows how 8C31 integrase system benefits in gene therapy. (1) One DNA vector that carries the desired therapeutic transgene expression cargo (promoter, transgene, reporter gene, polyA signal) and also the *attB* site that is recognized by 8C31 integrase enzyme and one DNA vector that carries the integrase encoding gene are co-delivered to human target cells. Both DNA vectors enter nucleus, the integrase is produced in the cytoplasm and redirected to the nucleus. (2) 8C31 integrase enzyme binds as a dimer to the *attB* site on the therapeutic DNA vector and bounds at pseudo *attP* sites present in the human genome. (3) A cut-and-paste recombination reaction occurs at the site of integration, which results in the insertion of the therapeutic transgene expression cargo into the chromosome of human cells at the pseudo *attP* site. This process provides long-term expression of the transgene and production of therapeutic protein in the desired target cells.

provides higher expression levels than those inserted randomly into the human genome that possibly causes either silencing, undesired side effects of over expression of the therapeutic gene, or oncogenesis/genotoxicity (Nafissi and Slavcev, 2014). In addition, bacteriophage 8C31integrase catalyzes unidirectional integration of therapeutic gene into only pseudo attP sites in the human genome and eliminates the risk of excision of inserted transgene (Chalberg et al., 2006),which is the most important advantage of this system over transposon systems. Transgenic animals generated by 8C31integrase-mediated sitespecific insertion never showed any cancer development (Calos, 2006), or the human ESC that been genetically modified by 8C31 integrase retained their ability to differentiate normally into all three germ layers (Thyagarajan et al., 2008). In animal models, for example, hereditary tyrosinemia type I and muscular dystrophy was treated by 8C31 integrase–mediated integration of fumarylacetoacetate hydrolase and dystrophin, respectively. In these studies, plasmid DNA carrying the attB site and the therapeutic transgene were co-delivered with plasmid DNA expressing integrase by hydrodynamic tail vein injection (Held et al., 2005) or intramuscular injection followed by electroporation (Bertoni et al., 2006).

A comprehensive study confirmed the biosafety of 8C31 integrase-mediated integration of therapeutic transgene into human umbilical cord lining epithelial cells (CLECs) as source for several stem-cell like cell types (Sivalingam et al., 2010).

Any plasmid DNA-based integration system can be combined with synthetic carriers to form nanoparticles that are very efficient in transfecting human cells and transferring the DNA cargo into the nuclei. Different studies were carried out to improve this system. Very recently, Oliveira and colleagues combined polymer-based nanoparticles with integrase technology to create 8C31-chitosan-mediated gene delivery and integration system. They have characterized the nanoparticles and significantly improved the transfection and integration efficacy in human embryonic kidney cells. They achieved higher expression levels of small and large integrated transgenes for over 10 weeks post transfection (Oliveira et al., 2015). Ease of transfection into mammalian cells using reliable synthetic formulations and variety of nanoparticles makes the DNAbased integrating systems superior compared to their viral counterparts, in particular in terms of immunocompatibility and safety with genotoxicity perspective.

## Homologous Recombination for Cell Therapy

Unlike viral vectors that usually integrate into the host chromosome in an uncontrolled fashion, one of the most policed methods of genome modification is to insert the therapeutic transgene into the recognized and desired location by homologous recombination (**Figure 4**). HR was defined as critical process involved in genetic diversity and repair of DNA double-strand breaks (DSBs) in pro- and eukaryotic cells (Kakarougkas and Jeggo, 2014). The process of HR is very important in maintaining the chromosome integrity and to protect and recover any open ended DNA caused by environmental assaults or invading DNA. This process involves the alignment of similar DNA sequences to form a mobile junction between four strands of DNA, termed "Holiday junctions," which are highly conserved in both prokaryotic and eukaryotic cells. HR needs energy and cofactors from host cells (Nafissi and Slavcev, 2014) and the presence of DNA DSB in mammalian cells is the essential step to stimulate HR (West, 2003). The frequency of integration by HR appears to be generally low, about 10−<sup>6</sup> , for most mammalian cells, but it was reported that this can be increased up to 100 fold by making DSBs at the target site by restriction enzymatic reactions such as CRISPRs, TALENs, and ZFNs. Some of the drawbacks of HR system is the general low frequency of HR in mammalian cells, providing the homology arms in the donor DNA vector carrying therapeutic gene, and the resources of enzymes to generate DSB in hosting target sequences (Pan et al., 2013; Yin et al., 2014b). Their sequence-specific binding modules can recognize unique nucleotides in the genome along with the fused endonuclease such as FolK1, Cas9, or ZFN to specifically induce a double-stranded break (DSB) in the chromosome. All endonuclease systems introduce DSBs in the target DNA and lead to recruitment of the cellular repair machinery, which can drive HR with dramatically higher frequency at the cleavage site in the presence of complementary sequence arms that are cloned into HR donor plasmid DNA vectors. These systems have been applied in repair of damaged DNA, gene disruption, gene insertion, gene correction and point mutagenesis, and chromosomal rearrangements (Doudna and Charpentier, 2014; Kim and Kim, 2014). Advanced genome engineering tools are now commercially available. For example, the TALENs, ZFNs, and CRISPR technologies are available from Addgene, Sigma-Aldrich, and Life Technologies.

## Zinc Finger Nucleases

As it was mentioned, the process of HR in human cells is indeed too inefficient to lead to a reliable therapeutic result and creating DNA breaks is the key step to increase efficacy of this process. Therefore, ZFNs have been widely used to create the genespecific DNA breaksand enhanced the HR efficacy to several orders of magnitude (Porteus, 2006, 2008). Zinc Finger DNA Binding Proteins (ZFPs) are a class of DNA binding proteins that naturally exist in prokaryotes. Through modification of the DNA recognition and binding function of ZFPs, it is possible to direct ZFPs to a target sequence of DNA, which enables the site-specific localization of the ZFPs in a designated sequence. ZFPs have been developed to either regulate gene expression by a turning on and off mechanism, or correct genes using endonuclease enzymes (Davis and Stokoe, 2010). ZFNs are artificial proteins composed of a DNA-binding protein and a nuclease protein that is fused to the first protein. Two ZFNs protein domains are required to bind to DNA, to dimerize and activate the nuclease domain, and to create a DSB (Porteus and Baltimore, 2003). A ZFN subunit encompasses three to six zinc-fingers arranged in a tandem repeat and a catalytic domain of an endonuclease enzyme, like FolK1, in such a way that a short linker connects

excision activity on the *loxP* sites through "Cre/LoxP recombination systems."

the two domains. At the ZFN target site located at the host genome, the two ZFN subunits are dimerized and the nuclease is activated and cuts the DNA. Precise ZFN-mediated gene targeting involves several steps as follows: (1) identify the full ZFN-binding site within the target GOI; (2) design a pair of ZFNs; (3) test the ZFN pair for activity; (4) identify a targeting construct to create the desired genomic modification; (5) co-transfer the ZFNs and the targeting vector to create the breaks and to insert the therapeutic transgene at the target site, respectively (Jamieson et al., 2003). ZFN derived HR was applied in different clinically relevant studies. In the field of neuroscience, engineered ZF protein transcription factors (ZFP TFs) has been used to improve expression of endogenous glial cell-derived neurotrophic factor (GDNF) in a rat model of Parkinson disease and resulted in improved functional neuroprotection in the brain (Laganiere et al., 2010). Also, Lipofectamine-mediated transfection into human embryonic kidney cells indicated that ZFNs can be used to specifically target the ROSA26 safe site of the chromosome in order to effectively establish new cell lines with predictable expression level of the desired therapeutic gene (Perez-Pinera et al., 2012). In a different study, lipid-based nanoparticles were applied to deliver HR donor plasmids carrying the homologous sequence to hepatitis B virus (HBV) and specific ZFNs into cells that have been infected by HBV virus. This study confirmed the pharmaceutical application of ZFNs toward specific cleaving the invading virus genome in infected human cells as a new therapeutic approach (Cradick et al., 2010).

Like any other integrating system, application of ZFNs are limited to the very precise recognition of the target site and the homology sequence, precise design and construction of the appropriate ZFN, risk of non-specific and random binding of the ZFNs into the undesired sites with similar but not perfect homology and consequent integration of therapeutic transgene into unwanted locations, genotoxicity, and finally, the risk of leaving open ended double strand breaks and its consequent risks such as mutation, genome rearrangement, or cell death (Wirt and Porteus, 2012). Several investigations carried on to improve the specificity of ZFNs in clinically relevant cells. Some of these studies applied lipid-based nanoparticles to target cells and deliver donor DNA vectors to the target site. For example, a modular assembly technique was described to find the most effective ZFNs after delivery (Kim et al., 2009), or the "OPEN (Oligomerized Pool ENgineering)" technology was developed to construct over 30 highly active and effective ZFN pairs to target different chromosomal regions in human cells by lipofection (Maeder et al., 2008; Fu et al., 2009).

## TALEN

The TALENs are pre-designed restriction enzymes that are capable of specific cutting of a desired DNA sequence. However, similar to ZF or any other nuclease-based insertion method, customer/project-based design of these restriction enzymes is necessary and needs precise understanding and computer-based analysis of the target cell and insertion site because the off-target cleavage of potential similar sequences causes too much damage to the host genome and usually cause cell death. Very recently, applying the "Golden Gate TALEN and TAL Effector Kit" from Addgene, human embryonic kidney and human iPSC cells were genetically modified efficiently with TALEN system (Cermak, 2011). The "Golden-Gate assembly system" was developed using over 400 different plasmid DNA vector carrying TAL effector and Folk1 encoding genes in order to make a TALEN library. These plasmids complexed with Lipofectamine or polyethylenimine were delivered to several human cells to monitor their activity (Kim et al., 2013). However, similar to ZFN, TAL effector proteins require protein engineering to bind to a desired DNA sequence. Other improved system has been introduced by System Bioscience Company, which facilitates the precise insertion of the gene of interest into the AAVS1 "safe site" of the human genome using the TALEN endonuclease technology. The AAVS1 site, located on chromosome 19, is a natural integration hotspot of Adeno-associated virus (AAV) type 2 (Hüser and Heilbronn, 2003). It was shown that AAVS1 site can be considered as a safe site for nuclease mediated HR and insertion of the therapeutic genes for stable and long-term expression (Sadelain et al., 2012). Recently, ZFNs and TALEN targeted to AAVS1 were also used to insert therapeutic genes into this location (**Figure 5**), such as insertion of the transgene encoding gp91 in iPS cells derived from patients with X-linked chronic granulomatous disease and restore the gp91 function. These patients suffer from defective gp91 expression and are not able to produce reactive oxygen species (Zou et al., 2011). Insertion of exogenous gp91 into the AAVS1 site is a novel pharmaceutical approach to cure this disease. In a different study, the activity of several exogenous and endogenous mammalian promoters on the expression levels of the transgenes were investigated applying plasmid-based TALEN strategy combined with nanoparticlemediated transfection to insert promoters for more transient expression and check expression of the downstream genes over time (Smith, 2008). More studies on the effect of epigenetic regulations on transgene expression level, effect of promoter silencing, and transcription factors have been done using TALEN technology. For example, TALEN plasmid and transgene plasmids were delivered into human cells using Lipofectamine and polyethyleneimine nanoparticles to investigate selective activation of the endogenous oct4 pluripotency gene via TALEN technology by activating reprogramming genes and inhibiting epigenetic factors to generate induced pluripotent stem cells (iPSCs) from somatic cells (Bultmann et al., 2012).

## CRISPR

CRISPR is the name of DNA sites in bacteria containing short sequences repeated multiple times within the genome of prokaryotes. These sequences produce adaptive immunity in bacterial cells against phage and plasmid infections and protect bacterial cells against invading viruses or foreign plasmids. The endonuclease encoding locus stores snippets of foreign sequence, which eventually would be transcribed into RNAs. These RNAs would be used as a guide based on sequences complementarity to introduce site-specific double-strand break in the target DNA or to cleave invading nucleic acids (Doudna and Charpentier, 2014). The endonuclease CRISPR-associated protein Cas9 cleaves DNA according to the sequence within an RNA duplex and creates site-specific double-strand break. CRISPRs are often associated with cas genes which code for endonucleases that perform various functions related to CRISPRs such as using a guide sequence within a RNA duplex and aligning it with a target DNA sequence. Post alignment, Cas9 generates a DSB in the target DNA, which provides the DNA open ends and initiate nonhomologous recombination or homologous recombination if the similar sequence has been provided by a plasmid DNA vector or any other endogenousnucleic acid (Doudna and Charpentier, 2014). Recently, the CRISPR-Cas system has been adapted as a gene editing technique in mammalian cells by applying a DNA vector carrying cas genes, therapeutic transgene, and specifically designed CRISPRs to precisely cut host human cell genomic DNA at a desired location (Mali et al., 2013). However, for few years limited progress has been made to improve Cas specificity in variety of human cells due to target cell genotype heterogeneity. Unlike other nuclease methods, the CRISPR-Cas system requires the design and synthesis of a guide RNA that simultaneously targets multiple genomic sites. In last few years, the CRISPR-Cas technology in combination with better delivery systems such as physical methods and more advanced nanocarriers dramatically improved its specificity and efficacy. The number of publications that applied this technology to edit the genome of human somatic cells, stem cells, and iPSC cells has been increased radically and several companies are offering commercially available kits and services to assist scientists with multidisciplinary backgrounds who are interested in applying this technique. The following are just few examples to demonstrate the increasing popularity of this technology among clinical and scientific researchers. CRISPR-Cas system has been used to (A) provide in-depth understanding of carcinogenesis processes through generating

FIGURE 5 | Schematic representation of ZFN and TALEN-mediated homologous recombination into AAVS1 safe site of human chromosome. The FokI endonuclease enzyme derived from the prokaryote *Flavobacteriumokeanokoites* was found to function as two separate domains - one binds DNA in a sequence specific manner and one cleaves in a sequence independent manner that is highly specific because it needs dimerization for endonuclease activity on the target DNA. Therefore, by fusing a FokI monomer to two ZFPs or TALEs, which bind to adjacent sequences at a safe site on the human chromosome such as *AAVS1* target site, it is possible to generate sequence specific DNA nuclease complexes that facilitate selective targeting and homologous recombination within human genomes (Davis and Stokoe, 2010). (1) Two ZFPs or TALEs are designed to specifically target *AAVS1* site in anticipated clinically relevant target cell. These proteins are fused to a FokI cleavage domain to facilitate their nuclease activity. The ZFN or TALEN are co-delivered with the HR donor plasmid DNA vector into the cell by a non-viral gene delivery technique to reach the nuclei. (2) Inside the nucleus, ZFNs or TALENs recognize specific DNA sequences on the target cell's chromosome. (3) Binding of two ZFNs or TALENs complex to the target site allows FoKI to dimerize. (4) ZFNs or TALENs complex creates a targeted chromosome break at the *AAVS1* site of the host cell, which facilitates HR between the donor DNA vector and homologous sequences. (5) After generating the DNA double strand break, the ZFNs or TALENs complex disassociate from target DNA. (6) The AAVS1 homology arms subcloned into the HR donor DNA vectors recognize the homolog sequences located on the host genome. (7) HR-mediated insertion of the therapeutic gene expression cargo occurs at the *AAVS1* site of the clinically relevant host cell.

tumor-associated genomic rearrangements by causing DSB and non-homologous recombination of the DNA open ends (Choi and Meyerson, 2014; Torres et al., 2014); (B) facilitate systematic analysis of gene functions in human cells by activation (Cheng et al., 2013) or inactivation of essential genes for positive or negative selection (Wang et al., 2014) or genes that play key roles in cell viability and cancer progression, reprogramming somatic cells to generate iPSC cells (Shalem et al., 2014), and large-scaled genetic screening for drug targets (Doudna and Charpentier, 2014); (C) facilitate gene therapy of genetic diseases by correcting genetic mutations involved in inherited disorders by HR (Wu et al., 2013); (D) produce genetically engineered animal models for different human diseases and drug discovery (Ma et al., 2014; Wang et al., 2014); (E) develop new antiviral treatments by making DSB in the provirus genomic DNA integrated into the infected host cell genome and deactivating the viral genes necessary for pathogenesis or reproduction, reassembly and release of new viruses (Ebina et al., 2013; Hu et al., 2014); and (F) facilitate precise ex vivo remodeling of stem/progenitor cells (Doudna and Charpentier, 2014).

## Conclusions

More advanced and safer techniques for genome editing have many potential applications in further understanding of neurodegenerative diseases, including (A) generating new cell lines to study basics of disease progression, cell growth, and differentiation; (B) screening, identification, and developments of novel therapeutics and their targets; (C) generating disease animal models; and (D) developing new candidates for cell replacement or cell rescue therapies (Yin et al., 2014a).

Neurotrophic factor therapeutic genome engineering is a highly coveted clinical aspiration of glaucoma therapy, but only recent technological advances in improved delivery techniques, the new generation of bacterial sequence-depleted DNA vectors carrying only the desired therapeutic transgene expression cargo, and the availability of modern genome editing techniques could provide the necessary tools for the modification of complex genomes and gene therapy in a targeted fashion. We previously highlighted the rationale for NTFs and their transgene delivery into cell candidates and discussed the significant effects of several NTFs on the support and protection of retinal cells and other cell candidates that might be potential therapeutic options for glaucoma neuroprotective gene therapy (Nafissi and Foldvari, 2015). Here, we highlighted the rationale of genetic modifications for glaucoma therapies either as the gene therapy targets or cell replacement therapy to provide longterm neurotrophin therapeutics on-site using new generations of safer and more efficient DNA vectors and modern genetic modification techniques.

In general, compared to cell replacement therapies to replace and integrate functionally active stem/progenitor cells into the retina, it was shown that cell-based neuroprotective therapies by prolonged secretion of NTFs is a more straightforward strategy to nourish and support dying RGCs in progressive glaucoma. In addition to supplying NTFs, cell-based neuroprotective therapies may also be able to promote RGCs survival directly through modifying gene expression in surrounding cells to enhance NTFs expression on-site (Madhavan et al., 2008). Genetically modified retinal progenitor cells, that express and secrete the NTFs, have been shown to provide neuroprotective support to degenerated or injured RGCs similar to the supportive role of NTFs expressing NSC or mesenchymal stem cells (MSCs) in regeneration of neurons (Akerud et al., 2001; Nomura et al., 2005; Zhu et al., 2005).

Thus, genetically modified NTFs secreting cell-based therapies provides new avenues to fight with the irreversible loss of RGCs associated with glaucoma. Currently we have efficient tools for permanent, reversible, or conditional genetic manipulations of almost any type of cells. The success of clinical trials in other disorders gives rise to similar approaches in the context of neurodegenerative diseases including glaucoma. Ultimately, a better understanding of cell development and stem cell biology at the molecular level, along with improvements in the techniques for non-viral gene delivery, advances in nanotechnology and nanomedicine for synthesis of more cell-friendly structures, and in-depth knowledge of gene modification systems at the

#### References


molecular level would encourage scientists to develop novel treatments for degenerative disorders. Non-viral gene delivery techniques such as electroporation (Helledie et al., 2008) nucleofection (Dwivedi et al., 2012; Madeira et al., 2013), as well as nanotechnology tools such as application of polysaccharide nanoparticles (Deng et al., 2013), cationic liposome nanoparticles (Madeira et al., 2010), calcium phosphate nanoparticles (Cao et al., 2011), and nanoneedles (Nakamura et al., 2002), in parallel with efficient and safe therapeutic transgene insertion techniques such as 8C31 integrase, transposons, and ZFN, TALEN, or the CRISPR/Cas9 nuclease systems (Gaj et al., 2013), are new tools for developing clinically approved cell-based therapeutics for neurodegenerative disorders like glaucoma. These systems are more reliable since they avoid the high risks associated with using viral vectors such as insertional mutagenesis and undesired immune rejection, provide life-long therapy by more policed insertion of the therapeutic gene into the desired site, target a broader range of disorders due to their capability to accommodate genes of different sizes, and finally, provides higher activity owing to their ability to target hard- to-transfect human cells. Before ex vivo modified cells can achieve therapeutic efficacy for clinical glaucoma treatment in human, many limitations remain to be overcome. A reliable and expandable source of retinal cells must be isolated, expanded, characterized and established. These cells must demonstrate appropriate engraftment within the retina. Even with the aforementioned challenges, the in vitro, ex vivo, and pre-clinical support for genetically modified cellmediated RGC repair is undeniable, and warrants the kind of robust investigation that is currently under way. These methods, if successful in human trials, could offer several advantages over traditional pharmacological approaches. While much work remains to address the safety issues for all of the above mentioned approaches, evidently, this is a multidisciplinary field which undeniably needs collaboration from different scientific domains to bring the bench results to the clinic.

## Acknowledgments

The authors thank Elham Mahootchi for her great help in graphic illustration of figures. This paper was supported by the Canadian Institutes of Health Research, the Natural Sciences and Engineering Research Council of Canada, the Canada Foundation for Innovation and the Ontario Research Fund. The generous support of the Canada Research Chairs Program the Canada Foundation for Innovation is also gratefully acknowledged (MF).

Benowitz, L. I., and Yin, Y. (2010). Optic nerve regeneration. Arch. Ophthalmol. 128, 1059–1064. doi: 10.1001/archophthalmol.2010.152

Benowitz, L., and Yin, Y. (2008). Rewiring the injured CNS: lessons from the optic nerve. Exp. Neurol. 209, 389–398. doi: 10.1016/j.expneurol.2007.05.025

Bertoni, C., Jarrahian, S., Wheeler, T. M., Li, Y., Olivares, E. C., and Calos, M. P. (2006). Enhancement of plasmid-mediated gene therapy for muscular dystrophy by directed plasmid integration. Proc. Natl. Acad. Sci. U.S.A. 103, 419–424. doi: 10.1073/pnas.0504505102


optic nerve of American Cocker Spaniel dogs with spontaneous glaucoma. Vet. Ophthalmol. 10, 12–19. doi: 10.1111/j.1463-5224.2007.00504.x


integration technique. Circulation 126(11 Suppl. 1), S20–S28. doi: 10.1161/CIRCULATIONAHA.111.084913


closed DNA minivector. BMC Infect. Dis. 14:P74. doi: 10.1186/1471-2334-14- S2-P74


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2015 Nafissi and Foldvari. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Nanoparticle-Based and Bioengineered Probes and Sensors to Detect Physiological and Pathological Biomarkers in Neural Cells

#### Dusica Maysinger <sup>1</sup> \*, Jeff Ji <sup>1</sup> , Eliza Hutter <sup>1</sup> and Elis Cooper <sup>2</sup>

<sup>1</sup> Department of Pharmacology and Therapeutics, McGill University, Montreal, QC, Canada, <sup>2</sup> Department of Physiology, McGill University, Montreal, QC, Canada

#### Edited by:

Ioan Opris, Wake Forest University School of Medicine, USA

#### Reviewed by:

Dambarudhar Mohanta, Tezpur University, India Ruud Hovius, École Polytechnique Fédérale de Lausanne, Switzerland Ruxandra Vidu, University of California, Davis, USA

> \*Correspondence: Dusica Maysinger dusica.maysinger@mcgill.ca

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 28 July 2015 Accepted: 30 November 2015 Published: 18 December 2015

#### Citation:

Maysinger D, Ji J, Hutter E and Cooper E (2015) Nanoparticle-Based and Bioengineered Probes and Sensors to Detect Physiological and Pathological Biomarkers in Neural Cells. Front. Neurosci. 9:480. doi: 10.3389/fnins.2015.00480 Nanotechnology, a rapidly evolving field, provides simple and practical tools to investigate the nervous system in health and disease. Among these tools are nanoparticle-based probes and sensors that detect biochemical and physiological properties of neurons and glia, and generate signals proportionate to physical, chemical, and/or electrical changes in these cells. In this context, quantum dots (QDs), carbon-based structures (C-dots, grapheme, and nanodiamonds) and gold nanoparticles are the most commonly used nanostructures. They can detect and measure enzymatic activities of proteases (metalloproteinases, caspases), ions, metabolites, and other biomolecules under physiological or pathological conditions in neural cells. Here, we provide some examples of nanoparticle-based and genetically engineered probes and sensors that are used to reveal changes in protease activities and calcium ion concentrations. Although significant progress in developing these tools has been made for probing neural cells, several challenges remain. We review many common hurdles in sensor development, while highlighting certain advances. In the end, we propose some future directions and ideas for developing practical tools for neural cell investigations, based on the maxim "Measure what is measurable, and make measurable what is not so" (Galileo Galilei).

Keywords: gold nanoparticles, nanosensors, microglia, neurons, quantum dots, Ca2+, MMP, caspases

## INTRODUCTION TO SENSORS AND THEIR APPLICATIONS IN NEUROSCIENCE

Recent advances in nanotechnology have provided neuroscientists with powerful new tools. Among these are some probes and nanosensors constructed with materials ranging from organic molecules to metallic nanostructures to engineered fluorescent proteins. Probes are defined as small devices used to explore, investigate or measure something by penetrating or being placed in the cells, cell lysate, and extracellular media. A sensor is an assembly required to detect and communicate a particular event: a device or biological structure which (i) recognizes an entity of interest (e.g., molecules, ions, or physical changes such as temperature) and (ii) transduces an event of recognition into a measurable signal. Recognition and signal transduction is followed by signal detection in the process of biosensing (see **Figure 1**). In case of nanosensors, the terms "probe" and "sensor" often overlap, because nanostructures are penetrating or "in-place" devices and usually serve as a recognition element, a transducer and even a signal amplifier at the same time.

Neural cells respond to dangers and noxious stimuli with a cascade of events involving diverse classes of molecules and ions. So far, probes and sensors have been designed to detect proteins such as signaling molecules and enzymes, ions (e.g., Ca2+, K+, Na+, H+, or pollutants such as Hg2<sup>+</sup> and Cd2+), simple molecules which are critically important for cell metabolism (e.g., glucose, lipids), DNA, changes in pH, redox, and neurotransmitters as well as morphological (e.g., shape and size of neuronal and glial soma, neurites and post-synaptic spines) and functional changes (e.g., action potentials, mitochondrial potential, inter and intracellular organellar communication). "Biofriendly" nanosensors seem to be suitable candidates for intracellular sensing since they are significantly smaller than the size of cells, and chemically inert so as not to interfere with cellular functions during measurements (Howes et al., 2014). However, only few nanosensors have been tested in neural cells. Recent review by Howes et al. (2014) provides a general overview of nanoparticle-based sensors, their use and limitations in biology.

In this review, we focus on several nanoparticle-based and bioengineered sensors mainly for proteases (e.g., metalloproteases and caspases) and biomolecules implicated in disrupted homeostasis in neural cells. To highlight the advantageous features of these nanoparticle-based tools, as well as to discuss critically some of their limitations, we have chosen a few examples from research on inflammatory processes in the nervous system (e.g., caspase-1). First, we provide a brief overview of nanostructures used as probes or components of nanosensors, then we discuss nanosensors and genetically engineered sensors for proteases and aromatase. We then highlight probes and sensors for ions and ion channels focusing on calcium, a principal regulator of many neuronal functions. We provide examples of neural stimulation using nanostructures. To emphasize the complexity of the sensing in the nervous system, we comment on glia as "natural biological sensors" and finally summarize current approaches and challenges in designing suitable nanostructured sensors to detect biomarkers under physiological and pathological conditions.

## NANOPARTICLE-BASED AND BIOENGINEERED SENSORS FOR NEURAL CELLS

Many neurological impairments are associated with different chemical and physical insults that disrupt cell homeostasis. Numerous attempts has been made to follow the progression of pathological processes non-invasively, but only a few established and commonly used bioengineered sensors are able to monitor biochemical and morphological changes in neural cells longitudinally. Moreover, no nanostructured materials are dedicated solely to the development of nanosensors for the detection and monitoring of changes specifically in neural cells, because: (1) noxious stimuli are deleterious to various cell types aside from neural cells, (2) cell responses to danger and harmful stimuli are often similar in different cell types, and (3) availability of primary human and animal neural cells is limited. Examples of organic and metal-based nanostructures for measurement of biomolecules in cells and tissues of the nervous system are provided in **Table 1**.

Most of the nanostructures shown in **Table 1** have to be internalized by cells to detect and measure the intracellular changes of biomolecules. Endocytosis and cooperative transmembrane penetration of nanoparticles are the ways by which nanoparticles can enter cells (Cleal et al., 2013; Yameen et al., 2014; Shang et al., 2014a,b; Beddoes et al., 2015; Kafshgari et al., 2015; Tan et al., 2015; Zhang et al., 2015b). Although endocytosis of nanoparticles (NP) has been studied for quite a while, the precise mechanisms are still not defined due to the complexities of these processes and technical problems associated with them. The emerging picture is that different cells engage in different, and sometimes complementary routes of internalization. This conclusion was derived from studies employing pharmacological agents, knockdown approaches using siRNA and knockout technologies taking advantage of selective deletion of the selected protein anticipated to be involved in the endocytic process (e.g., Iversen et al., 2011). To enhance the chance of nanoparticle entry into cells, various surface modifications were made including the attachment of cell penetrating peptides (Jones and Sayers, 2012; Onoshima et al., 2015). Penetration of lipid bilayers in membranes (cooperative transmembrane penetration) is considered an alternative pathway of NP entry aside from endocytosis (Zhang et al., 2015b). Translocation efficiency via this non-endocytic route depends on NP quantity, NP surface properties, NP aggregation and agglomeration state as well as the properties of cell membranes. It is conceivable that cooperative penetration contributes to the internalization of NP-based sensors. Simulation studies by Zhang et al. suggest that the particle quantity, NP surface properties and membrane structures are closely linked with NP-membrane forces and efficiency of NP penetration (Zhang et al., 2015b).

Traditionally, detection and quantification of intracellular analytes were achieved by employing fluorescent dyes (probes; Haugland, 2005; Lakowicz, 2006) and examples are listed in **Table 2**.

#### Analyte Nanoparticle type Cell/tissue type Detection principle References Oxygen QD Hippocampal slice The sensor is based on FRET between QDs and a fluorescent dye. The QDs are immobilized in a thin layer of polymer matrix, which is deposited either on glass (hippocampal slices are then placed directly on top of the film) or on a microelectrode (placed into the extracellular matrix of the CA1 stratum pyramidale). Luminescence intensity ratio between the QD and dye changes according to the O2 content in the artificial cerebral spinal fluid bathing the brain slices. Ingram et al., 2013 Oxygen Anionic NPs of acrylic co-polymer Primary neural cells, multi-cellular aggregates (3D spheroids) and cultured organotypic brain slices NPs are impregnated with a phosphorescent dye, internalized by endocytosis, and are transported to lysosomes. The phosphorescence lifetime of the dye correlates with the intracellular O2 concentration. Dmitriev et al., 2015 Reactive oxygen species AuNP Post-ischemic rat brains Fluorescein-labeled hyaluronic acids (HA) are immobilized on AuNPs. The probes are injected locally into the focal ischemic brain of a brain stroke animal model. When ROS degrades the HA, the fluorescence dye is released from the AuNPs and is unquenched. Hyun et al., 2013 Sodium PAMAM-CG Primary neurons A sodium dye is encapsulated in a PAMAM dendrimer nanocontainer. When loaded into neurons in live brain tissue, it homogenously fills the entire cell volume, including small processes. The fluorescence intensity correlates with sodium concentration. Lamy et al., 2012 Nitric oxide Carbon nanotubes Microdialysate from rat brain (in vivo) Hemin and multi-wall carbon nanotubes are covalently attached to chitosan; the chitosan is electrodeposited on the surface of carbon fiber microelectrodes. Exogenously applied NO is measured by square wave voltammetry in the rat brain in vivo. Santos et al., 2013 Ascorbate Carbon nanotubes Microdialysate from rat brain (in vivo) A glass carbon electrode modified with heat-treated single-walled carbon nanotubes (SWNTs) is capable of electro-oxidizing the ascorbic acid (AA). Brain microdialysate is directly delivered into a thin-layer radial electrochemical flow cell for the continuous measurement of AA concentration. Liu et al., 2008 Glucose AuNP Microdialysate from rat brain ssDNA modified AuNPs aggregate in the presence of glucose resulting in an absorbance peak shift (in vitro colorimetric detection). Jiang et al., 2010 Cysteine AuNP Microdialysate from the striatum of rat brain Cysteine causes the aggregation of citrate stabilized AuNPs, resulting in an absorbance peak shift (in vitro colorimetric detection). Qian et al., 2012 Lead Graphene quantum dot Cerebrospinal fluid of rats A rigid structure is formed between tryptophan and GQD-DMA conjugates in the presence of Pb2<sup>+</sup> (acting as a cross-linker). The resulting increase in fluorescence allows for the detection of Pb2<sup>+</sup> in brain microdialysate (in vitro measurement). Qi et al., 2013 Inducible Nitric Oxide Synthase AuNP Lysed A172 neuronal cell An electrode is modified with AuNPs and anti-iNOS antibodies. The attachment of iNOS causes changes in chronoamperometric measurements in a concentration dependent manner. Koh et al., 2011 Caspase-1 QD Glial cells QDs and a fluorescent dye are linked through a caspase-1 substrate peptide (FRET condition). In the presence of caspase-1 activity, FRET is lost and fluorescence ratios Moquin et al., 2013a

#### TABLE 1 | Examples of physiologically relevant molecules in neural cells or tissues measured by various nanoparticle-based probes and sensors.

(Continued)

change (in vitro measurements from cell lysates).

#### TABLE 1 | Continued


QD, Quantum dots; AuNP, Gold nanoparticles; PAMAM-CG, polyamidoamine dendrimer course grain.

These dyes are non-invasive probes and relatively simple-tofollow, but rapid bleaching, the requirement of organic solvents (dissolution of lipophilic dyes) or the unpredictable interactions with intracellular molecules often limit their usefulness. QDs, carbon nanomaterials and PEBBLEs are good alternatives to avoid some of the problems associated with fluorescent dyes.

### QDs

Among different nanotechnological products, quantum dots (QDs) attracted special attention because of their unique physicochemical and optical properties, some of which supersede certain qualities of fluorescent organic probes (Mattoussi and Cheon, 2009; Howes et al., 2014; Breger et al., 2015; Moloney et al., 2015; Silvi and Credi, 2015; Wegner and Hildebrandt, 2015). These properties include high fluorescent quantum yield, size-tunable emission and a broad absorption spectrum, ranging from ultraviolet to infrared wavelengths. QDs are composed of a semiconductor core (e.g., cadmium selenide (CdSe), cadmium telluride (CdTe), and are often capped with a shell [e.g., zinc sulfide (ZnS)] to improve core stability. There are also silica (Si)-based QDs which are highly stable and have long lifetime (Cooper et al., 2009). **Figure 2** shows the anatomy of the QDs and a comparison between their lifetimes with those of fluorescent dyes and cell autofluorescence. In addition, the narrow emission spectra and broad excitation spectra allow for simultaneous detection of several different analytes in vitro and in vivo.

The exceptional photostability of QDs, relative to organic fluorescent dyes, makes them more suitable for biomedical labeling and imaging applications, particularly for longitudinal (repeated) imaging (Lidke et al., 2004; Cui et al., 2007; Rajan et al., 2008; Fichter et al., 2010; Vermehren-Schmaedick et al., 2014; Kovtun et al., 2015; Vu et al., 2015). For example, after binding to their receptor, TrkB, the internalization and intracellular trafficking of QD-BDNF (quantum dots with brainderived neurotrophic factor ligand) was visualized. Generally, 2-h incubation with 1 nM QD-BDNF in the axon compartment led to 5–20 QD-BDNF-containing endosomes being transported in a 60 mm long axonal segment. It took ∼40 min for the first QD-BDNF to reach the cell body and the accumulation of QD-BDNF in the cell body was observed within 2 or 3 h. This method can be used to address whether peripheral neurons in diabetic animals have impaired axonal transport (Xie et al., 2012; Zhao et al., 2014). Nerve growth factor (NGF), an important trophic factor for the survival and function of peripheral sympathetic and sensory neurons, has also been studied by employing QDs (Cui et al., 2007; Echarte et al., 2007). Results from these studies showed that NGF is retrogradely transported in axons of the rat dorsal root ganglia. Neural and other cells use tunneling nanotubes for intercellular communication ranging from electrical signaling to the transfer of organelles (Wang et al., 2012; Austefjord et al., 2014; Tosi et al., 2014; Wang and Gerdes, 2015). Tunneling nanotubes mediated transport of functional mitochondria can reduce the impact of ultraviolet light-induced insults. QDs can also be transported through tunneling nanotubes as shown by Zhang's group. Their studies show that QDs move bi-directionally with a mean velocity of



1.23 um/s (He et al., 2010). The velocity of the QD movement along the nanotube varies from 4.27 um/s to 0.054 um/s (mean velocity = 1.23 ± 0.01 um/s). Although most of the QDs reversed their directions of movement every few seconds, more than 80% of the QDS moved along the nanotube toward one of the connected cells. The likely microtubule associated motors mediating bidirectional transport of QDs are kinesins and dyneins.

One concern in using QDs with a Cd-based semiconductive core as sensors for live imaging is the potential cytotoxicity of Cd2+. To mitigate this concern, we investigated other types of QDs as potential candidates for sensors to probe neural cells. We showed that Indium-gallium phosphate/ZnS QDs (InGaP-QDs), core sizes of 5.0 nm and a fluorescence emission maximum at 680 nm, have low toxicity when tested in primary nerve cell cultures and in PC12 cells (Behrendt et al., 2009); indicating that these QDs are promising candidates for live cell and ex vivo imaging in the far red light spectrum. The strong signal that we observed from the internalized InGaP/ZnS QDs suggests that these nanoparticles aggregate in the cytoplasm, but not in the nucleus. In primary neural cultures enriched with astrocytes and glia, InGaP/ZnS QDs were internalized most avidly in microglia, followed by astrocytes, and were barely detectable in neurons. Quantitative analyses of internalized InGaP/ZnS QDs at the organellar level indicated that these NPs were mainly present in lysosomes but not in mitochondria (Behrendt et al., 2009).

Interestingly, we found that the subcellular distribution of InGaP/ZnS QDs is altered by oleic acid, a common ingredient of our daily diet (Behrendt et al., 2009). This finding suggests that changes in membrane structures by fatty acids (endogenous or exogenous) modulate the uptake and distribution of nanostructures in neural cells.

#### PEBBLEs

Kopelman's team developed an interesting array of NP-based sensors called PEBBLEs (photonic explorer for biomedical use with biologically localized embedding; Sasaki et al., 1996; Clark et al., 1998; Lee et al., 2009). PEBBLEs are 1-1000 nm diameter nanoparticles that include both fluorescent analyte-sensitive indicator dyes and analyte-insensitive reference dyes (Lee and Kopelman, 2012b). As such, these sensors allow for ratiometric, reversible measurements and they are protected from interaction with the cellular environment. Two types of PEBBLEs are distinguished. Type 1 PEBBLE uses a single sensing entity, serving as both analyte recognizer and signal transducer, while in Type 2 PEBBLE the analyte recognizer and optical transducer are distinct. PEBBLEs have been developed to measure a number of physiologically relevant parameters, including ion concentrations (protons, calcium, copper, iron, magnesium, potassium, sodium, lead, zinc, chloride), small molecules (oxygen, singlet oxygen, peroxyl radical, hydrogen peroxide), enzymatic intracellular processes (apoptosis), and physical properties (temperature, electric field; Lee and Kopelman, 2012a). PEBBLEs have been used as sensors for intracellular pH and calcium concentration measurements in neural cells (Clark et al., 1999).

A good example of Zn ion sensor constructed as a PEBBLE (Type 2) is based on CdSe/ZnS QDs covalently linked with three different azamacrocycles, non-fluorescent Zn2<sup>+</sup> ligands: TACN (1,4,7-triazacyclononane), cyclen

(1,4,7,10-tetraazacyclododecane), and cyclam (1,4,8,11 tetraazacyclotetradecane; Ruedas-Rama and Hall, 2008). As the surface-conjugated azamacrocycles disrupt the radiative recombination process of the QDs, the QDs' fluorescence is quenched. The binding of Zn2<sup>+</sup> with the azamacrocycles switches on the QD emission, resulting in an increase in fluorescence intensity. Three zinc ion sensors based on CdSe-ZnS core-shell QDs showed a very good linearity in the range 5–500µM, with detection limits lower than 2.4µM and relative standard deviation ∼3%.

Although promising, one limitation of such zinc sensors is that interference from autofluorescence decreases their sensitivity. One-way to improve the intracellular sensitivity of the PEBBLEs is to avoid interference from cellular autofluorescence by using near infrared (NIR) fluorescent probes/reporters, twophoton excitation, and "MOON (modulated optical nanoprobe)" type PEBBLEs (Lee et al., 2009). MOONs are microscopic photonic probes that look like moons; one side appears bright and reports on the local microenvironment whereas the other side is dark. The MOONs rotate in response to thermal or magnetic fields (MagMOONs; Anker et al., 2005). The MOONs allow for sensitive chemical analyses where signal to background ratio can reach up to 4000-fold. Magnetofluorescent MOONs have been more recently developed by Bawendi's group (Chen et al., 2014a). Their studies show that after surface PEGylation, these fluorescent "super" nanoparticles can be magnetically manipulated inside living cells. PEGylation is the covalent conjugation of poly-ethylene glycol to polymers and drug molecules. PEGylation prolongs the circulation half-life of drugs, reduces the immunogenicity of molecules, and stabilizes nanoparticles (Ginn et al., 2014; Kolate et al., 2014).

## Carbon Nanomaterials (CNM)

Carbon nanomaterials have emerged as alternatives to QDs, molecular sensors, and bioengineered sensors (see **Table 1**). Carbon nanomaterials have unique electronic, magnetic and optical properties (Ding et al., 2014; Baptista et al., 2015; Wen et al., 2015; Zheng et al., 2015). In addition, carbon nanostructures such as graphenes, C-dots and nanodiamonds have relatively low toxicity, although there are some safety concerns regarding these materials (Bayat et al., 2015; Boyles et al., 2015; El-Sayed et al., 2015; Himaja et al., 2015; Jeannet et al., 2015; Lim et al., 2015; Misra et al., 2015; Pierrat et al., 2015; Qin et al., 2015b). To achieve the full potential of these structures as bionanosensors, further improvement is necessary. In particular, low sensitivity is a significant limitation (for example, detection of dopamine by graphene-based sensors). On the other hand, nanotubes seem to be useful tools for detecting DNA- drug interactions and could eventually become useful diagnostic aids in oncology (Health Quality, 2006). Developments in carbon nanomaterial based sensors were reviewed (Baptista et al., 2015).

An interesting example of a simple construct employing Cdots was used as a neuroanatomical retrograde tracer (Zhou et al., 2015). Cholera toxin B–carbon dot conjugates were taken up and retrogradely transported both in differentiated pheochromocytoma cells (PC12) and in vivo in Balb-c mice. Results from these studies suggested that cholera toxinBmodified C-dots were mainly taken up by dorsal root ganglia at the lumbar level L3–L5. Due to their superior properties (e.g., low toxicity, high signal intensity) over the Cd-containing QDs, C-dots are promising as multifunctional nanodevices for simultaneous multiple imaging modalities, drug delivery and sensing. However, currently available C-dot-based nanodevices have not yet reached this level of sophistication (Ding et al., 2014).

New fluorescent nanodiamonds are attracting considerable interest in the design of biological nanosensors due to their high resistance to photobleaching and low toxicity (Yu et al., 2005; Montalti et al., 2015). Compared to QDs and typical organic probes, nanodiamonds have high quantum yield and lifetime, although a limited tunability of emission spectrum and a larger emission width (Schirhagl et al., 2014). When tested in dorsal root ganglia cultures, nanodiamonds were highly fluorescent, but unfortunately, they also reduced growth cone extensions and neurite outgrowth (Huang et al., 2014). Similarly, when singlewalled carbon nanotubes were tested in dissociated cultures of the peripheral nerves (Belyanskaya et al., 2009), they were found to be particularly toxic to Schwann cells, although less so for neurons. Current information on carbon nanodiamonds and carbon onions suggests that these morphologies may be better suited for nanosensors than carbon nanotubes due to their lower toxicity (Ding et al., 2005; Schrand et al., 2007). The morphology of carbon-based NP plays a role in internalization and intracellular trafficking (Chu et al., 2015). Importantly, neither prickly nanodiamonds nor morphologically similar Aunanourchins are markedly cytotoxic (Hutter et al., 2010). Thus, both of these NP morphologies might be used for nanosensors to detect changes in cytoplasmic proteins under physiological and pathological conditions. Comparisons between different studies using such nanomaterials are difficult because experimental conditions including cell types, method of carbon nanomaterial preparation, and duration of exposure were different.

In summary, carbon-based nanosensors, particularly nanodiamonds and carbon onions seem to be promising nanomaterials for the construction of sensors to investigate neural cells under physiological and pathological conditions, once they are converted into fully biocompatible "biophils."

### AuNP

Gold nanoparticles (AuNP) have several unique properties that make them attractive as components in biological nanosensors (Dykman and Khlebtsov, 2012; Jin, 2014). These include: a high absorption coefficient (e.g., 40 nm gold nanospheres show an absorption cross-section 5 orders of magnitude higher than organic dyes), scattering flux (e.g., 80 nm gold nanospheres scatter light 5 orders of magnitude compared to some fluorescent dyes), luminescence and conductivity, the ability to enhance electromagnetic fields, quench (or enhance) fluorescence, and catalyze reactions. As such, these nanosensors are the most common (e.g., see **Table 1**).

Many shapes and various surface modifying molecules have been exploited for AuNP-based nanosensor construction (Vo-Dinh et al., 2013; Liu et al., 2015; Qin et al., 2015a; Zhang et al., 2015c). Gold nanorods, nanoshells and nanourchins are particularly attractive tools for in vivo applications, since their optical resonance lies in the near-infrared spectral window, away from the region of biomolecular excitation transitions, which precludes photochemical damage and allows for deeper penetration of light in living systems. It is important to note, however, that cellular internalization of NPs depends on the shape, size and surface properties of the NPs just as much as on the cell type. For example, when spherical, rod and urchin GNPs were added to primary hippocampal neurons and microglial cells, Au nanourchins were found to be taken up preferentially by microglia, while only Au nanorods were internalized by neurons (Hutter et al., 2010).

Characterization of AuNPs is essential for any sensor construction (Leifert et al., 2013; Conde et al., 2014). The most common techniques to determine the particle size and its distribution (core and shell) and the successful ligand attachment are electron microscopy, absorption spectroscopy, dynamic light scattering, asymmetric flow-field flow fractionation (Moquin et al., 2013b, 2015), and zeta potential measurements. Elemental analysis, thermogravimetric analysis/differential scanning calorimetry, nuclear magnetic resonance spectroscopy, infrared spectroscopy and X-ray photoelectron spectroscopy can provide some quantitative information about the ligand shell composition and functionality (Leifert et al., 2013).

To generate a sensor, the ligand is covalently attached to the AuNP surface by a dative metal-thiol bond (Howes et al., 2014). The bond dissociation energy for AuNP—S bonds is ∼40–50 kcal/mol, approximately half of that for a typical C—C or C—H bond. In contrast, AuNP—N (∼8 kcal/mol) and AuNP— COO—∼2 kcal/mol) dissociation energies are considerably weaker, comparable in strength to a hydrogen bond, and can easily be displaced (Muddineti et al., 2015).

AuNPs can function in conjunction with many traditional biological probes such as antibodies, nucleic acids, ligands and receptors, and are used for a highly sensitive and selective detection of various biomarkers (see **Table 1**). Some of these assays have already been commercialized (Nanosphere, Merck, BBInternational, etc.). Measurements revealing changes in the cellular environment components and enzymatic activities are important in the exploration of physiological and pathological processes such as neuroinflammation. Enzymes often activated in neural cells in inflammation are caspase-1, cyclooxygenases and matrix metalloproteinases (MMPs), among others. Nanosensors for MMPs based on AuNP and their brief description are included in **Table 1** and are discussed in Section Nanosensors for Proteases and Aromatase.

### Bioengineered Sensors

With genetically encoded fluorescent proteins, one can study dynamic changes in intracellular processes in live cells, and quantify the expressions of proteins which are difficult to measure by other means (Enterina et al., 2015). Fluorescent protein based biosensors are useful tools for the study of signaling processes in neurons (Shen et al., 2015). This approach was recognized as an invaluable contribution to biological investigations including those in neuroscience by the Nobel Prize in chemistry granted to Osamu Shimomura, Martin Chalfie and Roger Y. Tsien. Genetically encoded proteins are either used to label and image proteins of interest without fundamental changes in their properties (imaging tools) or to "sense" complex biochemical processes in living cells (Frommer et al., 2009; Oldach and Zhang, 2014). Some examples of genetically engineered sensors and their properties that can be exploited in investigating neural cells under physiological and pathological conditions are shown (**Table 3**).

Bioengineered sensors presented in **Table 3** are useful sensors for the detection of cell metabolites, ions, and enzymes. Similarly, bioengineered sensors provide the means of detection for different kinases and their participation in spatially restricted areas and their function in macro-compartments (i.e., dendritic spines) or micro-compartments (i.e., AKAP signalosomes; Willoughby et al., 2010; Yasuda, 2012). They can be used to determine enzymatic activities in real-time and in a cell specific manner. Transgenic expression of reactive oxygen species (ROS) generating proteins (RGPs) as fusions to native proteins allow for exertion of spatial and temporal control over ROS production and ROS signaling (Wojtovich and Foster, 2014).

### NANOSENSORS FOR PROTEASES AND AROMATASE

## AuNP-Based Sensors for MMPs

Matrix metalloproteinases (MMPs) are pleiotropic endopeptidases involved in a variety of neurodegenerative

TABLE 3 | Examples of genetically engineered biosensors for the detection of ions and enzymes in neural cells.


processes including neuroinflammation (Sonderegger and Matsumoto-Miyai, 2014; Yamamoto et al., 2015). They interact with a large number of substrates (Malemud, 2006). Most of the MMPs are synthesized as inactive latent enzymes. Conversion to the active enzyme is generally mediated by activator systems that include plasminogen activator or the pro-hormone convertase, furin. MMPs are active at physiological pH and they catalyze the normal turnover of extracellular matrix (ECM) macromolecules. The endogenous inhibitor of MMPs (TIMP-2) is constitutively expressed in microglia and markedly inhibited by proinflammogen lipopolysaccharide (LPS) treatment (Lee and Kim, 2014). Knockdown and overexpression experiments in microglial cell line BV2 suggest that endogenously expressed TIMP-2 plays an anti-inflammatory role. Overexpression or knockdown of TIMP-2 in BV2 cells leads to reciprocal expression and release of inflammatory biomarkers after LPS treatment. Overexpression of TIMP-2 in BV2 cells doubles the expression of the anti-inflammatory IL-10 and markedly decreases the expression of pro-inflammatory cytokines TNF-α, IL-1β, as well as nitric oxide. Conversely, siRNA knockdown of TIMP-2 in BV2 cells reduces the expression of IL-10 by ∼ 30% and increases ROS, nitric oxide, and TNF-α. An overexpression of TIMP-2 suppressed microglial activation through inhibition of the activity of mitogen-activated protein kinases (MAPKs) and transcription factor NF-κB. The results from these studies indicated an enhancement of the activity of anti-inflammatory Nrf2 and cAMP-response element binding protein (CREB) transcription factors in microglia with overexpressed TIMP-2. Microglia activation contributes to the degradation of extracellular matrix proteins by increasing metalloproteinases activities (Lively and Schlichter, 2013). In the transgenic 5X FAD mouse model of Alzheimer's disease (AD) the expression of MMP-2, MMP-9, and MT1-MMP was upregulated concomitantly with the tissue inhibitor of MMPs-1 (TIMP-1) and several markers of inflammatory/glial response (Py et al., 2014). Data from these studies suggest a regulatory interplay between MMPs and the amyloid precursor protein (APP). The role of astrocytes and MMP-9 in synaptic dysfunction has been reviewed (Kamat et al., 2014).

In fluorescence-based assays, including these for MMPs, AuNPs are employed as acceptors, quenching the emission of donor chromophores. Because of their large absorption cross section, AuNPs have a superior quenching efficiency in a broad range of wavelengths compared to other organic quenchers. Therefore, they can be used for studies in which donor-acceptor distances are expected to extend beyond 10 nm, or studies in which multiple dyes need to be quenched. Since they have no defined dipole moment, energy transfer takes place for any orientation of the donor relative to the surface of the AuNPs. For instance, even if the distance between AuNPs and fluorescent dyes are as large as 22 nm, the quenching efficiency can be as high as 95% (Mayilo et al., 2009).

An example of one such assay detects MMP-7. MMP-7 is an extracellular protease that exerts a broad range of biological functions including important roles in synaptic plasticity (Sonderegger and Matsumoto-Miyai, 2014). In the assay for MMP-7 detection, carboxy AuNPs (5 nm in core diameter) are used as both quenchers and metal chelators, and are strongly associated with the hexahistidine regions of dyetethered peptides in the presence of Ni(II) ions; this leads to fluorescence quenching of the dye by AuNPs. Upon adding MMP-7, the peptide is cleaved and the fluorescent intensity of the dye is efficiently recovered. The degree of dequenching is directly dependent on the MMP-7 concentration in a hyperbolic manner, ranging from as low as 10–1000 ng mL−<sup>1</sup> (Park et al., 2012).

MMP-2 is a constitutive protein found in the normal brain cardiovascular systems and glioma (Lebel and Lepage, 2014; Mittal et al., 2014; Wang et al., 2014, 2015c; Ruan et al., 2015; Yang and Rosenberg, 2015). Experimental evidence suggests that MMP-2 may contribute to early brain enhancement of the cytokine interleukin-1 beta concentrations in transient ischemia thereby promoting cortical neuron damage (Amantea et al., 2014). MMP-2 measurements in vivo can be achieved by exploiting near-infrared-fluorescent dye functionalized gold nanoparticles. The fluorescent dye is quenched until the protease cleaves its link to the AuNP and releases it. Tumors with high protease activity can be visualized by the near infrared fluorescence signals from gold nanoparticle probes upon MMP-2 activation (Lee et al., 2008).

MMP-9 has been recognized as a regulator of dendritic spine morphology. Dendritic spines are dynamic structures that change their morphology in response to various stimuli (Stawarski et al., 2014b). Spine remodeling occurs in many degenerative disorders and MMP-9 seems to play a critical role. Stawarsky et al. developed a biosensor to measure MMP-9 activity in living cell (Stawarski et al., 2014a). This biosensor can be used to monitor the changes of MMP-9 activity and correlate them with plastic changes of dendritic spines. Considering the accessibility of AuNP sensor to the spines, it is likely that AuNP sensors would be useful in assessing MMP-9 activity and be correlated with spine morphology and function. Extracellular proteolytic cleavage at synapses is executed by a relatively small number of peptidases which have a limited set of target proteins. MMP-9 is one of the peptidases which plays critical roles in synaptic structure and function. MMP-9 expression is enhanced by reactive oxygen species (ROS; e.g., post injury, high glucose; Hsieh et al., 2014). MMP-9 might play a dual role in epilepsy (Michaluk and Kaczmarek, 2007). It is also highly expressed in microglia in response to inflammatory stimuli (Gottschall et al., 1995) and can induce neuronal cell death (Murase and McKay, 2012; Gao et al., 2015). Analyses of MMP-9 deficient mice showed a significant reduction in neuronal damage in the hippocampus after transient global cerebral ischemia (Lee et al., 2004), but in dendritic spines conflicting results (enlargement and thinning) were reported. These controversial issues, as well as the roles of metalloproteinase 2 and 9 in the development, plasticity and repair of the nervous system were reviewed (Verslegers et al., 2013).

MMPs increase the permeability of the blood–brain barrier as part of the neuroinflammatory response in hypoxia–ischemia, multiple sclerosis and infection (Rosenberg, 2009). Recent studies have also implicated MMPs in the chronic neurodegeneration associated with vascular cognitive impairment, Alzheimer's disease, and Parkinson's disease. Therefore, both the identity of the active MMPs and their cellular origin could determine whether disease pathogenesis or regeneration occurs. Thus, synthetic MMP inhibitors might be valuable for treating some CNS diseases (Rosenberg, 2009).

#### Bioengineered Sensors for MMPs

Only recently an MMP-9 sensor has been developed using mCherry fusion protein to quantify intracellular protein and secreted protein in the extracellular medium (Duellman et al., 2015). This microplate reader-based mCherry fluorescence detection method had a wide dynamic range of 4.5 orders of magnitude and a sensitivity that allowed detection of 1–2 fmol fusion protein. The rate of secretion was calculated from the linear region of data from 8 to 24 h post-transfection. Comparison with the Western blot protein detection method indicated greater linearity, wider dynamic range, and a similar low detection threshold for the microplate-based fluorescent detection assay of secreted fusion proteins.

## AuNP and QD-Based Sensors for Caspases

Many neurological disorders are associated with inflammation (Cardoso et al., 2015; Crotti and Glass, 2015; De Felice and Lourenco, 2015; Franco and Fernández-Suárez, 2015; Freeman and Ting, 2015; Hayashi and Cortopassi, 2015; Hoogland et al., 2015; Maphis et al., 2015; Ward et al., 2015); consequently, enzymes implicated in inflammatory processes are likely good targets for therapeutic interventions (Py et al., 2014; Kaushal et al., 2015; Savard et al., 2015; Yang and Rosenberg, 2015; Wang et al., 2015a). The initiation of inflammatory responses involves the formation of cytosolic structures named "inflammasomes" (Martinon et al., 2002; Fang et al., 2015; Frank et al., 2015; Freeman and Ting, 2015; Guo et al., 2015; Szabo and Petrasek, 2015; Yang and Chiang, 2015). Inflammasomes are multiprotein complexes that enable the activation of pro-inflammatory caspases, mainly caspase-1 (Gross et al., 2011). Mechanism of caspase-1 and caspase-11 activity was reported (Kayagaki et al., 2015; Shi et al., 2015). Different caspases recognize different peptide sequences and cleave them. These caspasespecific peptides (substrates) can be incorporated in sensors as linkers between fluorogenic and/or quenching entities. As a consequence of enzymatic cleavage, there is a change in fluorescence intensity. Since the identification of sequences that are cleaved in various caspase substrates (McStay and Green, 2014; Kang et al., 2015; Parsons et al., 2015) and the development of synthetic substrates, these sensors became attractive to reveal enzyme-specific reactions in different cells, including neural cells, but mainly in lysed cells. Real-time measurement of caspase activity in live cells and animals is more challenging, but with the advancement of technology (particularly nanotechnology) it has become possible (Ai et al., 2008; Hutter and Maysinger, 2013; Moquin et al., 2013a). Examples of QD- based nanosensors for measuring enzymatic activities were reviewed (Hutter and Maysinger, 2013) and are given in **Table 1**.

The basic principles of enzymatic activity measurements using QDs and AuNP are changes in fluorescence intensities (due to the substrate cleavage) or shift in absorbance maximum (due to the change in aggregation status of nanoparticles), as illustrated (**Figure 3**).

FIGURE 3 | Nanosensor measurements based on absorbance and fluorescence. (A) Principle of a fluorescence-based nanosensor where a quantum dot (QD) is linked to quencher molecules through substrate linkers. An active protease cleaves off the link leading to dequenching (enhanced fluorescence proportional to protease activity). (B) The principle of an absorbance-based nanosensor using gold nanoparticles (AuNP) functionalized with cross-linked peptides [causing the aggregation of the AuNP which have an absorbance peak around 600–700 nm (OD)]. An active protease chews the substrate peptide, causing the disaggregation and a blue shift in the peak absorbance (500 nm).

To investigate how microglia respond to proinflammatory stimuli, our laboratory has developed a QD-based assay for caspase-1. Caspase-1 activity in this assay is determined by ratiometric measurements of fluorescence signals based on FRET (**Figure 4**; Moquin et al., 2013a). An attractive feature of our assay is that we could follow caspase-1 activity over time, because the signals from QDs are more stable than conventional organic fluorophores. Our sensor is suitable for measuring changes in caspase-1 activity at the single cell level.

Aside from QD, AuNPs could also be used for caspase sensor construction. The SPR peak of AuNPs is very sensitive to their environment, shifting toward the longer wavelengths (red shift) and broadening significantly upon the aggregation of AuNPs, i.e., the absorption peak of dispersed particles changes toward longer wavelengths when the particles are aggregated. This color change is easily measured by conventional spectrometers. The color shift can also occur in the reverse direction: breaking the AuNP aggregates into individual particles causes a blue shift in the absorption spectrum of dispersed particles. The basic principle of AuNP-based colorimetric assays is that the extent of aggregation/separation is proportional to the absorption peak shift, and therefore the signal is quantifiable. By monitoring the ratio of the area under the surface plasmon peak of aggregated AuNPs (spanning from 490 to 540 nm) and the area under that of dispersed particles (550–700 nm), it is possible to obtain a ratiometric quantification of enzymatic activity. Although rarely used, the power of this approach has provided an extremely high detection limit of 90 zeptograms/mL(10−<sup>21</sup> g/mL) for thermolysin (Laromaine et al., 2007). A similar approach could be used for sensors consisting of AuNPs assembled through protease cleavable peptides and dispersed in the presence of the active enzyme. Such a sensor could be easily applied to the measurement of enzymatic activities of various proteases, particularly those imbedded in plasma membranes with their active site exposed to extracellular environment.

#### Bioengineered Sensors for Caspases

In addition to nanoparticle-based sensor for caspases, there are several bioengineered sensors for this class of proteases (**Table 3** and **Figure 5**). The suitability of a FRET-based sensor for caspase-3 was demonstrated in 3D-cultures using breast cancer cells, but similar construct could be used for the determination of caspase 3-activity in neural cells (Anand et al., 2015). Caspase-3 has been implicated not only in cell death but also in synaptic failures in the absence of cell death (D'Amelio et al., 2011). The molecular mechanisms underlying synaptic failure are still incompletely understood, but studies by D'Amelio et al. identified a caspase-3-dependent mechanism that drives synaptic failure and contributes to cognitive dysfunction in Alzheimer's disease mouse model and possibly in humans (D'Amelio et al., 2011).

Multiple roles of caspase-1 in neuroinflammation have been reported in several animal models (Alfonso-Loeches et al., 2014; Freeman and Ting, 2015). For example, CNS human neurons express functional NLRP1 inflammasomes, which activate caspase-1 and subsequently caspase-6. Studies by LeBlanc (Kaushal et al., 2015) reveal a fundamental mechanism

linking intraneuronal inflammasome activation to caspase-1-generated interleukin-1-β-mediated neuroinflammation and caspase-6-mediated axonal degeneration. The basic principle of protein dimerization-based bioengineered sensors for caspase-1 and caspase-3 is illustrated (**Figure 5**).

red fluorescence in the nucleus with increased caspase activity.

Using sensors for caspase-1 and caspase-3 (**Figure 5**), we showed that human cells exposed to lipopolysaccharide markedly activated caspase-1, but not caspase-3. In turn, caspase-3, but not caspase-1, was activated when cells were exposed to staurosporine (Ding et al., 2015). These bioengineered sensors for caspase-1 and 3 can be a great help assessing the microenvironment-modifying agents and antineoplastic agents in glioblastoma and also in neurodegenerating CNS.

#### Aromatase Function and Measurements

Neural-active steroids play a critical role in the development of the central nervous system and in the maintenance of functional circuitries (Melcangi et al., 2011; Remage-Healey et al., 2011; Arevalo et al., 2015; Frankfurt and Luine, 2015; Hojo et al., 2015; Krentzel and Remage-Healey, 2015). Aromatase is the key enzyme which transforms testosterone into estradiol (Yague et al., 2010). The measurements of aromatase enzymatic activity are mainly indirect and based on radioactive measurements of tritiated water. There are currently no ways to measure aromatase activity in neural cells directly and non-invasively. Some neurological disorders, particularly in postmenopausal women are ascribed to inadequate estrogen concentrations in certain brain structures, e.g., hippocampus (Danilovich et al., 2003; Markham et al., 2005; Spence and Voskuhl, 2012; Daniel, 2013). Aromatase inhibition by letrozole is used in breast cancer therapy and there are reports that such therapeutic interventions can cause memory impairments in certain female populations; however, it is not clear what makes them more vulnerable to letrozole treatment (Zhou et al., 2007; Chang et al., 2013; Turnbull et al., 2015; Vierk et al., 2015). In this context, we have shown that in organotypic hippocampal cultures treated with letrozole, post-synaptic dendritic spines are reduced in number, resulting in dysfunctional neural circuitry (Chang et al., 2013). We propose a FRET-based assay for the measurement of aromatase enzymatic activity by using nanoparticles (**Figure 6**).

In theory, AuNPs and QDs can be combined in a FRET-based assay. A good example is the design to detect the activity of aromatase (**Figure 6**). In this strategy, the substrate of aromatase, testosterone is covalently conjugated to the surface of quantum dots ("QD-testosterone"). AuNPs are modified with a complex of thiol-terminated estradiol-binding aptamers ("AuNP-aptamer") and their fluorophore-labeled complementary DNA strand (reporter DNA; Alsager et al., 2015). In the presence of aromatase some of the QD-conjugated testosterone is converted to estradiol; then, upon addition of AuNP-aptamers, the fluorescently labeled reporter DNA is released and the aptamer binds to the estradiol, bringing the AuNPs into a close proximity of QDs. The fluorescence of QDs is quenched by the AuNP, while the fluorescence of the released reporter DNA is restored. In addition, the aggregation of AuNPs and QDs results in a very distinct change in the surface plasmon absorption spectrum of AuNPs. This elegant triple detection system (QDs, fluorophore and AuNP) allows for monitoring of the testosterone to estradiol conversion by three parallel ways (by QD quenching, fluorophore turn-on, NP aggregation). Such an assay is not yet available but it would be very useful in experiments and clinical studies addressing questions related to the role of aromatase in neurological disorders.

In summary, we have identified a number of nanoparticlebased sensors for enzymes relevant in the nervous system and we have highlighted both their advantages and limitations.

The most common sensors are those for caspases and MMPs. These enzymes are particularly relevant because of their roles in synaptic plasticity and neurodegenerative disorders. Being extracellular enzymes, MMPs are relatively more easily accessible by nanomaterials than intracellular enzymes such as caspases. MMPs, acting as extracellular matrix and spine "sculptors," play essential roles in neural cell functions; therefore, further indepth studies are warranted to unravel the intricate roles of MMPs, particularly their potential beneficial contributions in post-injury repair of the nervous system. Nanosensors that allow simple and reproducible measurements of MMPs enzymatic activities could facilitate these investigations. These assays would be of value for designing new therapeutic interventions in neurological disorders where abnormal regulation of these proteases contribute to neural malfunction and death. Analyses of various MMP substrates incorporated into nanoparticle-based sensors could certainly help delineate preferable specific substrates in neural cells and downstream pathways leading to beneficial or detrimental MMP-mediated functions. Finally, understanding MMP-mediated proteolysis in neural cells would go far beyond these cell types.

## SENSORS FOR IONS AND CELL METABOLITES

In this section, we will highlight several examples that incorporate genetically engineered or NP-based sensors. We will focus mainly on pH, O<sup>2</sup> and Ca2<sup>+</sup> measurements using nanosensors.

## Bioengineered Sensors for the Detection of Calcium Ions

Ever since Ringer's serendipitous observation that Ca2<sup>+</sup> caused the contractions of isolated hearts, physiological roles of this ion were intensively investigated (Ringer, 1883). The extracellular calcium concentrations are high (about 1 mM), whereas intracellular pools are low (100 nM). The most important calcium stores are the endoplasmic reticulum and mitochondria. Special temporal patterns of calcium are regulated by pumps, channels, and buffering proteins. Regulation of calcium signaling and the role of mitochondria in its regulation contributing to cell metabolism, and cell survival has been reviewed (Rizzuto et al., 2012).

In neurons calcium is in constant flux to facilitate neurotransmission and even modest alterations in calcium signaling can cause cellular stress and negatively impact cell function (Kawamoto et al., 2012). For instance, in Alzheimer's disease, neuronal calcium disturbances and abnormally regulated calcium signaling proteins are speculated to play a major role in causing mitochondrial dysfunctions (Supnet and Bezprozvanny, 2010). The variety of cellular responses to calcium signaling depends on the duration, subcellular location, and the amplitude of the change. Calcium signaling can occur in microdomains or macrodomains. Local increases in calcium concentration in neuronal presynaptic terminals triggers neurotransmitter release (Neher and Sakaba, 2008). Global changes such as elevated calcium levels can trigger autophagy, apoptosis amongst other processes in neural cells (Matute et al., 2006; Wojda et al., 2008; Brini et al., 2014). Activation of different signaling pathways depends on the extent of calcium influx; the strength of NMDA-dependent calcium influx determines whether long-term potentiation or depression occurs in hippocampal neurons (Lüscher and Malenka, 2012).

Currently, synthetic calcium flurophores such as Fluo-3 or Fura Red (see **Table 2**) are popular tools for quantifying calcium concentrations since they allow short term spatiotemporal monitoring of Ca2<sup>+</sup> concentration. However, these synthetic dyes require loading, have limitations related to variable dye entry and leakage from cells and thus are not suitable for long-term or in vivo monitoring of calcium levels (Kantner et al., 2015; Thomas and Oliver, 2015). Furthermore, these dyes are not tissue or organelle specific. Genetically encoded calcium indicators (GECI) are used as a non-invasive alternative to measure calcium levels; being genetically expressed, no dye loading step is required and the protein is constantly replenished which is favorable for long term live cell imaging. Another advantage of genetically encoded sensors is the potential for tissue or organelle targeting, which can be achieved by modifying the promoter region or tagging with a localization signal moieties. For example, the ER targeted GECI D1ER was used to measure ER calcium levels in primary hippocampal astrocytes (Williams et al., 2013).

GECI follow two design paradigms: (1) Single fluorescent protein based or (2) FRET-based sensor. Single fluorescent protein based sensors report calcium concentrations based on the intensity of a single fluorophore. The most popular single fluorescent protein based calcium sensors belong to the GCaMP family (Nagai et al., 2001). Recently, a family of ultrasensitive GCaMP calcium sensors have been developed that outperformed other sensors in cultured neurons and in zebrafish, flies and mice in vivo (Chen et al., 2013). GCaMP6 sensors can detect synaptic calcium transients in individual dendritic spines and have proven to be extremely useful for investigations on the organization and dynamics of neural circuits over multiple spatial and temporal scales. FRET-based sensors detect calcium concentration based on ratiometric measurements. An interesting ratiometric tripartite FRET-based Ca2<sup>+</sup> sensor is shown below (**Figure 7**).

FIGURE 7 | A tripartite calcium ion biosensor. The tripartite calcium biosensor relies on the expression of one plasmid-encoded polypeptide: a red, dimerization-dependent fluorescent protein (RFP) linked with calmodulin (CaM), a non-fluorescent partner protein "B," the calmodulin-binding domain of skeletal muscle myosin light chain kinase (M13), and a green, dimerization-dependent fluorescent protein (GFP). In the absence of intracellular calcium (Ca2+) increase, "B" binds with either RFP or GFP, and the ratio of red-to-green signal is in equilibrium. Upon an intracellular increase in calcium (Ca2+), CaM binds Ca2<sup>+</sup> and undergoes a conformational change, allowing it to bind M13. The binding of CaM with M13 causes a shift in the binding equilibrium of "B," and an increase in the red-to-green ratio of the biosensor. Thus, cells expressing the calcium biosensor display red and green fluorescence in equilibrium in the absence of intracellular calcium release. Following an increase in calcium levels, red fluorescence increases in the cytoplasm.

Maysinger et al. Sensors for Neural Cells

A recently optimized family of FRET-based ratiometric GECI "Twitch" showed greater signal amplitude and signal-to-background ratio compared to commercially available synthetic calcium dye Fura-PE3. The overall FRET of Twitch sensors upon calcium binding was improved from 10–20% ratio change initially to >1000%. As with other GECIs, a trade-off can be seen in Twitch sensors between high-affinity binding and fast response kinetics; sensors with lower affinity such as Twitch-5 (K<sup>d</sup> = 9.25µM, τ = 0.16 s) show relatively fast kinetics, whereas the higher affinity binding Twitch-3 (K<sup>d</sup> = 250 nM, τ = 1.5 s) shows slower kinetics. Twitch-2B was able to measure tonic action potentials in mouse visual cortex neurons in vivo, exhibited low cytotoxicity and high signal to noise ratio after 141 days post transfection, demonstrating its usefulness as an in vivo calcium sensor (Thestrup et al., 2014). Since the quantification of calcium depends on ratiometric measurement between a donor and acceptor molecule, FRET-based GECI are less sensitive to artifacts arising from variability in sensor expression, laser intensity, cell thickness, and are preferred for long-term in vivo studies. Detailed procedures describing the calibration, common pitfalls, and organelle specific quantification of calcium levels using FRET-based ratiometric GECIs' was recently published (Park and Palmer, 2015a,b,c).

Advantages of single fluorescent protein GECI are a defined narrow optical spectrum and smaller construct size. A photoactivatable GCaMP GECI was recently designed for selective activation of GECI inside individual cells (Berlin et al., 2015) providing a powerful tool for investigations of neuronal signaling and synaptic plasticity. Recently, an elegant approach was proposed for monitoring brain activity with protein voltage and calcium sensors (Storace et al., 2015). Results from their studies suggest that the voltage sensor ArcLight has certain advantages over GCaMP6 calcium sensors such as optical electrophysiology of mammalian neuronal population activity in vivo.

## Bioengineered Sensors for the Detection and Measurements of pH

In neural cells, as with most other cell types, pH is tightly regulated since both structure and function of cellular proteins depends critically on changes in pH, and changes in pH drive post translational protein modifications, such as phosphorylation. Phosphorylation of protein substrates changes their physical properties (e.g., charge), function (activation and inactivation) and intracellular fate (e.g., translocation from the cytosol to the nucleus and vice versa). Furthermore, detecting changes in H<sup>+</sup> is critical for the function of sensory neurons which use acid-sensing ion channels for nociception and mechanoreception (Omerbašic et al., 2015; Sluka and Gregory, 2015). For instance, extracellular acidification to a pH of 6.9 generates a rapid inward current from acid-sensing ion channels (Waldmann et al., 1997). Properties of acid-sensing ion channels are reviewed elsewhere (Gründer and Pusch, 2015; Krishtal, 2015).

To monitor changes in pH, researchers have turned to fluorescent dyes and fluorescent proteins (probes and sensors). For some studies, it is important to correlate changes in pH with changes in some other intracellular process, such as adenosine triphosphate (ATP) levels. For these studies, one can use dual probes with non-overlapping excitation and emission wavelengths (Tantama et al., 2011). For example, in diabetes research, it has been shown that one can use a GFP-ATP sensor in conjunction with a RFP- pH sensor to measure ATP and cytoplasmic pH simultaneously in glucose deprived cells. Dual biosensors that exploit fluorescent dyes are superior to single pH measurements, but they are more difficult to generate (Fisher and Campbell, 2014). Among the most common measurements of pH changes relate to the endosomal and lysosomal vesicles. The difference in pH between an early endosome and a lysosomes is enormous and can cover 2 pH units corresponding to approximately a 100-fold difference in proton (H+) concentration (Paroutis et al., 2004).

## Bioengineered Sensors for the Detection of Lactate

Neurons and other cells with high energy demand (e.g., heart) become functionally impaired when metabolic processes that normally produce ATP are disrupted. For example, the buildup of lactic acid can lead to neurodegeneration (Ruffin et al., 2014) but can also be neuroprotective when acting on GPR 81 or possibly related receptors (Lauritzen et al., 2014; Tang et al., 2014; Morland et al., 2015; Mosienko et al., 2015). L-lactate is produced by both neurons and astrocytes; moreover, there is a strong evidence that neurons use L-lactate as a supplementary fuel and signaling molecule, and genetically encoded fluorescence nanosensors exist to monitor energy metabolites, such as lactate (Sotelo-Hitschfeld et al., 2015). These sensors are valuable new tools to investigate the lactate pools in models such as dissociated astrocytes in cultures, cortical slices and even in vivo. The result obtained with these sensors show that astrocytes in vitro and in vivo maintain a cytosolic reservoir of lactate, which in response to plasma membrane depolarization, is immediately released into the extracellular space through a lactate-permeable ion channel. These findings support the roles for lactate in neuronal fueling and in gliotransmission (Sotelo-Hitschfeld et al., 2015).

## AuNP-Based Biosensors for Detection of Glucose

A good example of a simple AuNP-based colorimetric assay is that used to detect biological thiols (Ghasemi et al., 2015). The low-molecular-weight biological thiols show high affinity to the surface of AuNPs; this causes the replacement of AuNPs' shells with thiol containing target molecules, leading to the aggregation of the AuNPs through intermolecular electrostatic interaction or hydrogen-bonding. As a result of the predetermined aggregation, AuNPs' color and UV-visible spectra change. The principle of AuNP-based aggregation assays (with an example of protease detection) is illustrated in **Figure 3**. These AuNP-based colorimetric assays have been used to detect oxidative stress in neurons and non-neuronal cells (Kumar et al., 2013; Wang et al., 2015b). Glucose not only represents the primary energy source for the brain, but also plays important roles in synaptic transmission. In this assay, the aggregation of AuNPs was induced by glucose through cascade reactions involving glucose, H2O2, and <sup>∗</sup>OH (Jiang et al., 2010).

## NEURONAL AND GLIAL RESPONSES TO NANOSTRUCTURES

#### Neural Stimulation Using Nanostructures

Recent developments in NP technologies provide new approaches for recording and stimulating nerve cells. Among these are the incorporation of carbon nanotubes (CNTs) to improve implantable three-dimensional (3D) microelectrode arrays (MEA) to record nerve activity in large numbers of neurons in regional circuits (Gabay et al., 2007; Kim et al., 2014; Monaco and Giugliano, 2014). The advantage of using CNTs in the design of MEAs is that CNTs are chemically inert and stable. Furthermore, CNTs exhibit excellent electrical conductivity and, most importantly, biocompatibility with neurons (Lin et al., 2009; Gerwig et al., 2012; Musa et al., 2012; David-Pur et al., 2014; Samba et al., 2014). When embedded in polymer film, these CNT electrode arrays provide a flexible device to record activity that can be implanted in the brain.

In addition, promising new developments with NPs have demonstrated the feasibility of using heat generated from targeted NPs to control the activity of specific populations of neurons within a particular brain region. In one approach, neurons were stimulated with localized heat from gold nanorods (AuNRs) after irradiation with NIR laser light (Paviolo et al., 2014; Nakatsuji et al., 2015). Briefly, AuNRs are plasmonic nanoparticles that absorb minimally invasive NIR light and achieve highly localized photothermal heat generation (Paviolo et al., 2014). AuNRs can be targeted to the plasma membrane by coating them with a genetically cationized form of highdensity lipoprotein (HDL; Nakatsuji et al., 2015). Not only does this complex reduce the cytotoxicity of the AuNRs but it also localizes the generated heat to neurons to enable the activation of heat-sensitive membrane ion channels, such as transient receptor potential vanilloid (TRPV) family members (Tominaga et al., 1998). This interesting approach has been used to stimulate cultured mouse sensory neurons from dorsal root ganglia, neurons that express TRPV1 endogenously, yet, this method has not been applied to neurons in vivo. One limitation is the poor penetration of NIR in intact neural tissue.

An interesting alternative technique uses magnetic nanoparticles (MNPs) to generate heat. Briefly, MNPs convert alternating magnetic fields into biological stimuli by dissipating heat through hysteretic power loss. Low-radiofrequency alternating magnetic fields (100 kHz to 1 MHz) can penetrate into the body without substantial attenuation and thus enable signal delivery into deep brain regions. Moreover, radiofrequency (RF) magnetic fields can be applied remotely, allowing for noninvasive remote stimulation of neurons in awake behaving animals (Huang et al., 2010; Chen et al., 2015). In addition to neuroscience, targeted magnetic NPs are being investigated in cancer therapy; as well, this method can be applied to manipulate remotely signal transduction pathways and other cellular machinery (Bonnemay et al., 2015).

The feasibility of using magnetic hyperthermia to stimulate neurons was demonstrated using coated manganese ferrite (MnFe2O4) MNPs conjugated with streptavidin (Huang et al., 2010). These particles were targeted to the surface of cultured hippocampal neurons overexpressing TRPV1. When exposed to low RF magnetic fields, the MNPs generated sufficient heat to activate the TRP channels and depolarize neurons without causing cell damage (Huang et al., 2010). This work established magnothermal stimulation as an attractive non-invasive method to excite specific neurons.

There has been some concern, however, that MNPs conjugated with proteins could become internalized and/or reduce the effectiveness of targeting and heat dissipation in vivo. To overcome these potential issues, Fe3O<sup>4</sup> MNPs replaced MnFe2O<sup>4</sup> MNPs. Untargeted Fe3O<sup>4</sup> MNPs that have been optimized for efficient heat dissipation at clinically relevant alternating magnetic fields (Chen et al., 2015). These Fe3O<sup>4</sup> MNPs have high heating rates, and when exposed to therapeutically relevant frequencies, they can trigger widespread firing of cultured hippocampal neurons expressing TRPV1. An attractive aspect of magnetothermal stimulation is its ability to stimulate neurons in deep brain structures; a good example is use in stimulating neurons in the ventral tegmental area (VTA; Chen et al., 2015). Since VTA neurons do not express TRPV1 channels endogenously, neurons were infected with lentivirus expressing TRPV1 cDNAs, the region was injected with Fe3O<sup>4</sup> MNPs, and the animals were exposed to alternating magnetic fields. Even after 1 month of MNP injection, magnetic field stimulation triggered a significant increase in neural activity in the vicinity of the MNP injection site, as indicated by immediate early gene c-fos expression (Chen et al., 2015).This work demonstrates the feasibility of remote, wireless magnetothermal stimulation to activate neurons in deep brain areas.

Both NIR-AuNR and RF-activation of MNPs provide interesting approaches for stimulating neurons. However, this method can be applied only to neurons that express heatsensitive ion channels, either endogenously, such as peripheral sensory neurons, or after expression with virally-mediated gene transfer. To enhance these approaches, it might be attractive to adapt them for use in exciting new implantable wireless fluidic devices that have been developed for programmable in vivo pharmacology (Jeong et al., 2015), thereby providing an interesting alternative to established optogenetics techniques (Warden et al., 2014). Optogenetic approaches and their attractions and limitations have been extensively reviewed elsewhere (Williams and Deisseroth, 2013; Thompson et al., 2014; Fan and Li, 2015; Kale et al., 2015; Lüscher et al., 2015; Tonegawa et al., 2015; Webber et al., 2015).

#### Glial Response to Nanostructures

Collectively, glial cells (microglia and astrocytes) are equipped with sophisticated sensing, transducing and amplifying machinery that outperforms any artificial sensors. Microglia and astrocytes use toll–like receptors (TLRs) to sense pathogen signals, and those from nanoparticles. TLRs will recognize the "stranger" (e.g., nanoparticle) similar to the recognition of a pathogen (e.g., bacteria; Hanke and Kielian, 2011; Okun et al., 2011; Harry, 2013; Schaefer, 2014). TLR4 recognizes lipopolysaccharide (LPS) produced by Gram-negative bacteria and also nanoparticles associated with LPS (Lalancette-Hébert et al., 2010). TLR4 responds transiently to cerium oxide NPs (nanoceria). In contrast, unprotected ("naked") CdTe QDs cause a strong microglia activation leading to robust luciferase activation as shown in vivo in transgenic mice expressing luciferase driven under the control of glial fibrillary acidic protein promoter (Maysinger et al., 2007). Similarly, when microglia are exposed to AuNPs, the intensity and temporal pattern of the TLR2 responses varies with the configuration of the NP. For example, in transgenic mice, gold nanorods exert a bimodal activation of microglia in transgenic mice with TLR2 promoter-luciferase reporter (Hutter et al., 2010).

Once TLRs have "sensed" the nanoparticle and other danger signals, the transduction system becomes engaged. This system includes IkB kinases, MAP-kinases and a number of transcription factors such as NFkB, AP-1 and interferon regulator factor (IRF) families (Takeuchi and Akira, 2010). Following the initial detection and transduction stages, signals recruit highly inducible genes, such as cytokines (e.g., interleukine 1beta, tumor necrosis factor alpha and others), which serve as amplifiers of the inflammation program (Glass et al., 2010). Under pathological conditions, microglia commonly assume a macrophage phenotype; however, in the healthy central nervous system, microglia do not behave as macrophages and their precise functions are currently under intense investigation (da Fonseca et al., 2014; Chen and Trapp, 2015; Tremblay et al., 2015).

### CHALLENGES AND LIMITATIONS

#### Fluorescent Dyes

Fluorescent molecules ("dyes") have played a major role in intracellular sensing and imaging in neural and other cells, in large part, because they respond rapidly to stimuli, have high signal intensities, and enter neural cells noninvasively. Consequently, these molecules are considered the current standard method for the quantification of intracellular analytes (Haugland, 2005; Lakowicz, 2006); none the less, some limitations compromise their usefulness. These include: (1) Possible interference with cellular processes and cytotoxicity; (2) organic solvents required for the dissolution of lipophilic probes; (3) unpredictable cellular responses when dyes interact with intracellular constituents; (4) relatively rapid bleaching limits their usefulness for time lapse experiments; (5) Only few dyes allow for ratiometric measurements.

#### QDs

Although the photophysical advantages of QDs for experiments at the single cell level orin vivo experiments are greater than those of fluorescent dyes, the current generation of QD-based sensors has some drawbacks. For example: (1) the highest quantum yields are from QDs that contain toxic components, such as Hg; (2) QDs that emit signals in the near infrared spectrum are currently too large to enter neural cells unless their surfaces are adequately modified QD-based sensors, (3) and most QDs are not readily eliminated from the body and accumulate in liver and kidneys.

Ideas to reduce or prevent QD toxicity have been reviewed recently (Winnik and Maysinger, 2013). The extent of QD toxicity depends on its core composition, size, shape, surface coating, ligand arrangement, and charge (Hoshino et al., 2004; Jiang and Asryan, 2008; Verma and Stellacci, 2010; Kauffer et al., 2014). The mechanisms that account for the toxicity have not been fully resolved. They may involve the formation of reactive oxygen species (ROS) due to the degradation of the QD core and the release of free cadmium ions (Derfus et al., 2004), followed by oxidative stress and inflammation (Lovric et al., 2005; Manke et al., 2013). Or, exposure to QDs may cause cell growth inhibition and lipid peroxidation (Choi et al., 2007) as well as epigenetic and genetic changes (Choi et al., 2008; Stoccoro et al., 2013).

## AuNP and Other NPs

AuNP-based sensors and enzymatic assays are not widely used and exist mainly as basic research tools. Colorimetric assays have a good potential for high-throughput applications: they are robust, simple, inexpensive, and require minimal instrumentation. However, colorimetric measurements are not always easily adapted for complex biological environments, such as tissues or living organisms, due to the substantial interference from cellular macromolecules. The detection limits of AuNP-based assays are surprisingly low, suggesting remarkable sensitivity and are promising for high throughput screening of enzyme inhibitors and activators. For measurements of enzymatic activities employing AuNPs in cells, FRETbased assays are more suitable then colorimetric measurements (Freeman et al., 2013; Lindenburg and Merkx, 2014; Chou and Dennis, 2015; LaCroix et al., 2015; Shamirian et al., 2015).

The ratio of acceptor to donor fluorescence is usually used as a surrogate for actual FRET efficiency. FRET efficiency can also be inferred from the rate of photobleaching of the donor or acceptor. Both of these approaches are qualitative and difficult to quantitate because the concentrations of fluorophore at the specific intracellular site is unknown. A more sophisticated approach, one that overcomes the problem of inconsistent fluorophore concentrations is fluorescence life time imaging (FLIM), an approach that relies on lifetimes of fluorophores and QDs (Murakoshi et al., 2008; Ueda et al., 2013; Doré et al., 2014; Chen et al., 2014b; Datta et al., 2015; Kaur et al., 2015; Yellen and Mongeon, 2015).

With any nanoparticle-based sensor, of course, it is important to take into consideration the possibility of nonspecific binding and ligand exchange. For example, in the case of AuNPs, with thiolated ligands, the ligands can be exchanged with intracellular glutathione. To prevent ligand exchange and nonspecific binding, AuNPs can be PEGylated. The main challenge still is monitoring the enzymatic processes in vivo, in a longitudinal, real-time manner; for these purposes, cell type specific expression of genetically engineered proteins remains a promising alternative.

## LOOKING AHEAD TO NEW WAYS TO SENSE NEURAL CELLS UNDER PHYSIOLOGICAL AND PATHOLOGICAL CONDITIONS

Many pharmacological interventions in neurological disorders fail because the initiation of intervention started too late and the damage to the tissues was too extensive and irreversible. One of the reasons for such a failure is the lack of adequate tools to detect the changes in neural and other cells early enough when pathology can be arrested and deleterious consequences minimized. Nanosensors are beginning to change the diagnostic arena but they are still at an early stage of development. Nanoparticle-based and bioengineered sensors as well as fluorescent molecular probes should be used in a complementary manner. New nanoparticle-based lipase sensor sparked interest in development of sensors for other lipases and could be used to reveal the role of lipids in neurodegeneration (Tang et al., 2015).

As opposed to the well-studied and most frequently used spherical particles, many other particle morphologies have optical resonances in the near-infrared spectral window, allowing a deeper penetration within tissues and an absence of photochemical damage; these features are highly advantageous for intracellular or in vivo imaging applications. Only a few designs employing gold nanorods (Cheng et al., 2015; Park et al., 2015; Zhang et al., 2015a) or nanocages (Chen et al., 2005a,b) exploit the tunability of surface plasmon absorption peak and optimize the sensitivity of AuNP-based assay.

For two-photon luminescence (TPL) gold nanorods and other anisotropic morphologies should be exploited. TPL imaging, which is superior to dark-field imaging in terms of signal-tonoise ratio, is an attractive option for intracellular imaging, where a strong background signal often hampers the detection of low concentrations of AuNPs. TPL has been employed mainly for bioimaging (Vo-Dinh et al., 2013; Yellen and Mongeon, 2015), but it should be considered for detection of changes in enzymatic activities in neural cells under physiological and pathological conditions. Luminescent AuNPs with high quantum yields are attractive candidates to replace QDs or some organic fluorophores (Maysinger and Hutter, 2015). A "plasmonic resonance energy transfer (PRET)" can be exploited for measurements of enzymes engaged in disrupted redox homeostasis, neuroinflammation and protein shedding (cleavage of the ectodomain of membrane proteins; Altmeppen et al., 2013; Saftig and Bovolenta, 2015).

Nanoparticle-based and bioengineered sensors presented here were mostly tested in non-neural and some neural cells but their employability under true pathological conditions remain to be investigated. Currently available sensors require considerable improvements to provide reliable, reproducible and simple measurements of biomarker concentrations, duration of the processes and their precise location. The internalization of nanoparticle-based sensors still remains a challenge.

Nanosensing is a very dynamic field, and the employability of the new designs and the proportions of targeted application areas keep changing. In the near future, advances in nanotechnology and imaging techniques combined with electrophysiological recordings, could elucidate critical signaling players under physiological and pathological conditions thereby providing the way of more successful testing of new therapeutic interventions in neurological disorders. Combined approaches employing electrophysiology, bioengineering and nanotechnology could contribute to finding ways of getting "out of clutter and finding simplicity" (Albert Einstein).

### ACKNOWLEDGMENTS

The authors acknowledge Dr. Jack Diamond, an enthusiastic neuroscientist who inspired us to write this review. The editorial assistance by Ms. Linda Cooper is also acknowledged. Ms. Issan Zhang is acknowledged for two illustrations and reading the manuscript. Financial support was provided by Canadian Institutes of Health Research (CIHR, MOP-119425) and National Sciences and Engineering Research Council of Canada (NSERC, RGPIN-2015-04994).

## REFERENCES


of drug-induced apoptosis in a 3D breast tumor model. Biotechnol. Bioeng. 112, 1673–1682. doi: 10.1002/bit.25572


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer Ruxandra Vidu and handling Editor declared a current collaboration and the handling Editor states that the process nevertheless met the standards of a fair and objective review.

Copyright © 2015 Maysinger, Ji, Hutter and Cooper. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# **Mimicking Collective Firing Patterns of Hundreds of Connected Neurons using a Single-Neuron Experiment**

*Amir Goldental 1 †, Pinhas Sabo1 †, Shira Sardi 1, 2, Roni Vardi 2 † and Ido Kanter 1, 2\**

<sup>1</sup> Department of Physics, Bar-Ilan University, Ramat-Gan, Israel, <sup>2</sup> Gonda Interdisciplinary Brain Research Center and The Goodman Faculty of Life Sciences, Bar-Ilan University, Ramat-Gan, Israel

The experimental study of neural networks requires simultaneous measurements of a massive number of neurons, while monitoring properties of the connectivity, synaptic strengths and delays. Current technological barriers make such a mission unachievable. In addition, as a result of the enormous number of required measurements, the estimated network parameters would differ from the original ones. Here we present a versatile experimental technique, which enables the study of recurrent neural networks activity while being capable of dictating the network connectivity and synaptic strengths. This method is based on the observation that the response of neurons depends solely on their recent stimulations, a short-term memory. It allows a long-term scheme of stimulation and recording of a single neuron, to mimic simultaneous activity measurements of neurons in a recurrent network. Utilization of this technique demonstrates the spontaneous emergence of cooperative synchronous oscillations, in particular the coexistence of fast γ and slow δ oscillations, and opens the horizon for the experimental study of other cooperative phenomena within large-scale neural networks.

#### *Edited by:*

Mikhail Lebedev, Duke University, USA

#### *Reviewed by:*

Benjamin Lindner, Bernstein Center for Computational Neuroscience, Germany Wolfgang Kinzel, University of Würzburg, Germany

#### *\*Correspondence:*

Ido Kanter ido.kanter@biu.ac.il

† These authors have contributed equally to this work.

#### *Specialty section:*

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

*Received:* 29 October 2015 *Accepted:* 21 December 2015 *Published:* 20 January 2016

#### *Citation:*

Goldental A, Sabo P, Sardi S, Vardi R and Kanter I (2016) Mimicking Collective Firing Patterns of Hundreds of Connected Neurons using a Single-Neuron Experiment. Front. Neurosci. 9:508. doi: 10.3389/fnins.2015.00508 **Keywords: neuronal plasticity, neural networks,** *in-vitro***, neuronal response latency, neuronal response failures**

## **INTRODUCTION**

One of the fundamental goals in neuroscience is to understand the mechanisms underlying the emergence of time-dependent collective activities of neural networks (Silva et al., 1991; Gray, 1994; Contreras et al., 1997; Buzsaki and Draguhn, 2004; Buzsaki, 2006; Chialvo, 2010). This understanding will shed light on the way the brain reliably analyzes information and generates behavior (Klimesch, 1999; Basar et al., 2001; Wiest and Nicolelis, 2003; Kahana, 2006; Bollimunta et al., 2008; Fries, 2009; Giraud and Poeppel, 2012). The experimental accomplishment of this goal requires the following two advanced abilities. The first ability is to record from a large number of neurons over a period of seconds and minutes, which reflects the time scale of the collective network phenomena. The second ability is to know all network parameters, e.g., the network connectivity, synaptic delays and synaptic strengths (**Figure 1A**). Thus, the number of simultaneous measurements has to be in the order of the number of neurons and synapses (**Figure 1B**). Although the technology of electrophysiological measurements was significantly enhanced during the last decades, there is not yet such a technology which can record from thousands of individual neurons with a single-cell resolution (Marx, 2014) concurrently with realtime gathering of detailed network topology, including synaptic strengths and delays (Pastrana, 2012).

**FIGURE 1 | Illustration of the fundamental experimental difficulty. (A)** An illustration of a neural network. Synaptic strengths and synaptic delays are indicated by the brightness and length of the connections, respectively. The different properties of each neuron are indicated by different colors and shapes. **(B)** The knowledge of the current neuronal and synaptic properties requires an enormous number of measurements carried out by many devices (green), e.g., extracellular and intracellular electrodes, inserted in specific targeted spots in the network. **(C)** The large number of measurements and inserted devices may change the properties of the network. This is schematically exemplified by the difference between the shaded network [identical to the initial network in **(A)**] and the interfered network as a result of the measurements (front colored network).

It is impartial to assume that the implementation of an enormous number of measurements on the network will influence its activity, and as a byproduct will modify the network parameters. Hence, as a result of many measurements, the estimated network parameters will differ from either the original or from the actual ones (**Figure 1C**). All in all, this limitation puts in question the ability to experimentally pinpoint the quantitative interplay between the network properties and its functionalities. This limitation reminds the fundamental quantum measurement difficulties (Braginsky et al., 1995), where a measurement affects the state of the system. Although in this case there is no physical principle that prohibits an accurate measurement, the multimeasurements are expected to modify the network and induce unavoidable learning processes, preventing flawless real-time estimations.

We present and utilize a real-time experimental long-term single-neuron stimulation and recording scheme which allows the study of the collective firing activity of a recurrent neural network, given its synaptic strengths and delays. It extends previous attempts to understand network behavior from iterative stimulation and simulations of single cells (Reyes, 2003; Lerchner et al., 2006; Brama et al., 2014; Dummer et al., 2014). Hence, the robustness of the collective firing phenomena can be examined for different sets of synaptic delays and strengths. The presented experimental scheme serves as a mirror image of the reverse engineering methods (Csete and Doyle, 2002; Gregoretti et al., 2010), where the topology of the recurrent network is estimated from its activity and verifies recent simulations and theoretical results, which predicted similar cooperative oscillations in excitatory networks (Goldental et al., 2015).

## **MATERIALS AND METHODS**

#### **Experimental Procedures Animals**

All procedures were in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and the University's Guidelines for the Use and Care of Laboratory Animals in Research and were approved and supervised by the Institutional Animal Care and Use Committee.

### *In vitro* **Experiments Culture Preparation**

Cortical neurons were obtained from newborn rats (Sprague-Dawley) within 48 h after birth using mechanical and enzymatic procedures. The cortical tissue was digested enzymatically with 0.05% trypsin solution in phosphate-buffered saline (Dulbecco's PBS) free of calcium and magnesium, and supplemented with 20 mM glucose, at 37◦C. Enzyme treatment was terminated using heat-inactivated horse serum, and cells were then mechanically dissociated. The neurons were plated directly onto substrateintegrated multi-electrode arrays (MEAs) and allowed to develop functionally and structurally mature networks over a time period of 2–3 weeks *in vitro*, prior to the experiments. Variability in the number of cultured days in this range had no effect on the observed results. The number of plated neurons in a typical network was in the order of 1,300,000, covering an area of about 380 mm2. The preparations were bathed in minimal essential medium (MEM-Earle, Earle's Salt Base without L-Glutamine) supplemented with heat-inactivated

composed of 30 inter-stimulation intervals of 62.5 ms (16 Hz) and 30 inter-stimulation intervals of ∼111 ms (9 Hz) (black). The probability for an evoked spike (green

horse serum (5%), glutamine (0.5 mM), glucose (20 mM), and gentamicin (10 g/ml), and maintained in an atmosphere of 37◦C, 5% CO2, and 95% air in an incubator as well as during the electrophysiological measurements.

circles) indicates a fast adaptation, short memory, of the neuronal response probability.

#### **Synaptic Blockers**

All experiments were conducted on cultured cortical neurons that were functionally isolated from their network by a pharmacological block of glutamatergic and GABAergic synapses. For each culture 20μl of a cocktail of synaptic blockers was used, consisting of 10μM CNQX (6-cyano-7-nitroquinoxaline-2,3-dione), 80μM APV (amino-5 phosphonovaleric acid) and 5μM bicuculline. This cocktail did not block the spontaneous network activity completely, but rather made it sparse. At least 1 h was allowed for stabilization of the effect.

#### **Stimulation and Recording**

An array of 60 Ti/Au/TiN extracellular electrodes, 30μm in diameter, and spaced 500μm from each other (Multi-Channel Systems, Reutlingen, Germany) were used. The insulation layer (silicon nitride) was pre-treated with polyethyleneimine (0.01% in 0.1 M Borate buffer solution). A commercial setup (MEA2100-2x60-headstage, MEA2100-interface board, MCS, Reutlingen, Germany) for recording and analyzing data from two 60-electrode MEAs was used, with integrated data acquisition from 120 MEA electrodes and eight additional analog channels, integrated filter amplifier, and three-channel current or voltage stimulus generator (for each 60 electrode array). Mono-phasic square voltage pulses typically in the range of [−800, −500] mV and [60, 400] μs were applied through extracellular electrodes. Each channel was sampled at a frequency of 50 k samples/s, thus the changes in the neuronal response latency were measured at a resolution of 20μs.

#### **Cell Selection**

A neuron was represented by a stimulation source (source electrode) and a target for the stimulation—the recording electrode (target electrode). These electrodes (source and target) were selected as the ones that evoked well-isolated, well-formed spikes, and reliable response with a high signal-to-noise ratio. This examination was done with a stimulus intensity of −800 mV with a duration of 200μs using 30 repetitions at a rate of 5 Hz, followed by 1200 repetitions at a rate of 10 Hz.

#### **Data Analysis**

Analyses were performed in a Matlab environment (MathWorks, Natwick, MA, USA). The reported results were confirmed based on at least eight experiments each, using different sets of neurons and several tissue cultures. Action potentials were detected on-line by threshold crossing, using a detection window

of typically 2–10 ms following the beginning of an electrical stimulation.

#### **Implementation of the Mimicking Scheme**

The scheme is based on the neuronal short term memory (**Figure 2**), and the management of the stimulation history of each of the mimicked neurons, as well as the timings of their evoked spikes, is done in real-time as exemplified in **Figure 3**. A simplified version of the scheme is presented in **Figure 4** in the form of a flowchart as well as in the Supplementary Movie. A detailed description of the mimicking procedure is presented in the Appendices (Supplementary Material).

### **RESULTS**

When a neuron is stimulated repeatedly, the time-lag between a stimulation and its corresponding evoked spike, the neuronal response latency (NRL), stretches gradually (Wagenaar et al., 2004; De Col et al., 2008; Vardi et al., 2014, 2015; **Figure 2A** and Materials and Methods). Above a critical stimulation frequency, f*c*, which varies much among neurons (Vardi et al., 2015), this stretching terminates at the intermittent phase. This phase is characterized by large fluctuations around a constant NRL and by neuronal response failures, NRFs (**Figure 2A**). The non-zero fraction of NRFs is such that the average firing frequency is f*c*, independent of the stimulation frequency; hence, the neuron operates similar to a low pass filter (Vardi et al., 2015; **Figure 2B**). In addition to the preservation of the neuron's average firing frequency under periodic stimulations, the response failures were found to be statistically independent (Vardi et al., 2015; **Figure 2C**). Specifically, for inter-stimulation-intervals that are longer than the refractory period, the firing probability is independent of the neuron's firing history. In the general stimulation scenario, aperiodic stimulations, the statistics of

the NRFs were experimentally found to depend on the shortterm stimulation history of the neuron, which typically consists of several stimulations only (Vardi et al., 2015; **Figure 2D**). These effects might be an indirect result of some kind of spike-frequency adaptation (Benda and Herz, 2003) or a related mechanism.

The proposed experimental technique allows the mimicking of the activity of a neural network, given the features of the connections and the initial condition of the firing neurons. For the sake of simplicity, we first demonstrate the utilization of the proposed method using a diluted network with abovethreshold synapses and with uniform delays between neurons, τ. In such a case the history of a network appears as consecutive "snapshots" of the network separated by τ time-lags between them (**Figure 3A**). Each "snapshot" of the network defines which are the stimulated neurons and which neurons fire at that time. Specifically, each neuron in each snapshot belongs to one of the following three states: received a stimulation that results in an evoked spike, received a stimulation that results in a response failure, or did not receive stimulation at that time (**Figure 3B**). The neurons to be stimulated in the consecutive snapshot are determined according to the network connectivity (**Figure 3B**). For example, assume neuron A is pre-synaptic to neuron B and neuron A fires at time T, consequently neuron B is stimulated at time T + τ. Neurons in the network are stimulated either if their pre-synaptic neurons fired at the previous snapshot, or if they are stimulated by a stochastic noise, e.g., synaptic noise. An example is presented in **Figure 3C1**, given the network dynamics until the snapshot at time T + 4τ, three neurons will receive a stimulation at the next snapshot, T +5τ (**Figure 3C1**). The

#### **FIGURE 5 | Continued**

and 2 post- above-threshold synaptic connections, and all delays are set to 13 ms. Results, produced using a single-neuron experiment in vitro, indicate f<sup>δ</sup> ∼2.5 Hz oscillations which coexist with f<sup>γ</sup> ∼75 Hz oscillations (inset). **(B)** Similar to **(A)** where each neuron has randomly selected 50 pre- and 50 postbelow-threshold synaptic connections, and all delays are set to 15 ms. An above-threshold stimulation requires cooperation of at least four below-threshold stimulations. Results indicate f<sup>δ</sup> ∼0.8 Hz oscillations which coexist with f<sup>γ</sup> ∼65 Hz oscillations (inset). **(C)** A raster plot of a network consisting of 500 neurons where each neuron has randomly selected 2 preand 2 post- above-threshold synaptic connections, and delays are randomly selected from the uniform distribution U(8,12) ms. Results indicate f<sup>δ</sup> ∼1.3 Hz oscillations which coexist with spontaneous f<sup>γ</sup> ∼65 Hz oscillations, originated from 1/(average(τ+latency)) (inset). The rate is calculated from the number of spikes in a sliding window of 20 ms with a resolution of 0.1 ms. **(D)** The NRL of the mimicking neuron in **(A)** (response failures are denoted at NRL = 3 ms). The stimulation rate (upper orange curve) and firing rate (lower orange curve) are calculated using a sliding average of 2000 stimulations. The average stimulation rate is much higher than fc ∼5 Hz, indicating that the neuron is in the intermittent phase, which is characterized by large fluctuations of the NRL and response failures which lead to a firing frequency around fc ∼5 Hz.

goal now is to determine whether these three neurons will fire, based on their short-term stimulation history. This task is done experimentally *using a single mimicking neuron (in vitro or in vivo)* (see Supplementary Movie and Materials and Methods) and is based on the following two steps:

*The mimicking step*: The current responsiveness, response susceptibility to stimulations, of a neuron from the network is mimicked by the enforcement of its short-term stimulation history on the mimicking neuron (see Supplementary Movie), e.g., three last stimulations at **Figures 3C2-4**. After the completion of this step, the mimicking neuron will have the same responsiveness as the mimicked neuron in the current state of the network (**Figure 2D**).

*The responsive test*: τ [ms] after the termination of the first step, the mimicking neuron is stimulated. In case of an evoked spike, we conclude that the mimicked neuron in the network fires and this event is noted in the current snapshot.

For each stimulated neuron in the snapshot these two steps are repeated sequentially in *real-time* (**Figure 3C5**), using the same mimicking neuron, until the responsiveness of all neurons in the current snapshot is determined (**Figure 3C6**). After the snapshot at time T + 5τ was completed, the procedure is repeated to determine the state of the next snapshot, T + 6τ (**Figure 3C6**), and so on.

The massive management of the stimulation history of each neuron in the network as well as their spike timings is done in *real-time* (Materials and Methods), demanding faster operations in at least two orders of magnitude than the time scale of τ.

The realization of the proposed real-time method is first demonstrated for an excitatory network consisting of 500 neurons, using a cultured mimicking neuron (Appendix A in Supplementary Material). Each neuron has 2 pre- and 2 postsynaptic connections, all are above-threshold and randomly chosen, τ = 13 ms, with additional external stimulation Poissonian noise with a rate of 1 Hz (Materials and Methods). The network dynamics over ∼2 s indicates δ oscillations of ∼2.5 Hz which coexist with γ oscillations of ∼75 Hz (**Figure 5A**), which trivially originates from the resolution 1/τ (**Figure 3C**).

This prototypical real-time technique is realized in a more realistic biological network, consisting of sub-threshold synapses as well (Appendix B in Supplementary Material and **Figure 5B**). The excitatory network consists of N=500 neurons, where each neuron has *50 pre- and 50 post-synaptic connections* with τ = 15 ms, with an additional 1 Hz Poissonian noise (stimulations). An above-threshold stimulation requires the firing of at least four pre-synaptic neurons, or a stimulation originated from the noise. Results indicate δ oscillations of ∼0.8 Hz which coexist with γ oscillations of ∼65 Hz (**Figure 5B**), which again originates from the resolution 1/τ.

The generalization of the proposed real-time technique for networks with a continuous distribution of connection delays requires a complicated procedure, since the scheme of discrete time snapshots (**Figure 3B**) is not valid in this case. The advanced procedure requires the management of the stimulation history and the timings of the evoked spikes of all neurons in a continuous time manner. The mimicking process per neuron is similar, however, technically the complexity of the algorithm is enhanced since the constraint of specific simulations and firing times is released and occur in continuous time. Utilization of the continuous scheme indicates the coexistence of δ and spontaneous γ oscillations (Appendix C in Supplementary Material with **Figure 5C** and Appendix D in Supplementary Material with **Figure 6**), where the periods of collective firing are slightly smeared as a result of continuous connection delays.

In the case where all connection delays are equal to τ, the GCD of loops of such random networks is expected to be equal to τ (Kanter et al., 2011; Vardi et al., 2012a,b). In such a case, neurons will fire in synchrony every τ [ms], therefore forming γ oscillations with frequency of 1/τ (**Figures 5A,B**). On the other hand, in case of random continuous connection delays, the GCD vanishes and no synchrony is expected beyond the δ oscillations. Our results clearly indicate that the non-trivial high frequency synchrony is dominated by the average delay, i.e., the *spontaneously originated* γ *oscillations* have the frequency of 1/(average delay) as also observed in simulations (Goldental et al., 2015). The distribution of the connection delays affects only the quality of the synchrony (**Figure 5C**).

Mimicking the dynamical behavior of a network consisting, for instance, of thousands of neurons over several seconds requires real-time stimulations and recordings of the mimicking neurons over several hours. Specifically, the real-time duration of the experiment is equal to the number of stimulations occurred dynamically in the network, multiplied by the time it takes to mimic a neuron. For illustration, in **Figure 5A**, a network of N = 500 neurons is mimicked for 2 s. Since, fc = 5 Hz each one of the neurons in the network was mimicked approximately (2 s) · (2fc) = 20 times. The mimicking of one stimulated neuron in the network, requires approximately 0.4 s. Hence, the total real-time of the experiment with a single mimicking neuron is expected to be 20 · N · 0.4 s = 4000 s, which is indeed close to ∼3700 s (**Figure 5D**). During this period, the mimicking neuron remains in the intermittent phase

as indicated by the large fluctuations of the NRL and the response failures which resulted from the high stimulation rate (**Figure 5D**).

#### **DISCUSSION**

The presented experimental results verify recent simulations and theoretical work which predicted such oscillations (Goldental et al., 2015). The experimental scheme presents more reliable evidence since it takes into account biological time dependent fluctuations in the responsiveness of neurons and variations in the neuronal critical frequency, as opposed to the simulations and theory.

Currently, there are some limitations to the proposed mimicking method, which is based on short-term neuronal dynamics. Long-term effects and synaptic plasticity are ignored, however they are not expected to dominate the dynamics of the network within several seconds (**Figures 5**, **6**). It might be possible to introduce synaptic dynamics, excitatory and inhibitory, to the mimicking process by stimulating and recording from coupled neurons through synaptic connections, using patch clamp technique (Debanne et al., 2008). Currently this kind of dynamics is simplified to excitatory electrical pulses.

Experimental difficulties arise when the mimicked network is composed of thousands of neurons. Primarily, the experimental time scales linearly with the size of the network, hence it is expected to exceed several hours. Preliminary results (not shown) indicate that it is possible to mimic a network of thousands of neurons, however sailing toward much larger systems is in question. A possible bypass to this obstacle, and a way to mimic more heterogeneous networks, with several types of neurons, is to implement several mimicking neurons in parallel, however it will require the realization of a much more complicated experimental scheme.

The idea of mimicking network dynamics using a single neuron was previously demonstrated for feed-forward networks (Reyes, 2003), where the parameters of the network are adjusted to control the activity, and the mimicking process does not take into consideration short-term neuronal plasticity. This work, on the other hand, examines recurrent random networks, where the parameters are independent of the stability of the firing rates. In addition, the average delay between successive layers in a feedforward network is irrelevant for the dynamics, since it only shifts the time of the activity by a constant. In contrast, in recurrent networks, the exact delay times are important since each neuron is affected by many delay loops. Hence, the implementation of the mimicking process of a recurrent network consisting of short delays of several milliseconds is a challenge. Additionally, as a result of the many loops, each neuron is revisited many times through the dynamics. Hence, the mimicking process is done many times per neuron, and keeping the network parameters fixed is essential to describe the dynamical properties of the recurrent network.

#### **REFERENCES**


The proposed real-time experimental method can also be used to mimic the firing patterns of large recurrent neural networks *in vivo*, based on long-term scheme of stimulation and recording of a single neuron *in vivo* (Brama et al., 2014). Nevertheless, the real-time management of the *in vivo* mimicking process, where delays are several milliseconds only, is still an experimental challenge.

The presented experimental technique to use a long-term experiment on a single node in order to mimic the parallel activity of a large scale network is applicable to a variety of networks with propagation delays, where the nodes exhibit a finite or no memory of the preceding conditions. Thus, this technique is expected to be relevant to a wide range of networks that play a key role in other fields such as physics, biology and economics.

#### **AUTHOR CONTRIBUTIONS**

AG developed the theoretical framework and compared experimental results to simulations. PS developed and designed the interface for the real-time experiments. RV and SS prepared the tissue cultures and the experimental materials. All authors performed the experiments and analyzed the data. IK supervised all aspects of the work. All authors discussed the results and commented on the manuscript.

#### **ACKNOWLEDGMENTS**

This research was supported by the Ministry of Science and Technology, Israel.

#### **SUPPLEMENTARY MATERIAL**

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnins. 2015.00508


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Copyright © 2016 Goldental, Sabo, Sardi, Vardi and Kanter. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Nanomaterial-Enabled Neural Stimulation

#### Yongchen Wang<sup>1</sup> and Liang Guo2, 3 \*

*<sup>1</sup> Department of Biomedical Engineering, The Ohio State University, Columbus, OH, USA, <sup>2</sup> Department of Electrical and Computer Engineering, The Ohio State University, Columbus, OH, USA, <sup>3</sup> Department of Neuroscience, The Ohio State University, Columbus, OH, USA*

Neural stimulation is a critical technique in treating neurological diseases and investigating brain functions. Traditional electrical stimulation uses electrodes to directly create intervening electric fields in the immediate vicinity of neural tissues. Second-generation stimulation techniques directly use light, magnetic fields or ultrasound in a non-contact manner. An emerging generation of non- or minimally invasive neural stimulation techniques is enabled by nanotechnology to achieve a high spatial resolution and cell-type specificity. In these techniques, a nanomaterial converts a remotely transmitted primary stimulus such as a light, magnetic or ultrasonic signal to a localized secondary stimulus such as an electric field or heat to stimulate neurons. The ease of surface modification and bio-conjugation of nanomaterials facilitates cell-type-specific targeting, designated placement and highly localized membrane activation. This review focuses on nanomaterial-enabled neural stimulation techniques primarily involving opto-electric, opto-thermal, magneto-electric, magneto-thermal and acousto-electric transduction mechanisms. Stimulation techniques based on other possible transduction schemes and general consideration for these emerging neurotechnologies are also discussed.

Edited by:

*Ioan Opris, University of Miami, USA*

#### Reviewed by:

*Daniel A. Wagenaar, University of Cincinnati, USA Yoonsu Choi, University of Texas Rio Grande Valley, USA Polina Anikeeva, Massachusetts Institute of Technology, USA*

> \*Correspondence: *Liang Guo guo.725@osu.edu*

#### Specialty section:

*This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience*

Received: *17 November 2015* Accepted: *15 February 2016* Published: *07 March 2016*

#### Citation:

*Wang Y and Guo L (2016) Nanomaterial-Enabled Neural Stimulation. Front. Neurosci. 10:69. doi: 10.3389/fnins.2016.00069* Keywords: nanotechnology, nanomaterial, neural stimulation, non-contact, noninvasive

## INTRODUCTION

Neural stimulation is an essential technique for restoring lost neural functions and correcting disordered neural circuits in neurological diseases (Hassler et al., 2010). For example, it has exciting applications in the restoration of auditory, visual, bladder and limb functions and the treatment of Parkinson's disease, tremor, dystonia, epilepsy, depression and obsessive-compulsive disorder (Cogan, 2008). Conventional electrode-based, electrical neural stimulation is limited by the strong attenuation of electric fields through tissues and thus often requires surgical placement of the electrodes in an intimate contact to the target neural tissue (Cogan, 2008; Huang et al., 2010). Therefore, it faces challenges such as long-term biocompatibility of the implanted electrodes and surgery-induced trauma (Marin and Fernandez, 2010). Noninvasively applied electrical stimulation, however, suffers from an even poorer spatial resolution, requires a higher power and can cause complications to the intermediate tissues (Histed et al., 2013; Menz et al., 2013).

To address these challenges, noninvasive neural stimulation techniques use light, magnetic fields or ultrasound to directly stimulate neurons in a contactless way (Ueno et al., 1988; Gavrilov et al., 1996; Wells et al., 2005). These techniques have a temporal resolution of milliseconds, but are constrained by a poor spatial resolution (Bolognini and Ro, 2010; Menz et al., 2013). For example, transcranial magnetic stimulation only achieves a spatial resolution at the millimeter scale (Ro et al., 1999; Bolognini and Ro, 2010). The spatial resolution of acoustic neural stimulation highly depends on the ultrasound frequency (Clement et al., 2005; Menz et al., 2013). A relatively high spatial resolution can be achieved for retinal stimulation when a high-frequency ultrasound is used, but stimulation of deep neural tissues such as in the brain requires a low frequency for deep tissue penetration, which leads to a low spatial resolution (Menz et al., 2013).

Noninvasive or minimally invasive neural stimulation techniques that can be spatially resolved at a near cellular level are greatly desired for clinical diagnosis and treatment of neurological diseases as well as neuroscience studies (Menz et al., 2013). To pursue a minimally invasive neural stimulation technique with a significantly improved spatial resolution, nanomaterials of unique properties are explored as mediators to convert a wirelessly transmitted primary stimulus to a localized secondary stimulus at the nanomaterial-neuron interface, as shown in **Figure 1**. Additionally, nanomaterials are easy to be surface-modified and bio-conjugated for cell-specific targeting, can be delivered by injection, and can match to the dimensions of subcellular components, such as those of the neuronal membrane and ion channels (Winter et al., 2005; Lugo et al., 2012).

Common primary stimuli also employ light, magnetic fields or ultrasound, which are converted by the nanomaterial to a localized secondary stimulus, primarily electric fields or heat. Localized electric fields stimulate a neuron by perturbing its local transmembrane potential and activating voltage-gated ion channels (Catterall, 1995). Localized heat stimulates a neuron through two proposed mechanisms: the thermal effect on the cell membrane (1) changes the membrane capacitance and/or (2) activates temperature-gated ion channels of the family of transient receptor potential vanilloid (TRPV) channels (Albert et al., 2012; Shapiro et al., 2012; Paviolo et al., 2014b).

This class of nanomaterial-enabled neural stimulation schemes includes, but is not limited to, opto-electric transduction via quantum dots (QDs; Winter et al., 2001, 2005; Gomez et al., 2005; Pappas et al., 2007; Molokanova et al., 2008; Lugo et al., 2012; Bareket et al., 2014), opto-thermal transduction via gold nanomaterials (Paviolo et al., 2013, 2014a, 2015; Eom et al., 2014; Yong et al., 2014; Yoo et al., 2014; Carvalho-de-Souza et al., 2015), magneto-electric transduction via magneto-electric nanoparticles (Yue et al., 2012; Guduru et al., 2015), magnetothermal transduction via superparamagnetic nanoparticles (Huang et al., 2010; Stanley et al., 2012; Chen et al., 2015), and acousto-electric transduction via piezoelectric nanomaterials (Ciofani et al., 2010; Marino et al., 2015). These schemes are categorized in **Table 1** based on their primary stimulus and reviewed in this paper.

## NANOMATERIAL-ENABLED OPTICAL STIMULATION

Optogenetics genetically inserts photosensitive ion channels into a neuron's membrane and modulates the neuronal activity using a blue light (Boyden et al., 2005). This technique has an impressive spatiotemporal resolution and cell-type specificity. However, due to the limited tissue-penetrating capability of the blue light, this method is usually invasive, requiring the implantation of a light

#### TABLE 1 | Transduction schemes of nanomaterial-enabled neural stimulation.


source close to the target tissue (Zhang et al., 2010; Jacques, 2013). Noninvasive infrared light is used to directly stimulate neurons without genetic or chemical pre-modification (Wells et al., 2005, 2007). However, the responsivity and sensitivity of this technique need to be further improved (Peterson and Tyler, 2014). Integrating nanomaterials as mediators into optical neural stimulation can help to achieve this goal and also improve the spatial specificity, energy efficiency and safety by using a light source of a significantly lower power (Eom et al., 2014). Opto-electric and opto-thermal stimulations enabled by QDs and gold nanomaterials respectively are two primary types of nanomaterial-enabled optical stimulation techniques and are reviewed below.

#### Opto-Electric Stimulation Enabled by QDs

QDs are semiconducting nanoparticles with a diameter from 2 to 6 nm (Algar et al., 2010). Their opto-electric transduction property endowed by quantum confinement makes them suitable as mediators for optical neural stimulation (Winter et al., 2001, 2005; Gomez et al., 2005; Pappas et al., 2007; Molokanova et al., 2008; Lugo et al., 2012; Bareket et al., 2014). Such QD-neuron interfaces have been reviewed in the class of optical neural stimulation techniques elsewhere (Bareket-Keren and Hanein, 2014; Thompson et al., 2014).

At their excitation wavelengths, optically excited QDs generate dipole moments and electric fields (Wang and Herron, 1991; Winter et al., 2001, 2005). Theoretical simulation revealed the possibility of their opto-electric transduction to create adequate localized electric fields to activate voltage-gated ion channels and excite neurons (**Figure 1B**; Winter et al., 2005; Lugo et al., 2012). Two strategies were used to construct QD-neuron interfaces (Bareket-Keren and Hanein, 2014): the first bound QDs to a neuron's membrane via antibodies or peptides (Placement II in **Figure 1A**; Gomez et al., 2005; Winter et al., 2005); the second immobilized QDs on a substrate and cultured neurons on top (Placement I; Winter et al., 2005; Pappas et al., 2007; Molokanova et al., 2008; Lugo et al., 2012; Bareket et al., 2014).

An active QD-neuron interface was explored by directly binding antibody- or peptide-conjugated QDs to the neuron's

membrane (Winter et al., 2001; Gomez et al., 2005). However, a stable interface for opto-electric transduction was not achieved due to internalization of QDs and nonspecific targeting (Gomez et al., 2005; Bareket-Keren and Hanein, 2014). A subsequent attempt to avoid the internalization problem by tethering QDs to a substrate to create a film only achieved short-term stability (Winter et al., 2005). In a further pursuit of using a QD film to interface with neurons, neuroblastoma NG108 cells were activated to fire action potentials by a photocurrent generated from layer-by-layer assembled, multiplayer films of HgTe QDs (Pappas et al., 2007). In another study, illumination induced membrane depolarization in both nonexcitable and excitable cells and triggered action potentials in NG108 cells and primary hippocampal neurons (Molokanova et al., 2008). Interfaces were also built between a CdTe QD film and prostate cancer LnCap cells, a CdSe QD film and cortical neurons, and a CdSe QD probe and cortical neurons (Lugo et al., 2012). Upon illumination, responding cells were depolarized or hyperpolarized, and action potentials were evoked in the depolarized cortical neurons.

However, the stimulation efficiency and reliability on these QD films still need further improvement (Pappas et al., 2007; Lugo et al., 2012). Even in the best case, only a small portion (e.g., 11%) of the cells was excited (Pappas et al., 2007). Additionally, some neurons were depolarized, whereas others hyperpolarized; and the responses varied considerably among measurements (Lugo et al., 2012). These were improved by composite films through chemically conjugating CdSe/CdS core-shell semiconducting nanorods to carbon nanotubes (Bareket et al., 2014). These films were used to stimulate a chick retina lacking developed photoreceptors under a pulsed light at a wavelength of 405 nm. Their in vitro biocompatibility and stability were good for up to 21 days. However, the excitation wavelength was only suitable for

superficial stimulation due to limited tissue penetration (Jacques, 2013).

There are also a few other challenges associated with QD-enabled, opto-electric neural stimulation: (1) the strong cytotoxicity of QDs is a concern, particularly when a thin coating is used to achieve an active QD-neuron interface (Derfus et al., 2004; Gomez et al., 2005; Winter et al., 2005; Pappas et al., 2007); (2) the stability of the QD-neuron interface is limited by internalization of QDs via endocytosis (Gomez et al., 2005); and (3) the feasibility of such stimulation schemes needs to be tested in vivo.

## Opto-Thermal Stimulation Enabled by Gold Nanomaterials

In order to generate localized heat to stimulate neurons, microparticles were used as optical absorbers to convert light to heat (Migliori et al., 2012; Farah et al., 2013). The optothermal transduction of gold nanomaterials due to localized surface plasmon resonance makes them particularly suitable as optical absorbers for neural stimulation (**Figure 1C**; Paviolo et al., 2013, 2014a, 2015; Eom et al., 2014; Yong et al., 2014; Yoo et al., 2014; Carvalho-de-Souza et al., 2015). Upon irradiation at the resonant frequency, electrons in gold nanomaterials oscillate and collide, generating and dissipating heat (Roper et al., 2007; Cao et al., 2014). The use of gold nanorods for optical neural stimulation was also reviewed elsewhere (Paviolo et al., 2014b).

Gold nanorods coated with silica were used to stimulate nongenetically modified rat auditory neurons in vitro (Placement I and IV; Yong et al., 2014). Illuminated by a pulsed laser at a resonant wavelength of 780 nm, these gold nanorods activated nearby neurons with a linear correlation to the duration of the laser pulse. It was also found that internalized gold nanorods promoted neurite outgrowth and induced a Ca2<sup>+</sup> influx in NG108-15 cells under continuous and pulsed irradiation respectively, both at a near-infrared resonant wavelength of 780 nm (Placement IV; Paviolo et al., 2013, 2014a, 2015).

In vivo optical stimulation of non-genetically modified rat sciatic nerves via gold nanorods was also demonstrated (Placement I; Eom et al., 2014). Illuminated by a pulsed laser at a near-infrared resonant wavelength of 980 nm, sciatic nerves with injected gold nanorods were nearly six times more responsive to fire compound action potentials with a threshold three times lower than the null control. Therefore, the power and exposure duration of the laser stimulus could be greatly reduced, significantly decreasing the risk of tissue damage.

Gold nanoparticles were also used for in vitro and ex vivo opto-thermal neural stimulation (Carvalho-de-Souza et al., 2015). Gold nanoparticles were conjugated to ligands and specifically targeted to ion channels in the neuron's membrane (Placement III). Upon illumination with light pulses at a visible wavelength of 532 nm, the generated heat depolarized rat dorsal root ganglion neurons and mouse hippocampal slice neurons to fire action potentials. These ion channel-bound gold nanoparticles showed good washout resistance.

For these opto-thermal neural stimulations, internalization of gold nanorods is still a challenge, causing inconsistency, variability and short-term cytotoxicity (Paviolo et al., 2013; Yong et al., 2014). It was reported that an increased pulsed laser irradiance reduced the Ca2<sup>+</sup> influx induced by internalized gold nanorods (Paviolo et al., 2014a). Inhibitory effects on hippocampal, cortical and olfactory bulb neurons were also observed with gold nanorods electrostatically bound to the neuron's membrane (Placement II; Yoo et al., 2014). Temperature-sensitive inhibitory TREK-1 channels were assumed responsible. Therefore, another challenge is to diverge the different effects in a specific stimulation scheme, so that the neuronal responses can be precisely controlled.

## NANOMATERIAL-ENABLED MAGNETIC STIMULATION

The weak interaction between magnetic fields and tissues enables magnetic fields to penetrate deep into tissues (Huang et al., 2010). However, neural stimulation using magnetic fields usually requires converting the fields to a localized secondary stimulus (Huang et al., 2010). This can be enhanced by magnetoelectric nanoparticles via magneto-electric transduction and superparamagnetic nanoparticles via magneto-thermal transduction. These two nanomaterial-enabled magnetic neural stimulation schemes are reviewed below.

## Magneto-Electric Stimulation Enabled by Magneto-Electric Nanoparticles

Magneto-electric nanoparticles, usually made of multiferroics, show a strong magneto-electric coupling and can convert magnetic fields to electric fields due to the magneto-electric effect (Fiebig, 2005). Based on this effect, an idea of using magnetoelectric nanoparticles to control voltage-gated ion channels for neural stimulation was proposed (Kargol et al., 2012). Theoretical analysis justified the possibility for deep brain stimulation (**Figure 1D**; Yue et al., 2012). A proof-of-concept in vivo study in mice was conducted using magneto-electric CoFe2O4-BaTiO<sup>3</sup> core-shell nanoparticles under a low-intensity magnetic field to modulate deep brain circuits (Guduru et al., 2015). More research is still needed to assess its feasibility with mechanistic specificity and long-term in vivo biocompatibility.

## Magneto-Thermal Stimulation Enabled by Superparamagnetic Nanoparticles

Widely used superparamagnetic nanoparticles can convert alternating magnetic fields to localized heat via magneto-thermal transduction (Laurent et al., 2008), enabling the development of magneto-thermal neural stimulation techniques (**Figure 1E**; Huang et al., 2010; Stanley et al., 2012; Chen et al., 2015). Streptavidin-conjugated superparamagnetic manganese ferrite (MnFe2O4) nanoparticles were targeted to the biotinylated peptide of a genetically engineered anchor protein in the membrane of neurons expressing the temperature-gated TRPV1 ion channels (Placement II; Huang et al., 2010). Upon application of a radio-frequency magnetic field, highly localized heating via magneto-thermal transduction induced a Ca2<sup>+</sup> influx through the TRPV1 ion channels, depolarized the neurons to fire action potentials in vitro, and triggered thermal avoidance in worms.

A more specific ion-channel targeting strategy was also implemented in a mouse xenograft model by tethering nanoparticles directly to the TRPV1 ion channels (Placement III; Stanley et al., 2012). 6x-His epitope tag-inserted TRPV1 ion channels were genetically inserted in the cell membrane, and 6x-His epitope tag antibody-conjugated iron oxide nanoparticles were specifically targeted to these ion channels and heated under a radio-frequency magnetic field. The localized heat activated the TRPV1 ion channels and induced a Ca2<sup>+</sup> influx into the cells faster than in the above work (Huang et al., 2010).

To improve the temporal resolution for neuronal activation and realize a long-term in vivo stimulation feasibility, untargeted superparamagnetic iron oxide nanoparticles were dispersed in the vicinity of TRPV1-expressing human embryonic kidney HEK-293FT cells, dissociated hippocampal neurons and neurons at the ventral tegmental area of mice (Placement I; Chen et al., 2015). Upon applying an alternating magnetic field, the magneto-thermally generated heat induced a Ca2<sup>+</sup> influx in the HEK-293FT cells, stimulated hippocampal neurons to fire action potentials, and activated the neurons at the ventral tegmental area of mice to have an enhanced expression of c-fos, achieving a stimulation response with a latency of 5 s after the onset of the magnetic field. And the stimulation in the mouse model remained effective for at least 1 month thanks to the good biocompatibility, stability and decreased endocytosis of extracellularly dispersed nanoparticles.

Magneto-thermal neural stimulation enabled by superparamagnetic nanoparticles can achieve a uniform stimulation of the target cell population due to the uniform expression of TRPV1 ion channels in these cells across the tissue (Huang et al., 2010; Chen et al., 2015). Although both opto-thermal and magneto-thermal stimulations use heat as the localized secondary stimulus, only the latter has employed genetic modifications to the target neurons for specific targeting and TRPV1 ion channel expression (Huang et al., 2010; Stanley et al., 2012; Chen et al., 2015). The safety of TRPV1 ion channels is, however, concerned, due to their high Ca2<sup>+</sup> permeability, and thus temperature-gated Na<sup>+</sup> ion channels are desired as the target (Knöpfel and Akemann, 2010).

## NANOMATERIAL-ENABLED ACOUSTIC STIMULATION

As a wirelessly transmitted primary stimulus, ultrasound interacts with tissues weakly and can penetrate deep into soft tissues with minimal energy absorption (Tyler, 2011). It can also be focused at a submillimeter resolution (Gavrilov et al., 1996; Menz et al., 2013; Ibsen et al., 2015). Ultrasound has been directly applied to stimulate both the peripheral and central neural systems, but these techniques are limited by a low energy efficiency and mechanistic non-specificity (Gavrilov et al., 1996; Tyler, 2011; Legon et al., 2014).

## Acousto-Electric Stimulation Enabled by Piezoelectric Nanomaterials

Piezoelectric nanomaterials can convert ultrasound waves to electric fields via acousto-electric transduction due to their piezoelectricity (Wang and Song, 2006; Wang et al., 2007). Such an acousto-electric transduction may facilitate neural stimulation by a low-intensity ultrasound (**Figure 1F**; Ciofani et al., 2010; Marino et al., 2015). Neurite outgrowth of PC12 and SH-SY5Y cells was promoted by internalized piezoelectric boron nitride nanotubes under ultrasound stimulation (Placement IV), implying a stimulating effect of the acousto-electric transduction (Ciofani et al., 2010). Piezoelectric barium titanate nanoparticles were electrostatically attached to SH-SY5Y cells and induced Ca2<sup>+</sup> and Na<sup>+</sup> influxes in an ultrasonic field (Placement II; Marino et al., 2015). The possibility of neural stimulation via the acousto-electric transduction of piezoelectric nanoparticles was also theoretically justified (Marino et al., 2015). These works showed the promise in using piezoelectric nanomaterials to facilitate noninvasive acoustic neural stimulation. However, direct evidence for neuronal activation has not been established yet. More work is needed to establish the feasibility of this stimulation technique, particularly with primary neurons and in animal models.

## FUTURE DIRECTIONS

Nanomaterial-enabled neural stimulation is an emerging class of neurotechnologies, with numerous exciting breakthroughs in the past decade. As a powerful enabling tool, nanomaterials can be either applied alone or combined with other approaches including synthetic biology to facilitate innovative neural stimulation schemes. These new techniques not only allow non- or minimally invasive neural stimulation of a high spatial resolution and cell specificity, but also improve the safety by significantly reducing the required power of the primary stimulus (Huang et al., 2010; Eom et al., 2014).

Nanomaterials of other transduction mechanisms, such as magneto-mechanical, acousto-mechanical, and optooptical transductions, are also worth considering for potential development of additional neural stimulation schemes. Magnetomechanical transduction via magnetic nanoparticles can convert magnetic fields to localized mechanical forces to activate mechanosensitive ion channels such as the TREK-1 channels (Hughes et al., 2005, 2008; Dobson, 2008). Nanomaterialenabled, acousto-mechanical transduction may be combined with the recently developed sonogenetics (Ibsen et al., 2015) to improve the activation efficiency of genetically inserted membrane mechanosensitive ion channels. Opto-optical transduction via upconversion luminescent nanoparticles, which convert a long-wavelength light to one of a shorter wavelength, may provide a noninvasive alternative to the implanted laser in optogenetics by converting deep penetrating near-infrared light to localized visible light for activating photosensitive ion channels (Jacques, 2013; Berry et al., 2015).

To select the primary and secondary stimuli, several factors are considered. The primary stimulus needs to penetrate tissues deeply, be easy to focus at an appropriate spatial resolution and be safe for long-term exposure. The secondary stimulus needs to be selected according to an adequate expression of the target ion channels in the neuron's membrane. For example, it is not necessary to genetically modify the target neurons with voltagegated ion channels to use electric fields as the secondary stimulus, whereas, to use heat, the TRPV1 ion channels may need to be genetically inserted into the membrane of target neurons (Huang et al., 2010; Stanley et al., 2012; Chen et al., 2015). Additionally, placement of the nanomaterials (see **Figure 1A**), which is crucial to the stability of the nanomaterial-neuron interface (Winter et al., 2001, 2005; Gomez et al., 2005; Bareket-Keren and Hanein, 2014; Chen et al., 2015), should be considered in conjunction with possible pre-modification to the target neurons (Huang et al., 2010; Stanley et al., 2012).

This diverse class of nanomaterial-enabled neurotechnologies is still in their early stages of development, with many having only been validated in vitro. To move forward, many issues including biocompatibility, stability, consistency, efficiency and reliability will need to be addressed (Gomez et al., 2005; Pappas et al., 2007; Yong et al., 2014; Chen et al., 2015). For a significant period of

#### REFERENCES


time, these neurotechnologies will be used primarily as scientific tools for in vitro and/or in vivo studies. Clinical application is promising, but remains very challenging due to concerns on the safety of nanomaterials, viral vectors for gene delivery, and genetic modification to the target neurons (Manilla et al., 2005; Maynard et al., 2006).

### AUTHOR CONTRIBUTIONS

YW and LG analyzed the relevant published work, designed the perspective and structure, and wrote the manuscript.

#### ACKNOWLEDGMENTS

This work was supported by startup funds generously provided to LG by the Department of Electrical and Computer Engineering and Department of Neuroscience at The Ohio State University. YW was supported by the 2015 OSU Center for Cognitive and Brain Sciences Summer Graduate Research Award and the 2015-2016 OSU HHMI MED Into GRAD Scholars Award.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Wang and Guo. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Micro- and Nanotechnologies for Optical Neural Interfaces

#### Ferruccio Pisanello<sup>1</sup> \*, Leonardo Sileo<sup>2</sup> and Massimo De Vittorio<sup>1</sup> \*

<sup>1</sup> Center for Biomolecular Nanotechnologies, Istituto Italiano di Tecnologia, Lecce, Italy, <sup>2</sup> Dipartimento di Ingegneria dell'Innovazione, Università del Salento, Lecce, Italy

In last decade, the possibility to optically interface with the mammalian brain in vivo has allowed unprecedented investigation of functional connectivity of neural circuitry. Together with new genetic and molecular techniques to optically trigger and monitor neural activity, a new generation of optical neural interfaces is being developed, mainly thanks to the exploitation of both bottom-up and top-down nanofabrication approaches. This review highlights the role of nanotechnologies for optical neural interfaces, with particular emphasis on new devices and methodologies for optogenetic control of neural activity and unconventional methods for detection and triggering of action potentials using optically-active colloidal nanoparticles.

#### Edited by:

Ruxandra Vidu, University of California Davis, USA

#### Reviewed by:

Ileana Hanganu-Opatz, Center for Molecular Neurobiology Hamburg, Germany Lohitash Karumbaiah, The University of Georgia, USA

#### \*Correspondence:

Ferruccio Pisanello ferruccio.pisanello@iit.it; Massimo De Vittorio massimo.devittorio@iit.it

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 02 September 2015 Accepted: 15 February 2016 Published: 08 March 2016

#### Citation:

Pisanello F, Sileo L and De Vittorio M (2016) Micro- and Nanotechnologies for Optical Neural Interfaces. Front. Neurosci. 10:70. doi: 10.3389/fnins.2016.00070

Keywords: nanotechnology, optogenetics, nanoparticles, neural interfaces, optical fibers

## INTRODUCTION

The activity and interconnections of the billions of neurons in the human brain determine the function of our senses, dictate our motor choices, form memories, and guide behavior. Understanding, monitoring and manipulating neural activity with high spatial and temporal resolution in vivo and on a large number of neurons is mandatory for a deeper knowledge of neural circuitry, and to shine light on causal relations between neurons or between neurons and behavior.

New strategies and technologies to systematically monitor thousands of functional links that each neuron forms with other neurons are being developed. Nanoscience and nanotechnology can play a key role in developing new ideas and experimental approaches to create detailed maps of the human and mammalian brain (Nanotechnology and Neuroscience, 2014). Among new approaches for neuroscience, optical methods are very promising for both recording and manipulating neural activity. A major breakthrough in this respect has been the advent of optogenetics (Boyden et al., 2005), relying on the genetic expression of exogenous light-gated ion channels and ion pumps to control neuronal activity, allowing unprecedented causal manipulation of specific neural circuits.

In optogenetic experiments light of visible wavelengths is shined onto single neurons or, feasibly, large brain regions, to activate or inhibit specific classes of neurons, while simultaneously recording electrophysiological data or monitoring behavior in freely moving animals. However, due to induced tissue damage and light scattering and absorption, light delivery in vivo and in deep brain regions of animal models is still very challenging, and far from being effective (Stujenske et al., 2015). Both acute and chronic optical implants need to meet several requirements for different experiments: site specific light delivery or uniform illumination of large brain volumes, low physical, and thermal tissue damage, biocompatibility, high fidelity and minimized photoelectric artifacts, high switching speed, and tunable wavelength (Pisanello M. et al., 2014; Warden et al., 2014; Grosenick et al., 2015). Additionally, optical methods can be exploited for simultaneous light collection for all-optical manipulation and monitoring of neural activity, by using light-based genetically encoded neural activity indicators (GEAIs), such as fluorescent Ca2<sup>+</sup> indicators or voltage sensitive dyes (VSD; Cui et al., 2013, 2014; Gunaydin et al., 2014).

This review encompasses the latest developments and technologies in the field of light delivery and possible approaches for optical monitoring of neuronal activity in vivo. A variety of different nanotechnologies and optical methods applied to neuroscience are presented, including active implanted LEDs, passive arrayed waveguides, nanomachined tapered fibers, and self-organized colloidal nanostructures.

## TOP-DOWN FABRICATION PROCESSES FOR MULTIPOINT OPTOGENETIC STIMULATION

Optogenetics—"the combination of optics and genetics to achieve gain or loss of functions of well-defined cellular events in specific cells of living tissue" (Deisseroth, 2011)– is widely adopted in the central nervous system on animal models to modulate neural activity and to regulate release of specific neurotransmitters (Adamantidis et al., 2007; Aravanis et al., 2007; Petreanu et al., 2009; Lin et al., 2013). This is achieved through the use of specific transmembrane proteins, called opsins, which respond to light by generating a flow of ions across the cellular membrane, acting as light-gated ion channels. An example of opsin used to trigger action potentials is Channelrhodopsin 2 (ChR2), a non-specific cation channel used to depolarize the neuron (Nagel et al., 2003). Inhibition of neural activity can instead be achieved by using Halorhodopsin (Halo) and Archaerhodopsins (Arch), light-driven ion-pumps used to hyperpolarize the cell, therefore inhibiting the generation of action potentials by reducing the probability of supra-threshold events (Nagel et al., 2003; Fenno et al., 2011; Tye and Deisseroth, 2012). ChR2, Halo and Arch, as well as many other membrane proteins, can be delivered into the brain by means of transfection approaches such as in utero electroporation, viral transfection or transgenic crossing, all allowing for gene delivery only to molecularly-defined classes of neurons (Han, 2012). This latter is the main advantage of optogenetics with respect to electrical stimulation of neural activity: light can be used to modulate electrical activity only of genetically-defined neural sub-populations without affecting nearby neurons of a different type, still allowing for post-synaptic effects to take place. After its first use in mammalian neurons in 2005 (Boyden et al., 2005), optogenetics is now adopted in several animal models from Caenorhabditilis elegans to primates and, in particular, in mice and rats to study functional connectivity of specific classes of neurons and to identify their particular role in neural diseases and disorders.

In this framework, a crucial aspect is represented by technologies to deliver light into the brain. The neural tissue is, indeed, a highly scattering medium and microscopy-based techniques are still restricted to the shallower layers of the cortex (Warden et al., 2014). Standard experimental protocols based on the implantation of a fiber stub with a flat-cleaved end are limited by the single illumination spot and the small volume excited at the fiber tip, since the delivered light power is strongly attenuated after a few hundreds of micrometers (Aravanis et al., 2007; Yizhar et al., 2011). To interact with deeper brain structures, new generations of implantable devices are being developed (Pisanello M. et al., 2014; Warden et al., 2014; Grosenick et al., 2015). In this context, top-down nanotechnology fabrication processes are allowing for unprecedented functionalities and integration processes, which resulted in a minimized damage to the brain tissue during implantation and, simultaneously, to the possibility of optically control a wider brain volume. As well, micro, and nanotechnologies have been exploited for realizing multifunctional devices, which recently overtook the classic concept of "optrode" (e.g., a device for simultaneous optical control and electrical monitoring of neural activity Grosenick et al., 2015) and can now integrate microfluidic systems for in situ drug delivery (Canales et al., 2015; Jeong et al., 2015a) or other devices such as temperature sensors or photodetectors (Kim et al., 2013).

## Implanted µLED

A very promising strategy to bring light to deep brain regions consists in using micro light emitting diodes (µLED) implanted directly in the target area. A straightforward and multi-purpose implementation of that was presented in 2013 in Kim et al. (2013) McCall et al. (2013). Kim et al. developed a method to realize releasable gallium nitride (GaN) µLEDs on a sapphire substrate (only 6.5µm thick), which were then moved to thin plastic strips hosting multiple and independently addressable emitters. This was a constituent component of a layered implant that can be customized depending on the experimental needs, and can incorporate also other electrical elements such as platinum electrodes for extracellular recording or for electric stimulation, platinum temperature sensors, local heaters and microscale photodetectors (**Figures 1A,B**). The so-obtained stack is implanted via a releasable microneedle, extracted after the surgery. Driving electronics stays instead outside the skull and, interestingly, the system allows for straightforward wireless operation (**Figure 1C**) and, as very recently shown in Jeong et al. (2015a), for the integration of wireless-driven drug delivery systems. Blue light emitted by the µLEDs was used to stimulate dopaminergic neurons in the ventral tegmental area of untethered mice behaving in a complex environment containing sites for dopamine rewards, preferred by the ChR2-transfected animals during the experiment. This manuscript of Kim and coworkers represented a boost for related technologies, and in last 2 years other interesting approaches were suggested for µLEDbased stimulation in other regions of the mouse brain. This is the case of the auditory pathway toward the brain (Hernandez et al., 2014). Hernandez et al. implanted a µLED to stimulate auditory brain stream responses via optogenetic excitation of spiral ganglion neurons in the mouse cochlea, showing that optical stimulation allows for a better frequency resolution with respect to classical monopolar electrical approach. Although this experiment was realized with a single µLED, the possibility to integrate multiple emitters on flexible shafts for multipoint stimulation has been recently demonstrated (Goßler et al., 2014). This technology exploits laser-lift-off to transfer from the original

stimulation devices. (A–C) A wireless system consisting of µLEDs on a flexible shank. (D) A µLEDs device for site-selective stimulation of mouse (Continued)

#### FIGURE 1 | Continued

neocortex. (E) Monte-Carlo simulations of the light radiation pattern from a single µLEDs implanted in the scattering tissue. (F) A 3D set of silicon oxynitride waveguides for custom optogenetic stimulations of defined points in a 3D fashion. (G) Multipoint-emitting optical fibers for stimulation of multiple brain regions with a single and tapered optical fiber. (H) Multifunctional polymeric fibers. (I) ZnO-based multipoint optical arrays for simultaneous optical control and electrical recording of neural activity. (J) Array of implantable optical fibers coupled to µLEDs on a flexible polyamide cable. (K,L) Array of tapered SU-8 waveguides coupled with µLEDs and electrodes for extracellular readout of neural activity. (A–C) are reproduced with permissions from Kim et al. (2013). Panels (D,E) are reproduced with permissions from McAlinden et al. (2015). Panel (F) is reproduced with permissions from Zorzos et al. (2012). Panels (G–G3) are modified from Pisanello F. et al. (2014). Panel (H) is reproduced with permissions from Canales et al. (2015). Panel (I) was reproduced with permissions from Lee et al. (2015). Panel (J) is reproduced with permission from Schwaerzle et al. (2015). Panels (K,L) are reproduced with permission from Kwon et al. (2015).

sapphire substrate the GaN µLEDsto polyamide films, having the proper mechanical properties to follow the cochlear curvature. Furthermore, multipoint stimulation with µLEDs on sapphire shanks was also demonstrated for site-selective stimulation of neocortical circuits, with Monte-Carlo simulations used to evaluate the broadening of emitted light induced by tissue scattering (McAlinden et al., 2015; **Figures 1D,E**).

Together with the advantages of a straightforward implementation on flexible devices, suitability for wireless operation and the possibility to integrate high-density emission points on a single shaft with a plurality of electrodes for extracellular recording, GaN-based µLED technology is still facing important challenges. In particular, the heat at the surface of the emitter poses an upper limit to the duration of light delivery stimuli. In McAlinden et al. (2013) this aspect was analyzed by a finite element model and Mc Allinden et al predicted that a conservative limit of 0.5◦C increase of tissue temperature would not be reached until a pulse duration of ∼200 ms (at 350 mW/mm<sup>2</sup> ; McAlinden et al., 2013). By virtue of the multiple functionalities integrated in their flexible device, Kim et al. instead directly measured the temperature rise with a platimum sensor and measured temperature variations below 0.12◦C for 10 ms-long pulses up to 20 Hz and power density up to ∼15 mW/mm<sup>2</sup> (Kim et al., 2013) The minimum distance between multiple emitters, instead, depends strongly on the presence of a dielectric material at the emitters/tissue interface and it is limited by tissue scattering and the Lambertian emission profile of the µLEDs (McAlinden et al., 2015).

#### Waveguides-Based Implants

Another promising strategy to deliver light into the brain is represented by waveguides-based devices, which have recently seen the definition of important routes toward viable multipoint optogenetic stimulation. With respect to µLED-based approaches, they have the main advantages of keeping the light sources outside the tissue, thus avoiding direct heating induced by implanted electronics and to be able to change the delivered light wavelength based on experimental needs. In 2010 Zorzos et al. proposed a multipoint-emission device obtained by top-down fabrication process (Zorzos et al., 2010), with aluminum-coated multiple silicon oxynitride waveguides ending with corner mirrors to direct light laterally with respect to the implant direction. Each waveguide was coupled to external sources with single mode fibers, for overall outcoupling efficiency ranging from 23 to 33%. The same authors, in 2012, extended this technology to 3D arrays of independently addressable light emission points with multiple shanks arranged in three-dimensions using a micro-fabricated baseplate holder (**Figure 1F**; Zorzos et al., 2012). Each shank contains several light emission points in the same configuration of Zorzos et al. (2010), each of which can be independently addressed by external coupling systems such as digital micromirrors chips coupled to a microlenses array or galvanometric mirrors and a f-theta lens. In the same year, another approach was proposed by Abaya et al. based on a 3D array of sharpened waveguides allowing for stimulation at two different depths (Abaya et al., 2012). The SiO2 waveguides are realized through a dicing process, defining first the pyramidal shape of the shanks with a bevel blade and then the vertical pillars by deep kerfs. HF-based etching is then employed to thin the shanks and, finally, an annealing step is used to relieve internal stress and reduce the surface roughness.

Although these methods allow for a dynamic reconfiguration of the stimulation geometry during the experiment, a sever limit to viable in vivo implementation is represented by the pronounced implant cross section. A solution for that was proposed in 2014, exploiting the photonic properties of tapered optical fibers (Pisanello F. et al., 2014). The device is composed by a tapered optical fiber with a sub-micrometer tip diameter, with the tapered region covered with a gold layer to keep light confined into the waveguide. Light is allowed to outcouple into the brain through optical windows realized in the gold coating by Focused Ion Beam milling (Sileo et al., 2015), thus allowing optogenetic control of neural activity only at specific sites along the taper (**Figure 1G**). The active window can be selected by modifying the light coupling angle at the other end of the fiber, therefore injecting into the waveguide different subsets of guided modes (Pisanello et al., 2015) and allowing up to three independent stimulation points on a 1-mm-long segment of the taper (**Figure 1G**). The in vivo application of this technology was shown in both mouse motor cortex for layer-selective stimulation of GABAergic neurons and in the striatum of awake and headrestrained mice, and optrodes were realized by placing the nanostructured optical fiber beside a linear electrodes array for extracellular recording (Pisanello F. et al., 2014). On the other hand, with respect to standard optical fibers, multipoint emitting optical fibers need a higher injection power to achieve effective optogenetic control of neural activity and the total efficiency depends on the distance between the active window and the taper tip. The integration of a linear electrodes array beside the fiber, moreover, increases the invasiveness of the device.

Very recently, a series of approaches have allowed for integrated multipoint stimulation and multipoint electrical readout of neural activity (Grosenick et al., 2015). Canales et al. (2015) have developed a set of multifunctional devices based on a polymeric technology for simultaneous drug delivery, optogenetic control and extracellular recording (see representative images in **Figure 1H**). Together with multiple integrated functionalities, these fibers are able to better match the brain mechanical properties by virtue of the combination of different flexible materials, including poly(etherimide), poly(phenylsulfone), polycarbonate, and cyclic olefin copolymer and conductive poly-ethylene (Canales et al., 2015). Lee et al. (2015), instead, have recently proposed a new system based on optically transparent and electrically conductive ZnO semiconductor. As schematically shown in **Figure 1I**, the device is composed by a matrix of ZnO waveguides coated with Parylene-C up to ITO-coated tips. This configuration allows for a strong reduction of photoelectric artifacts induced by direct electrode illumination, and thereof to spatially match light delivery stimuli and electrical readout.

It is important to highlight that the main limitation of waveguides-based multipoint stimulation devices relies on the need to tether the animal to an optical bench, for coupling with the proper light injection system. For some of these approaches, animal movement and the resulting fiber bending and stretching can potentially lead to crosstalk between the different channels, and can generate inhomogeneous light delivery, in particular in the case of multimodal waveguides (Cui et al., 2013, 2014).

## Coupling of µLED with Implanted Optical Waveguides

Hybrid approaches are instead represented by integrated technologies to couple light emitted from µLEDs into implanted waveguides. This was achieved on a flexible polyimide ribbon cable by Schwaerzle et al. using a silicon housing to align implantable optical fibers with the light sources and subsequent fixation with UV-curable adhesive fixation (Schwaerzle et al., 2015; **Figure 1J**). Very recently, Kwon et al. also developed a technique to couple µLEDs with microfabricated microneedles in a wireless-driven implant for multisite and bilateral stimulation of the rat visual cortex (Kwon et al., 2015; **Figure 1K**). It is composed by two arrays of SU8 tapered waveguides covered by a stack of ITO/Parylene-C/Gold/Parylene-C to allow for simultaneous electrical recording of neural activity in the proximity of the light delivery site (**Figure 1L**). The waveguides are realized on a Polydimethylsiloxane substrate and are then aligned and coupled to the µLEDs arrays placed on a polyamide cable. The result is an integrated device that, using a capacitorbased stimulator system, can be controlled via an inductive link with up to 32 bidirectional channels. If, on one hand, these methods greatly combine the advantages of implanted µLEDs and waveguide-based approaches, their main limitation is still represented by the highly divergent radiation pattern of the light sources, which do not allow for straightforward and repeatable coupling efficiency (Schwaerzle et al., 2015).

## Future Challenges: Integrated Light Collection-Delivery Systems and Long Term Experiments

Although, the technology development of last years mainly focused on devices for stimulation or inhibition of neural activity, a crucial aspect remains the possibility to simultaneously monitor neural activity. Most of the above-described devices can integrate electrodes for extracellular recording, and some of them allow for mapping of neural activity with single or multiple light-delivery sites, whose state of the art has been recently reviewed by Grosenick et al. (2015). However, in the same way classical electrical neural stimulation excite all the cells within the stimulated region, electrical readout cannot select for specific classes of neurons. This is instead possible with genetically encoded neural activity indicators (GEAIs), such as fluorescent Ca2<sup>+</sup> indicators or voltage sensitive dyes (VSD). These probes respond to a variation of neural activity by changing their fluorescence intensity and are widely adopted in microscopy techniques in vivo to monitor electrical activity of cortical neural circuits (Svoboda et al., 1997; Kuhn et al., 2008; Warden et al., 2014). Deep brain regions, however, are widely not accessible for microscopy and the most common technique to collect light emitted from GEAIs remains the use of large core optical fibers, and are limited to a single and relatively small volume of the neural tissue (Cui et al., 2013, 2014; Gunaydin et al., 2014). The development of new techniques to efficiently collect light from sub-cortical regions is thus essential to boost the development of viable and integrated all optical bidirectional neural interfaces. Nevertheless, it is important to highlight that electrical and optical monitoring of neural activity are complementary strategies, since some electrical signals such as local field potentials have not yet an optical counterpart.

Another important challenge is represented by the possibility of using multipoint optical and optoelectronic neural interfaces in long-term experiments in untethered animals. Concerning electrical readout of neural activity, chronic implants suffer from a high variability and limited longevity of their electrical performances, as a result of failures related to a combination of biological responses of the tissue, materials stability, and mechanical properties of the device (James et al., 2013; Prasad et al., 2014). The biological aspects include (but are not limited to) damage to the blood brain barrier, inflammation responses and increased astroglial activity (Kozai et al., 2010, 2012, 2014; Saxena et al., 2013). Material failures are known to include corrosion, cracking and degradation of the insulating layer (Abhishek et al., 2012; Gilgunn et al., 2013; Prasad et al., 2014). From the point of view of the mechanical properties of the implant, most of the technologies targeting deep brain regions are based on hard materials that do not match with the softness of the brain, and therefore can hardly follow its natural movements (Hyunjung et al., 2005; Jeyakumar et al., 2005). This has increased the demand for new approaches based on flexible optoelectronic and conductive polymers (Jeong et al., 2015b), able to bend and flex to take into account pulsations and volume changes of the tissue over time [an example are the low bending stiffness multifunctional fibers described in Canales et al. (2015)]. Although very recent reports show that also the mechanical mismatch within a device can cause failures (Kozai et al., 2015), a straightforward integration between planar technologies for optical neural interfaces and flexible electronics (Kim et al., 2013; McCall et al., 2013; Goßler et al., 2014; Jeong et al., 2015a), or conceptually new approaches as syringe-injectable flexible devices (Liu et al., 2015) are therefore needed to create multifunctional devices that can match with the mechanical properties of the brain. Moreover, devices for untethered animals experiments would greatly benefit of wireless communication systems to trigger the spatiotemporal configuration of light stimuli and to retransmit the recorded data. Some approaches based on radiofrequency links (Kim et al., 2013) or optical wireless communication (Jeong et al., 2015a) have already been proposed, but next generation will have to focus on integrated systems for duplex communication allowing also for wireless and real-time optical and electrical readout of neural activity.

## Nanoparticles for Optical Modulation and/or Readout of Neural Activity

Together with the new technological frontiers opened by nanotechnologies for multisite light delivery, there is a widespread agreement that the quantum properties of nano-sized materials, together with their small size and the high surface to volume ratio, can be employed to investigate alternative strategies for building next generation of optical neural interfaces. This is the case of colloidal nanoparticles, extremely small structures produced via wet-chemistry which exploit their ultra-small size to enhance quantum effects (Pellegrino et al., 2005; Carbone and Cozzoli, 2010). In particular, the reduced size often results in an enhanced sensitivity on the electromagnetic properties of the surrounding environment, and this specific feature has been exploited in two very recent works to stimulate neural activity by using plasmonic gold nanoparticles (João et al., 2015) and to evaluate the possibility of optically monitoring neural activity exploiting charge carriers dynamic in semiconductor nanocrystals (Marshall and Schnitzer, 2013).

In the case of plasmonic gold nanoparticles (PGNPs), the authors of João et al. (2015) suggested the use of a plasmonmediated high absorption at green wavelengths to generate a localized heating and therefore to trigger action potentials, as schematically represented in **Figure 2A**. Carvalho-de-Souza et al. developed a technique to functionalize the surface of the neurons with spherical gold nanoparticles of about 20 nm in diameter, having a maximum plasmonic absorption band at 523 nm. When irradiated by green light, the particles convert the absorbed energy in local heating, inducing a variation in the membrane capacitance (Shapiro et al., 2012), rather than acting on temperature-dependent ion channels (Stanley et al., 2012), and a subsequent fast cell depolarization. If threshold potential is reached, voltage-gated channels open and the action potential is triggered (João et al., 2015). The functionalization is obtained with primary antibodies targeting different proteins on the cell membrane and secondary antibodies conjugated on the particle surface, but also exploiting a synthetic version of the Ts1 neurotoxin, which selectively binds to voltage-gated sodium channels. This latter was tested in both dorsal root ganglion neuronal cultures and hippocampal brain slices, suggesting the generality of the approach. Although, this technique does not need any genetic treatment of the tissues of interest and represents an important complement to existing technologies, it still has important challenges to face, in particular if compared to optogenetics. Indeed, one of the main

peculiarities of optogenetics is the possibility to inhibit neural activity, while PGNPs can be used only for stimulation and, moreover, the particles are allowed to last into the organisms for a fixed period of time, reasonably much shorter than the almost permanent expressions of light-gated ion channels (João et al., 2015).

Another promising, but still only theoretical, application of colloidal nanoparticles is the readout of neural activity by using semiconductor nanocrystals (NCs). NCs are nanometer-sized semiconductors that present quantized levels in both valence and conduction band, rather than the quasi-continuum of states allowed in bulk semiconductors. A photon absorbed by a NC generates a bound electron-hole (e-h) pair, which can recombine following radiative (e.g., emitting a photon at lower energy) or non-radiative channels. For NCs made of II-VI semiconductors, emitted light is in the visible spectral range and, at physiological conditions, it has a well-defined single-peak emission spectrum with a ∼30 nm-large Gaussian distribution around the peak wavelength. When specific types of nanoparticle are inserted into an electric field, however, the e-h pairs polarize along the direction of the applied field, and both the emission wavelength and the lifetime of the excited state change accordingly (Galland et al., 2011). Marshall et al. suggested to exploit this feature to optically read out action potentials (Marshall and Schnitzer, 2013). The model system relies in a NC, either spherical or elongated, placed into the lipid bilayer in order to have the particle sensitive to the highest possible voltage variation across the cell membrane (**Figure 2B**). When the membrane potential changes, the radiative lifetime changes as well, and the fraction of excitation events leading to non-radiative recombination is modified accordingly. This directly leads also to a variation of the fluorescence intensity and, overall, to the possibility to sense action potential by monitoring the fluorescence intensity of the NCs in a way very close to standard approaches used for voltage sensitive dyes. This technique have, potentially, strong advantages if compared to the use of other fluorescent indicators of neural activity, which are still facing the challenge of a limited signal to noise ratio and pronounced photobleaching. On the contrary, NCs have a strong absorption cross-sections at both one and two photons, high emission intensity, and can support longer excitation times before bleaching, making them ideal from the point of view of fluorescence stability (Spinicelli et al., 2009; Galland et al., 2011; Pisanello et al., 2013). However this remains a theoretical proposal and its implementation is mainly limited by the absence of techniques to viably and stably localize semiconductor NCs across the cell membrane of neurons.

## CONCLUSIONS AND PERSPECTIVES

Nanotechnology, exploiting surface and bulk nano and micromachining and self-organized nanochemistry, has dramatically improved the capability to produce devices that deliver, shape, and collect light with high spatial and temporal resolution. In the past decade all those technologies have been applied to ICT applications and for biosensing assays, with many materials and methods derived from the telecom and nanoelectronic industry.

Since the advent of different in vivo biophotonic methods, such as optogenetics, and new nanobiosensing approaches and clinical applications, there has been an increasing need of optical tools and convergence of integrated nanophotonic technologies toward lifescience. However, biological tissue are dispersive media and they do not allow a straightforward propagation and control of light inside organs and, specifically, in the brain.

In this paper we have shown how the scientific community is facing this challenge. So far light delivery to deep brain regions cannot be achieved without being extremely close or in contact with the region of interest, and the proposed approaches always rely on waveguides, optical fibers or wired LEDs mounted on rigid or flexible supports. At the current stage, optogenetics allow to target genetically defined classes of neurons, while the recently developed multipoint devices add the feature of controlling closely spaced neurons belonging to the same class also in deep brain regions, complementing the high spatial resolution obtained with microscopy in the firsts cortical layers.

#### REFERENCES


Moreover, the integration of multipoint extracellular recording of neural activity is allowing for unprecedented spatial resolved stimulation/readout in free-moving animals (Grosenick et al., 2015) and the integrated drug delivery systems improved spatial matching of viral injection, light delivery and extracellular readout (Canales et al., 2015; Jeong et al., 2015a). Although, most of these technologies were not yet used for novel biological insights, it is clear that a new generation of approaches better able to interface with the extreme complexity and diversity of brain topology and connectivity will shortly represent a key for neuroscientists to answer long-standing questions about brain functional connectivity. However, new approaches exploiting a combination of top-down and bottom up fabrication methods, nanophotonics (nanoplasmonics, quantum optical nanoantennas, etc) and new biological and neurophysiological methods are still needed. The final target is the control and distribution of light over thousands of single neurons, or even at sub-cellular level, and their wireless or waveguide-less manipulation and monitoring.

#### AUTHOR CONTRIBUTIONS

All authors listed, have made substantial, direct and intellectual contribution to the work, and approved it for publication.

pathways during action initiation. Nature 494, 238–242. doi: 10.1038/nature 11846


silicon-based neural probes for laminar recording. Biomaterials 37, 25–39. doi: 10.1016/j.biomaterials.2014.10.040


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Pisanello, Sileo and De Vittorio. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Corrigendum: Micro- and Nanotechnologies for Optical Neural Interfaces

#### Ferruccio Pisanello<sup>1</sup> \*, Leonardo Sileo<sup>1</sup> and Massimo De Vittorio1, 2 \*

<sup>1</sup> Center for Biomolecular Nanotechnologies, Istituto Italiano di Tecnologia, Lecce, Italy, <sup>2</sup> Dipartimento di Ingegneria dell'Innovazione, Università del Salento, Lecce, Italy

Keywords: nanotechnology, optogenetics, nanoparticles, neural interfaces, optical fibers

#### **A corrigendum on**

#### Edited and reviewed by:

Ruxandra Vidu, University of California, Davis, USA

> \*Correspondence: Ferruccio Pisanello ferruccio.pisanello@iit.it Massimo De Vittorio massimo.devittorio@iit.it

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 05 September 2016 Accepted: 29 September 2016 Published: 14 October 2016

#### Citation:

Pisanello F, Sileo L and De Vittorio M (2016) Corrigendum: Micro- and Nanotechnologies for Optical Neural Interfaces. Front. Neurosci. 10:468. doi: 10.3389/fnins.2016.00468 **Micro- and Nanotechnologies for Optical Neural Interfaces**

by Pisanello, F., Sileo, L., and De Vittorio, M. (2016). Front. Neurosci. 10:70. doi: 10.3389/fnins.2016.00070

After the publication of this paper Pisanello et al., we have noticed that there was an error in the institutional affiliations of some of the authors. In particular, Leonardo Sileo, who appeared affiliated to "Dipartimento di Ingegneria dell'Innovazione, Università del Salento, Lecce, Italy", should be affiliated instead only to "Center for Biomolecular Nanotechnologies, Istituto Italiano di Tecnologia, Lecce, Italy." Massimo De Vittorio, who appeared affiliated only to "Center for Biomolecular Nanotechnologies, Istituto Italiano di Tecnologia, Lecce, Italy" should be affiliated also to "Dipartimento di Ingegneria dell'Innovazione, Università del Salento, Lecce, Italy." Correct affiliations appear below the title of this corrigendum.

### AUTHOR CONTRIBUTIONS

All authors contributed equally to this work.

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Pisanello, Sileo and De Vittorio. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Can Nanofluidic Chemical Release Enable Fast, High Resolution Neurotransmitter-Based Neurostimulation?

#### Peter D. Jones \* and Martin Stelzle

BioMEMS & Sensors, Natural and Medical Sciences Institute at the University of Tübingen, Reutlingen, Germany

Artificial chemical stimulation could provide improvements over electrical neurostimulation. Physiological neurotransmission between neurons relies on the nanoscale release and propagation of specific chemical signals to spatially-localized receptors. Current knowledge of nanoscale fluid dynamics and nanofluidic technology allows us to envision artificial mechanisms to achieve fast, high resolution neurotransmitter release. Substantial technological development is required to reach this goal. Nanofluidic technology—rather than microfluidic—will be necessary; this should come as no surprise given the nanofluidic nature of neurotransmission. This perspective reviews the state of the art of high resolution electrical neuroprostheses and their anticipated limitations. Chemical release rates from nanopores are compared to rates achieved at synapses and with iontophoresis. A review of microfluidic technology justifies the analysis that microfluidic control of chemical release would be insufficient. Novel nanofluidic mechanisms are discussed, and we propose that hydrophobic gating may allow control of chemical release suitable for mimicking neurotransmission. The limited understanding of hydrophobic gating in artificial nanopores and the challenges of fabrication and large-scale integration of nanofluidic components are emphasized. Development of suitable nanofluidic technology will require dedicated, long-term efforts over many years.

Keywords: nanofluidic, nanopore, microfluidic, neurotransmitter, neurotransmission, chemical neuroprosthesis, hydrophobic gating, artificial synapse

## INTRODUCTION

Neuroprostheses are becoming increasingly important for treatment of neurological disorders (Borton et al., 2013). Cochlear implants have restored hearing to hundreds of thousands of patients (Shannon, 2012). Deep brain stimulation has helped more than 100,000 patients suffering from Parkinson's disease and may treat numerous additional diseases (Lozano and Lipsman, 2013). Retinal prostheses restore basic visual percepts in patients with previously-untreatable conditions such as retinitis pigmentosa (Zrenner, 2013). Stimulation of the somatosensory cortex has been proposed to restore touch and proprioception in patients with prosthetic limbs (Bensmaia, 2015); a first demonstration in humans was recently reported and is expected to be published soon (Sanchez, 2015).

#### Edited by:

Ruxandra Vidu, University of California, Davis, USA

#### Reviewed by:

Giuseppe D'Avenio, Istituto Superiore di Sanità, Italy Vassiliy Tsytsarev, University of Maryland School of Medicine, USA

> \*Correspondence: Peter D. Jones

peter.jones@nmi.de

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 02 January 2016 Accepted: 18 March 2016 Published: 31 March 2016

#### Citation:

Jones PD and Stelzle M (2016) Can Nanofluidic Chemical Release Enable Fast, High Resolution Neurotransmitter-Based Neurostimulation? Front. Neurosci. 10:138. doi: 10.3389/fnins.2016.00138

Current neuroprostheses address the nervous system by electrical stimulation. Other potential modalities include optical and chemical stimulation. A clinical trial investigating optogenetic treatment of retinitis pigmentosa in humans has begun (RetroSense Therapeutics, 2015; Bourzac, 2016), but challenges of low photosensitivity and targeting of specific retinal cells must be overcome and ethical issues and safety of human genetic modification must be addressed (Barrett et al., 2014). Chemical stimulation has been proposed (Iezzi and Finlayson, 2011) and supported by preliminary experiments (Finlayson and Iezzi, 2010; Inayat et al., 2014). Although, implantable drug delivery systems have been used clinically for decades (Penn and Kroin, 1985) and neural probes capable of drug delivery have been demonstrated (Frey et al., 2011; Altuna et al., 2013; Pongrácz et al., 2013), the distinct concept of functional chemical stimulation capable of mimicking neurotransmission remains inaccessible. The absence of suitable technology is the primary roadblock, which will be discussed in more detail below.

### Neuronal Signaling at Chemical Synapses

Neurons communicate with each other by the release of chemical neurotransmitters from the axonal terminal of presynaptic cells into the synaptic cleft (some neurons also use electrical synapses) (Purves et al., 2004). Release occurs by exocytosis of neurotransmitter-containing vesicles, triggered by an influx of calcium ions through voltage-gated channels which open upon depolarization of the axonal terminal by action potentials. Recognition of these chemical signals by specific, spatiallylocalized receptors in the postsynaptic membrane causes postsynaptic neurons to change their transmembrane potential in spatially-restricted areas, such as the synaptic cleft. Local changes of transmembrane potentials propagate across the cellular membrane, allowing integration of multiple synaptic inputs, and triggering successive neurotransmitter release to downstream neurons.

The specificity of chemical signaling relies on recognition of neurotransmitter by postsynaptic receptors. Prevention of continued stimulation of postsynaptic neurons requires removal of neurotransmitters from the extracellular volume. Most neurotransmitters are recycled by uptake into presynaptic neurons or degraded enzymatically.

Physiological concentrations of neurotransmitters cover at least seven orders of magnitude, from nanomolar to more than 100 mM (Featherstone, 2010). However, quantification of both synaptic and ambient extracellular concentrations is challenging (Scimemi and Beato, 2009; Sun et al., 2014). Peak synaptic glutamate concentrations are estimated to be in the low millimolar range (0.5–5 mM) (Featherstone, 2010). Nanomolar concentrations of ambient extracellular glutamate were predicted (Zerangue and Kavanaugh, 1996) and measured (Herman and Jahr, 2007). Other measurements revealed a wide range (25 nM to 10 µM), but higher concentrations are believed to result from damage due to the measurement technique (Sun et al., 2014).

### Electrical Stimulation

Electrical neurostimulation depolarizes cells by extracellular voltage gradients generated by the spread of current injected into tissue by electrodes (Durand, 2000). Efforts have been made to target specific membrane areas or neuronal processes to modulate neuronal signaling by triggering or inhibiting release of neurotransmitters. However, electrical neurostimulation does not mimic physiological neurotransmission as the stimulating electric fields produce unspecific polarization of cells and act in large volumes compared to the size of synapses and neuronal processes. Electric fields may stimulate any cellular structures, depending on their spatial arrangement with nearby electrodes. This contrasts with neurotransmission, in which receptors on local areas of cell membranes respond to specific chemical signals. The main advantage of electrical stimulation compared to chemical stimulation is that the electrical field can be switched within microseconds at arbitrary locations. Technology to inject chemical signals with similar precision and speed does not exist, and chemical signals must rely on other mechanisms for removal.

Neuroprostheses rely on technology adapted from the microelectronics industry. Although modern nanofabrication achieves sub-10-nm resolution, extracellular neurostimulation electrodes maintain dimensions of tens of micrometers. The challenge of injecting sufficient current to excite neurons must be met while avoiding high voltages and dangerous side reactions (Merrill et al., 2005).

Preclinical experiments predicted a resolution limit of tens of micrometers for electrical stimulation of the retina (Stett et al., 2007). Intracortical microstimulation excites neurons hundreds of micrometers away from electrodes, although producing distinct percepts with closely-spaced electrodes has not been investigated (Bensmaia, 2015). Current retinal prostheses approach the predicted limits, with the Alpha IMS prosthesis (Retina Implant AG, Reutlingen, Germany) having 1500 electrodes with 70 µm pitch. The best artificial visual acuity demonstrated until now is 0.037 (Snellen acuity of 20/546), corresponding to a spatial resolution of 126 µm on the retina (Stingl et al., 2013). In comparison, visual acuity achieved by the fovea of healthy human retinas relies on photoreceptor cells with a pitch of ∼5 µm (Hirsch and Curcio, 1989).

New electrode materials such as conducting polymers stimulate at lower voltages (Gerwig et al., 2012; Samba et al., 2015), which may enable the safe operation of smaller electrodes. Investigations with complex electrode configurations (Lorach et al., 2015) and guidance of neurons into features on artificial devices (Adekunle et al., 2015) are attempting to address these issues. However, a recent review emphasizes that electrical stimulation is not expected to restore visual acuity approaching normal vision (Eiber et al., 2013).

Recently, electrical recording and stimulation of single cells has been achieved by nanoscale electrodes (Angle et al., 2015), which may penetrate the cell membrane (Qing et al., 2013) or be engulfed (Hai et al., 2010). Reproduction of these results in vivo will be necessary to evaluate their potential for use in neuroprostheses. Although promising for high resolution stimulation, such methods would continue to rely on unphysiological mechanisms.

### Progress Toward Chemical Stimulation

Researchers have long attempted to mimic neurotransmission. Neurotransmitter release from micropipettes by pressure injection or iontophoresis is an important technique for neurochemical investigations (Lalley, 1999). Such methods confirmed that delivery of glutamate produces physiological responses in retinas, supporting the concept of chemical-based prostheses (Finlayson and Iezzi, 2010; Inayat et al., 2014). Communicating in the extracellular language of neurons allows chemical stimulation to mimic real synaptic neurotransmission (Murnick et al., 2002; Finlayson and Iezzi, 2010; Iezzi and Finlayson, 2011; Inayat et al., 2014).

As with electrical neuroprostheses, advanced chemical stimulation devices would require dense integration by microor nanofabrication, which cannot be realized with micropipettes. Microfluidic devices for parallel chemical stimulation from an array of sites have been demonstrated, which release chemical stimuli through apertures addressed by buried channels (Peterman et al., 2004; Scott et al., 2013). Microfluidic devices were also proposed following pipette-based investigations (Inayat et al., 2014). A critical shortcoming of microfluidic devices is their failure to suppress leakage by diffusion. Although diffusion can be countered by applying ionic currents or withdrawal of the liquid, actively countering leakage of multiple channels would be prohibitively complex. Rather, blocking diffusion requires disrupting the aqueous phase with a phase barrier.

Actuation of a physical barrier is challenging in microfluidics. The classical Quake valve has been produced as small as 6µm but requires multilayer soft lithography in monolithic structures (Araci and Quake, 2012). Solid barriers have been electrochemically opened but are not reversible (Chung et al., 2008). Air has been used to interrupt diffusion, but required a macroscopic gap which cannot be miniaturized or implanted (Zibek et al., 2010).

A further weakness of microfluidic stimulation is its low spatial resolution. Low spatial resolution hinders temporal resolution due to the so-called proximity effect (Iezzi and Finlayson, 2011). When diffusion is the primary transport mechanism, increasing distances cause exponential dilution and slowing of chemical signal transmission (e.g., latency of 1 s for 33 µm; Iezzi and Finlayson, 2011). Microfluidic or pipette-based release distant from target neurons clearly showed the latency of signal transmission (Peterman et al., 2004; Scott et al., 2013; Inayat et al., 2014). However, targeted iontophoretic stimulation of neuronal structures from ∼100 nm pipette tips supports the possibility of fast chemical stimulation (Murnick et al., 2002).

No reported microfluidic technology can control single chemical release sites, and microfluidic structures provide insufficient resolution for communication with neurons. The challenge of controlling of chemical release sites for high resolution chemical release cannot be realized by microfluidics. Mimicry of physiological neurotransmission will require nanofluidic technology. The following section will discuss how nanopores could achieve synapse-like release.

#### NANOFLUIDIC CHEMICAL RELEASE

The emphasis on nanofluidics is not surprising upon reviewing the physical mechanisms of neurotransmission. Neurotransmission relies on vesicular release of neurotransmitters, orchestrated by complex systems of molecular biology in the presynaptic neuron (Purves et al., 2004). Vesicles, 50 nm in diameter, contain up to 40,000 neurotransmitter molecules (Van der Kloot, 1991). Vesicular release occurs at specialized regions of synapses called active zones, with diameters of 200–500 nm (Südhof, 2012). Action potentials drive release of vesicles as fast as 2000 Hz (Kaeser and Regehr, 2014); although not sustainable, this rate can provide an upper estimate of release rates. The product of these values suggests an upper approximation for neurotransmitter release at an active zone of ∼10−<sup>16</sup> mol/s. A helpful comparison: if each molecule is a monovalent ion, this represents a current of 10 pA. The dimensions of active zones and their central role in neurotransmission make them an intriguing target for mimicry with an artificial nanopore device.

Replacement of the presynaptic neuron by an artificial device will require the capability of sustained release of at least the rate achieved by active zones. Real connections between devices and biology are never ideal, so higher release rates may be required to overcome reduced proximity. Release rates calculated below predict that nanofluidic elements can achieve physiologicallyrelevant release rates.

A notable property of synaptic release is its absolute nature. Vesicles enclose neurotransmitter molecules, and chemical transport across the cell membrane is strictly regulated. This contrasts with the limited gating capabilities achieved in artificial devices. Progress in nanofluidics has focused on nanopores: channels narrower than 100 nm fabricated perpendicularly through thin membranes. Exploitation of the influence of surface properties can control transport of ions or fluid, although electrostatic or steric effects have not achieved absolute shut-off (Taghipoor et al., 2015). A suitable technique to prevent diffusion requires a barrier to interrupt the aqueous phase. Hydrophobic gating could provide such a mechanism. Reports of relevant phenomena are discussed below to provide a perspective on what may be possible.

#### Nanofluidic Release Rates

An upper limit for vesicular release at single active zones was estimated above to be 10−<sup>16</sup> mol/s, and provides a reference for discussion of nanofluidic release rates (**Figure 1A**). Another useful value is obtained from microiontophoresis, which can selectively stimulate single synapses to produce physiological responses (Murnick et al., 2002). Typical currents range from a few nanoamperes up to 100 nA (Lalley, 1999). Although direct quantification is challenging (Herr et al., 2008), converting 1 nA directly to monovalent ions is 10−<sup>14</sup> mol/s (**Figure 1A**).

Diffusion and pressure may drive chemical release from a nanopore. Further discussion and derivation are available in the Supplementary Material. The release rate by diffusion, in mol/s, is

$$
\pi\_D = \frac{\pi c\_0 D d^2}{4L},
$$

with source concentration c0, diffusivity D, and nanopore diameter d and length L. This expression assumes a linear concentration gradient along the length of the nanopore.

The release rate by applied pressure is

$$m\_P = \frac{\pi c\_0 \Delta P d^4}{128 \eta L},$$

with pressure 1P and dynamic viscosity η. This expression assumes Hagen–Poiseuille flow without slip and neglects depletion or accumulation at the pore ends.

Example solutions are shown in **Figure 1A** for nanopores of various diameters, showing that nanopores can release chemicals at physiologically-relevant rates. This example uses a length of 500 nm, which would be sufficient to integrate electrodes or other functional elements. A 100 mM source is below synaptic vesicle concentrations. The release rates are sufficient to imitate synaptic transmission, given an appropriate method to turn release on and off.

### Nanofluidic Gating

Hydrophobic gating achieves absolute disruption of chemical transport by control of a vapor-phase barrier (**Figure 1B**). Although ubiquitous in cell membrane protein pores (Aryal et al., 2015), achieving similar effects in artificial nanopores has proven more challenging. Spontaneous nucleation of bubbles is prohibited in pores larger than a few nanometers in diameter, regardless of length (Lefevre et al., 2004; Guillemot et al., 2012). Top-down nanofabrication cannot achieve sufficient precision to mimic these dimensions. For example, hydrophobic gating in protein pores is affected by sub-nanometer changes in dimensions or single amino acid residue substitutions (Yoshimura et al., 1999; Beckstein and Sansom, 2003; Birkner et al., 2012).

Reversible hydrophobic gating in wider artificial nanopores has been demonstrated in response to applied electric fields (Powell et al., 2011; Smirnov et al., 2011). The reversibility demonstrated in these larger pores relies on trapping of bubbles within the pores (Smirnov et al., 2010), and their removal makes wetting irreversible. Deliberate bubble trapping by constrictions or surface chemistry has been suggested but not yet demonstrated (Smirnov et al., 2010; Guillemot et al., 2012). A hydrophobic liquid could provide an alternative to vapor bubbles; this concept has been demonstrated in macroscopic nanoporous membranes (Hou et al., 2015) but not in single nanopores. Reversibility could be achieved by generating bubbles by plasmonic heating (Li et al., 2015), Joule heating (Nagashima et al., 2014), or electrolysis (Chen et al., 2015).

As nanofluidic effects arise from surface properties, precise nanometer-scale control of structure and surface will be necessary to achieve desired functions. A leading example of artificial nanopore fabrication is a wafer-scale process for sub-20-nm-diameter pores with integrated electrodes (Bai et al., 2014). While most applications use homogenous surface chemistry, for example by silane-based modification (Miles et al., 2013), the molecular topography of such surfaces must not be overlooked (Fadeev and McCarthy, 1999). Long-term applications will require sufficient stability of nanopores' structure and surface chemistry. Silicon nitride pores widen due to decomposition and dissolution (Rollings et al., 2015). In biological environments, protein adsorption presents further challenges (Yusko et al., 2011). Nanopores are usually investigated in isolation, and integration of arrays of nanopores remains a challenge. A recent report integrated five pores in individual microfluidic channels (Tahvildari et al., 2015).

These reports hint at what may be possible, while emphasizing the limited robustness and poor understanding of these effects. Robust and reversible nanopore electrowetting has not been demonstrated, and the mechanisms of electric-field-induced wetting of nanopores are not known. However, a generic mechanism can be envisioned (**Figure 1B**). A hydrophobic barrier will block chemical transport by formation of a phase barrier of vapor or a hydrophobic liquid. The Young–Laplace law explains the resistance to wetting of hydrophobic nanopores (Lee and Karnik, 2010), and has been verified in pores as narrow as 2.6 nm (Lefevre et al., 2004). A stimulus will form a water channel across the hydrophobic barrier, allowing chemical transport. This may be achieved by modulating the pore's surface energy or by applying sufficient pressure or voltage to force water into the pore. Dewetting may occur spontaneously or may be driven, for example by heating. For applications in neuroprotheses, control of many nanopores simultaneously must be achieved (**Figure 1C**).

## PROPAGATION OF CHEMICAL SIGNALS

Propagation of chemical signals occurs throughout the nervous system with diverse spatial and temporal scales (Syková and Nicholson, 2008; Vizi and Lendvai, 2008; Rusakov et al., 2011). Neurotransmitters diffuse across the synaptic cleft (∼20 nm) faster than 1 µs; diffusion also drives volume transmission over larger distances at time scales of minutes or longer. Extracellular diffusion can be studied by iontophoretic injection and electrochemical detection of a tracer molecule (Nicholson et al., 1979). The diffusion equation can be modified to consider extracellular volume fraction and tortuosity, but such approximations are invalid in specific micro- or nanoscale geometries. Moreover, uptake and enzymatic reactions influence extracellular chemical signals.

The synapses of cone photoreceptor cells demonstrate the potential complexity of nanoscale chemical signal transmission (Regus-Leidig and Brandstätter, 2012). Neurotransmitter release from these neurons addresses multiple postsynaptic cells, whose responses depend on their spatial proximity to the active zone. Postsynaptic dendrites localized at the active zone receive rapid high neurotransmitter concentrations, while cells which contact the photoreceptors farther from the active zone receive smoother, lower concentrations. The responses of these postsynaptic neurons correlate with the nanoscale propagation of neurotransmitter.

The proximity effect is a challenge for chemical stimulation: diffusive transport to larger distances requires exponentially longer times and leads to exponential dilution (Iezzi and Finlayson, 2011). The dimensions of microfluidic chemical stimulation devices suggest a limit of seconds to minutes. However, the chemical communication mechanisms of the brain prove the capability of delivering chemical information over diverse spatial and temporal dimensions. Slow volume transmission across large distances (Syková and Nicholson, 2008) could provide a target for neuromodulatory chemical neuroprostheses. Fast, high resolution stimulation requires intimate contact with cells (Murnick et al., 2002).

Chemical signal propagation from nanopores may be illustrated by analytical solutions to the diffusion equation (Crank, 1975). Expressions for instantaneous or continuous release from a point source are described in the Supplementary Material and illustrated in **Figure 2**. **Figure 2B** shows the rapid rise and fall of a chemical impulse, with micrometer-scale resolution and millisecond-scale time course. **Figure 2C** shows the propagation of constant diffusion from a nanopore with diameter of 50 nm and length of 500 nm, and release by diffusion only (4·10−<sup>16</sup> mol/s). High concentrations rapidly establish near the source, while spread to larger distances is slower. Concentrations at larger distances are limited by continuous diffusion to a steady-state limit. Long release times cannot produce concentrated signals at large distances from a nanopore.

The analytical solution for continuous release also provides an estimate of potential density of independent nanopores. **Figure 2D** illustrates this for an array of nanopores with a pitch of 10 µm. At close distances, the signals are clearly resolved, while larger distances obscure the individual signals. This analytical solution provides only an estimate, without considering interference of the neighboring pores. Improved accuracy of time-varying chemical release from multiple nanopores could be obtained by numerical simulation, which could permit inclusion of cellular structures (Hepburn et al., 2012).

## CONCLUSION/OUTLOOK

The clinical success of electrical neuroprostheses has proven the possibility to treat neurological disorders with artificial devices, while also revealing limitations of addressing complex physiology with comparatively simple electrical methods. Neurotransmitterbased stimulation could enable the ultimate neuroprosthesis by addressing neurons with their own chemical language.

The challenges to be overcome must not be underestimated. The field of microfluidics remains immature in comparison to microelectronics (Becker, 2009; Whitesides, 2011). Nanofluidics has remained especially unexplored due to limitations of fabrication (Whitesides, 2011). Neurotransmitter-based stimulation will require continued developments of microfluidics and nanofluidics. The development of nanoelectrodes (Angle et al., 2015) may provide an indication of the challenges involved in interfacing nanoscale devices with biological systems. However, while electrode development may benefit from expertise of the nanoelectronics industry, similarly mature expertise in nanofluidics does not exist.

A robust nanofluidic gating mechanism is required. Hydrophobic gating may provide the mechanism, although a robust realization of this effect may be different from what has been reported until now. Reliable fabrication and operation of thousands of individual pores must be demonstrated. Large scale integration of individually controlled nanopores with microfluidic and electrical control will be necessary. Certainly, many generations of technology development will precede the realization of the goal of chemical neuroprostheses.

Some challenges can be predicted. Long-term operation in aggressive biological environments will be necessary for the goal of neuroprostheses. The requirement for intimate proximity with target neurons may require advanced biochemical functionalization of the device or release of neurotrophic factors to encourage the neurons' acceptance of an artificial device. Removal of chemical signals must be investigated to avoid excitotoxicity, although surrounding cells may accomplish this, for example by the widely-expressed excitatory amino acid transporters in the retina (Iezzi and Finlayson, 2011).

The complexity of the brain raises an important question: Do we understand the brain well-enough to rationally stimulate it with high resolution chemical signals? Stimulation of well-understood structures including the retina or sensory cortex should be possible (Iezzi and Finlayson, 2011; Bensmaia, 2015). However, deep brain stimulation protocols continue to rely on trial-and-error optimization with patient feedback (Kringelbach et al., 2007). Neurons integrate inputs of up to hundreds of thousands of synapses (Stuart and Spruston, 2015). If nanopores are envisioned to replace single synaptic inputs, an artificial device cannot be expected to interface with complex brain areas. However, as "megascience" efforts turn their focus to neuroscience, coming years may see an acceleration of our understanding of the brain (Grillner, 2014) which may reveal new possibilities for treatment of neurological disorders by artificial chemical stimulation.

### AUTHOR CONTRIBUTIONS

PJ and MS developed the concepts described in this work. PJ wrote the manuscript. PJ and MS revised the work and approved the final version.

#### ACKNOWLEDGMENTS

This work was supported by the European Commission (7th Framework Programme, Marie Curie Initial systems Training Network "NAMASEN," contract n. 264872) and the VW Foundation under their Experiment! Program. We thank Prof. Dr. Dieter P. Kern, Alexandra Wal, Dr. Paolo Cesare, and Prof. Dr. Elke Guenther for helpful discussions. We are grateful for critical feedback on the manuscript from Dr. Günther Zeck and Dr. Alfred Stett. Graphs were produced with matplotlib (Hunter, 2007) and the Anaconda Python distribution (Continuum Analytics). Figures were composed with Inkscape. Python scripts

#### REFERENCES


and full details to reproduce the figures are included in the Supplementary Material.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: http://journal.frontiersin.org/article/10.3389/fnins. 2016.00138

topography contributes to contact angle hysteresis. Langmuir 15, 3759–3766. doi: 10.1021/la981486o


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Jones and Stelzle. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Corrigendum: Can Nanofluidic Chemical Release Enable Fast, High Resolution Neurotransmitter-Based Neurostimulation?

#### Peter D. Jones \* and Martin Stelzle

BioMEMS & Sensors, Natural and Medical Sciences Institute, University of Tübingen, Reutlingen, Germany

Keywords: nanofluidic, nanopore, microfluidic, neurotransmitter, neurotransmission, chemical neuroprosthesis, hydrophobic gating, artificial synapse

#### **A corrigendum on**

#### Edited by:

Ruxandra Vidu, University of California, Davis, USA

> Reviewed by: Denis Scaini, University of Trieste, Italy

> > \*Correspondence: Peter D. Jones peter.jones@nmi.de

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

> Received: 27 April 2016 Accepted: 07 July 2016 Published: 21 July 2016

#### Citation:

Jones PD and Stelzle M (2016) Corrigendum: Can Nanofluidic Chemical Release Enable Fast, High Resolution Neurotransmitter-Based Neurostimulation? Front. Neurosci. 10:341. doi: 10.3389/fnins.2016.00341

#### **Can Nanofluidic Chemical Release Enable Fast, High Resolution Neurotransmitter-Based Neurostimulation?**

by Jones, P. D., and Stelzle, M. (2016). Front. Neurosci. 10:138. doi: 10.3389/fnins.2016.00138

We cited a publication by Scott et al. which was not the correct reference. The correct reference is "A microfluidic microelectrode array for simultaneous electrophysiology, chemical stimulation, and imaging of brain slices" (Scott et al., 2013). We apologize for the mistake.

## AUTHOR CONTRIBUTIONS

All authors listed, have made substantial, direct and intellectual contribution to the work, and approved it for publication.

#### REFERENCES

Scott, A., Weir, K., Easton, C., Huynh, W., Moody, W. J., and Folch, A. (2013). A microfluidic microelectrode array for simultaneous electrophysiology, chemical stimulation, and imaging of brain slices. Lab Chip 13, 527–535. doi: 10.1039/c2lc40826k

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Jones and Stelzle. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Peripheral Neuron Survival and Outgrowth on Graphene

#### Domenica Convertino1,2, Stefano Luin<sup>1</sup> , Laura Marchetti <sup>2</sup> \* and Camilla Coletti <sup>2</sup> \*

<sup>1</sup> NEST, Scuola Normale Superiore, Pisa, Italy, <sup>2</sup> Center for Nanotechnology Innovation @NEST, Istituto Italiano di Tecnologia, Pisa, Italy

Graphene displays properties that make it appealing for neuroregenerative medicine, yet its interaction with peripheral neurons has been scarcely investigated. Here, we culture on graphene two established models for peripheral neurons: PC12 cells and DRG primary neurons. We perform a nano-resolved analysis of polymeric coatings on graphene and combine optical microscopy and viability assays to assess the material cytocompatibility and influence on differentiation. We find that differentiated PC12 cells display a remarkably increased neurite length on graphene (up to 27%) with respect to controls. Notably, DRG primary neurons survive both on bare and coated graphene. They present dense axonal networks on coated graphene, while they form cell islets characterized by dense axonal bundles on uncoated graphene. These findings indicate that graphene holds potential for nerve tissue regeneration and might pave the road to novel concepts of active nerve conduits.

#### Edited by:

Mario I. Romero-Ortega, University of Texas at Dallas, United States

#### Reviewed by:

Petra Scholze, Medical University of Vienna, Austria Carlos Vicario-Abejón, Consejo Superior de Investigaciones Científicas (CSIC), Spain Yael Hanein, Tel Aviv University, Israel

#### \*Correspondence:

Laura Marchetti laura.marchetti@iit.it Camilla Coletti camilla.coletti@iit.it

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 10 October 2017 Accepted: 03 January 2018 Published: 22 January 2018

#### Citation:

Convertino D, Luin S, Marchetti L and Coletti C (2018) Peripheral Neuron Survival and Outgrowth on Graphene. Front. Neurosci. 12:1. doi: 10.3389/fnins.2018.00001 Keywords: graphene, neuron culture coating, peripheral DRG neuron, PC12, differentiation

## INTRODUCTION

A specific feature of peripheral nerves is the ability to spontaneously regenerate after traumatic injuries. In the presence of important gaps where an end-to-end suture is not possible, a surgical approach is used, where nerve conduits (generally, autografts, or allografts) are used as bridges between the nerve stumps and provide physical guidance for the axons (Faroni et al., 2015). However, they present limitations in functional recovery and other disadvantages, e.g., size mismatch and increasing healing time for autografts, and rejection and disease transmission for allografts (Daly et al., 2012). A promising alternative is represented by tissue engineered nerve grafts, that have shown to improve regeneration, reduce scar formation and increase the concentration of neurotrophic factors (Gu et al., 2014; Faroni et al., 2015). Among materials that can be used for the guide production, silicon stimulates excessive scar tissue formation thus lacking long-term stability, while some other natural polymers, such as collagen and chitosan, lack adequate mechanical and electrical properties (Tran et al., 2009; Fraczek-Szczypta, 2014; Pinho et al., 2016). In recent years, new materials have been suggested as alternative candidates for tissue engineering applications. In particular graphene and other carbon-based nanomaterials have been proposed in life-science applications and nerve tissue regeneration (Fraczek-Szczypta, 2014; Kostarelos and Novoselov, 2014; Ding et al., 2015).

Graphene is a monolayer of sp2-hybridized carbon atoms arranged in a two-dimensional honeycomb lattice that was first isolated in 2004 from graphite (Novoselov et al., 2004). The increasing research interest in graphene is due to its incredible properties: high electron mobility (also at room temperature), superior mechanical properties both in flexibility and strength, high thermal conductivity and high area/volume ratio (Lee et al., 2008; Castro Neto et al., 2009).

**109**

Furthermore, its biocompatibility and chemical stability make it ideally suited for biomedical applications (Bitounis et al., 2013).

Several studies have used graphene-based materials as biocompatible substrates for growth, differentiation and stimulation of different cell types, including neural cells (Agarwal et al., 2010; Li et al., 2011; Park et al., 2011; Bendali et al., 2013; Sahni et al., 2013; Tang et al., 2013; Bramini et al., 2016; Defterali et al., 2016; Fabbro et al., 2016; Guo et al., 2016; Rauti et al., 2016; Veliev et al., 2016). Polymer-coated graphene was shown to enhance the differentiation of neural stem cells (NSC) into neurons (Park et al., 2011), influencing their passive and active bioelectric properties (Tang et al., 2013; Guo et al., 2016). In addition, coated graphene-based materials were found to accelerate neurite sprouting and outgrowth of mouse hippocampal neurons (Li et al., 2011) and PC12 cells (Agarwal et al., 2010). A number of studies have also analyzed the effect of uncoated graphene-based materials on neural cells. Defterali et al. showed that uncoated thermally reduced graphene favored neural stem cells differentiation (Defterali et al., 2016). Neuron synapse formation and activity were not affected by graphene produced by liquid phase exfoliation (Fabbro et al., 2016), while an impairment of excitatory transmission was observed in primary neurons following a chronic exposure to graphene oxide flakes (Bramini et al., 2016; Rauti et al., 2016). Bare graphene was shown to be biocompatible, sustaining neuron survival and neurite outgrowth (Bendali et al., 2013; Sahni et al., 2013; Veliev et al., 2016), although the presence of defects may reduce the neural affinity, preventing cell attachment (Veliev et al., 2016). To date, most biomedical studies have investigated graphene covalent-functionalized forms such as graphene oxide (GO) and its chemical reduction known as reduced graphene oxide (RGO), or liquid phase exfoliated graphene (Agarwal et al., 2010; Bitounis et al., 2013; Bramini et al., 2016; Defterali et al., 2016; Fabbro et al., 2016; Rauti et al., 2016; Liu et al., 2017). These graphene-like structures have altered electronic structure and physical properties due to the variable fraction of sp2 and sp3 hybridized carbon atoms. With respect to those graphene-based materials, pristine graphene offers enhanced electrical and tribological properties and most notably an excellent electrical conductivity thus prospecting advantages for nervous system regeneration applications. Indeed, it has been demonstrated that conductive materials can enhance the electric field produced by the cell, influencing cell bioelectric properties (Guo et al., 2016). Electrical stimulation can also enhance and directs neurite outgrowth (Schmidt et al., 1997; Meng, 2014) and can accelerate axonal elongation (Fraczek-Szczypta, 2014). Neural conductive interfaces for neural regeneration application usually exploit conductive polymers, such as polyethylenedioxythiophene (PEDOT) and polypyrrole (PPy), or composite materials whose conductivity depends on the inclusion of graphene or carbon nanotubes (CNTs) (Schmidt et al., 1997; Deng et al., 2011; Pinho et al., 2016). Recently, graphene and carbon nanotubes (CNTs) have been successfully used to improve recording and electrical stimulation of neurons (Keefer et al., 2008; Kuzum et al., 2014) and surprisingly neural microelectrode arrays (MEAs) fabricated using graphene performed better than gold and indium tin oxide (ITO), in terms of signal-to-noise ratio (SNR) (Rastegar et al., 2017).

To date, the interaction between pristine graphene and peripheral neural cells has been investigated only in two studies (Lee et al., 2012; Hong et al., 2014), which suggest a positive effect on neurite outgrowth and proliferation when using graphene coated with fetal bovine serum (FBS). However, in both studies bare glass is used as control, thus the effect on the results of FBS coating, which per se is not a traditional coating for neural cells (Sun et al., 2012), is not investigated. No detailed study has yet examined the homogeneity and quality of the coatings typically adopted in neuronal culture. Predicting how polymeric surface coatings distribute onto graphene, due to its hydrophobicity and extreme flatness, is by no means trivial; furthermore, understanding how nerve cells can sense graphene under extracellular-matrix-like coatings is crucially important for possible in vivo applications. Overall, this lack of studies on pristine graphene leaves other carbon-based materials such as carbon nanofibers (CNF), carbon nanotubes (CNT), GO and rGO to star in its play (Ku et al., 2013; Fraczek-Szczypta, 2014; Ding et al., 2015; Liu et al., 2017).

In this work we investigate the potential of graphene as a conductive peripheral neural interface. We select epitaxial graphene obtained via thermal decomposition on silicon carbide (SiC) (Starke et al., 2012) as the ideal substrate for such investigations. In fact, epitaxial graphene on SiC combines high crystalline quality, scalability, thickness homogeneity and an extreme cleanliness. Graphene is used as a substrate for two cellular models: (i) PC12 cells, a non-neuronal cell line that is able to differentiate upon Nerve Growth Factor (NGF) stimulation and constitutes a widely-used model for peripheral sympathetic neurons (Greene and Tischler, 1982); (ii) dorsal root ganglion (DRG) sensory neurons, which are used as a model to study regenerative axon growth (Chierzi et al., 2005). The homogeneity and quality of a number of polymeric coatings typically adopted for neuronal culturing is investigated, and the most suitable ones are identified and adopted for the reported cultures. Furthermore, DRG neurons are also interfaced with bare graphene to assess their interaction with graphene per se, in the absence of a coating. Optical microscopy is used to investigate neurite length, number and differentiation, while viability assays are used to assess cytocompatibility. We compared results on monolayer graphene on SiC (G) with the ones on 4 possible control substrates: hydrogen etched SiC (SiC), gold coated glass coverslip (Au), glass coverslip (Glass) and polystyrene plate (well). The latter, being routinely used in cell culture procedures, was used as classic control. SiC controls were implemented since graphene was grown directly on such substrates, which display a good biocompatibility (Saddow et al., 2011) and present prospects for neural implants (Frewin et al., 2013). Finally, glass coverslips were coated with a very thin layer of gold to mimic the graphene layer grown on SiC. We used gold substrates as conductive controls, as gold, together with platinum (Pt, especially its porous form Pt-black), titanium nitride (TiN) and iridium oxide (IrOx), is typically interfaced with neurons in the fabrication of biomedical electrodes (Kim et al., 2014; Obien et al., 2015); Pt-Black, TiN, and IrOx are useful for the increased effective surface (Aregueta-Robles et al., 2014).

## MATERIALS AND METHODS

## Substrates Preparation and Characterization

Graphene on SiC was prepared by adopting a technique which allows to obtain quasi-free standing monolayer graphene (QFMLG) (Riedl et al., 2009). Briefly, buffer layer graphene was obtained via thermal decomposition of on-axis 4H-SiC(0001) performed at 1,250◦C in argon atmosphere. QFMLG was obtained by hydrogen intercalating the buffer layer samples at 900◦C in molecular hydrogen at atmospheric pressure (Bianco et al., 2015). The controls adopted in the experiments were: (i) Hydrogen etched SiC(0001) dices (the same substrates where graphene was grown) were cleaned with HF to remove the oxide layer, and hydrogen etched at a temperature of 1,250◦C as previously reported (Frewin et al., 2009). (ii) Gold coated glass coverslips were obtained by thermally evaporating on the coverslips, previously cleaned with oxygen plasma, a 2 nm titanium adhesive layer and a 4 nm thin gold layer. (iii) Bare glass coverslips were treated overnight with 65% nitric acid (Sigma-Aldrich). (iv) Polystyrene 48-well plates (Corning). The dimensions of all the substrates were about 6 × 6 mm<sup>2</sup> . The topography of the samples as well as the graphene number of layers and quality were assessed by both AFM and Raman spectroscopy (Figure S1). Before cell culture, all substrates were sterilized by 30 min immersion in 96% ethanol and then rinsed several times with deionized (DI) water.

#### Surfaces Functionalization

Samples were coated with different polymeric solutions suggested for the targeted cell cultures and AFM analyses were performed to investigate the morphology of such coatings on graphene and the controls. The following solutions were tested: 100µg/ml Poly-L-lysine (PLL) solution in water (Sigma-Aldrich), 200µg/ml Collagene Type I (Sigma-Aldrich) in DI water, 30µg/ml Poly-Dlysine (PDL) (Sigma-Aldrich) in PBS, 30µg/ml PDL and 5µg/ml laminin (Life Technologies) in PBS. The samples were incubated with the coating solution at 37◦C for 1, 4, and 12 h and rinsed three times in DI water before analyzing their topography via AFM. AFM was performed in tapping mode on samples with and without the polymeric coating, over several areas up to 10 × 10µm wide. AFM micrographs were analyzed using the software Gwyddion 2.45.

#### PC12 Cell Culture

PC12 cells (ATCC <sup>R</sup> CRL-1721TM) were maintained in a humidified atmosphere at 37◦C, 5% CO2 in RPMI 1640 medium supplemented with 10% horse serum, 5% fetal bovine serum, 1% penicillin/streptomycin and 1% L-glutamine (Gibco). Cells were plated at ∼40–60% confluency onto the substrates previously coated with 100µg/ml Poly-L-lysine solution (PLL) in water (Sigma-Aldrich). Differentiation was achieved using two different procedures: (1) direct addition of 50 ng/ml NGF (Alomone Labs) in complete cell medium after seeding; (2) a 5–6 days priming with 15 ng/ml NGF in complete medium, followed by seeding on the substrates with 50 ng/ml NGF in RPMI medium supplemented with 1% horse serum, 0.5% fetal bovine serum, 1% penicillin/streptomycin and 1% L-glutamine. In both cases, 2/3 of the medium was renewed every 2–3 days. With the second procedure an improved differentiation was observed. The cells were observed at different time points using an inverted microscope equipped with a 20×/40× magnification objective (Leica DMI4000B microscope). Typically, 10 fields per sample were acquired to perform morphometric analysis of PC12 differentiation. Three parameters were measured as previously reported (Marchetti et al., 2014): (i) the percentage of differentiated cells (Diff), determined counting the number of cells with at least one neurite with a length equal to or longer than the cell body diameter; (ii) the average number of neurites per cell in the field (av. neurites/cell); (iii) the mean neurite length measuring the longest neurite of each differentiated cell in the field (length). The calculated values of Diff, Av. neurites/cell and Length are reported in **Figure 2**. Cell viability was assessed with the Cell counting Kit-8 assay (CCK-8, Sigma-Aldrich), based on quantification of WST reduction due to the metabolic activity of viable cells. Samples were prepared according to the manufacturer's instructions and measured at the GloMax <sup>R</sup> Discover multiplate reader (Promega). The results are reported as % over the polystyrene well, considered as control. All the experiments were repeated at least twice independently.

### DRG Cell Culture

Rat Embryonic Dorsal Root Ganglion Neurons (R-EDRG-515 AMP, Lonza) cells were maintained in a humidified atmosphere at 37◦C, 5% CO2 in Primary Neuron Basal Medium (PNBM, Lonza) supplemented with L-glutamine, antibiotics and NSF-1 (at a final concentration of 2%) as recommended by the manufacturer. Neurons were plated on the substrates previously coated with a PBS solution of 30µg/ml Poly-D-lysine (Sigma-Aldrich) (PDL) and 5µg/ml laminin (Life Technologies). The medium was always supplemented with 100 ng/ml of NGF (Alomone Labs). Since 24 h after seeding, 25µM AraC (Sigma-Aldrich) was added for inhibition of glia proliferation. Half of the medium was replaced every 3–4 days. Neurons were observed at different time points using an inverted microscope (Leica DMI4000B microscope).

## Statistical Analysis

For all the experiments, we performed two independent cultures with two biological duplicates each. For the morphometric analysis of the PC12 cells, for each substrate we analyzed at least 200 cells (nc = number of cell) from selected fields (nf = number of field) of the four replicates (two biological duplicates per culture) obtained with a 40× objective (Au: nf = 17, nc = 203; Glass: nf = 33, nc = 1,106; G: nf = 42, nc = 877; SiC: nf = 35, nc = 1,004; well: nf = 37, nc = 724). For the DRG neurons we analyzed nf fields using a 40× objective for a total of nc cells for each substrate (day 1: Au, nf = 13, nc = 67, Glass: nf = 14, nc = 75; G: nf = 13, nc = 29; SiC: nf = 12, nc = 35; day 2: Au, nf = 16, nc = 89, Glass: nf = 13, nc = 100, G: nf = 12, nc = 34, SiC: nf = 11, nc = 37). The number of cells analyzed (nc) is the total pool of the four experiments. All data are expressed as the average value (mean) ± standard error of the mean (SE) unless stated otherwise. Data were analyzed by using Origin Software and nonparametric Kruskal–Wallis test with Dunn's multiple comparison test were used for statistical significance with <sup>∗</sup>p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001.

#### RESULTS AND DISCUSSION

## Polymeric Coating of Epitaxial Graphene and Control Substrates

NGF-induced neurite outgrowth of PC12 cells is favored by their adhesion on a substrate. This is typically achieved by coating the dish surfaces with polymers such as poly-L-lysine or biologically derived collagen (Greene and Tischler, 1982). We applied a water solution of both these coatings to all substrates adopted for our cultures and analyzed by AFM the quality and homogeneity of the coatings after different incubation times, i.e., 1, 4, and 12 h. **Figures 1A,B** show AFM phase and topography micrographs for the two different coatings and different incubation times on a graphene substrate. Clearly, the Poly-L-Lysine (PLL) coating presents better homogeneity with respect to Collagen Type I coating for which networklike aggregates can be detected (**Figures 1D,E**). On the other hand, PLL tends to form a homogeneous carpet of spots of 1– 2 nm (no aggregates) independent from the incubation time. We also analyzed the same coatings on SiC, gold and glass surfaces. On SiC, PLL and Collagen presented analogous topographies (Figures S2a,b). Due to the higher surface roughness of gold and glass substrates (presenting rms roughnesses of about 1 nm comparable to the features of the polymeric layer), no conclusions about the quality of the coating could be drawn (Figures S3a,b). However, presence of the coating was confirmed by the variation in the hydrophilicity observed with contact angle measurements (Figure S3c). Hence, for the PC12 cells cultured in this work, a PLL coating with an incubation time of 4 h was adopted.

The same characterization was performed for the polymeric coatings typically suggested for DRG neurons, i.e., PBS solution of Poly-D-Lysine (PDL) alone and PDL with laminin. **Figure 1C** shows the AFM topography and phase images taken for PDL/laminin coated graphene substrates for the three different incubation times (i.e., 1, 4, and 12 h). Also in this case, after the coating, an increased roughness was observed for all time points and in particular the formation of a network-like structure was consistently observed (**Figure 1F**). PDL alone coating gave rise to a similar net (Figure S4b). In order to exclude the effect of PBS, we dissolved the same polymeric amount in DI water and after 4h incubation we observed similar structures (Figure S4a). To check if the different molecule arrangement of PLL and PDL on graphene was dependent on their concentration, we tested also a PDL coating solution in DI water with the same concentration used for PLL (100µg/ml). We obtained structures similar to the ones observed for the lower PDL concentration (Figure S4c). On SiC no network formation was observed with or without laminin (Figures S2c,d). The stability of the coating was confirmed for

all the probed incubation times. In this case, PDL with laminin coating (with an incubation time of 4 h) was selected to carry on the following DRG culture experiments in order to mimic the extracellular matrix.

Interestingly, the coating solutions distributed differently on graphene and SiC, despite their similar morphologies before the coating, with nanometric terraces and comparable roughness (Figures S5a,b). All polymeric coatings exhibited similar distributions on SiC, while there were significant differences between the coatings on graphene. The dissimilar arrangement of the coatings on the substrates can be reasonably ascribed to the different hydrophilicity of graphene and SiC (Oliveros et al., 2011). As shown by the contact angle measurements reported in Figure S5c, graphene is in any instance (pre and post-coating) more hydrophobic than SiC. The contact angle estimated for graphene was 95.8◦ ± 1.3◦ while it was 38.3◦ ± 7.2◦ for SiC, in agreement with literature (Coletti et al., 2007; Wang et al., 2009; Oliveros et al., 2011). SiC hydrophilicity likely facilitated

(for all the various coatings adopted) a homogenous adhesion of molecules. The network-like structures often revealed by our analysis on graphene indicate that such pristine hydrophobic surfaces are less prone to be homogenously coated, an important aspect that should be considered in future works when studying cell cultures on graphene.

## Neurite Outgrowth of PC12 Cell on Graphene

We first investigated the effect of graphene on PC12 cells. **Figure 2A** reports typical optical micrographs obtained for PC12 cells cultured at day 5 (in the presence and absence of NGF) and at day 7 (with NGF) on the different substrates. The analyses conducted at day 5 evidence that almost no differentiation took place in the absence of NGF, while a significant neurite outgrowth occurred on all substrates upon NGF treatment.

Selected morphometric parameters describing the differentiation process were quantified at day 5 and are reported in **Figures 2B–D**: the percentage of differentiated cells in the fields (Diff), the average number of neurites per cell (av. neurites/cell) and the length of the longest neurite per differentiated cell (length). This analysis showed that 50% of the cells on graphene differentiate with a mean neurite length of 52.3µm (**Figures 2B,C**). Remarkably, the average length was significantly longer on graphene than on glass (∗∗∗), well (∗∗∗) and SiC (<sup>∗</sup> ) by 27, 22, and 13%, respectively. The percentage of differentiation on graphene was better than on glass (<sup>∗</sup> ), while the average number of neurites per cell was lower on graphene than on the control well (∗∗∗). These results indicate that PC12 cells grow longer neurites on graphene, with a neuronal differentiation that is comparable to that obtained for the standard control wells. Differently from reference (Hong et al., 2014), we did not observe increased PC12 proliferation on graphene, which could be due to the effect of the FBS coating used in that study. Furthermore, we found that at day 7 living PC12 cells forming neurite networks were present on all the substrates. To better assess graphene cytocompatibility, the viability of undifferentiated PC12 cells was assessed after 3, 5, and 7 days of culture and no statistically significant differences were observed between graphene and the other substrates (**Figure 2E**). These data are in agreement with previous observations that graphene induces neurite sprouting and outgrowth of hippocampal neurons due to an overexpression of growth-associated protein-43 (GAP-43) (Li et al., 2011). Also, Lee et al. showed an induced neurite outgrowth of human neuroblastoma (SH-SY5Y) cells on graphene, probably mediated by focal adhesion kinase (FAK) and p38 mitogen-activated protein kinase (MAPK) cascades and upregulation of genes involved in neurogenesis (NFL, nestin and MAP2) (Lee et al., 2015). Both the studies excluded a neurogenic effect from substrate topography and wettability. Thus, we speculate that also for PC12 cells, graphene surface chemistry and electrical conductivity can specifically increase neurite length during differentiation.

## DRG Primary Neurons on Graphene

Next, we investigated the effect of graphene on primary neurons using dorsal root ganglion (DRG) cells while using the same controls adopted in the previous culture. As motivated in section Polymeric Coating of Epitaxial Graphene and Control Substrates, all the samples were coated with PDL/laminin. **Figure 3A** shows typical optical microscopy images obtained at 1, 4, 9, and 15 days of culture. Starting from day 4, we observed numerous processes and an increase in the cell body area (Figure S6) and in the neurite length (**Figure 3A**). Neurons were observed on all the

DRG neurons cultured on gold (Au), glass coverslip, graphene (G) and SiC coated with Poly-D-lysine and laminin (30µg/ml PDL and 5µg/ml laminin in PBS, 4 h incubation) at different days of culture. Scale bar: 50µm. (B) Axon length quantification at 24 and 48 h after cell seeding. We analyzed nf fields for a total of nc cells for each substrate (day 1: Au, nf = 13, nc = 67, Glass: nf = 14, nc = 75; G: nf = 13, nc = 29; SiC: nf = 12, nc = 35; day 2: Au, nf = 16, nc = 89, Glass: nf = 13, nc = 100, G: nf = 12, nc = 34, SiC: nf = 11, nc = 37) and data are reported as mean ± SE. (C) DRG neurons on bare gold and graphene at day 10. Scale bar: 100µm.

substrates up to 17 days of culture. We observed that both at day 1 and day 2 the average axon length was higher on graphene than on the other substrates (**Figure 3B**). This observation confirms the trend reported for PC12, although in this case no statistical significance was retrieved. Axonal length was not quantified for longer culturing times due to the highly dense network forming after day 2 (see day 9 and 15 in **Figure 3A**).

Given that neuronal growth was previously reported also for non-coated graphene (Wang et al., 2011; Bendali et al., 2013; Sahni et al., 2013; Defterali et al., 2016; Fabbro et al., 2016; Veliev et al., 2016; Keshavan et al., 2017), we tested also the bare substrates to observe their effect on the neurons. Differently from non-coated glass, where they did not survive, DRG neurons could be nicely cultured on non-coated graphene and gold up to 17 days. On coated graphene neurons distributed homogeneously on the entire samples (**Figure 3A** and Figure S7a), while on uncoated graphene neurons formed small interconnected cell islets already after 24h from seeding (Figure S7a). After 2-3 days of culture, we observed neurites sprouted from the islet toward the substrate, and at longer times neurons formed cell bodies aggregates and neurite bundles (**Figure 3C** and Figures S7a,c), probably due to a reduced neural adhesion in the absence of coating, as previously observed for retinal ganglion cells (Bendali et al., 2013) or cortical neurons (Sahni et al., 2013). We rarely observed neurite bundles on coated graphene, while they were present on uncoated graphene already after 2 days of culture and they increased in size with time (Figure S7d). Cell body area was comparable with the one on coated graphene. Higher cell body area on uncoated graphene was observed starting from day 4, but the values did not differ significatively (Figure S7b).

In order to improve adhesion and neuron homogeneous distribution, the surface modification with an hydrophilic coating turned out to be useful (Li et al., 2011; Keshavan et al., 2017). Moreover, as previously suggested, the coating could mask the presence of surface inhomogeneity and defects that affect neural adhesion (Veliev et al., 2016).

Concerning material stability issues, it should be noted that graphene showed a good stability and remained intact during the entire culturing period, as revealed by Raman measurements after cell removal (Figure S8).

## CONCLUSION

This work provides novel data about the use of graphene as a substrate for peripheral neuron cultures. We chose to use graphene on SiC because, thanks to its high quality and cleanliness, it allowed us to examine the graphene effect on peripheral neurons with fewer concerns for contaminations and crystalline quality that may affect neuron adhesion (Veliev et al., 2016).

We use the PC12 cell line as a consolidated model for peripheral sympathetic neurons and show that such cells grow well on graphene with an increased neurite length (up to 27%) at 5 days of differentiation when compared to controls. Remarkably, graphene performs better than gold, which we used as conductive control. Culture of DRG neurons also shows a positive outcome on graphene: neurons survive both on bare and coated graphene until day 17, with a dense axon network that is comparable to the control substrates. In order to investigate graphene influence on axonal outgrowth, further studies are necessary, e.g., using compartmentalized chambers (Taylor et al., 2005). The obtained results confirm the potential of graphene as an active substrate in conduit devices for nerve guidance: it would allow the transmission of electrical signals between neurons and make external electrical stimulation feasible to enhance axon regeneration. While for many biomedical applications graphenebased materials with higher roughness might be desirable, in specific cases when high transparency and electrical conductivity are required, pristine highly crystalline graphene might be the

ideal choice (Kuzum et al., 2014; Reina et al., 2017). It should be noted that flexibility is a requirement in neural regeneration that cannot be met by using graphene on SiC. To use graphene as neural interface other graphene production methods, such as chemical vapor deposition (CVD), could be more suitable.

### AUTHOR CONTRIBUTIONS

DC, LM, and CC conceptualized the study, DC performed the experiments. All the authors discussed the results. The manuscript was written through contribution of all authors.

#### REFERENCES


#### ACKNOWLEDGMENTS

The authors would like to thank Fabio Beltram and Giovanni Signore from NEST-Scuola Normale Superiore for fruitful discussion.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnins. 2018.00001/full#supplementary-material


substrate for inducing neurite outgrowth. Biochem. Biophys. Res. Commun. 460, 267–273. doi: 10.1016/j.bbrc.2015.03.023


intercalation. Phys. Rev. Lett. 103, 1–4. doi: 10.1103/PhysRevLett.103. 246804


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Convertino, Luin, Marchetti and Coletti. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Capacity, Fidelity, and Noise Tolerance of Associative Spatial-Temporal Memories Based on Memristive Neuromorphic Networks

Dmitri Gavrilov <sup>1</sup> \*, Dmitri Strukov <sup>2</sup> and Konstantin K. Likharev <sup>3</sup>

<sup>1</sup> Department of Electrical and Computer Engineering, Stony Brook University, Stony Brook, NY, United States, <sup>2</sup> Department of Electrical and Computer Engineering, University of California, Santa Barbara, Santa Barbara, CA, United States, <sup>3</sup> Department of Physics and Astronomy, Stony Brook University, Stony Brook, NY, United States

We have calculated key characteristics of associative (content-addressable) spatial-temporal memories based on neuromorphic networks with restricted connectivity—"CrossNets." Such networks may be naturally implemented in nanoelectronic hardware using hybrid memristive circuits, which may feature extremely high energy efficiency, approaching that of biological cortical circuits, at much higher operation speed. Our numerical simulations, in some cases confirmed by analytical calculations, show that the characteristics depend substantially on the method of information recording into the memory. Of the four methods we have explored, two methods look especially promising—one based on the quadratic programming, and the other one being a specific discrete version of the gradient descent. The latter method provides a slightly lower memory capacity (at the same fidelity) then the former one, but it allows local recording, which may be more readily implemented in nanoelectronic hardware. Most importantly, at the synchronous retrieval, both methods provide a capacity higher than that of the well-known Ternary Content-Addressable Memories with the same number of nonvolatile memory cells (e.g., memristors), though the input noise immunity of the CrossNet memories is lower.

Keywords: spatial-temporal memories, associative memories, nanoelectronics, neuromorphic networks, memristors, CrossNets, capacity, noise tolerance

## INTRODUCTION

Associative spatial-temporal memories (ASTM), which record a time sequence of similarly-formatted spatial patterns, and then may reproduce the whole sequence upon the input of just one of these patterns (possibly, contaminated by noise), are valuable parts of cognitive systems. Indeed, we all know how a few overheard notes trigger our memory of an almost-forgotten tune. Such observations have been confirmed by detailed neurobiological studies of "episodic memories", apparently localized in the hippocampus—see, e.g., the recent review by Eichenbaum (2013). Another example (which also gives a very natural language for the description of spatial-temporal patterns, used in this paper), is a reproduction of a movie, triggered by the input of its one, possibly incomplete or partly corrupted, frame. More generally, multi-dimensional associative memories may be used for a broad range of cognitive tasks—see, e.g., Imani et al. (2017) for recent literature.

#### Edited by:

Mikhail Lebedev, Duke University, United States

#### Reviewed by:

Martin Ziegler, Christian-Albrechts-Universität zu Kiel, Germany Christoph Richter, Technische Universität München, Germany

\*Correspondence:

Dmitri Gavrilov dmitri.gavrilov@stonybrook.edu

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 27 December 2017 Accepted: 12 March 2018 Published: 28 March 2018

#### Citation:

Gavrilov D, Strukov D and Likharev KK (2018) Capacity, Fidelity, and Noise Tolerance of Associative Spatial-Temporal Memories Based on Memristive Neuromorphic Networks. Front. Neurosci. 12:195. doi: 10.3389/fnins.2018.00195

The recent fast progress of mixed-signal nanoelectronic hardware, in particular of hybrid memristive circuits. Such circuits, which are based on nanoelectronic crossbars, with a memristive device (for example, a metal-oxide memristor) at each crosspoint (see e.g., the reviews by Likharev, 2008; Yang et al., 2013), may enable ASTMs with extremely high speed and energy efficiency. One option here is to use the so-called Ternary Content-Addressable Memory (T-CAM) architecture—(see, e.g., Pagiamtzis and Sheikholeslami, 2006). Indeed, as was discussed by Alibart et al. (2011), the memristive version of such a memory requires just two crosspoint devices per cell. As a result, the total number n of such devices in an associative memory holding Q spatial patterns ("frames"), of N bits ("binary pixels") each, is just 2NQ, i.e., is only twice larger than that necessary for the usual binary resistive memory, with no noise correction ability (It will be more convenient for us to discuss it operation in section Comparison with T-CAM).

In this paper, we will show that these hardware costs may be reduced even further using the hybrid neuromorphic networks with "CrossNets" architecture (see, e.g., Fölling et al., 2001; Türel et al., 2004; Likharev, 2011; Merrikh-Bayat et al., 2015; Adam et al., 2017), in which continuous-state memristive synaptic devices work together with CMOS-implemented neural cells see **Figure 1**. If the voltages V<sup>j</sup> , developed by the neural cells and applied to the crossbar input lines, are not too large (for typical metal-oxide memristors, below ∼1 V), they do not alter the preset states of the crosspoint devices, and the crossbar, with the virtual-ground condition Vout ≈ 0 enforced on its output lines, performs a multiplication of the vector of these voltages by the matrix of synaptic weights:

$$I\_i = \sum\_{j=1}^{M} \omega\_{ij} V\_j,\tag{1}$$

where I<sup>i</sup> is the output current (which serves as an input signal for the recipient neural cell), M is the cell connectivity, and the synaptic weight wij is, in the simplest case, proportional to the Ohmic conductance Gij of the corresponding device (A modification of the relation wij ∞ Gij, beneficial for practical implementation, will be discussed in section Readout Options). Hence the memristive crossbar, with continually and precisely adjustable crosspoint devices, can perform, on the physical level, the neuromorphic network's most common inferencestage operation, which is the main bottleneck at their digital implementation. As a result, the intercell communication delays in nanoelectronic CrossNets may be reduced to just few nanoseconds, and their energy efficiency may approach that of the human cerebral cortex.

The global connectivity of a limited number N of neuron cells, with M = N – 1, may be implemented by placing the cells peripherally, around a single N × N crossbar. However, for most real-world applications, such global connectivity is redundant, and an area-distributed interface between a memristive crossbar and an array of CMOS-implemented neurons may be used to provide the desired restricted connectivity graph. For example, the very natural "InBar" interface topology (Türel et al., 2004) may ensure the connectivity of each neuron with all other neurons in its vicinity with a shape approaching that of a square m × m, so that M = m<sup>2</sup> – 1 < N, see **Figure 2** (For practically interesting cases, 1 << M << N). Such shape of the connectivity domain is very convenient for the discussion of the CrossNet ASTM (though not necessary for its physical implementation), and will be used in this paper.

**Figure 3** shows the basic idea of operation of the memory. Just as in **Figure 2**, the neural cells are mapped on a rectangular grid,

FIGURE 1 | The equivalent circuit of the simplest memristive crossbar that can provide adjustable, nonvolatile coupling between neural cells.

each cell corresponding to one B/W pixel of all movie frames. At the movie recording stage, for each pair of sequential frames, the synaptic weight connecting two pixels, within their connectivity domain, is strengthened if the two pixels have the same value (1 or 0) in both frames, and is weakened in the opposite case. For example, in the case of **Figure 3**, where the pixels of a certain value (say, 1) are placed on gray background, the weights wij and w<sup>i</sup> ′ j ′ (symbolized by solid lines) are strengthened, while the weights w<sup>i</sup> ′ <sup>j</sup> and wij′ (dashed lines) are weakened.

If the recording procedure has been efficient, then at the readout (also called the "retrieval") stage, the activation of pixels by those of an input frame leads to a correct sequential activation of the following frames of the movie, even if the input frame is either incomplete, or partly corrupted by noise. **Figure 4** shows an example of such operation. A set of 4 different movies, with 25 frames each, obtained by applying edge detection to grayscale movies showing running humans, was recorded into a simulated ASTM of the same size (N = 120 × 160 = 19,200), with connectivity M = 31 × 31 – 1 = 960. The left column shows three frames of one of the original movies: the initial Frame 1, an intermediate Frame 8, and the final Frame 25. The right column shows the brightness-coded snapshots of the spikes at the retrieval, triggered by the first frame of this particular movie (Three middle snapshots correspond to the same recorded Frame 8, separated by very small time intervals; their difference will be explained in section Readout Options below).

One can see that the movie retrieval is almost perfect (The same network, without any change of synaptic weights wij, gives an equally fair reproduction of any of 3 other movies, when triggered by its frame). However, such faithful retrieval is only possible when the total number Q of the frames does not exceed a certain number Qmax, called the memory capacity. This limitation, Q < Qmax is due to the fact that different frame pairs typically impose contradictory requirements on the same synaptic weight wij.

The general idea of such operation of the ASTM is not quite new. Its software aspects were repeatedly discussed starting from the 1960s—see, e.g., Grossberg (1969). A review of the initial work, mostly for the firing-rate networks, may be found in section 3.5 of Hertz et al. (1991). This idea was revitalized (Gerstner et al., 1993) at the advent of spiking network research, and in this context, discussed in quite a few publications—see, e.g., the reviews (Kremer, 2001; Wörgötter and Porr, 2005), and later papers (Yoshioka et al., 2007; Brea et al., 2011; Nguen et al., 2012; Kabasov et al., 2013; Yu et al., 2016). However, to the best of our knowledge, the key issue of the ASTM capacity was addressed only in the Ph.D. thesis by Wills (2004), for a very specific readout timing model, very inconvenient for hardware implementation (The capacity calculated in that work is also substantially lower than for the best readout methods described below).

The objective of this work was a detailed study of the recording and readout methods, which would enable the highest capacity of the CrossNet ASTM. In the next section, we start with the discussion of the readout options, in particular the most critical issue of readout timing, and proceed to the definition of four most plausible ways of data recording. The following section contains the results of analysis of the proposed methods, including calculation of the corresponding capacity-vs.-fidelity tradeoffs. The best two recording methods are discussed in more detail, with emphasis on their immunity to the input frame corruption and crosspoint device variability. The section ends with the comparison of the performance of ASTM with that of the memristive T-CAM suggested by Alibart et al. (2011). Finally, in the Discussion section we summarize our results, and discuss prospects of experimental implementation of ultrafast CrossNet ASTM.

#### METHODS FOR RECORDING AND RETRIEVING DATA FOR ASTM

#### Readout Options

As follows from the above qualitative description of the memory, it is quite suitable for the asynchronous spiking mode of operation—see, e.g., Gerstner and Kistler (2002). In this mode, the input of all initial frame's "active" pixels (say, equal to 1) triggers simultaneous spikes Vj(t) at the outputs of the corresponding neural cells. As a result of their action on the memristive crossbar, all other cells of the system receive input pulsesIi(t) described by Equation (1). In some of the cells (ideally, all and only those corresponding to the active pixels of the next frame), the input pulses promote the action potential beyond the spiking threshold, causing them to fire. This new series of spikes triggers spiking in the next cell set, corresponding to the active pixels of next frame, etc.

We have carried out extensive numerical experiments with this readout mode, using the simple leaky integrate-and-fire (LIF) model of the cells (Gerstner and Kistler, 2002), and the following shape of the spike:

$$g(t) = C \sin\left(\frac{t}{\pi}\right) \exp\left(-\frac{t}{2\pi}\right), \quad t \ge 0, 1$$

where constant C is used for scaling the amplitude of the spike and time constant τ is selected based on the desired spike duration. The particular spike shape and the values of C and τ are not critical for the effects discussed in this paper.

The simulations have shown the following very interesting behavior, illustrated by **Figure 4**. In the absence of global synchronization, the cells corresponding to active pixels of a frame (besides the initial one) do not necessarily fire simultaneously, because of the previously accumulated individual action potentials, which are practically random. Only the initial frame 1, triggered by the simultaneous input spikes, is reproduced perfectly. In the typical intermediate frame 8, the spikes are spread in time—see the readout snapshots 8A−8C, made with small time intervals between them. If the number Q of the recorded frames is much smaller than the memory capacity Qmax (for a particular recording method), this "jitter" of the spikes is almost negligible. As Q is increased, the jitter also increases, but the spikes belonging to each frame remain clustered in time, with the cluster width not exceeding the average distance between the frames, and time-averaged contents of each frame is still reproduced correctly, i.e., the jitter does not accumulate—see, e.g., the much later frame 25 in **Figure 4**. Only when Q approaches Qmax, the spiking time clusters are getting blurred, and the reproduced movie eventually degrades into noise.

We believe that the observed effect may be rather interesting for theoretical neuroscience, and has to be studied in more detail. However, we could not help noticing that the elementary global timing (synchronization) of all spikes of each frame kills this jitter, simultaneously increasing the memory capacity Qmax rather dramatically. Such global timing may be achieved without much hardware overhead—for example, just by a periodic simultaneous lowering of the firing thresholds of all the cells, with a time period somewhat larger than the characteristic time of the RC-transient in the crossbar.

Because of this, the qualitative results presented in the balance of this paper are for the globally-synchronous readout mode. In order to analyze this mode, we have used the following simple model (which blurs the difference between spiking and firing-rate operation): for each time period, corresponding to the reproduction of one frame of the movie, the voltages V<sup>j</sup> and currents I<sup>i</sup> in Equation (1) are considered constant, with each neural cell providing a static threshold activation function Vi (q+1) <sup>=</sup> <sup>f</sup>(I<sup>i</sup> (q) ), where the upper index is the frame number (q = 1, 2, . . . , Q). The activation function was taken in the simple form

$$V\_i^{(q+1)} = V\_0 \text{sgn } I\_i^{(q)},\tag{2}$$

where V<sup>0</sup> is a constant coefficient, selected for convenience of implementation, and influencing only the scaling of synaptic weights (In our simulations, we set V0=1 without loss of result generality). The relation (2) implies the zero-centered operation mode, in which V<sup>i</sup> , I<sup>i</sup> , and wij may be either positive or negative. This mode may be naturally implemented in differential CrossNets, in which the jth cell contribution to the current input of the ith cell is the sum of currents through two crosspoint devices (both with positive conductances G), fed by equal voltages of opposite polarities:

$$I\_i = \sum\_{j=1}^{M} \left( G\_{ij}^+ V\_j - G\_{ij}^- V\_j \right)^2$$

so that the effective synaptic weight wij ∞ Gij <sup>+</sup> – <sup>G</sup>ij <sup>−</sup> may have an arbitrary sign (Türel et al., 2004); such mode is also convenient for the compensation of the temperature dependence of memristor conductances—see, e.g., Prezioso et al. (2015).

#### Recording Methods

At the first stage of our work, we have explored the tradeoff between the movie retrieval fidelity (in terms of the probability of the correct readout, in a statistical ensemble of random frames) and the network capacity Qmax, for four most natural methods of movie recording, temporarily assuming perfect hardware operation.

#### The Hebb Rule

Conceptually, the most straightforward recording method is using the Hebb rule in its simplest form (Amari, 1972):

$$\boldsymbol{w}\_{ij} = \frac{1}{Q} \sum\_{q=1}^{Q} s\_i^{(q+1)} s\_j^{(q)},\tag{3}$$

where s<sup>j</sup> (q) = ±1 are the symmetrized values of the B/W pixels of the qth frame. This rule evidently corresponds to the verbal description of the weight setup discussed in the Introduction. It is suitable for in situ recording in hardware, using the spike-timedependent plasticity (STDP)—see, e.g., Markram et al. (2012). For that, the network is exposed to a periodic sequence of external signals, each period corresponding to the one recorded frame, with each frame pixel signal acting on an individual network's neuron. When the frame are applied, each "active" binary pixel is causing the respective neuron to fire. Under the effect of these spikes, the STDP performs either reinforcement or weakening of synaptic weights based on the timing of the spikes of the connected neurons. Practical details of such implementation of the STDP with memristive crossbars are discussed by Prezioso et al. (2016).

For simulation, this method is also the simplest one, giving an explicit expression for each synaptic weight.

#### Quadratic Programming

The Hebb rule, described by Equation (3), does not guarantee the perfect recording, because the contributions into wij, given by each pair of frames, may be (and typically are) mutually contradictory. Better performance may be expected from imposing the minimal requirement forI<sup>i</sup> (q) to have the same sign as the proper next frame's pixel s<sup>i</sup> (<sup>q</sup> <sup>+</sup> 1):

$$\left(s\_i^{\{q+1\}}\right)\_{I\_i}(q) \propto \sum\_{j=1}^{M} s\_i^{\{q+1\}} s\_j^{\{q\}} \,\w\_{\vec{w}} > 0, \quad \text{for } i = 1, 2, \ldots N,\tag{4}$$

for each pixel of every frame.

According to the algebra basics (see, e.g., p. 13 in Bertsekas, 1995), the system of NQ inequalities (4) for 2N(M – 1) > NQ binary weights wij only defines a multi-dimensional polygon in the weight space, so that for getting a unique solution for the weight set, it must be complemented with some reasonable additional conditions. With this goal, we have first tried several available algorithms of the linear programming (Vanderbei, 2014); however, they typically lead to growth of the width of the synaptic weight distribution, especially strong at Q →Qmax. Such a broad distribution is rather inconvenient for the hardware implementations, in which the range of possible crosspoint device conductances G is always limited—see, e.g., Merrikh-Bayat et al. (2015). We have achieved much better results by using the quadratic programming (Best, 2017), in which Equation (4) is complemented with the requirement of the smallest norm of the vector of synaptic weights wij. The calculations have been performed using the MATLAB's function quadprog().

This recording method is computationally rather intensive, requiring CPU times approximately two orders of magnitude larger than the Hebb rule (3). Also, since the quadratic programming is a global optimization algorithm, we are not aware of any it's possible in situ analog-hardware implementation without involving a very significant digital-circuit (i.e., essentially ex-situ) overhead.

#### Analog Gradient Descent

The next natural recoding method is an iterative algorithm similar to the well-known delta-rule of the feedforward perceptron training, describing the gradient descent of the quadratic error function (see, e.g., section 5.4 in Hertz et al., 1991):

$$
\Delta w\_{i\bar{j}} = -\eta s\_{\bar{j}}^{(q)} s\_{i}^{(q+1)}.\tag{5}
$$

Here η is a small training rate (in simulations we used the value η = 0.001, though its choice does not affect the results much), and ε is the error of the previous prediction of the next frame's pixel:

$$\varepsilon\_i^{(q+1)} = \sum\_{j=1}^{M} w\_{ij} s\_j^{(q)} - s\_i^{(q+1)}.\tag{6}$$

Since the weight updates are computed and applied separately for each pair of the consecutive movie frames, this recording algorithm is effectively implementing the stochastic gradient descent optimization with the batch size of one frame pair. The number of times the complete movie is applied to the ASTM represents the number of training epochs.

#### Discrete Gradient Descent

We have found that the analog gradient descent method may be improved by rounding the sum a<sup>i</sup> (q) to the closest of ±1:

$$a\_i^{(q+1)} = roundd\left(a\_i^{(q)}\right) - s\_i^{(q+1)}, \qquad a\_i^{(q)} = \sum\_{j=1}^{M} w\_{ij} s\_j^{(q)}$$

thus limiting the error ε<sup>i</sup> (q+1) to the set of values {−2, 0, 2}. However, preliminary testing showed some limitations of this approach, which tends to result in synaptic weights that produce very small values of a<sup>i</sup> (q) , and hence unnecessarily increase the system's sensitivity to noise.

This deficiency may be easily eliminated by introducing a small gap with the range [–D, +D] and modifying the optimization procedure, so that the values of a<sup>i</sup> (q) are "pushed" outside the gap in the process of reducing the cost function. For the case of binary pixels with values ±1, the resulting expression for the error takes form

$$s\_i^{(q+1)} = s\_i^{(q+1)} - s\_i^{(q+1)},\tag{7}$$

where the integer S depends not only on the current prediction of the output pixel, as in Equation (6), but also on its proper value:

$$S\_i^{\{q+1\}} = \text{sgn}\left(\sum\_{j=1}^M \boldsymbol{w}\_{ij} s\_j^{\{q\}} - D s\_i^{\{q+1\}}\right). \tag{8}$$

Simulations showed that changing the width of the gap D (before training) may be used to proportionally scale all the weights of the network, without changing its performance. Therefore, the selection of D can be based solely on the implementation convenience. In our simulations we chose D = 1. The weights are updated according to Equation (5) with training rate η = 0.005.

#### Simulation Procedures

The ASTM models discussed in this paper represent complicated nonlinear systems that are difficult to evaluate using analytical methods. Therefore, with one notable exception discussed in section Hebb Rule below, we had to use numerical simulations to estimate the performance characteristics of the system, including first the ASTM capacity, and then its sensitivity to pixel and weight noise.

All simulations have been performed on a square lattice of N × N neural cells. In order to mitigate the effects of large but finite size N, we have used the usual cyclic boundary conditions on both pairs of opposite sides of the square (equivalent to wrapping the network on a thorus).

In order to exclude the effects of hardly-controllable pixel correlation in real-life B/W movies, such as shown in **Figure 4**, the memories were evaluated on movies composed of fully random frames. This approach may be further justified by taking into account that in order to be recorded in a binary memory like ours, an analog or multi-bit (say gray-scale or color) pixel needs to be represented by many binary pixels, whose correlation, averaged over the whole connectivity domain, is vanishingly small. Besides the data shown in **Figure 9**, the duty cycle of each frame, i.e., the percentage of "active" (say white) pixels was 50%.

Each memory readout trial starts with exposing it to a randomly selected frame of the recorded movie, and then using Equation (1) to sequentially recover all the remaining frames (Since each recorded movie formed a closed loop, the readout continued up until the input frame was reached again). If this last frame virtually matched the input one, the recovery was considered successful. The memory capacity was determined as the maximum length of the movie that could be recorded and read out with a 1% fidelity.

The sensitivity to pixel noise was evaluated by "flipping" a certain percentage of random pixels in the initial frame. In this case, the final frame of each readout attempt was compared with the uncorrupted version of the input one. The sensitivity to pixel noise, was characterized with the percentage of failures to recover the movie correctly, as a function of the corrupted input bit number and the movie length (Typically, the memory either reproduces the movie perfectly after a few first frames, or completely corrupts it).

The sensitivity to weight noise was calculated similarly, except that Gaussian noise was added to each synaptic weight before each readout attempt. In this case, the memory may be capable of reproducing the movie correctly, but with small percentage of wrong pixels in each frame. This is why, in order to evaluate the noise sensitivity, we have set a certain threshold on the percentage of errors in the final frame, used to decide whether the readout is successful. The results presented below are for the thresholds of 1 and 3%.

Due to the stochastic nature of these numerical experiments, getting accurate results requires their averaging over large number of simulation experiments. Each such experiment used for the memory capacity evaluation, included using a new, randomly generated movie, with just one readout run, starting with a random frame (The number of experiments used to obtain each data point is specified in figure captions). The noise sensitivity evaluations were based on 10 series of experiments, with each series using a unique movie and 100 attempts to recover the movie, starting from a random frame, after adding a new random noise pattern—to either the initial frame or to synaptic weights. This procedure provided 1000 data points for each final (average) point shown below.

## RESULTS

#### Hebb Rule

Since the Hebb Rule gives an explicit expression (3) for the synaptic weights, the resulting capacity-to-fidelity tradeoff may be readily evaluated analytically, assuming that all the binary pixels in the whole movie are random and uncorrelated. Indeed, let us assume that in a frame number q, all M cells within the connectivity domain of an ith cell have correct values: V<sup>j</sup> (q) = V0s<sup>j</sup> (q) . Then plugging Equation (3) (with the summation index replacement q→q') into Equation (1), we may calculate the normalized product of the signal I<sup>j</sup> arriving at the j th cell, by the sign of its correct value, s<sup>i</sup> (<sup>q</sup> <sup>+</sup> 1), in the next frame:

$$\frac{Q}{V\_0} I\_i^{(q)} s\_i^{(q+1)} = \sum\_{j=1}^{M} \sum\_{q'=1}^{Q} s\_i^{(q+1)} s\_i^{(q'+1)} s\_j^{(q')} s\_j^{(q)}.\tag{9}$$

Due to the independence of different pixels, the sum of MQ terms in the right-hand part of Equation (9) has only M terms (all with q = q') always equal to +1, while all other terms have an equal probability to equal either +1 or−1. At M, Q >> 1, the sum of these M(Q – 1) random terms has a Gaussian probability distribution with a zero statistical average, and the variance equal to M(Q – 1) ≈ MQ. As a result, the probability of the negative sign of the whole sum (Equation 9), i.e., of an error of the ith pixel in the (q + 1)st frame, is

$$p \approx \frac{1}{\sqrt{2\pi M Q}} \int\_{M}^{\infty} \exp\left(-\frac{x^2}{2M Q}\right) dx \equiv \frac{1}{2} \text{erfc}\sqrt{\frac{M}{2Q}},\tag{10}$$

where erfc(x) = 1 – erf(x) is the complementary error function. Note that this result is similar to that for the Hopfield networks with the similarly sharp activation function (see, e.g., section 2.2 in Hertz et al. (1991), and similarly restricted connectivity (Türel et al., 2004), because for this calculation, the addition of 1 to the upper indices in the right-hand part of Equation (9) is not important.

In the most important limit of small error probability, Equation (10) is reduced to

$$p \approx \sqrt{\frac{Q}{2\pi M}} \exp\left\{-\frac{M}{2Q}\right\} < <\text{ I, for } 1 < $$

At larger p, we need to take into account the induced errors, i.e., the effect of an error in a q th frame on the error probability in the (q + 1)st frame. At Q, M >> 1, such a calculation may be performed analytically using the mean-field approach, similar to that used for the calculation of the Hopfield network's capacity see, e.g., section 2.5 in Hertz et al. (1991). However, since the main focus of this work was on other recording methods, giving better results (see below) we have opted for the simple numerical simulation of the readout. These numerical experiments have shown that at the retrieval process, the fraction of incorrect pixels per frame rapidly approaches some stationary, equilibrium value p; these values are plotted by points in **Figure 5** for several N and M. For all the simulated cases, the normalized memory capacity Qmax/M is virtually independent of these parameters—just as in Equations (10) and (11).

For the practically interesting fidelity range (p ≤ 1%), the corrections due to induced errors are not important, and the numerical results, with a good accuracy, are described by Equations (10) and (11). In particular, for the 99% fidelity (p = 0.01), Qmax ≈ 0.18M. Such low capacity is not too surprising, given the well-known result Qmax ≈ 0.14M for the Hopfield networks with the similarly restricted connectivity (Türel et al., 2004), and the similar activation function.

#### Quadratic Programming

The simulations have shown that with the growth of the number Q of the recorded frames, the correct retrieval degradation is different from that at the Hebb-rule recording. Namely, the number of wrong pixels in each retrieved frame is typically very small, but when a few errors appear, they almost immediately lead to a complete corruption of the remaining frames of the movie. As a result, the system's fidelity violation is better characterized by the probability p of the movie corruption, measured on a large statistical ensemble of different movies (again, with completely random and independent pixels).

**Figure 6** shows the p so defined as a function of the same ratio Q/M as in **Figure 5**. The results, which were obtained by first recording and then replaying randomly generated movie for each simulation run, show that for a reasonable fidelity (say, p = 1%),

FIGURE 5 | The probability of pixel retrieval error in the ASTM using the Hebbian recording (Equation 3), as a function of the normalized number Q of the recorded frames. Lower curve: Equation (10). Dashed curve: Equation (11), valid only for small probability of pixel errors. Upper points: numerical simulation results, which automatically take into account the induced errors.

the network capacity Qmax, averaged over two simulated cases (M = 440 and M = 960), is (1.75 ± 0.05)M, i.e., is almost an order of magnitude higher than for the Hebb-rule recording.

Note that for the case of global connectivity (M = N – 1), this number is close to the theoretical capacity maximum Qmax = 2(N – 1) of the usual (spatial) associative memory, based on a recurrent neuromorphic network (Gardner and Derrida, 1988).

#### Analog Gradient Descent

The data recording was performed by iteratively updating weights according to Equations (5) and (6) until the optimization algorithm converged to global minimum of prediction error. We assumed that the solution is near the global minimum when the magnitudes of all errors (Equation 6) drop below 0.1. Preliminary testing showed that setting a more stringent criterion did not improve the capacity or noise sensitivity of the network. In cases when the algorithm was not converging, iterations were stopped after 10<sup>5</sup> epochs.

The numerical simulation has shown that the movie retrieval dynamics is qualitatively similar to that for the quadratic programming (see the previous subsection): an increase of the number Q of the recorded frames leads to an increase of the probability p of the total corruption of the retrieved movie. **Figure 7** shows a typical dependence of this probability on the ratio Q/M; it indicates that the memory's capacity is approximately two times lower than that for the quadraticprogramming recording; for p = 1%, Q ≈ 0.97M [Similarly to the shown QP results, in **Figure 7** (and **Figure 8** below) each simulation run involved recording and replaying randomly generated movie].

In hindsight, such relatively poor results might be anticipated. Indeed, the algorithm (Equations 5, 6) forces the network outputs to approach the exact integer values s<sup>i</sup> (<sup>q</sup> <sup>+</sup> 1) of the next pixels, while for the successful movie retrieval, it is only necessary for it to have its sign correct—see Equation (2). As the result, the unnecessary changes of the weights interfere with the substantial ones, and hinder the iterations' efficiency.

## Digital Gradient Descent

The ASTM recording procedure was organized similarly to that for Analog Gradient Descent, except that in the iterative optimization algorithm weight updates are determined by Equations (5), (7), and (8). **Figure 8** shows the probability of the retrieved movie corruption as a function of the normalized

gradient-descent recording (The lines are only guides for the eye). N = 101 × <sup>101</sup> <sup>=</sup> 10,201, <sup>M</sup> <sup>=</sup> <sup>21</sup> <sup>×</sup> 21 – 1 <sup>=</sup> 440, <sup>η</sup> <sup>=</sup> <sup>10</sup>−<sup>3</sup> . The recording iterations (Equation 5) were stopped either after 10<sup>5</sup> epochs, or when the magnitude of all errors (Equation 6) dropped below 0.1. The error bars represent the standard deviation of the mean based on 100 simulations.

number Q of the recorded frames, for several values of parameters N and M. The results imply that the capacity-tofidelity tradeoff is almost as good as that available from the (much less convenient) quadratic programming; for example, at p = 1%, Qmax ≈ (1.67 ± 0.02) M, depending on M and N.

Note that all the CrossNet ASTM capacity results, shown in **Figures 5**–**8**, are for completely random binary (B/W) pixels, i.e., for the 50% probability for each pixel to have a certain value (±1). If this probability is either lower or higher, the capacity is even larger—see, e.g., the results shown in **Figure 9**.

For very sparse patterns (with either d << 1 or 1 – d << 1), even higher capacity may be possible using a natural modification

FIGURE 8 | The probability of the retrieved movie's corruption at the discrete gradient-descent recording described by Equations (5), (7), and (8) (The lines are only guides for the eye). η = 0.01; D = 1. The recording iterations were stopped when the errors εi (q+1), defined by Equation (7), reached 0 for all <sup>i</sup> and q. The error bars represent the standard deviation of the mean, based on 1,000 simulations for M = 440 and 500 simulations for M = 960.

of the recording rules suggested for usual (spatial) associative memories—see, e.g., pp. 52–53 in Hertz et al. (1991). At the software implementations of the memories, these rules are sometimes applied to dense patterns (with d ∼ ½) as well, using their mapping on sparse ones. At the hardware implementation, however, such approach would require an impracticable increase of the necessary resources.

#### Immunity to Noise and Device Variability

To summarize the previous section, two of the methods we have studied, stand out of the competition: the first method, based on the quadratic programming, due to the largest memory capacity, and the second one (based on a discrete version of the gradient descent approach) due to its local nature, enabling hardware implementation of the recording, with a minimal involvement of peripheral circuitry—at a very competitive capacity. These two methods have been chosen for a more detailed study, namely a numerical evaluation of the CrossBar ASTM's immunity to the noise contamination of the input frame, and of its tolerance to random deviations of the synaptic weights from the optimal values calculated at the recording (Such deviations are currently the largest challenge for large-scale applications of metal-oxide memristors and other species of these prospective devices, fabricated using various technologies). Random deviations of weights were simulated by adding random deviations to the original weights before each movie retrieval attempt. The deviations were random and independent, obeying the Gaussian distributions with zero mean, and a relative r.m.s. value r.

The results of these calculations are presented, respectively, in **Figures 10**, **11**. The plots in **Figure 10** show, for example, that if the number Q of frames recorded into an ASTM, by either of the two methods, is at 25% of its maximum capacity (Q = 200), it may recognize the input frame with ∼10% corrupted pixels, but if Q is increased to 400, i.e., to 50% of Qmax, the input noise tolerance drops sharply, to only ∼10−<sup>3</sup> of the pixels.

On the contrary, as **Figure 11** shows, filling of memory has smaller effect on its tolerance to fluctuations of memristive device conductance. For example, if the ASTM with the quadratic programming recording (**Figure 11A**) is filled to 25% of its maximal capacity, its operation is not hindered by weight fluctuations with ∼38% relative r.m.s.. If additional data is written to such memory, so that it is filled to 50% of capacity, weight fluctuations with the r.m.s. above ∼15% cause movie corruption. For the discrete gradient descent recording (**Figure 11B**), the fluctuation tolerance is ∼20% for the 25% memory fill, and ∼5% for the 50% memory fill.

It is important to note that these results characterize not an instant, but rather a gradual suppression or amplification of the input noise. For example, **Figure 12** shows the number of wrong pixels in N = 10,201-pixel frames for 10 simulated movie retrievals, for a system with the cell connectivity M = 440, with Q = 250 frames recorded using the discrete gradient descent method. The plots show that all 500 input errors (which were independent for each retrieval attempt) eventually disappeared in 8 cases, but led to a full movie corruption in two cases (These data are a small part of a set of 1,000 movies, which gave the point

with f = 500/10,201 ≈ 0.049 and the error probability ∼0.2—see the point marked in **Figure 10B**).

bars represent standard deviation of the mean based on 1,000 simulations (10

#### Comparison With T-CAM

randomly generated movies, 100 tests per movie).

The results shown in **Figures 10**, **11** need to be compared with those for the main competitor of the CrossNet ASTM, the T-CAM circuits already mentioned in the Introduction. **Figure 13** shows a 2×3-cell fragment of the memristive T-CAM (Alibart et al., 2011). It is a rectangular matrix of cells, with two binary-state crosspoint devices (plus two diodes) per cell, with each bit stored in the complimentary binary states (ON and OFF) of these two devices, whose order encodes the bit.

Using the same movie language, the N binary pixels of each frame are stored in one row of such ASTM, so that the storage of Q frames requires Q rows. Before the movie retrieval, the row lines are pre-charged to the same voltage V0. The retrieval

is induced by feeding each pair of column lines with voltages {V0, 0}, in the order dictated by the value of the corresponding binary pixel of the input frame. If the bit recorded in a cell corresponds to the input bit (i.e., if the input voltage V<sup>0</sup> > 0 corresponds to the ON state of the corresponding crosspoint device, with a high conductance, while the input voltage 0, to the OFF state, with its very low conductance), the feed does not result in a noticeable current through the cell. As a result, if a recorded frame exactly matches the input one, the row line's voltage stays high. On the contrary, if some bits of a recorded frame are different from those of the input frame, the corresponding row line discharges, with the rate proportional to the number of misfit bits, i.e., to the Hamming distance between these two-bit strings. The discharge rates of all rows are

1000 simulations (10 randomly generated movies, 100 tests per movie).

FIGURE 12 | The input pixel noise suppression by a CrossNet with N = 101 × 101 = 10,201; M = 21 × 21 – 1 = 440 in the process of movie retrieval, as simulated for 10 independent random noise patterns. The recording of the movie with Q = 250 frames was performed using the discrete gradient descent method.

compared by the comparator C, and the row with the slowest rate is assumed to carry the requested frame. After the choice of the row has been made, the whole movie may be played out without any further input (This design may be readily generalized to more than two dimensions—see, e.g., Imani et al., 2017).

The fact that this circuit requires n = 2NQ memristive devices (besides the diodes and the peripheral circuits including the multi-input comparator) may be represented by saying that the frame capacity of the T-CAM with n devices is

$$Q\_{\text{max}} = \frac{n}{2N}.\tag{12}$$

This value should be compared with the result Qmax ≈ 1.75M = (7/8)M at 1% fidelity for the CrossNet ASMT discussed in this paper (for the two best recording methods). Since in that memory, with the differential encoding of the synaptic weights, the total number of crosspoint devices is n = 2MN, that result may be rewritten as

$$Q\_{\text{max}} \approx \frac{7}{8} \frac{n}{N},\tag{13}$$

i.e., the capacity (Equation 12) of the T-CAM with the same number of devices is a factor of 7/4 lower.

If the frame of N binary pixels, submitted to T-CAD for the recognition, has some number (say, fN) of corrupted pixels, there is a chance that its Hamming distance from a wrong recorded frame will be lower than that from the correct frame, so that the memory will recall that wrong frame. Since the Hamming distance between two random strings, of N >> 1 bits each, obeys the Gaussian distribution with the mean N/2 and the variance N/4, the probability of such an error is

$$p = \frac{1}{\sqrt{2\pi \ (N/4)}} \int\_0^{\tilde{N}} \exp\left[-\frac{\left(k - N/2\right)^2}{2\left(N/4\right)}\right] dk$$

$$= \frac{1}{2} \left\{ \text{erf}\left[\left(2f - 1\right)\sqrt{\frac{N}{2}}\right] - \text{erf}\left(-\sqrt{\frac{N}{2}}\right) \right\}.\tag{14}$$

According to this formula, at N >> 1 the error is extremely small until the fraction f of the pixels in the input frame approaches 50% very closely—by the distance of the order of 1/(2N) <sup>1</sup>/<sup>2</sup> << 1. Hence, the noise immunity of the T-CAM is higher than that of the CrossNet ASMT—cf. **Figure 10**.

The crosspoint device fluctuation tolerance of the T-CAM is also higher than that in the CrossNet ASTM. In order to characterize it, we should take into account that the Ohmic conductance G of real-life memristors is non-vanishing even in the OFF state. Hence the voltage decay rate in the line corresponding to the perfect fit to the input frame (**Figure 13**) is NV0GOFF > 0. On the other hand, the average rate of a misfit line discharge is NV0GON/2, with an r.m.s. fluctuation scaling as √ N << N. Hence an error due to the worst-case (simultaneous) fluctuations of the conductances appears only at

$$(G\_{\rm OFF})\_{\rm max} > \frac{1}{2} \left( G\_{\rm ON} \right)\_{\rm max} \tag{15}$$


#### CONCLUSION

Our calculations have shown that hybrid CMOS/memristor circuits with the CrossNet architecture may be indeed used as ASTM, especially if operated in the synchronous mode, with the global external timing of all neural cells. Of the studied information recording methods, two gave the best results for the capacity-fidelity tradeoff and noise tolerance: one using the quadratic programming approach, the second one based on a discrete version of the gradient descent method (The latter method, while providing a slightly lower capacity, is more convenient for the hardware-based recording).

With any of these recording methods, the CrossNet ASTMs may be more hardware-saving than the alternative, T-CAM circuits of the same capacity, by offering higher data recording density per memristor, though the input noise immunity and memristor variability tolerance of the CrossNet ASTM are lower. It is important to note that CrossNet ASTM's capacity increases naturally, without any modifications to the network, for more realistic cases of correlated frames (see **Figure 9**). On the other hand, T-CAM implementations would have to rely on coding and/or compression algorithms, which might have substantial implementation overhead and inferior information capacity.

One more challenge for the experimental implementation of the CrossNet ASCM is the still immature technology of memristor hybridization with underlying CMOS circuits (Chakrabarti et al., 2017). However, the field of possible applications of our results is much broader than the memristorbased networks. For example, they are fully applicable to CrossNet-like circuits using floating-gate memory cells, with analog data recording, as synapses—see, e.g., Hasler and Marr (2013). At the industrial-grade implementation of such cells, they may be quite comparable with memristors in size, and provide almost similar speed and energy efficiency. The recent fast progress of experimental work in this direction (Guo et al., 2016; Merrikh Bayat et al., in press) gives every hope that the CrossNet ASTMs based on such technology may become valuable components of future ultrafast cognitive hardware systems.

#### AUTHOR CONTRIBUTIONS

DG performed numerical studies of ASCM. KL and DS performed the theoretical analysis and supervised the project. All coauthors participated in numerous discussions of the results.

#### FUNDING

This work was supported by a DoD MURI program via the AFOSR Award #FA9550-12-1-0038, and by the ARO under Contract #W91NF-16-1-0302. A part of numerical calculations was performed using supercomputer facilities of the DoD's HCPMP program.

#### ACKNOWLEDGMENTS

The authors are grateful to E. A. Feinberg for stimulating discussions, and to reviewers for useful remarks.

## REFERENCES


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Gavrilov, Strukov and Likharev. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Multisite Attenuated Intracellular Recordings by Extracellular Multielectrode Arrays, a Perspective

Micha E. Spira\*, Nava Shmoel, Shun-Ho M. Huang and Hadas Erez

Department of Neurobiology, The Alexander Silberman Institute of Life Science, The Charles E Smith Family and Prof. Joel Elkes Laboratory for Collaborative Research in Psychobiology, The Harvey M. Kruger Family Center for Nanoscience, Hebrew University of Jerusalem, Jerusalem, Israel

Multielectrode arrays (MEA) are used extensively for basic and applied electrophysiological research of neuronal- and cardiomyocyte-networks. Whereas immense progress has been made in realizing sophisticated MEA platforms of thousands of addressable, high-density, small diameter, low impedance sensors, the quality of the interfaces formed between excitable cells and classical planar sensor has not improved. As a consequence in vitro and in vivo MEA are "blind" to the rich and important "landscape" of sub-threshold synaptic potentials generated by individual neurons. Disregarding this essential fraction of network signaling repertoire has become the standard and almost the "scientific ideology" of MEA users. To overcome the inherent limitations of substrate integrated planar MEA platforms that only record extracellular field potentials, a number of laboratories have developed in vitro MEA for intracellular recordings. Most of these novel devices use vertical nano-rods or nano-wires that penetrate the plasma membrane of cultured cells and record the electrophysiological signaling in a manner similar to sharp intracellular microelectrodes. In parallel, our laboratory began to develop a bioinspired approach in-which cell biological energy resources are harnessed to self-force a cell into intimate contact with extracellular gold mushroom-shaped microelectrodes to record attenuated synaptic- and action-potentials with characteristic features of intracellular recordings. Here we describe some of the experiments that helped evolve the approach and elaborate on the biophysical principles that make it possible to record intracellular potentials by an array of extracellular electrode. We illustrate the qualities and limitations of the method and discuss prospects for further improvement of this technology.

Keywords: electrophysiology, multielectrode-array, action-potential, synaptic-potentials, hippocampus, cardiomyocytes, Aplysia, skeletal- myotubes

## INTRODUCTION

Multi-electrode arrays (MEA) are extensively used for basic and applied electrophysiological research of in vivo and in vitro neuronal and cardiomyocyte networks (Obien et al., 2014; Seymour et al., 2017). The core technology and concepts of contemporary MEA goes back half a century to the pioneering studies of Wise et al. (1970, in vivo) and Thomas et al. (1972, in vitro).

Edited by:

Mikhail Lebedev, Duke University, United States

#### Reviewed by:

Ahmed El Hady, Princeton University, United States Alexander Dityatev, Deutsche Zentrum für Neurodegenerative Erkrankungen (DZNE), Germany

> \*Correspondence: Micha E. Spira spira@cc.huji.ac.il

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 27 December 2017 Accepted: 16 March 2018 Published: 10 April 2018

#### Citation:

Spira ME, Shmoel N, Huang S-HM and Erez H (2018) Multisite Attenuated Intracellular Recordings by Extracellular Multielectrode Arrays, a Perspective. Front. Neurosci. 12:212. doi: 10.3389/fnins.2018.00212

**129**

Whereas immense progress has been made over the last 50 years in realizing sophisticated in vitro and in vivo MEA platforms of thousands of addressable, high-density, small diameter low impedance sensors (for example, Berdondini et al., 2005, 2009; Amin et al., 2016; Jäckel et al., 2017; Jun et al., 2017; Viswam et al., 2017), the quality of the interfaces formed between cultured neurons or cardiomyocytes and the substrate integrated planar sensor has not improved in any significant manner. In fact, this interface remains the weakest link in the chain of electrical coupling between cultured cells and planar electrodes. As in the past, current day in vitro and in vivo planar sensor based MEAs are limited to recordings of extracellular field potentials (FPs) generated by propagating action potentials (APs). These electrodes are "blind" to the rich and important "landscape" of sub-threshold inhibitory, excitatory and electrotonic synaptic potentials generated by individual neurons. As a result, neurons that do not fire APs are not noted. Since in some brain areas and possibly also in cultured neuronal networks, a fraction of the cells fire at very low rates or do not fire at all, their potential contribution to information processing and integration go undetected and ignored (Shoham et al., 2006; Epsztein et al., 2011; Barth and Poulet, 2012). This disregarding of an essential fraction of the signaling repertoire in information processing and plasticity has become the standard and practically speaking the "scientific ideology" for many MEA users. This is the case despite unequivocal documentation that meaningful subthreshold signaling between neurons plays critical roles in neuronal network computations (Spira and Hai, 2013). This irrational neglect and standard reflect a failure of the neuroscientific community to communicate the need to develop novel electrophysiological technologies that will enable to record the entire brain-signaling repertoire. The outstanding successes in harnessing chronic in vivo FP recordings to functionally interface CNS neuronal activity and prostatic devices, to link disrupted groups of neurons, operate cochlear and retinal implants validate the immense importance and contribution of extracellular MEA to basic and applied brain research. However, extracellular recordings of FP-spatiotemporal patterns are not sufficient in themselves to understand how neuronal networks composed of excitatory and inhibitory neurons with heterogeneous membrane properties and synapses compute and undergo activity dependent plastic changes. Likewise, the popular use of planar MEA to extract the mechanisms of information processing, or screen toxins and drugs by recording FPs do not provide sufficient information vis a vis the mechanism by which a network operate and drugs exert their effects. Recall that alterations in membrane excitability or changes in excitatory or inhibitory synaptic efficacies may produce undistinguishable changes in the firing patterns of excitable cells networks.

Cognizant of the biophysical source for the inherent limitations of substrate integrated planar MEA platforms to only record extracellular FPs and the immense contribution expected from the development of multisite long-term "intracellular capable" MEA-platforms to brain research, a small number of laboratories have elegantly begun to develop in vitro MEA for intracellular recordings (for a review (Spira and Hai, 2013; Angle et al., 2015). The majority of these devices use passive or active (transistorized) nano-rods or nano-wires that are designed to penetrate the plasma membrane of cultured cells and record the electrophysiological signaling in a way similar to classical sharp intracellular microelectrodes. Because the diameter of the nano-rods (wires) is in the range of 50–500 nm, the penetration of the sensors through the plasma membrane does not damage the cells. This approach has been successfully used to record both attenuated APs and synaptic potentials from neurons and cardiomyocytes (Tian et al., 2010; Angle and Schaefer, 2012; Duan et al., 2012; Gao et al., 2012; Robinson et al., 2012; Xie et al., 2012; Angle et al., 2014; Lin and Cui, 2014; Lin et al., 2014; Qing et al., 2014; Dipalo et al., 2017; Liu et al., 2017). Recently the laboratory of H. Park (Abbott et al., 2017) reported the simultaneous recordings of attenuated APs from hundreds of cultured primary cardiomyocytes.

Along with the exciting development of cell-penetrating nano-structures technologies, our laboratory began to test a different approach in which micrometer-sized, extracellular gold mushroom-shaped microelectrodes (gMµEs) record attenuated synaptic and APs with characteristic features of intracellular recordings (Spira et al., 2007; Hai et al., 2010a,b; Fendyur and Spira, 2012; Spira and Hai, 2013; Rabieh et al., 2016; Shmoel et al., 2016). In this manuscript we first describe some of the experiments that helped develop this approach and clarified the biophysical principles that enable to record attenuated intracellular potentials by an array of extracellular electrodes. Next we briefly illustrate the qualities and limitations of the intracellular recordings obtained by the extracellular gMµE-MEA, and finally briefly discuss prospects for further improvement of this technology.

#### THE BIOPHYSICAL PRINCIPLES THAT ENABLE CELL-NONINVASIVE EXTRACELLULAR ELECTRODES TO RECORD INTRACELLULAR POTENTIALS

The mode and quality of the electrical coupling between an excitable cell and a substrate integrated planar electrode are defined by the spatial relationships and electrical properties of the living cells, the sensing device and the space between the two (**Figure 1A**). A simplified analog electrical circuit of a cultured neuron adhering to a substrate integrated planar electrode is depicted in **Figure 1A**.

Changing the quantitative relationships between the seal resistance, junctional membrane properties and the electrode impedance will therefore change the mode (extracellular to intracellular), quality (time derivative of the AP to Ohmic recordings) and amplitude (from microvolts to tens of mV) of the recorded potentials (Fromherz, 2003; Shmoel et al., 2016). Along with advances in material science, significant progress has been made over the years in the fabrication of low impedance planar electrodes. Despite extensive attempts to increase the seal resistance increasing the adhesion of the cells to the substrate in practice, the seal resistance has remained relatively low, and sufficient to shunt a large fraction of the currents generated by propagating APs to the bath ground. To the best of our

FIGURE 1 | (Ai) Schematic drawing of a cell (green) residing on a planar sensing electrode (yellow) and the space between them (white). (Aii) An analog electrical circuit of the cell-electrode interface. In the model, the cell's surface area is subdivided into a non-junctional membrane (njm, red) that faces the grounded culture medium, and a junctional membrane (jm, blue) that faces the electrode. Each of these membrane compartments is represented by a resistor and capacitor in parallel Rnjm, Cnjm, Rjm, and Cjm, respectively. The cleft between the cell and the sensor is represented by a resistor (the seal resistance-Rs). The electrode is represented by a resistor and capacitor (Re, Ce, respectively). The electrical coupling coefficient (CC) between a cell and a recording device is defined as the ratio between the voltage recorded by the device (electrode-amplifier) and the voltage generated across the plasma membrane of an excitable cell (Velect/Vcell). The square pulse in between the ground jm and njm is a voltage calibration pulse (Bi,Bii) schematic illustrations of downward displacement of a cell to increase the seal resistance (mp-a fire polished pipette to exert the pressure, µe-intracellular recording electrode). (Biii) Concomitant alterations in the extracellularly recorded FPs (upper traces- black) and intracellular APs (lower traces- red) during the displacement of the neuron's cell body toward the planar electrode. From left to right, initially the increased FPs amplitude is not associated with changes in the intracellular AP amplitude. Thereafter (4th trace), the extracellular FP recorded by the planar electrode transformed to intracellular recordings of an AP. This is associated by reduction in the AP amplitude recorded by the µE. Release of the mechanical pressure led to the reversal of the process (traces 6 and 7). (Biv) Super positioning of the intracellular recorded APs with the sharp electrode and the extracellular planar electrode (multiplied by 8). Note that although the amplitudes of the two APs are different the shapes of the APs are similar (Biii,Biv modified with permission from Cohen et al., 2008). (C) A scanning electron microscope image of a gMµE. (D) a latex bead engulfed by a cultured Aplysia neuron. (E) An electron microscope image of a thin section showing a gMµE (black) tightly engulfed by a cultured PC12 cell.

knowledge, no studies have reported successful attempts to locally increase the conductance of the membrane patch facing substrate integrated planar electrodes for long durations.

Two studies have however demonstrated the exceptional potential to generate effective electrical coupling between cells and integrated planar electrodes. That of Jenkner and Fromherz (1997) using isolated leech neurons followed by that of Cohen et al. (2008) using Aplysia neurons. Both studies demonstrated that when a neuron's cell body, or an axon, is mechanically displaced downwards toward the surface of a planar sensor, the contacting surface area between the cell and the substrate increases gradually (**Figure 1B**). This gradual increase is accompanied by increased FP amplitude (recorded by the planar electrode) but without a significant change in its shape (**Figure 1B**). Concomitant intracellular recordings of the APs by a sharp glass electrode revealed that the intracellular spike amplitude and shape were not altered. This increased amplitude of the extracellular FP was attributed to increased Rseal due to the increased contact area between the neuron and the sensor, and possibly also to reduction in the average cleft width (**Figure 1B**). Surprisingly, further increases in the mechanical pressure transformed the recorded extracellular FP (by the planar sensor) into positive monophasic attenuated APs with characteristic features of classical intracellular recordings (**Figure 1B**). The transformation of an extracellular FP into an intracellular attenuated AP was associated with decreased amplitude of the intracellularly recorded AP by the sharp glass electrode. This indicated that stretching the neuron's plasma membrane against the substrate led in addition to the increased Rseal to a transformation of the coupling from capacitive to Ohmic apparently due to the increase in the conductance of the membrane facing the planar electrode. Releasing the mechanical pressure led to the reversal of all the parameters, including the amplitude of the intracellularly recorded AP.

## CELL BIOLOGICAL ENERGY RESOURCES TO INTIMATELY INTERFACE CELLS AND MICROELECTRODES

The use of external force to mechanically manipulate the tip of an electrode into contact with a cell (as is done in classical sharp or patch electrodes) or to manipulate a cell into contact with a flat electrode (as described above) cannot be adopted across the board for multisite, long-term, electrophysiological recordings. We hypothesized that an alternative way to achieve this goal would be to engage cell biological energy resources to self-force a cell into intimate contact with an electrode. To that end our laboratory began to examine this possibility by deviating from the "traditional" planar MEA fabrication technology to the use of 3D electrode structures that mimic neuronal structures in terms of their shape and dimensions. Our initial hypothesis (Spira et al., 2007) was that combining 3D electrodes with an appropriate surface functionalization could "trick" neurons or muscle cells to "believe" that the 3D micrometer sized electrodes were neighboring biological entities "that should" be contacted and subsequently actively internalized by endocytotic mechanisms (Spira et al., 2007; Hai et al., 2009a,b).

The specific shape and dimensions of the 3D electrodes was chosen by mimicking the structure and dimension of "spines" that extend from the dendrites of vertebrate neurons (Tonnesen and Nagerl, 2016). Since the 3D electrodes we fabricated bear a resemblance to mushrooms (**Figure 1C**) and because the term spine-shaped was not familiar to the non-biology members of the MEA community, we referred in subsequent studies to the 3D electrode as gold mushroom-shaped microelectrodesgMµEs. In addition, to facilitate the adhesion of the cells to the electrode surface and facilitate the gMµE engulfment we covalently functionalized the gMµE by a cysteine terminated arginine-glycine-aspartic acid (RGD) repeat peptide (Spira et al., 2007).

We tested the hypothesis that neurons, "un-professional phagocytes," can actively engulf micrometer size gMµEs and thereby significantly increase the physical contact between the electrode surface and the cell membrane in two steps. First, because of economic considerations of time and cost, we examined whether cultured Aplysia neurons and rat hippocampal cells could endocytose functionalized micrometer size latex beads. These tests proved positive (**Figure 1D**) and also supported the hypothesis that the chemical functionalization of the beads enhanced the phagocytic activity. Follow-ups by transmission electron microscopic examination of the ultrastructural relationships formed between different cultured cell types and dense arrays of gMµEs (**Figure 1E**) demonstrated that cultured neurons, primary cardiomyocytes, striated muscle fibers and non-excitable cells (NIH/3T3, CHO, PC-12, H9C2) engulfed gMµEs by forming a reduced cleft width and an increased contact area. Nonetheless, the gMµEs clearly maintained their extracellular position with respect to the cell's plasma membrane. Complementary confocal imaging of Aplysia neuron-gMµE hybrids revealed that the formation of the tight physical contact involved the restructuring of the sub-membrane actin skeleton around the gMµE stalk. Quantitative estimates of the seal resistance formed by the neurons and the gMµE suggested that the configuration indeed improved the seal resistance as compared to substrate integrated planar electrodes (Hai et al., 2009a,b, 2010a,b; Ojovan et al., 2015).

## ELECTROPHYSIOLOGICAL RECORDINGS BY gMµE

A comparison of the electrophysiological recording characteristics obtained by gMµE based-MEA using different cell types provides a better understanding of the strength and limitations of this approach. Using cultured Aplysia neurons, we found that within 48–72 h of plating, gMµE functionalized by the RGD-repeat engulfment-promoting-peptide recorded attenuated APs and post synaptic potentials (PSPs) with the characteristic features of intracellular recordings (**Figure 2A**). Since the gMµE are tightly engulfed by the neurons plasma membrane but remains outside of it, we refer to this mode as "IN-CELL recordings" by extracellular electrodes. This was done to emphasize the way it differs from genuine intracellular recording in which the electrode tip forms direct Ohmic contact with the cytosol. The amplitudes of the IN-CELL recorded APs generated by a single large Aplysia neuron (∼80µm in diameter) that contacts a number of gMµEs (spaced 20µm apart) were not identical and ranged from ∼2 to 30 mV. This reflects the variability in the seal resistance formed between the neuron and the various gMµEs (Hai et al., 2010a) and possible differences in the gMµEs' impedances (**Figure 2Ai**). PSPs of up to 5 mV were recorded from Aplysia neurons that were well coupled to the gMµEs as indicated by the high IN-CELL

FIGURE 2 | (Ai) Differences in the levels of the seal resistance formed between a single Aplysia neuron residing on 8 gMµEs (insert) leads to differences in the IN-CELL recorded APs amplitudes. A cultured Aplysia neuron was intracellularly stimulated to fire 3 consecutive APs (red trace). Simultaneous recordings of these APs by 8 gMµE (blue traces) revealed differences in the IN-CELL recorded amplitudes. (Aii) Synaptic- and action-potentials recorded by an extracellular gMµE. Stimulation of a presynaptic neuron by an intracellular sharp electrode (red) lead to the generation of excitatory post synaptic potentials (blue) recorded by a gMµE. Note the summation of the EPSPs to reach firing threshold (Ai,Aii modified with permission from Hai et al., 2010a). (B) Spontaneous activity recorded by a gMµEs from a cultured hippocampal neuron. (Bi) Control spontaneous APs firing recorded by a gMµE, (Bii) 10 min after the application of 10µm GABAzin the firing pattern was changed. (Biii) Enlargements of the records in (Bi). (Biv) Enlargement of the right box in (Biii,Bv) enlargement of the left box in (Biii). The low amplitude long duration potentials in (Biii) (left box) have the features of excitatory post synaptic potentials. The fast spikelets indicated by \* could be either dendritic spikes or the firing of electrically coupled neurons. As the dynamics and amplitudes of the potentials shown in (Bv) are not altered by GABAzin it is unlikely that they reflect barrage of FPs generated by remote neurons. (C) Comparison of intracellular recorded potentials to IN-CELL recordings from cultured myotubes obtained by gMµE electroporation. (Ci,Cii) Depict simultaneous extracellular FPs recordings by a gMµE (Ci) and intracellular recordings by a sharp electrode (Cii). The recordings (Ci,Cii) revealed identical firing patterns and similar qualitative alterations in the amplitudes of the recorded action potentials. (Ciii) Electroporation of the myotube changed the recording mode by the gMµE from extracellular to the IN-CELL. Note that although the IN-CELL recorded amplitude of the APs is about an order of magnitude lower than that of the intracellular electrode APs, the shape of the recorded potentials are identical. Also, note that both the gMµE (Ciii) and the intracellular sharp electrode (Civ) recorded subthreshold potentials in between the APs (C modified with permission from Rabieh et al., 2016). (D) IN-CELL recordings from cultured human cardiomyocytes. Three seconds before electroporation 4. Thirty-three and sixty-three seconds after electroporation.

recorded AP amplitude associated with the recorded PSPs (Hai et al., 2010a) and **Figure 2Aii**). IN-CELL recordings of APs by gMµE that did not undergo chemical functionalization were not generated spontaneously but could be induced to IN-CELL recorded by the delivery of electroporating voltage pulses through the recording gMµE (Hai and Spira, 2012). Using the analog electrical circuit model, we estimated that to account for the experimental results we had to assume that IN-CELL recordings are only possible when in addition to an increased seal resistance, the junctional membrane conductance was increased with respect to the nonjunctional membrane. The mechanism underlying this increase was not investigated but is likely due to the recruitment of voltage independent ionic channels to the junctional membrane or to the formation of nanopores within the confined region of the junctional membrane. Both mechanisms can be triggered by the curvature of the membrane along the gMµE cap and stalk (Petrov et al., 1989; Zhao et al., 2017).

The studies conducted using Aplysia neurons revealed that the neuron-gMµE junction was stable for approximately 2 weeks and that the neurons-gMµE hybrid configuration did not alter the passive and active membrane properties of the neurons and their synaptic functions (Hai et al., 2009b).

The results obtained using primary cultured rat embryonic hippocampal neurons differed in a number of ways from those of Aplysia neurons. Unlike Aplysia neurons that are isolated from mature nerve systems and thus fire full-blown APs at the time of culturing and establish functional synaptic contacts within 12–24 h after plating, neurons derived from embryonic rats are not fully differentiated and require 10–14 days in culture to differentiate morphologically, express excitable membrane properties and form functional synapses. Contrasting with cultured Aplysia neuron cultures, the variability in the shapes and amplitude of the IN-CELL recorded potentials from cultured hippocampal neurons was significantly larger, and ranged from biphasic extracellular field potentials with amplitudes in the range of 100 µV (very much like those recorded by substrate integrated planar electrodes), to positive monophasic IN-CELL recorded APs of up to 5 mV (Shmoel et al., 2016, and **Figure 2B** here). The large variability in the recording mode (extracellular to IN-CELL) from hippocampal neurons can be attributed to two factors. Because of the relatively small diameter of hippocampal neuron cell-bodies, the probability of a given neuron to be optimally positioned with respect to a gMµE (spaced at 100µm) and fully engulf it is lower than for the large diameter Aplysia neurons (Ojovan et al., 2015). Thus, the probability for hippocampal neurons to generate a sufficiently large seal resistance is smaller than for Aplysia neurons. In addition, unlike in Aplysia neurons even in cases in which a high seal resistance was formed between a neuron and the gMµEs, as indicated by the amplitude of the recorded potential, the resistance of the membrane patch facing the gMµE (Rjn) could remain high. Under these conditions the electrical coupling between the neuron and the gMµE remained (partially or fully) capacitive rather than Ohmic and the AP shape appeared to be similar to the time derivative of an intracellular AP (Figure 4 in Fromherz, 2003; Shmoel et al., 2016).

Whereas in cultured Aplysia neurons IN-CELL recorded post synaptic potentials can be identified by their shape, amplitude and temporal relationships with evoked presynaptic APs, the unequivocal identification of IN-CELL recorded PSPs from hippocampal neurons is by far more complex. The complication stems from the fact that the shape of FPs generated by spontaneous bursts that are remote from the recording site is not easily distinguishable from barrages of PSPs (Shmoel et al., 2016 and **Figure 2B** here). Nevertheless, as outlined in the legend to **Figure 2B**, and quantitatively estimated (Ojovan et al., 2015), PSPs can be recorded by extracellular gMµEs when the neurongMµEs coupling coefficient is high and the Jmem conductance is sufficiently low.

Interestingly, contrary to what we found in Aplysia the electrical coupling levels and mode of recording formed between hippocampal neurons and gMµEs were not improved by functionalizing the gMµE with the RGD repeat peptide. Ultrastructural observations by our lab and others have revealed that hippocampal neurons (and cardiomyocytes) engulf and tightly interface gMµEs functionalized by poly-L-lysine or polyethylene-imine/laminin, suggesting that the 3D structure in itself (mushroom-, or a nanopillar-shaped protrusion) is sufficient to trigger the engulfment (Santoro et al., 2013, 2014; Zhao et al., 2017). It is conceivable that the expected effects of the RGD repeat peptide on the junctional membrane conductance of hippocampal neurons is not expressed by the time the hippocampal network matures its electrophysiological functions (10–15 days after plating). The peptide layers at the gold electrode surface could be degraded by enzymes secreted by the cells or by hydrolysis. Attempts to achieve Ohmic contact between gMµEs and the neurons by electroporation were unsuccessful. We observed that hippocampal neurons- gMµE hybrids can be stable for periods of up to 10 days but this issue requires further examination. We have not yet tested whether gMµEs alter the physiological properties of the neurons or the network.

Studies on the use of gMµE arrays to record from contracting cardiomyocytes and striated muscle fibers (**Figures 2C,D**) revealed that both cell types engulf gMµEs functionalized by laminin, this making it possible to record high signal to noise ratio, extracellular FPs generated by individual cells for a number of days (**Figure 2Ci**). In a fraction of the experiments, striated muscles spontaneously formed IN-CELL recording configurations that allowed recordings of APs with amplitudes of 5–10 mV (Rabieh et al., 2016). In both muscle cell types, a short electroporation pulse delivered through a gMµE converted extracellular recordings to IN-CELL recordings (**Figures 2Cii,D**). The shape and duration of the recorded IN-CELL APs were similar to those recorded intracellularly (**Figure 2C**). Interestingly as in the cases of IN-CELL recordings generated by electroporation of cultured Aplysia neurons, the increased junctional membrane conductance of the electroporate cardiomyocytes and striated muscles spontaneously recovered, leading to a reversal of the IN-CELL recording mode to extracellular. Electroporation induced in-cell recordings followed by recovery could be repeated a number of times over a few days.

## CONCLUSION

A diverse series of experiments demonstrated that the concept of attenuated intracellular recordings by extracellular MEA (IN-CELL recordings) can be applied successfully to various neurons, contracting cardiomyocytes and striated muscles under in vitro conditions. Thus, the application of the method should be a "game changer" in terms of enhancing our understanding of the physiological mechanisms underlying in vitro excitable cell network computations, and long-term changes such as in learning and memory. Planar MEA platforms suffer from low signal to noise ratio and low source resolution (Seymour et al., 2017). These drawbacks are solved by tedious spikedetecting, spike-sorting and signal averaging techniques which rely on estimated parameters. The IN-CELL technique may help circumvent this hurdle.

The recent development of biotechnologies that use induced human pluripotent stem cells taken from healthy subjects and patients, and in vitro drug screening for the development of personalized medicine will benefit as well since the IN-CELL recording method makes it possible to extract significantly more information with respect to planar MEA.

#### REFERENCES


There are three focused methodological issues that should nevertheless be addressed before the in vitro IN-CELL approach can be used in practice. These are: (a) the development of reliable methods to accurately position and maintain the cultured cells in close proximity to the target gMµE, (b) the development of methods to locally increase the junctional membrane conductance without inducing cell-repair mechanisms that recover the jmem conductance, and (c), the reduction of gMµE impedance.

#### AUTHOR CONTRIBUTIONS

All authors listed contributed to the writing of the manuscript and approved it for publication.

## ACKNOWLEDGMENTS

Parts of the research reported in this publication was supported by the National Institute of Neurological Disorders and Stroke of the National Institutes of Health under Award Number U01NS099687. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Spira, Shmoel, Huang and Erez. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Force-Mediating Magnetic Nanoparticles to Engineer Neuronal Cell Function

Trevor J. Gahl and Anja Kunze\*

*Department of Electrical and Computer Engineering, Montana State University, Bozeman, MT, United States*

Cellular processes like membrane deformation, cell migration, and transport of organelles are sensitive to mechanical forces. Technically, these cellular processes can be manipulated through operating forces at a spatial precision in the range of nanometers up to a few micrometers through chaperoning force-mediating nanoparticles in electrical, magnetic, or optical field gradients. But which force-mediating tool is more suitable to manipulate cell migration, and which, to manipulate cell signaling? We review here the differences in forces sensation to control and engineer cellular processes inside and outside the cell, with a special focus on neuronal cells. In addition, we discuss technical details and limitations of different force-mediating approaches and highlight recent advancements of nanomagnetics in cell organization, communication, signaling, and intracellular trafficking. Finally, we give suggestions about how force-mediating nanoparticles can be used to our advantage in next-generation neurotherapeutic devices.

#### Edited by:

*Ioan Opris, University of Miami, United States*

#### Reviewed by:

*Devrim Kilinc, Institut Pasteur de Lille, France Takaki Miyata, Nagoya University, Japan*

> \*Correspondence: *Anja Kunze anja.kunze@montana.edu*

#### Specialty section:

*This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience*

Received: *13 March 2018* Accepted: *18 April 2018* Published: *15 May 2018*

#### Citation:

*Gahl TJ and Kunze A (2018) Force-Mediating Magnetic Nanoparticles to Engineer Neuronal Cell Function. Front. Neurosci. 12:299. doi: 10.3389/fnins.2018.00299* Keywords: intracellular forces, nanomagnetics, nanoparticles, neurons, cell guidance, cell communication, cell polarity

## KEY SENTENCE

Quantitative intracellular force interrogation is needed to push the field of neuro mechanobiology into the next level.

## INTRODUCTION

Forces inside a cell, called intracellular forces, play an important role during the regulation of a wide range of cellular processes like membrane protrusion, (Ji et al., 2008) cell migration and spreading, (Galbraith and Sheetz, 1998; Pita-Thomas et al., 2015), or transport of intracellular organelles (Svoboda and Block, 1994; Visscher et al., 1999; Klumpp and Lipowsky, 2005; Kunze et al., 2017). These forces can be generated by the cell itself or be imposed on the cell through a force-mediating object or changes in the extracellular environment. The force-mediating object can manipulate cell structures inside the cytosol or outside of the cell at the cell membrane. We will refer here to forces being operated outside of the cell as extracellular forces. In both cases forces can promote or block healthy cell function depending on the magnitude, the direction, the duration, the rate, and the frequency of application.

In the brain, forces are widely associated with traumatic brain injury, where a physical change in the extracellular environment imposes sheer on the brain cells that can lead into damages at the neuronal cell network (Bigler, 2001; Matthew Hemphill et al., 2015), or induce inflammatory neurodegenerative signals (Maneshi et al., 2014). Recent technical advances in capturing mechanical aspects of brain cells in culture have revealed insights into different magnitudes of forces and their impact on regulating brain cell function like calcium signaling (Calabrese et al., 2002; Tay et al., 2016a; Tay and Di Carlo, 2017), neurite elongation (Bray, 1984; Kunze et al., 2015; Pita-Thomas et al., 2015), or vesicle movement (Ahmed et al., 2012; Kunze et al., 2017). These force-mediated cell functions let us hypothesize that forces in the brain may not only cause lesions, far more, they may be used to our advantage in next-generation neurotherapeutic devices.

To integrate force stimulation into therapeutics or diagnostics, however, comes with challenges. How should we design nextgeneration therapeutic devices to stimulate deep brain tissues without inducing undesired cell effects through high-magnitude forces in more superficial brain tissues? To be able to answer this question, a deeper understanding of the force range affecting single brain cell function and promoting healthy cell communication without generating unintended side effects is required. Furthermore, force-mediated stimulation of cell signals can trigger a variety of intracellular and intercellular downstream processes, inside the stimulated cell, but also on surrounding cell neighbors and tissues, and needs to be better understood for therapeutic applications.

The purpose of this review is to (i) provide an overview about forces at the subcellular scale, (ii) discuss how they can be used to interfere with mammalian cell function, (iii) highlight recent technical advances that allow us to manipulate and interfere with intracellular forces, and (iv) show what needs to be done to advance nanomagnetic force stimulation into a clinical setting. First, we will discuss the variety of forcemediated cellular responses which has been poorly-linked to specific defined magnitudes of forces. For instance, it remains unclear how much force is needed to displace a whole cell body and does the magnitude of force correlate with cell size during migration. Since the magnitude of force is not the only parameter impacting cell function; rate of change, duration, and frequency of force application should also be considered. In a variety of studies, however, the magnitude of force is the only reported or considered parameter. Thus, our review will elaborate on different magnitude ranges of forces used to impact cell signaling, function, communication, and morphology. Furthermore, we will discuss these magnitude ranges specifically for brain cells and provide an overview of force-mediated changes in neuronal cell function. Second, this review we will put a special focus on force-delivery through magnetic nanoparticles which we call nanomagnetic forces. These nanomagnetic forces are mechanical forces induced through a magnetic field gradient on nanometer-sized magnetic particles. Independently on the magnetization properties (ferromagnetic, superparamagnetic), or the magnetic materials of the particle core (magnetite, hematite, migmatite, iron oxide) or the coating materials (silica, dextran), or the functional groups (chitosan, starch, amines, antibodies) a nanomagnetic force is considered as a force acting at the subcellular level within nanometer dimensions. Nanomagnetic forces can be operated inside and outside of cells depending on their geometric and chemical surface properties (Calatayud et al., 2014; Tay et al., 2016a). Because of recent technical advancements, we can operate nanomagnetic forces in parallel through arrays of high magnetic field gradients and apply them to thousands of cells at the same time (Tseng et al., 2012; Kunze et al., 2015; Murray et al., 2016). Third, we will state our opinion about what needs to be done to translate nanomagnetic force stimulation into next-generation neurotherapeutic devices.

## THE FORCE-MEDIATING TOOLBOX

Forces may act on a mammalian cell or can be generated by a cell. In both cases, forces affect the extracellular or the intracellular environment. To capture and manipulate extracellular or intracellular forces, an object needs to bind, or to enter the mammalian cell to translate a pulling or pushing force on the cellular structure (**Figure 1**). Dimensions of this object should be chosen in the sub micrometer range to gain high subcellular precision. Thus, most force-related cell applications employ functionalized nanoparticles which is the first tool in the force-mediating toolbox. The second tool is a probe generating a field gradient which imposes a force on the particle and allows the user to control force parameters like direction, magnitude, or frequency.

Technically, we can quantify forces exerted by a cell through traction force microscopy, (Sniadecki et al., 2007; Style et al., 2014; Kilinc et al., 2015), atomic force microscopy (Baumgartner et al., 2003; Elkin et al., 2007; Kuznetsova et al., 2007; Kirmizis and Logothetidis, 2010; Azeloglu and Costa, 2011), or laser ablation (Campàs, 2016). These methods capture changes in cell shape formation, force dynamics of filopodia at growth cones, at the terminal end of neurites, or reveal shootin1– cortactin interactions within the promotion of traction forces at growth cones at high subcellular precision in single cells (Chan and Odde, 2008; Kubo et al., 2015). These methods, however, are not capable to mediate force stimulation. If it is desired to control and operate magnitude and direction of forces at cells, negative pressure or shear stress can be applied on the cell membrane through microchannels, micropipettes, or micro indenters (Fass and Odde, 2003; Huang et al., 2004; Franze, 2013; Campàs, 2016). The dimensions of the channel and micropipette are the determining factor for precision. Alternatively, optical, magnetic, thermal, or electric tweezers are tools that allow for direct force manipulation depending on the physical properties of the force-mediating object (Thoumine et al., 2000; Baumgartner et al., 2003; Jeney et al., 2004; Neuman and Nagy, 2008; Kilinc et al., 2015; Allen Liu, 2016; Tay et al., 2016b; Timonen and Grzybowski, 2017). An external magnetic, optical, or electrical field is required to direct and accelerate the internalized object (**Figure 2A**). While optical and electrical fields may impact other cell processes, only magnetic fields are transparent to cells. Although, magnetic field gradients for nanomagnetic force manipulation based on permanent magnetic fields (PMF) in combination with micronsized magnets, or electromagnetic alternating magnetic fields (AMF) are tough to design, they are often the preferred force-mediating toolbox. Lately, technical advancements targeted the cell-by-cell time-consuming data acquisition of magnetic

tweezer. The advancement came through microfabricating up to 10,000 parallelized arrays of magnetic field gradient. The fabrication approach based on permalloy, borrowed from solid state devices, was integrated into cell culture chips of the size of few millimeters (Tseng et al., 2012; Kunze et al., 2015, 2017; Murray et al., 2016; Tay et al., 2016a). Acoustic tweezers are also used to move and modulate intracellular trafficking. Standing acoustic field can be created using ultrasonic waves which causes the objects to feel acoustic radiation force. This force is used to move objects to acoustic pressure nodes and antinodes (Chen et al., 2014). This method has advantages over optical tweezers such as causing less damage to organelles while applying more force.

In the case of an extracellular force stimulus, the forcemediating object can either work as a mechanical cue (**Figure 1A**), or as a membrane actuator (**Figure 1B**), or actively target transmembrane proteins. Distinct force applications at the cell membrane can be achieved through selective surface coatings on the nanoparticle (Calatayud et al., 2014; de Castro et al., 2018). Through forces acting on the cell membrane, cell protrusion can be initiated to start growing an axon in neurons (Fass and Odde, 2003; Ji et al., 2008; Betz et al., 2011; Diz-Muñoz et al., 2013; Pita-Thomas et al., 2015; Bidan et al., 2018). Extracellular forces at higher magnitudes can elongate neurites or growth cones (Bray, 1984; Zheng, 1991; Suter and Miller, 2011; Kilinc et al., 2015; Ren et al., 2018), guide cell displacement and migration, (Kunze et al., 2015; Doolin and Stroka, 2018; Van Helvert et al., 2018) or open membrane channels to interfere with neuronal cell communication (McBride and Hamill, 1993; Martinac, 2004; Morris and Juranka, 2007; Reeves et al., 2008; Beyder, 2010; Sanjeev Ranade et al., 2015; Tay et al., 2016a).

When it comes to the intracellular space, forces are involved in molecular motor transport (Svoboda and Block, 1994; Visscher et al., 1999; Klumpp and Lipowsky, 2005), the formation of cytoskeletal structures like actin filaments and microtubules (Dogterom and Yurke, 1997; Brangwynne, 2006) and the local signaling of proteins (**Figure 1C**) (Kosztin et al., 2002). To study forces involved in the intracellular space with high precision, we need the force-mediating object to enter the cell through phago-, pino-, or endocytosis. The uptake mechanism of the force-mediating object highly depends on the cell type, the metabolic state of the cell and particle properties like shape, size and surface functionality (Lesniak et al., 2012; Tay et al., 2016c; Suarato et al., 2017). From previous studies with optical tweezers, we know that single molecular motor and cytoskeleton filament forces are within the lower pico-Newton range, e.g., kinesin motors stall between five to six pico-Newtons (Svoboda and Block, 1994; Visscher et al., 1999). The ability to precisely operate intracellular forces, however, is still a challenge due to low experimental through-put and limited targeting specificity of the force-mediating object inside mammalian cells. Future work is required to systematically generate a map of force ranges considering magnitudes, duration and rate of application in relation with cell specific effects to precisely specify force sensitivity in brain cells.

### NEURONAL CELL FORCE SENSITIVITY

To mechanically engineer cellular effects, it is essential to know the exact force range for (a) the desired cell effect and (b) the different force-mediating tools. Based on theoretical and empirical observations (**Figure 2B**), cells are sensitive over three

controlled through physical concepts based on electric field gradients = electric, optical tweezers = optical, acoustic tweezers = ultrasound, magnetic tweezers = magnetic, thermal tweezers = thermal, or mechanical actuation = mechanical. (Jeney et al., 2004; Neuman and Nagy, 2008; Mosconi et al., 2011). (B) Minimum reported force values required for different cell effects based on computational or experimental models including membrane rupture (MR), (Almeida and Vaz, 1995) microtubules stretching (MS), (Dogterom and Yurke, 1997; Brangwynne, 2006) kinesin motor stalling (KS), (Svoboda and Block, 1994; Visscher et al., 1999) actin stalling (AS), (Footer et al., 2007) protein polymerization (PP), (Footer et al., 2007) ion channel opening and thermal fluctuation (IO). (Meister, 2016) (C) Overview of experimentally reported force ranges indicating significant changes in neuronal cell function and morphology induced through a force stimulus. The list highlights general reported mechanical sensitivity for neurons (GS), (Zablotskii et al., 2016a,b) force mediated calcium induction (CI), (Calabrese et al., 2002; Matthews et al., 2010; Maneshi et al., 2014; Tay et al., 2016a) force mediated cell migration/displacement (CM), (Kunze et al., 2015) force mediated modulation of vesicle motion (VM), (Kunze et al., 2017) force mediated protein positioning (PO), (Kunze et al., 2015) force mediated axon towing and stretching (AT), (Bray, 1984; Suter and Miller, 2011) and force mediated filopodia/growth cone stretching (FS). (Franze, 2013).

distinct ranges of force magnitude (Footer et al., 2007; Kenry and Lim, 2016; Meister, 2016; Zablotskii et al., 2016a). The three force ranges can be separated into nano-Newton forces (1–1,000 nN), pico-Newton forces (1–1,000 pN), and femto-Newton forces (1–1,000 fN). Within these force ranges a single event at the subcellular space can be the opening of a calcium channel. An experimentally-based smallest magnitude of 200 fN has been reported to open a force-sensitive TREK-1 ion channel with a 250 nm-diameter particle in auditory hair cells, or kidney fibroblast-like cells (Howard and Hudspeth, 1988; Hughes et al., 2008; Meister, 2016). Magnitude of forces between 150 pN and 5 nN acting on mechanosensitive ion channels, via integrincytoskeleton coupling using a 4.5 µm-diameter particle, triggered calcium influx in endothelial cells (Matthews et al., 2006, 2010). Mechanical sensitive calcium channels can also be found in primary neurons, where significant increases in calcium influx were observed for neurons similar to endothelial cells above 150 pN.(Matthews et al., 2006; Tay et al., 2016a; Tay and Di Carlo, 2017). The gap between the reported femto- and pico-Newton magnitude of forces to stimulate calcium influx can be attributed to the difference between a delivering vs. acting force. While a larger particle can deliver a higher force to the cell, or specifically to the integrins, or to the ion channels, it also can act on or be targeted to a higher number of ion channels at a cell surface than a smaller particle. The actual minimal amount of force required to stimulate the opening of an ion channel remains then similar. This observation suggests that cells which are exposed to higher magnitude or rate of forces should show a stronger cellular effect. For the calcium channel opening and protein displacement this effect has been demonstrated within a distinct force interval (Kunze et al., 2015; Tay and Di Carlo, 2017). Observations of neuronal cell behavior across a larger range of magnitudes of forces, however, have revealed that above certain force thresholds cells change their entire response. While forces in the lower pico-Newton range interfere with cell functioning; forces in the higher pico-Newton range may induce cell transformative effects that impact the cell morphology or may break through the cell membrane.

**Figure 2B** highlights the most important minimum cellular force thresholds above which major changes in cellular responses were reported. In the upper femto-Newton range, 200 fN are necessary to open a single ion channel in a cellular membrane and to overcome thermal fluctuation effects (Howard and Hudspeth, 1988; Dobson, 2008; Hughes et al., 2008). At least 200 fN to 500 fN have been reported to induce actin polymerization leading to a minimal stalling force threshold for actin polymerization of 1 pN (Tyler, 2012). To stop the motion of a single kinesin motor, 5.6 pN are required (Svoboda and Block, 1994; Visscher et al., 1999). In the same range, at 5 pN, stalling forces for single microtubules have been reported (Tyler, 2012). Increasing intracellular forces from the lower to the higher pico-Newton range show a different picture. Bundles of microtubules in combination with microtubules-associated proteins can withstand up 100 pN (Tyler, 2012). Above this threshold the cell morphology starts to change. Subcellular structures like microtubules and lipid membranes appear to destabilize and to transform into a fluidic state allowing the cell to deform their cellular membrane without rupturing it (Diz-Muñoz et al., 2013; Pita-Thomas et al., 2015). The next reported force threshold occurs in the nano-Newton range, around 25 nN, which is the maximum force a cell membrane can withstand before it ruptures (Almeida and Vaz, 1995). The rupture, however, can be very local and may be reversible. In this case live cell nanosurgery becomes possible (Obataya et al., 2005; Praveenkumar et al., 2015). The list of reviewed force thresholds has its limits when it comes to spatiotemporal changes. A lower force threshold might be possible for certain cell effects when applied just long enough, or when operated faster than currently possible.

Focusing on brain cells the force sensitivity range is much more limited. Although most cell-generic force thresholds apply for neurons, force-mediated cellular effects are only reported in the pico-Newton range (**Figure 2C**). Operating and controlling forces across the whole cell sensitivity range, quickly limits our toolbox to magnetic tweezers and systems based on the current technical state of art. The specific neuronal cell sensitivity, however, promotes optical and magnetic tweezers.

## ORGANIZING CELL TISSUE CONSTRUCTS WITH NANO-GUIDES

Replicating the filigree organization of biological tissues has been the focus of many studies during the last 20 years (Butler et al., 2000; Goldberg et al., 2007; Pampaloni et al., 2007; Sakar and Baker, 2018). Biological tissues are highly organized constructs which consist of a diverse range of cell types, which are assembled into layers of heterogenous cell densities to perform different function. Integrated in the constructs is a densely branched vascular system that provides oxygen and nutrition.

Most tissue engineering studies report on manipulating the chemical and mechanical properties of the extracellular environment to trigger a desired cell response, e.g., local cell organization, cell orientation, cell migration, and cell network formation, which then potentially leads to the desired tissue organization and physiology. In contrast to engineering the extracellular environment, direct positioning of cellular bodies allows us to engineer cell tissues from the bottom to the top. In this context, magnetic gradients and forces are utilized to collect and assemble mammalian cells to specific local positions on a plane in a controlled manner (**Figure 3A**) (Tanase et al., 2005; Ino et al., 2007, 2008; Rampini et al., 2015; Zablotskii et al., 2016a,b). The orientation of the magnetic field poles and the magnitude of the magnetic force are the most determining parameters to either assemble, orient or direct cell position (**Figures 3A–C**). To attract nonadherent cells to specific places on a surface or in suspension the cell body needs to incorporate magnetic materials or needs to attach to magnetic guides. The position of the magnetic poles then attracts the magnetic guides allowing the cells to locally attach to the surface, to assemble into a tissue, or form multi-layered spheroid clusters in connection with additional cell neighbors and cell layers (Ino et al., 2007; Ito and Kamihira, 2011; Lee E. et al., 2014). Using this concept, Marcus et al. positioned rat pheochromocytoma PC12 cells on multi-pole arrays (**Figures 3D,G**) (Marcus et al., 2016). When cell bodies already adhered to their growth surface, applying magnetic gradients, and forces can orient the cellular morphology (**Figures 3B,E**). Specifically, fine tuning the magnetic field gradient forces provides the possibility to orient primary cortical neuron cell growth in the lower pico-newton range (**Figure 3E**) and to induce cell migration in the higher pico-newton range (**Figures 3C,F**) (Kunze et al., 2015). PC12 cell orientation was also reported for aligned magnetic nanoparticle guides in a twopole magnetic field (**Figure 3H**) (Riggio et al., 2014). Subsequent seeding of cells over the unoccupied cell regions would allow for co-culturing of different cell types cell-by-cell or layer-by-layer. Magnetic gradient forces were also reported to support Schwann cell migration through astrocyte-rich cell regions (**Figure 3I**) (Xia et al., 2016; Huang et al., 2017). What remains unclear is how easy magnetic-guided tissue engineering can be applied to primary neurons. Above mentioned studies focused on PC12 cells which are neuron-like cells, or non-neuronal brain cells known to show differences in particle endocytosis in comparison to primary neurons (Pinkernelle et al., 2012).

Another important aspect of tissue growth and assembly is the time-varying relation between cell migration and function. Throughout the development of newly forming tissues, the individual cells must adapt to changes in the biochemical and biomechanical environment and decide to leave, stay, or modulate their environment. Little is known about how fast cells respond to biomechanical changes, or what happens if the cell fails to do so. Magnetic field gradients can be switched on and off, either through an external electrical current or through removing the externally applied permanent magnetic field. Combining magnetic field gradients with time-varying cell assays seems to be a versatile way to study the adaption of cell tissue functionality in changing environments. Overall, utilizing magnetic gradients and magnetic forces are an attractive method to assemble and grow cells into complex constructs and to further investigate

neuron cultures show a shift of intracellular markers toward left oriented magnetic gradient forces. w/o, no magnetic field; w/, with magnetic field. Reproduced with permission from Kunze et al. (2015). (F) Primary cortical neurons dissociated from rat brain tissues (E18) were cultured on poly-l-lysine surfaces and exposed to fMNPs after being 24 h in culture. These neurons grow and form neurite networks under magnetic fields and start migration toward magnetic field poles under strong magnetic forces (> 250 pN). Scale bar = 12µm. Reproduced with permission from Kunze et al. (2015), Copyright © 2015, American Chemical Society. (G) Histogram plots of accumulated neuron-like cells which were cultured above the single and two pole patterns, respectively. (H) Orientation index extracted from PC12 that were observed to align in parallel to magnetic field orientation after being cultured with fMNPs. f-MNP-M+, fMNPs with magnetic field; f-MNP-M−, fMNPs without magnetic field. Reproduced with permission from Riggio et al. (2014), Copyright © 2014, Elsevier Inc. (I) Schwann cells migrate into astrocyte-rich region under an oriented magnetic field gradient after internalizing PEI-fMNPs (PEI-SPIONs). White arrow indicates direction of magnetic pole. Scale bar = 100µm. Reproduced with permission from Xia et al. (2016), Copyright © 2016, Dove Medical Press Limited.

time-varying changes in the cellular environment and their effect on cell function.

## MODULATING CELL COMMUNICATION WITH NANOMAGNETIC FORCES

Neuronal cells propagate information based on ionic sodium and potassium signals, which can be electrically monitored. Calcium signals play an important role in this ionic signaling and signal propagation mechanism (Rasmussen, 1970). Through precisely activating Ca2<sup>+</sup> channels, calcium influx can be stimulated, protein function can be post-translationally modified, or gene transcription can be induced (Berridge et al., 1998, 2000; West et al., 2001). Most recently, calcium influx was remotely controlled through heat induction or force manipulation using alternating or permanent magnetic field, respectively (**Figure 4**; Calabrese et al., 2002; Maneshi et al., 2014; Bonnemay et al., 2015; Tay et al., 2016a; Tay and Di Carlo, 2017). Both approaches have been proven to be beneficial to induce calcium signals in a confined area, or at distinct subcellular compartments. Heat-mediated calcium influx occurs when an alternating magnetic field is applied in conjunction with an overexpression of TRPV+ channels in neurons (**Figure 4A**), resulting in locally increased calcium concentrations in primary hippocampal neurons (**Figure 4B;** Chen et al., 2015). However, the effect of sustained heating over more than a few minutes needs to be further demonstrated. Nimpf and Keays provided in this context further critical comments about limitations and reproducibility's of heat mediated magnetogenetic approaches (Nimpf and Keays, 2017).

While modulating cell communication through heat is applicable to the in vivo environment of neurons, the host organism must be genetically modified. In contrast to heat induction, mechanical forces can bend the cell membrane and interrogate associated calcium channels (**Figure 4B**) (Matthews et al., 2006, 2010). Tay et al. demonstrated an average increment of calcium influx by 20 % for magnetic nanoparticles imposing forces above ∼200 pN at the cell membrane (**Figure 4B**) and a 10 % increase for forces operating inside the cell in primary

Reproduce with permission from Chen et al. (2015), Copyright © 2015, American Association for the Advancement of Science. (D) False color heat maps show changes in fluorescently-labeled intracellular calcium influx in primary cortical neurons (E18, rat) with and without fMNPs and with and without PMF stimulation.

cortical neurons (Tay et al., 2016a; Tay and Di Carlo, 2017). Additionally, Hughes et al. have demonstrated the selective activation of ion channels via magnetic nanoparticles (Hughes et al., 2008). Magnetic nanoparticels were introduced to TREK-1 transfected COS-7 cells and by placing a rare earth magnet ∼1.5 cm away from the cells, a magnetic field of ∼80 mT was applied with a field gradient of ∼5.5 Tm−<sup>1</sup> . The results indicated that channel activation occurred at ∼0.2 pN per particle when using 250 nm particles (Hughes et al., 2008). The difference in forces magnitude between the two studies may be due to differences in membrane targeting, or due to differences in the sensitivity of the optical vs. electrophysiological probing method. While Tay et al. used nanomagnetic forces to bend the membrane and to mechanically activate N-type calcium channels, Hughes et al. specifically targeted the magnetic particles in their study to the mechanosensitive TREK-1 ion channel. Alternatively, the magnetic field can also be operated either to induce torque (Hudspeth et al., 2000; Mosconi et al., 2011). or to induce tensile stretch on mammalian cells to stimulate ion channels and cell communication (Lee J. et al., 2014). Recently, the torque approach has been used in conjunction with confocal microscopy to image force responses in living cells (Zhang et al., 2017). The approach has been further expanded upon by Chen et al. through the integration of a multi-pole electromagnet that allows for control of both the twisting direction as well as the magnetic strength (Chen et al., 2016).

Reproduce with permission from Tay et al. (2016a), Copyright © 2016, American Chemical Society.

While multiple studies have examined the usage of magnetic forces for channel activation in vitro translating nanomagnetic force stimulation in vivo still needs to be shown and will require accurate operation and positioning of magnetic field gradients in the body. Using magnetic implants based on current chip technology, or electromagnetic micro needles (Matthews et al., 2004) opens the possibility to operate calcium communication inside the brain through mechanical stimuli, however, it will remain an invasive procedure.

#### COMPARTMENTALIZING INTRACELLULAR PROTEINS

Separating intracellular organelles and proteins into distinct compartments within a cell is a critical event during cell differentiation, cell mitosis, cell signaling, and to establish functional cell polarity in neurons (Bradke and Dotti, 1997, 2000; Bentley and Banker, 2016; Hansen et al., 2017). Compartmentalizing the location of proteins in the cytosol can be effectively altered though the application of subcellular forces. Mechanically manipulating the position of proteins can be controlled through endocytosed magnetic nanoparticles within magnetic field gradients (Pan et al., 2012; Bonnemay et al., 2013; Etoc et al., 2013, 2015; Kunze et al., 2015; Hughes and Kumar, 2016; Ducasse et al., 2017; Liße et al., 2017; Monzel et al., 2017). The force range to establish a specific protein gradient, however, should leave the tension at the cell membrane at a homeostatic level. This homeostatic level at the cell membrane is a balance between intracellular structural forces and extracellular adhesive forces keeping the cell membrane intact and the cell morphology at a constant shape. Keeping the cell membrane at a homeostatic constant level is highly essential for healthy functioning of cells, tissues, and organs (Smith, 2010). In contrast, impaired homeostatic levels were reported to correlate with cancer cell formation, and dysfunctional cell behavior (Dityatev et al., 2010; Gilbert and Weaver, 2017).

Different cell types, however, develop different cell morphologies (**Figures 5A–C**). While epithelia cells keep their cell membrane uniformly distributed around the nucleus, neurons grow their tangibles heterogeneously and far away from the nucleus, which results in a more complex cell morphology. The resulting level of cell membrane homeostasis may then differ between the different cell types. Applying a localized force stimulus for protein sorting on intracellular compartments will also put the cytoskeleton, the cellular membrane, and protein clusters under tensions. Hence, redistributing proteins based on magnetic or optical gradients in morphologically complex cells might require a higher force gradient, or longer force application than in spherical cells. To avoid damaging the cell membrane, or blocking intracellular transport through narrowed cellular features, nanomagnetic force amplitudes need to be adapted to the homeostatic cell level and should be ideally uniform across the entire cell, which is technically a challenge. Thus, magnetically sorting proteins in cells where cell morphologies ranges from a less to a more complex architecture will require special care regarding the application of force amplitudes.

In spherical cell-like liposomes and Xenopus laevis eggs asymmetric spots of microtubule fibers were assembled through positioning RanGTP proteins conjugated to superparamagnetic nanoparticles under the operation of magnetic field gradients (**Figures 5D,G**) (Bonnemay et al., 2013; Hoffmann et al., 2013; Ducasse et al., 2017). The authors reported operating nanomagnetic forces in the femtonewton range below the thermal fluctuation threshold (Ducasse et al., 2017). In HeLa cells, protein organization was spatially and temporally altered within a range of few femtonewtons up to 30 pNs depending on the magnetic particle size (**Figure 5E,H**; Etoc et al., 2015). In NIH 3T3 cells, Levskaya et al. regulated the actin cytoskeleton dynamics through local activation of Rho-family GTPases proteins (**Figure 5J**; Levskaya et al., 2009). The assembly of proteins was enabled through a genetically encoded light-control system which was operated within the pico-Newton range. In rat cortical neurons, cell morphology is more complex and proteins are more polarized than in other mammalian cell types, nanomagnetic forces between 4.3 and 70 pN have been shown to sort Tau proteins around a 180◦ axis (**Figures 5F,I**; Kunze et al., 2015). Overall, magnetic forces have been probed to modulate protein gradients across a variety of cell morphologies, what remains unclear is how strong does the cell morphology interferes with the formation of force-mediated protein gradients. The studies, we mentioned, suggest a spectrum of required force ranges for subcellular protein assembly and redistribution that might depend on the complexity of the cellular morphology, but also on particle functionalization and endocytosis pathways. Thus, a variety of investigations are required to better understand how intracellular protein sorting can be linked to cell disease, functionality, growth and death.

## MODULATING INTRACELLULAR TRAFFIC

Vesicle dynamics are a key component of transporting molecules inside cells to distinct subcellular sites for proper cell growth, signaling, and maintenance of homeostasis. Perturbating intracellular vesicle dynamics helps us to better understand the role of vesicle transport in a variety of diseases mechanism and propagation. Conventionally, vesicle dynamics were altered through genetically modified signaling pathways, or biochemically inhibiting transport dynamics. A comprehensive review about these approaches is provided by van Bergeijk et al. (2016) A genetically, or chemically independent approach to vesicle transport is through the application of mechanical forces. Two distinct methods employ mechanical tension on transport dynamics of vesicles in neuronal cells (Siechen et al., 2009; Ahmed et al., 2012, 2013; Kunze et al., 2017). The difference between the two methods are depicted in **Figure 6**. Both methods apply mechanical forces, either through substrate stretching outside of the cell (**Figure 6A**; Ahmed et al., 2010, 2013) or through nanomagnetic forces inside the cell (**Figure 6B**; Kunze et al., 2017).

Depending on the application of the extracellular force via a stretchable cell culture platform (**Figure 6C**), the stretch, or buckle will result in a uniform elongation, or compression of the whole cell body in adherent cells including their cellular compartments. Constantly stretched in vivo axons in Drosophila embryonic motor neurons (where only the embryonic body was fixed to the platform) accumulated synaptotagminlabeled vesicles in the axonal tip in the absence of Ca2<sup>+</sup> (**Figure 6G**). After the tension was removed, the effect persisted for at least 30 min. Because vesicles are constantly transported forward (anterograde = toward the synapse) and backward (retrograde = away from the synapse), the stretch-mediated effect was reported to be more dominant in the forward then in the backward transport (**Figure 6D**). Additionally, it was discovered that compressive strain along Drosophila motor neuron axons did not increase synaptic vesicle accumulation and decreased tension in Aplysia neurons which resulted in disrupted motion of large dense core vesicles. This effect persisted in Aplysia neurons for at least 15 min after standard tension was restored (Ahmed et al., 2012). One possible explanation for this effect can be the developmental state of the axonal tip. Although the in vivo Drosophila neuron had established a neuromuscular synapse, the in vitro Aplysia neurons was in a pre-developmental synaptic state. This finding opens the questions if force-mediated vesicle dynamics highly depends on the developmental state of the cellular compartment and its subcellular cytosolic composition.

Intracellularly, vesicle dynamics can be interfered through nanomagnetic forces localized on a magnetic cell culture platform (**Figure 6E**). This method makes use of internalized magnetic nanoparticles which are encapsulated in membraneoriginated lipid vesicles (**Figure 6H**). Through the application

bar = 10µm. HTL-sMNPs, HaloTag-ligand-silica-based magnetic nanoparticles. (F) Primary cortical neurons with superparamagnetic nanoparticles re-assemble Tau proteins toward the magnetic field gradient when exposed to a permanent magnetic field. Scale bar = 16µm. (G) Microtubules nucleation position of RanGTP-magnetic nanoparticles (Ran-NPs) without (Off) and with (On) magnetic forces. (D,G) Reproduced with permission from Bonnemay et al. (2013), Copyright © 2013, American Chemical Society. (H) Surface intensity plot shows correlation between nanoparticles (HTL-sMNPs = sMNPs) and protein assembly (HT-eGFP) in transfected HeLa cells dropping away from the magnetic tip. (E,H) Reproduced with permission from Etoc et al. (2015), Copyright © 2015, American Chemical Society. (I) Histogram plot for the nanomagnetic force range were protein assembly was significant different from its native distribution. (F,I) Adapted from Kunze et al. (2015), Copyright © 2015, American Chemical Society. (J) Force-mediated local activation actin cytoskeleton dynamics through dragging Rho-family GTPases proteins. Reproduced with permission from Levskaya et al. (2009), Copyright © 2009, Springer Nature.

of nanomagnetic forces on chip, (**Figure 6F**) the motility and transport direction of lipid vesicles in primary cortical neurons was either stalled or re-directed, even against insulin-mediated chemical signals (**Figure 6I**; Kunze et al., 2017). In addition to magnetic forces, optical tweezer platforms provide a similar force range as nanomagnetic forces and have been probed

movement pattern under magnetic forces. (I) Extracted vesicle tracks without (no M) and with (w M) nanomagnetic forces. (C,D,G) Reproduced with permission from Ahmed et al. (2012,2013), Copyright © 2012, Royal Society of Chemistry. (E,F,H,I) were adapted and reproduced with permission from Kunze et al. (2017), Copyright © 2017, Royal Society of Chemistry.

for transporting and positioning of recycling endosomes and perixomes and RAB11 vesicles in COS-7 cells and primary hippocampal neurons (Harterink et al., 2016). The specificity of positioning organelles with optical tweezers, however, requires the knowledge of expressing tunable, light-controlled interacting protein tags in cells of interests (Strickland et al., 2012).

### MAGNETOFECTION

Magnetofection is a transfection technique in which an external magnetic field is utilized to improve delivery of nucleic acids attached to MNPs into cells. This technique was originally conceived by Plank et al. (2003). Recently, a study conducted by Smolders et al. compared the efficiency of magnetofection to other transfection methods using a microglial cell line. They found that Glial-Mag magnetofection of BV2 cells greatly outperformed standard chemical transfection methods; calciumphosphate precipitation, X-tremeGENE, and Lipofectamine 2000 with an efficiency of 34.95% compared to 0.34, 3.30, and 12.51% respectively (Smolders et al., 2018). In contrast to this study, however, Katebi et al. found that a static magnetic field reduces the uptake of exogenous oligonucleotide by rooster spermatozoa (Katebi et al., 2016). They observed that when primary spermatocytes were incubated in exogenous oligonucleotide solution with MNPs, the uptake was increased, however, when the static magnetic field was applied, a significant decrease in uptake occurred (Katebi et al., 2016). This indicates that the application of a static magnetic field may prove detrimental to different cell types and that further research should be conducted. Of particular interest would be what effect may the static field have on primary neuronal cells from different origins.

In contrast to static magnetofection, the method has been further developed to use oscillating magnetic fields. Fouriki et al. found that the application of an oscillatory field increased fluorescence intensity of transfected human embryonic kidney cells (H292) (Fouriki et al., 2010). This method has been applied, with frequency dependent efficiency, to rat astrocytes as well as neural stem cell in suspensions (Pickard and Chari, 2010; Adams et al., 2013). Adam et al. demonstrated a two-fold increase in transfection efficiency on neural stem cell suspensions at 4 Hz with no effect on cell viability, number, marker expression or differentiation profiles, indicating a safe transfection method for neural stem cells (Adams et al., 2013).

While the application of nanomagnetic forces has been well demonstrated in increasing transfection efficiency, further research needs to be done on the applications of oscillating magnetic fields. Current research is indicative of a frequency dependent component of oscillatory magnetofection that it may be possible to optimize. Furthermore, magnetically targeting specific individual cell types within a cell population is specifically interesting to study disease models in vitro. The spatial limitation of magnetofection, however, currently remains a challenge, because the externally applied macro magnetic fields will always impose a spatial magnetic gradient across the entire cell culture platform. The effect on other cells types within the same culture currently remains unknown. Thus, increasing spatial resolution and specificity of magnetic gradients down to single cell levels can be the focus of a variety of future studies.

## FUTURE CHALLENGES AND PERSPECTIVE

The purpose of this review was to highlight emerging applications of nanomagnetic forces and related concepts such as magnetic field effects and differences between permanent and alternating magnetic field stimulation on mammalian cell behavior. We have discussed several advantages of nanomagnetic force stimulation over other force-mediating methods, however, we do need to acknowledge that our current understanding of nanomagnetic force stimulation has its limits. While magnetic field gradient can penetrate tissues, organs, or the human body it currently remains challenging to operate nanomagnetic forces in a controlled and precise manner through three-dimensional tissue constructs. Furthermore, the response of cells to nanomagnetic force stimulation is limited in time. In the following, we would like to outline our opinion about how studies involving nanomagnetic force stimulation can address (i) spatiotemporal limitations of end-point experiments and (ii) bring this technology away from the bench and integrate it into mechanically-mediated diagnostics, pharmaceutical cell assays, and neurotherapeutics.

#### Spatio-Temporal Response

Current studies about cell-based nanomagnetic force stimulation compare cell effects based on endpoint measurements or based on short time-windows of several minutes, as in the case of calcium stimulation. It means that our current knowledge about nanomagnetic force stimulation in biological systems stems from either several minutes of live-cell experiments, or few day endpoint experiments (24 h and more) without access to capture time-related intermediate data. From the endpoint measurements, we can conclude how nanomagnetic forces interfere with cells and which down- or upstreaming cell signals get activated or inhibited. How cells, however, adjust temporally over a period of days or months to a potential nanomagneticbased treatment requires a better understanding of the spatialtemporal relation between the force stimulus and the cellular, tissue and organ response. Systematic long-term experiments, where cell growth and behavior are constantly monitored using either optic, or electric measurements without interfering with the experimental setting, would allow us to learn more about spatio-temporal response of nanomagnetic forces stimulation.

Future nanomagnetic force-mediating studies may reveal new properties about the link between the force-mediating object and the cellular response. **Figure 7A** depicts two potential mechanism how the nanoparticle may translate the force stimulus to the cellular structure based on a direct or an associative link. The link between the nanomagnetic force stimulus and the subcellular object (organelle, cell membrane, cytoskeleton) impacts the time lag for the cellular response. If the nature of this link between the nanoparticle and the cellular structure is direct, the cellular response should be seen almost immediately. After a force stimulus, the cell would need to at least interpret this stimulus in situ, if not triggering downstream signals immediately. In contrast, an associative link contains a storing capacity. The unloading of this capacity may or may not follow within the same time lag as for the direct link. It is more likely, however, that the storing capacity of the associative link triggers a cellular response within minutes, hours, or days. Thus, the time-lag will be an important parameter to better understand mechanotransduction and translational approaches in nanomagnetic force stimulation. Furthermore, Ricca et al. suggests within the context of mechanotransduction to use clearly defined extracellular mechanical cues as input signals to elucidate between an active and a passive input (Ricca et al., 2013). Within the context of our bound or associative nanoparticle which can be controlled through engineered surface coatings, the passive input can be modeled through an associative link and would show a delayed cellular response in comparison to the active, bound link. Concerning neurotherapeutic approaches, this delayed effect will be either desired, controlled, or prevented. Therefore, a deeper understanding of the spatio-temporal aspects of nanomagnetic force stimulation is essential to prepare this approach for further translational studies.

#### Neurotherapeutics

Our literature review focused on current single and multi-cell applications using nanomagnetic forces and related magnetic actuation concepts where we see a potential for translational applications regarding neurotherapeutics. In this last section, we want to provide to the reader an overview with the diverse potential of nanomagnetic force stimulation in translational research, neurotherapeutics and patient-specific prognostics (**Figure 7B**). In the previous section, we have outlined a fundamental question regarding the temporal response of nanomagnetic force stimulation, which needs to be answered for a variety of translational applications, nevertheless, we assume that this knowledge will be available in > 10 years. To truly realize the potential of nanomagnetic force stimulation, we need to go beyond single cell analysis and ask how nanomagnetic force stimulation will impact cell networks, specifically connected

neuronal cell circuitries. The next step toward neurotherapeutics is to incorporate nanomagnetic force stimulation into neural tissue engineering, (Goldberg et al., 2007; Ito and Kamihira, 2011) into delivery mechanism of biopharmaceuticals across the blood brain barrier, (Thomsen et al., 2015) or into axon elongation strategies for repairing spinal cord injuries (Kilinc et al., 2016). The potential of adding magnetic force stimulation to tissue engineering lays in the properties of the nanoparticles to modulate cell mechanics (Septiadi et al., 2018) and to induce controlled forces within extracellular constructs to switch between different mechanical properties through turning on and off the magnetic field (Zhang et al., 2016). The latter approach is beneficial to squeeze drugs out from a scaffold for a controlled duration during mechanically-force triggered drug delivery (Zhang et al., 2016). The transport of biopharmaceuticals through the blood brain barrier can further be promoted through magnetic force applications in combination with magnetoliposomes (Thomsen et al., 2015). Adding nanomagnetic forces stimulation to neural grafts for spinal cord repair can be an alternative to optogenetic approaches (Bryson et al., 2016; Kilinc et al., 2016). Finally, the differential uptake of magnetic nanoparticle into different brain cell types can be used to either selectively target and sort specific brain cell types, or to build controlled patterns of brain cells for artificial neural tissues. Further translation of nanomagnetic force stimulation into brain issues and neurotherapeutics will also require a systematic understanding of brain cell functionality through metabolomics and proteomics (Holle et al., 2018). Last, magnetic nanoparticles, which are the core of nanomagnetic forces are already common in cell sorting for cancer-based diagnostics, however, there is plenty of room to come up with new methods to integrate nanomagnetic forces into mechanicallymediated diagnostics and neuro- therapeutics based on protein chaperoning, separation, and on-chip cell technology.

### AUTHOR CONTRIBUTIONS

TG: Reviewed literature and wrote parts of the review paper; AK: Supervised, organized, reviewed literature and wrote this review paper. Both authors revised the manuscript.

## ACKNOWLEDGMENTS

The authors would like to thank all Kunze Neuroengineering Lab members for scientific discussions and the members of the scientific writing group at MSU for feedback on the manuscript. We also would like to thank the reviewers for their comments and time to improve our manuscript.

### REFERENCES


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Gahl and Kunze. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Recent Advances in the Therapeutic and Diagnostic Use of Liposomes and Carbon Nanomaterials in Ischemic Stroke

Lorena F. Fernandes, Gisele E. Bruch, André R. Massensini\* and Frédéric Frézard\*

Departamento de Fisiologia e Biofísica, Instituto de Ciências Biológicas, Universidade Federal de Minas Gerais, Belo Horizonte, Brazil

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Ana-Maria Zagrean, Carol Davila University of Medicine and Pharmacy, Romania Eleonore Fröhlich, Medizinische Universität Graz, Austria Sebastian Cerdan, Consejo Superior de Investigaciones Científicas (CSIC), Spain

#### \*Correspondence:

André R. Massensini massen@icb.ufmg.br; massen@ufmg.br Frédéric Frézard frezard@icb.ufmg.br

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 28 April 2018 Accepted: 13 June 2018 Published: 05 July 2018

#### Citation:

Fernandes LF, Bruch GE, Massensini AR and Frézard F (2018) Recent Advances in the Therapeutic and Diagnostic Use of Liposomes and Carbon Nanomaterials in Ischemic Stroke. Front. Neurosci. 12:453. doi: 10.3389/fnins.2018.00453 The complexity of the central nervous system (CNS), its limited self-repairing capacity and the ineffective delivery of most CNS drugs to the brain contribute to the irreversible and progressive nature of many neurological diseases and also the severity of the outcome. Therefore, neurological disorders belong to the group of pathologies with the greatest need of new technologies for diagnostics and therapeutics. In this scenario, nanotechnology has emerged with innovative and promising biomaterials and tools. This review focuses on ischemic stroke, being one of the major causes of death and serious long-term disabilities worldwide, and the recent advances in the study of liposomes and carbon nanomaterials for therapeutic and diagnostic purposes. Ischemic stroke occurs when blood flow to the brain is insufficient to meet metabolic demand, leading to a cascade of physiopathological events in the CNS including local blood brain barrier (BBB) disruption. However, to date, the only treatment approved by the FDA for this pathology is based on the potentially toxic tissue plasminogen activator. The techniques currently available for diagnosis of stroke also lack sensitivity. Liposomes and carbon nanomaterials were selected for comparison in this review, because of their very distinct characteristics and ranges of applications. Liposomes represent a biomimetic system, with composition, structural organization and properties very similar to biological membranes. On the other hand, carbon nanomaterials, which are not naturally encountered in the human body, exhibit new modes of interaction with biological molecules and systems, resulting in unique pharmacological properties. In the last years, several neuroprotective agents have been evaluated under the encapsulated form in liposomes, in experimental models of stroke. Effective drug delivery to the brain and neuroprotection were achieved using stealth liposomes bearing targeting ligands onto their surface for brain endothelial cells and ischemic tissues receptors. Carbon nanomaterials including nanotubes, fullerenes and graphene, started to be investigated and potential applications for therapy, biosensing and imaging have been identified based on their antioxidant action, their intrinsic photoluminescence, their ability to cross the BBB, transitorily decrease the BBB paracellular tightness, carry oligonucleotides and cells and induce cell differentiation. The potential future developments in the field are finally discussed.

Keywords: nanocarrier, liposomes, carbon nanotubes, graphene, fullerenes, stroke, nanobiosensor, imaging

## INTRODUCTION

fnins-12-00453 July 4, 2018 Time: 17:6 # 2

Stroke is one of the leading causes of death and disability worldwide. According to the most recent meetings of the American Heart Association the global prevalence of stroke was 42.4 million in 2015. Ischemic stroke was 24.9 million, and hemorrhagic stroke was 18.7 million in the entire world. Stroke incidence in the United States each year shows that approximately 610,000 people experience a first stroke attack and 185,000 are recurrent attacks (87% are ischemic and 10% are intracerebral hemorrhage strokes, whereas 3% are subarachnoid hemorrhage strokes cases). Direct and indirect costs with stroke accounted 40.1 billions of dollars between 2013 and 2014 (Benjamin et al., 2018).

Stroke is defined by a sudden decrease of blood supply to the brain tissue. It can result in the manifestation of symptoms like numbness of face, arms and legs, confusion, aphasia, among others. Signs and symptoms will depend on the affected area and can have an outcome ranging from complete patient recovery to severe neurological deficits and death (Feuerstein and Wang, 2000; Casals et al., 2011; Kyle and Saha, 2014). It can be divided in two different categories: ischemic and hemorrhagic. Ischemic stroke or cerebral ischemia is characterized by disruption of blood flow to the brain as a result of an obstruction in the arteries that irrigate a certain area of the brain. This obstruction can be caused by a clot or a plunger (Luo et al., 2017). The hemorrhagic stroke is less common and divided into intracranial, when a rupture of an artery causes extravasation of blood to the cerebral parenchyma (Mayer and Rincon, 2005); or subarachnoid, in which the blood leakage is usually caused by some trauma or ruptured aneurysm (Rafii and Hillis, 2006).

In this review, we will focus on ischemic stroke and discuss the current diagnostic and therapeutic challenges. Despite the great incidence, the severity of the outcomes and the high economic costs of ischemic stroke, tissue plasminogen activator (t-PA) is the only treatment approved by the Food and Drug Administration (FDA). However, due to safety concerns such as the risk of cerebral hemorrhage after treatment with t-PA, the number of patients who can actually use this drug is very low (Adeoye et al., 2011). In addition, even when blood flow is restored, secondary damage caused by reperfusion can be observed in brain tissue, mainly because of the production of deleterious substances, such as reactive oxygen species (ROS) and inflammatory cytokines (Iadecola and Anrather, 2011). In the case of ischemic stroke and also many other central nervous system (CNS) disorders, the current treatment and development of new therapies have been limited by the following factors: (1) ineffective delivery of most CNS drugs into the brain parenchyma because of the blood brain barrier (BBB), hindering preventive treatment or the rescue of areas not yet totally affected (Alyautdin et al., 2014); (2) poor stability or toxicity of a large number of drugs after systemic and/or oral administration; (3) insufficient understanding of the physiopathology of the disease; (4) difficulty to translate good results from pre-clinical studies to the clinic.

In this scenario, nanotechnology emerges with innovative tools for therapeutic, diagnostic and theranostic purposes. These tools include, for instance, imaging techniques, implants, sensors, biomarkers, drug development and carrier systems and biomaterials (Hong et al., 2015). Regarding treatment and diagnosis of ischemic stroke, nanomaterials can act in at least three different forms: (1) as a drug carrier, avoiding drug degradation and unspecific binding to sites where it may exert toxicity and facilitating its passage across the BBB (Fukuta et al., 2016); (2) through its specific characteristics and resulting pharmacological actions, for instance as an antioxidant (Lee et al., 2011); (3) as a diagnostic tool or building blocks in the development of sensors for several biomolecules (from ROS to neurotransmitters) (Kruss et al., 2013).

After a brief description of the physiopathology and the current limitations of the diagnosis and therapy of ischemic stroke, this review highlights the recent progress achieved in the use of two important types of nanostructures: liposomes as first generation of drug nanocarriers, but still the most advanced and studied; carbon nanomaterials recently investigated as neuroprotective agents, membrane-permeable and permeabilizing agents, cell and drug-carrier systems and innovative biosensors for mechanistic evaluation in experimental models. These two types of nanostructures were selected for comparison in this review because of their very distinct characteristics and ranges of applications. Liposomes represent a biomimetic system, with composition, structural organization and properties very similar to biological membranes. On the other hand, carbon nanomaterials, which are not naturally encountered in the human body, exhibit new modes of interaction with biological molecules and systems, resulting in unique pharmacological properties.

## PHYSIOPATHOLOGY OF ISCHEMIC STROKE

After cerebral ischemia, there is an interruption of blood supply to the brain, which induces a shortage of oxygen and glucose delivery, compromising the production of ATP by oxidative phosphorylation (Deb et al., 2010). A shift toward anaerobic glycolysis also occurs which results in the lowering of brain tissue pH (Anderson et al., 1999; Liu et al., 2011). As neurons lack energy store, they are vulnerable to this reduction in oxygen and glucose (Krol et al., 2013). The reduction in ATP levels causes energy imbalance and consequently the cells have difficulty to maintain ionic homeostasis. Membrane potential is lost and neurons and glia depolarize increasing the levels of Na<sup>+</sup> and Ca2<sup>+</sup> entering the cell (Katsura et al., 1994). This depolarization leads to a release of excitatory amino acids to the extracellular space. At the same time, energy dependent processes, such as presynaptic reuptake of excitatory amino acids, are disrupted resulting in excitotoxicity by accumulation of glutamate and subsequent Ca2<sup>+</sup> overload (Kyle and Saha, 2014). Glutamate acting continuously at its NMDA and AMPA receptors evokes Na<sup>+</sup> and Cl<sup>−</sup> inward, water flowing along with the ions resulting in cytotoxic edema (Dirnagl et al., 1999). Also, peri-infarct depolarizations can spread energy imbalance through brain cells (Kunz et al., 2010).

The increase in intracellular Ca2<sup>+</sup> concentration initiates a series of cytoplasmic and nuclear events that result in cellular damage, such as the production of ROS. ROS accumulation, particularly after reperfusion, will cause mitochondrial (Da Silva-Candal et al., 2017) and cellular membrane damage (Dugan and Choi, 1994). Intracellular signaling pathways triggered during excitotoxicity also stimulate the production of pro-inflammatory mediators (Dirnagl et al., 1999). All of these events happening together, after ischemia and also during reperfusion, will contribute to the progression of tissue damage and culminate with cellular death by either necrosis or apoptosis (Kunz et al., 2010).

The changes described above do not homogenously affect the ischemic territory. In the lesion core, where there is a reduction of approximately 80% of blood flow (Hossmann, 1994), permanent damage happens minutes after the ischemic event and cells are rapidly killed. In between this region and the normal tissue is the penumbra where there is still salvageable tissue, being thus the region of interest when it comes to neuroprotection (Dirnagl et al., 1999).

The events described so far, such as excitotoxicity, oxidative stress and inflammation, also affect the BBB structure. The BBB plays a vital role in regulating the traffic of fluid, solutes and cells at the blood-brain interface and maintaining the microenvironment homeostasis (Jiang et al., 2017). BBB is composed of brain endothelial cells (BEC) whose main particularity, when compared to other endothelial cells found in the organism, is the existence of an increased number of tight junctions connecting adjacent cells, resulting in a decrease of paracellular transport (Abbott et al., 2006). Thus, BEC form a barrier that protects the brain against possible toxic substances (Azad et al., 2015). Ischemia/reperfusion damage will act on BEC causing structural disruption of tight junctions and contributing to BBB increased permeability and dysfunction (Jiang et al., 2017).

Therefore, following an ischemic stroke, blood-borne cells, chemicals and fluid extravasate into brain parenchyma (Keaney and Campbell, 2015). The water and ion homeostasis of the brain is also disrupted, leading to cerebral vasogenic edema (Rosenberg, 1999). Infiltrating leukocytes exacerbate inflammatory responses and aggravate brain injury (Huang et al., 2006). Besides all the detrimental consequences of BBB disruption, one potential benefit is that it may facilitate therapeutic agents to reach the brain (Jiang et al., 2017).

### CURRENT LIMITATIONS IN THE DIAGNOSIS AND THERAPY OF ISCHEMIC STROKE

The reduced permeability of BBB to most drugs, including imaging contrast agents, macromolecular compounds, nucleic acids and proteins (Hawkins and Davis, 2005; Gabathuler, 2010; Essig et al., 2012; Tam et al., 2016) represents a major obstacle to the development of safe and effective diagnostic and treatment strategies for ischemic stroke (Mouhieddine et al., 2015). Even knowing that during ischemia/reperfusion there is a rupture in the BBB structure, this opening is transitory and the amount of drug that can actually reach the tissue is often not enough to produce the desired effect, especially in the case of macromolecular drugs (Chen and Gao, 2017; Merali et al., 2017). In this context, there has been much interest in the design of nanoparticles capable of carrying therapeutic and diagnostic agents across the injured and normal BBB and targeting the ischemic tissue and the penumbra region of the damaged tissue (Kafa et al., 2015; Mendonça et al., 2015; Fukuta et al., 2016; Zhao Y. et al., 2016).

Essentially two neuroimaging techniques are routinely used for the diagnosis and surveillance of patients suspected of acute ischemic stroke: X-ray computed tomography (CT) and multimodal magnetic resonance imaging (MRI) (Wang et al., 2016).

Although CT is widely available and less expensive, it is not sensitive for detecting ischemic stroke and distinguishing new events. Indeed, less than a third of patients with brain ischemia exhibits characteristics from CT findings within 3 h of symptoms onset (Chalela et al., 2007). Moreover, the fact that strokelike symptoms may also be present in a wide range of other non-vascular diseases like epileptic seizures and migraine, makes difficult the accurate diagnosis and decision for administration of a thrombolytic drug (Mouhieddine et al., 2015).

Conventional MRI is a vital and versatile imaging tool in clinical practice, offering advantages for the assessment of acute stroke, especially with diffusion-weighted imaging (DWI) (Chalela et al., 2007; Latchaw et al., 2009; Merino and Warach, 2010). However, the efficiency of detection of an ischemic stroke within 3 h of symptoms onset is around 70% (Chalela et al., 2007).

Although a more accurate diagnostic of ischemic stroke can still be achieved through the combined use of CT and MRI (Brazzelli et al., 2009), there is a great need for a more sensitive technique. In this scenario, much effort has been devoted to the search for serum biomarkers, new imaging methods and the improvement of contrast agents through nanotechnology (Mouhieddine et al., 2015; Wang et al., 2016).

Regarding therapeutic options for acute ischemic stroke, only two are available: thrombolysis and mechanical thrombectomy, both strategies focusing on reperfusion therapy. Thrombolysis is the leading treatment for acute stroke, performed by a pharmacological intervention using recombinant tissue plasminogen activator or t-PA, the only available treatment approved by the US FDA (Chen and Gao, 2017). However, a small number of stroke victims can actually use t-PA treatment since it has a narrow time window of treatment efficacy on the eligible patients (4–5 h of symptom onset), and an increased risk of intracerebral hemorrhage (Khandelwal et al., 2016).

Mechanical thrombectomy utilizes a surgical process to mechanically restore the blood flow in large cerebral arteries promoting a more effective recanalization than thrombolysis (Chamorro et al., 2016). Although the mechanical procedure has shown some advantages when compare to t-PA treatment alone (Albers et al., 2018) it is not available in every hospital, reducing the number of patients who can receive endovascular treatment with mechanical thrombectomy (Chamorro et al., 2016).

Moreover, even when recanalization is successful in installing reperfusion, there is still left sequels that can impair patients life quality, as reperfusion itself can induce ROS production and tissue damage (Manzanero et al., 2013). Aiming to protect brain tissue against ischemia/reperfusion damages, neuroprotective strategies have been studied to extend neurons survival, increase the therapeutic window, and induce neurological repair improving functional outcomes (George and Steinberg, 2015).

Several promising neuroprotective drug candidates have been identified in rodent models. However, most of the strategies established in animal models has failed in clinical trials (George and Steinberg, 2015). The poor methodological rigor in some preclinical studies and the inappropriate use of animal models when simulating the patient conditions might have contributed to the failure in translation of rodents results into clinical success (Feng and Belagaje, 2013).

Moreover, the complexity of the physiopathological events that take place during cerebral ischemia and the still limited knowledge about the molecular and cellular mechanisms involved and their spatial and temporal occurrence represent an obstacle to the identification of effective tissue biomarkers and the design of powerful nanotechnology-based strategies for diagnosis and therapy (Da Silva-Candal et al., 2017).

In summary, the diagnosis and therapy of ischemic stroke need to be improved and, in that sense, a critical step seems to be the safe and effective delivery of substances to the ischemic region (Mouhieddine et al., 2015). In this context, the design of tailored nanoparticles that could ameliorate the drug delivery and the existing imaging techniques with higher sensibility, specificity and spatial and temporal resolution could bring great benefits to stroke diagnosis and treatment. Along with that, the development of reversible, biocompatible and direct nanobiosensors for long term use in animal and tissue models of brain ischemia would be of great interest to provide knowledge about the pathophysiology of stroke and contribute to the development of more specific and effective treatments.

As illustrated in **Figure 1** and discussed in details below, liposomes and carbon nanomaterials have been extensively studied as nanomaterials for application against ischemic stroke and several different strategies have emerged with promising results.

## LIPOSOMES AS DRUG NANOCARRIERS

Liposomes are the first generation of drug nanocarriers (Budai and Szógyi, 2001). They are vesicles composed of concentric lipid bilayers, which are separated by water compartments. They are constituted of natural or synthetic lipids with amphiphilic nature: a polar head group covalently attached to one or two hydrophobic hydrocarbon tails (Frézard, 1999). These lipids are usually biocompatible and biodegradable and resemble those found in biological membranes (Masserini, 2013; Lamichhane et al., 2018). Liposomes can be found in the composition of several pharmaceutical products already marketed for treatment

of cancer, fungal infections, and as immunoadjuvant in vaccines (van Hoogevest and Wendel, 2014).

Liposomes are classified according to their size and number of lamellae: small unilamellar vesicles (SUV) with a size up to 100 nm and one bilayer; large unilamellar vesicles (LUV) with a size of more than 100 nm and one bilayer; and multilamellar vesicles (MLV) that can reach several µm and are made of multiple concentric bilayers (Masserini, 2013).

As illustrated in **Figure 2**, liposomes are highly versatile, allowing accommodation of hydrophilic drugs in the aqueous space or hydrophobic drugs in the lipid bilayer and surface modifications to control their interaction with biological environments and fate (Li et al., 2017). Once the drug is inside the liposome, it is protected against physiologically occurring events, such as enzymatic degradation, immunological and chemical inactivation and fast plasma clearance, resulting in enhancement and prolongation of its action. Another common benefit of liposome encapsulation is the decreased drug exposure of healthy tissues, reducing the side effects of the treatment (Bozzuto and Molinari, 2015). Thus, because of all of the characteristics discussed above, liposomes represent the most well-studied system for drug delivery to treat CNS disorders (Vieira and Gamarra, 2016).

Liposomes have been widely studied for use in stroke therapy, as shown in **Table 1**. Imaizumi et al. (1990) performed one of the precursor studies with liposomes in an animal model of focal ischemic stroke. Once superoxide dismutase (SOD), a free radical scavenger, has a short half-life and is unable to cross the BBB, this enzyme was administered in the liposome-encapsulated form via jugular vein in rats submitted to middle cerebral artery occlusion (MCAO). Brains of animals injected with the liposome formulation showed increased levels of SOD and a reduced infarct volume (Imaizumi et al., 1990).

cationic and/or anionic charges; and specific targeting ligands for BEC or

Blood brain barrier is one of the major issues in the treatment of stroke. It is crucial that liposomes can cross the BBB and it is also desirable that they remain in the bloodstream for a long period of time, so as to deliver a sufficient quantity of drug to the brain tissue (Moghimi et al., 2001). The circulation time of liposomes can be enhanced through reduction of particle size or modification of their surface by polyethylene glycol (PEG) (Fukuta et al., 2014; Bozzuto and Molinari, 2015; Wang et al., 2015; Campos-Martorell et al., 2016; Chen and Gao, 2017). It has been reported that PEGylated liposomes can accumulate into ischemic brain regions after a stroke event (Fukuta et al., 2016), presumably because of the disruption of the BBB and the enhanced vesicle permeation and retention at the ischemic site. Fukuta et al. (2014) have shown that PEGylated liposomes detection increased in the ischemic region over time period, even when there was little blood flow in the affected area. In another work, Ishii et al. (2013) evidenced that a single injection of PEGylated liposomes containing the immunosuppressant FK506 at a low dosage in MCAO rats significantly reduced cerebral cell death and ameliorated motor function deficits. Interestingly, liposome encapsulation of FK506 was found to decrease the drug toxicity that has been responsible for the failure of clinical trials of this drug against stroke (Ishii et al., 2013)

The surface charge is another factor that can influence the accumulation of liposomes into the brain. Campos-Martorell et al. (2016) showed that neutral and negatively charged PEGylated liposomes administered intravenously were captured to a lower extent by the liver and lungs in comparison to cationic vesicles. As consequence, the neutral and anionic liposomes showed prolonged circulation time in the blood and higher uptake in the ischemic region (Campos-Martorell et al., 2016). In another study, non-PEGylated liposomes with different surface charges were given through intracarotid injection as an attempt to achieve early brain deposition. The brain accumulation of these liposomes following intra-arterial injection was more effective from cationic vesicles than anionic or neutral ones, possibly due to the electrostatic interactions between the cationic liposomes and negatively charged cell surface, enhancing nanoparticle uptake by adsorptive-mediated endocytosis (Joshi et al., 2014).

An additional option to further increase the transport of drugs across the BBB is the coupling of specific ligands to the liposome surface in order to target surface proteins constitutively expressed at the BBB, such as low-density lipoprotein receptor, glucose transporter (GLUT1), transferrin or insulin receptors (Gabathuler, 2010; Ying et al., 2010; Spuch and Navarro, 2011). This kind of surface modification can increase liposome uptake by the BEC through receptor-mediated transcytosis, resulting in greater amount of drug that reaches the brain (**Figure 1**) (Wang et al., 2015). Following this approach, Zhao Y. et al. (2016) investigated a dual targeting strategy for ischemic stroke treatment, using liposomes containing a neuroprotectant (ZL006) with their surface decorated with a transferrin receptorderived peptide ligand (T7) to improve the passage across the BBB and a stroke-homing peptide (SHp) to target the ischemic region. T7&SHp-P-L/ZL006 liposomes decreased the infarct volume, neurological deficit, and histopathological severity in the MCAO rat model (Zhao Y. et al., 2016).

ischemic tissue receptors.


Zhang et al. (2018) reported that liposomes with Cyclo(Arg-Gly-Asp-D-Phe-Cys) (cRGD) covalently coupled to the membrane surface can bind to the activated platelets while not to the resting platelets. The authors demonstrated that the cRGD liposomes containing urokinase could improve the thrombolytic efficacy by almost fourfold over free uroquinase, showing potential for treatment of ischemic stroke (Zhang et al., 2018).

Among the neuroprotective drug candidates for ischemic stroke, those that can activate the ACE2-Ang-(1-7)-Mas axis of the Renin Angiotensin System (RAS) are of special interest (Bennion et al., 2015b). The heptapeptide, angiotensin-(1-7) [Ang-(1-7)], final endogenous effector of ACE2-Ang-(1-7)-Mas pathway, has shown neuroprotection in several models of ischemic stroke (Zhang et al., 2008; Mecca et al., 2011; Bennion et al., 2015a, 2018). On the other hand, the rapid in vivo metabolism of the peptide through proteolytic inactivation results in a short plasma half-life (less than 1 h) and biological actions, limiting its therapeutic potential. Furthermore, the high molecular weight and hydrophilic character of Ang-(1- 7) prevent its absorption across biological barriers, such as BBB. Interestingly, it was shown that encapsulation of Ang-(1- 7) into PEGylated liposomes prolonged the peptide biological action from 8 min to 5 days following microinjection into the rostral ventrolateral medulla (RVLM) of normotensive rats (Silva-Barcellos et al., 2004). This data strongly suggests that such liposome formulation of Ang-(1-7) may find application in the treatment of ischemic stroke.

The intranasal route has recently gained, attention once it is a region that has free access to the brain tissue by the olfactory and trigeminal nerve and it also constitutes a non-invasive alternative route (Zhao Y.Z. et al., 2016). Besides that, intranasal drug delivery provides decreased systemic exposure of the drug and limited degradation of therapeutics (Meredith et al., 2015). In this context, the use of nanoparticles also brings significant benefits, such as increase of mucoadhesion providing sustained and controlled drug release, enhanced drug deposition at olfactory epithelium and improved nasal drug absorption (Illum, 2000; van Woensel et al., 2013). In addition of promoting the accumulation of drug into the brain, liposomes can decrease the mucosal drug toxicity that may be unleashed when intranasal administration is done chronically (Mainardes et al., 2006). It is also noteworthy that cationic liposomes were particularly effective in prolonging the residence time at the nasal cavity and assisting transport of proteins across nasal mucosa (Zheng et al., 2015). Furthermore, PEGylated liposomes enhanced the drug bioavailability and showed greater residence time when compared to the conventional non-PEGylated liposomes (Khan et al., 2017). As another strategy, Zhao Y.Z. et al. (2016) prepared gelatin-cored liposomes encapsulating basic fibroblast growth factor (bFGF), a potential protective substance for patients with stroke. The bFGF-liposomes applied intranasally in a rat model of cerebral ischemia/reperfusion improved bFGF accumulation in brain tissues and promoted functional recovery of the animals (Zhao Y.Z. et al., 2016).

Nanotheranostics is the field that combines at the same time nanoparticle-mediated therapy and the study of nanoparticles localization and fate usually through imaging (Mouhieddine et al., 2015). Agulla et al. (2014) have identified the heat shock protein-72 (HSP72) as a suitable biomarker for the peri-infarct region and described the development of anti-HSP72 stealth immunoliposomes labeled with gadolinium and containing the neuroprotectant citicoline. The viability of this nano-platform for the diagnostic by MRI and therapy of cerebral ischemia was established in an animal model (Agulla et al., 2014). In a more recent study, Liu et al. (2016) proposed citicolineliposomes as a prototype theranostic system, after evidencing that their accumulation can be detected in the ischemic brain using Chemical Exchange Saturation Transfer (CEST)-MRI. Citicoline is a precursor of phosphatidylcholine that has a cytosine in its structure, which permits its detection by CEST-MRI. The more abundant expression of vascular cell adhesion molecule 1 (VCAM-1) on inflamed vessels in the ischemic brain led these authors to further design anti-VCAM-1 immunoliposomes that promoted greater accumulation of citicoline in the brain of ischemic rats, in comparison to non-targeted liposomes. However, the uptake of these targeted liposomes in the rat brain after ischemic injury was approximately fourfold lower and showed a more dispersed distribution after intravenous administration, when compared to intra-arterial injection.

After observing the increased expression of paired immunoglobulin-like receptor B (PirB) in the ischemic hemisphere of mice 24 h post-MCAO, Wang et al. (2018) constructed anti-PirB immunoliposome probe with a nearinfrared fluorophore that was successfully applied to in vivo imaging for upregulated PirB region in a cerebral ischemic stroke model. These authors also used soluble PirB ectodomain, as a therapeutic reagent encapsulated in liposomes, showing a significant functional recovery in a model of ischemic stroke.

Wen et al. (2012) have investigated PEGylated liposomes for brain drug targeting and imaging, after encapsulation of apomorphine in the internal aqueous compartment and quantum dot (QD) in the liposomal membrane. As main advantages over conventional organic fluorophores, semi-conductor-based QDs have narrow band emissions together with large ultraviolet absorption spectra, which enable multiplex imaging under a single light source. The QD- and drug-containing liposomes were found to accumulate into the brain of mice, as evidenced by QD fluorescence imaging and drug quantification in the tissue, showing potential for imaging and treating brain disorders (Wen et al., 2012).

Several concerns will need to be overcome before engineered liposomal nanocarriers can be incorporated into everyday clinical practice. The ideal nanoparticles for use as contrast agents must be detectable at reduced doses, should cross the BBB and target a time-accurate biomarker. It must also be non-toxic and preserve cellular functions. Hence, contrast agents associated to liposomes should be carefully chosen, or eventually optimized, so as to avoid adverse effects. For instance, a recent work emphasized the potential toxicities of QDs due to their specific metallic composition, the risk of release of toxic metal ions after prolonged exposure of biological systems and several undefined factors of nanoparticles themselves (Wang and Tang, 2018). Thus, much effort is still needed regarding the achievement of biocompatible QDs through surface modification and the understanding of their mechanisms of toxicity, before they can be considered for use in humans. Considering that MRI is routinely used in the clinic for diagnosis of stroke, the translational potential of liposome-based MRI contrast agents seems currently greater than that of fluorescence-based imaging agents.

It is also important to take into account that once the nanocarrier has reached its target, the encapsulated drug still has to be released to exert its pharmacological action. Three distinct mechanisms can contribute to the in vivo release of the encapsulated drug from liposomes: (i) the spontaneous simple diffusion of the drug across the liposome bilayer; (ii) the endocytosis of liposomes by cells, their degradation by lysosomal phospholipases and the subsequent release of the drug in the cytosol or the extracellular medium after exocytosis; (iii) the induction of drug release from liposomes by specific stimuli, such as external magnetic field or focused ultra-sound, variations in temperature or pH, depending on the characteristics of liposomes (Vieira and Gamarra, 2016). Since there is a decrease in the pH at the ischemic region (pH<6.75) (Anderson et al., 1999), it would be interesting to explore pH-sensitive liposomes to improve the release of neuroprotective agents or specific image markers into the ischemic region. In agreement with this proposal, pHsensitive liposomes have already shown good results in the delivery of anticancer drugs in the more acidic environment of tumors (Ferreira et al., 2013). Such liposomes were first prepared from the mixture of dioleoylphosphatidylethanolamine, a lipid with a carboxylic acid head group and a PEGylated

lipid (Simões et al., 2004). More recently, Pacheco-Torres et al. (2015) incorporated an engineered ion channel in the liposome membrane, which could discriminate physiologically relevant minor pH changes with an unprecedented precision of 0.2 pH unit, and release the drug accordingly in the tumor site.

### CARBON-BASED NANOMATERIALS

## General Physicochemical Characteristics

Carbon nanomaterials (CNMs) have been the subject of intense research during the last 30 years due to their unique properties related to the quantum confinement of the electrons movement at discrete energy levels in the nanometric structure. The most studied allotropes are the sp<sup>2</sup> hybridized carbons (**Figure 3**): fullerenes (zero-dimensional, 0D), carbon nanotubes (CNTs, one-dimensional, 1D) and graphene (two-dimensional, 2D), although structures such as carbon dots and nanodiamond have been reported (Xu et al., 2004; Mochalin et al., 2012). Fullerenes (C60) were first described when Kroto et al. (1985) reported the production of a stable cluster consisting of 60 carbon atoms by vaporization of graphite using laser irradiation.

Carbon nanotubes, tubular carbon structures with nanometric diameter, gained attention of the scientific community with the publication of Iijima (1991) reporting the synthesis of "helical microtubules of graphitic carbon" by arc-discharge evaporation method. Graphene, a two-dimensional single layer of carbon atoms, was reported by Novoselov et al. (2004), who described the preparation of monocrystalline graphitic films of few atoms thickness (including single layer) by mechanical exfoliation of highly oriented pyrolytic graphite. Nowadays these nanomaterials are widely employed in different fields including among others: communication, energy, military, aerospace, and nanomedicine (De Volder et al., 2013). However, for practical applications it is imperative to modify the surface of these materials to allow integration with the desired medium.

For biological applications, surface modifications (covalent or non-covalent) allow better dispersion of the nanomaterial in physiological medium and ensure biocompatibility. In a wider perspective, this approach can confer functional characteristics to the CNMs. The conjugation of the nanomaterials with different

molecules, such as polymers, proteins, different DNA sequences or other specific compounds, can generate different functions for in vivo applications, such as sensors, biomarkers, drug carrier for therapeutics (**Figure 4**). The purification of the CNMs is an earlier step that is crucial for elimination of residues or by-products, such as metal particles of catalysts used in the synthetic procedure amorphous carbons and other impurities that may exert toxicity.

## Potential Therapeutic Applications

Some of the main advantages of CNTs, when compared to other drug nanocarriers, are their high specific area, allowing binding of multiple copies of different molecules onto their surface, and their propensity to cross different biological barriers, entering the cytoplasm either passively through 'nanoneedle' mechanism or through endocytosis, which allows the transport and delivery of drugs and macromolecules (Tîlmaciu and Morris, 2015). The mechanism of uptake of CNTs appears to vary with functionalization, length, diameter, number of walls, and concentration of CNTs. Indeed, the unique 'nanoneedle' transport mechanism has been reported for certain types of single-walled carbon nanotubes (SWCNTs), whereas multiwalled carbon nanotubes (MWCNTs) are more likely to be endocytosed because of their larger diameter (Kafa et al., 2015; Mehra and Palakurthi, 2016). The intrinsic spectroscopic properties of CNTs, such as Raman and photoluminescence, afford additional advantages for real-time monitoring of drug delivery efficacy in vitro and in vivo (Lamprecht et al., 2012).

After SWCNTs given orally in a rodent Alzheimer's disease model were found to successfully deliver acetylcholine into the brain (Yang et al., 2010), CNTs started to be considered as potential tools in the treatment of CNS diseases. Other works have established the direct ability of CNTs to cross the BBB. Using an in vitro co-culture BBB model (primary porcine brain endothelial cells, PBEC), Kafa et al. (2015) brought evidence of the permeation of functionalized multi-walled carbon nanotubes (f-MWCNTs), more specifically MWCNT-NH<sup>3</sup> <sup>+</sup>, across the cell monolayer via energy-dependent transcytosis. In another study, f-MWCNTs were shown to enter into the brain via endothelium, following systemic injection in rodents, with accumulation not just in endothelium but also in brain parenchyma (Costa et al., 2016).

Carbon nanotubes can be used either for their own action on the nervous tissue or their ability to carry other drugs (**Table 2**). Lee et al. (2011) reported for the first time the neuroprotective effects of SWCNTs (without any therapeutic molecule), by showing that pretreatment of rats with MCAO ischemic brain injury through intra-ventricular administration of amine-modified single-walled carbon nanotubes (a-SWCNTs) could protect neurons and enhance the recovery of behavioral functions. Evidence was obtained that the Akt pathway and the maintenance of cell-to-cell interactions (higher levels of N-Cadherin) contributed to the protective action of a-SWNTs (Lee et al., 2011).

In addition to the neuroprotective action of the nanotube itself in a model of ischemia, this nanomaterial was also effective when impregnated with progenitor cells. An in vivo study has shown



that hydrophobic carbon nanotubes (HPCNT) impregnated with subventricular zone neural progenitor cells (SVZ NPCs) could repair damaged neural tissue following stroke. HPCNT-SVZ NPCs transplanted rats exhibited improved behavior and reduced volume and area of infarct cyst, compared to experimental control. The majority of the transplanted HPCNT-SVZ NPCs collectively broadened around the ischemic injured region and the SVZ NPCs differentiated into mature neurons, attained the synapse morphology (TUJ1, synaptophysin), and decreased microglial activation (CD11b/c [OX-42]). This study pioneered the concept that CNTs can improve stem cell differentiation, leading to heal stroke damage (Moon et al., 2012).

Al-Jamal et al. (2011) have confirmed that CNTs have also a potential to deliver siRNA into the brain, through demonstration of functional recovery in endothelin-1 stroke murine model after stereotaxic administration of siRNA (for caspase 3 silencing) complexed with amino-functionalized MWNT.

Fullerene derivatives were repeatedly reported as neuroprotective in in vitro and in vivo stroke models (Huang et al., 2001; Lin et al., 2002; Fluri et al., 2015; Vani et al., 2016). Hexasulfobutylated fullerene (C60 FC4S), when injected intravenously before and during MCAO in Long-Evans rats, promoted the increase of nitric oxide content and the decrease of LDH levels and total volume of infarction, presumably through action as free radical scavenger (Huang et al., 2001). Intracerebroventricular infusion of carboxyfullerene in rats submitted to MCAO stroke attenuated cortical infarction and prevented the elevation of lipid peroxidation and depletion of GSH level induced by transient ischemia/reperfusion. However, adverse effects and death were observed in some cases (Lin et al., 2002). In accordance with these works, Vani et al. (2016) showed that polyhydroxylated fullerene or fullerenol (OH-F) derivatives protected rat brain cells against ischemia/reperfusion injury and inhibited brain oxidative/nitrosative damage in a MCAO model, acting as a potent scavenger of free radicals. Also, Fluri et al. (2015) reported that fullerenol and glucosaminefullerene conjugate (GlcN-F) led to a reduction of cellular damage and inflammation after stroke. In this case, fullerenol

worked as a radical scavenger and the glucosamine derivative reduced inflammation (Fluri et al., 2015).

Another recently discovered nanomaterial, graphene, exhibited unique biological actions, with potential application in stroke treatment. In the study conducted by Mendonça et al. (2015) reduced graphene oxide (rGO) was found to reach the thalamus and hippocampus of rats following systemic injection. The entry of rGO involved a transitory decrease in the BBB paracellular tightness, as evidenced by extravasation of vital Evan's Blue stain into the brain (Mendonça et al., 2015). Importantly, the rGO-induced transitory opening of the BBB seems not to cause major deleterious effects. Although stroke leads to a disruption of the BBB, we cannot control the duration and extent of this phenomenon. The temporary permeabilization of the BBB caused by rGO may allow intentional enhancement of brain uptake of delivery systems for diagnostic or therapeutic purposes. Thus, rGO may be used to allow a controlled therapeutic window for carrying drugs into the ischemic site.

In order to fully address the potential of CNTs for the treatment of brain ischemia, one should also consider their toxicity, its possible causes and ways to overcome it. When used in their pristine state, directly after synthesis, CNTs contain impurities, aggregate into bundles in aqueous media and strongly interact with biomolecules, leading to severe toxic effects. Nonetheless, when purified and surface-functionalized, their toxicity is drastically decreased. It is now well established that characteristics of CNTs such as degree and type of functionalization, purity, shape, stability, surface reactivity and agglomeration state exert a marked influence on their biological and toxic effects (Grabinski et al., 2007; Johnston et al., 2010; Zhang et al., 2010, 2011). Several studies also support the model that the main mechanism of toxicity of these nanomaterials is the induction of oxidative stress (Murray et al., 2009; Pichardo et al., 2012; Shvedova et al., 2012; Liu et al., 2014; Weber et al., 2014). The use of dense functionalization, biocompatible polymers and highly purified materials generating aqueous-stable dispersions was found effective to minimize this toxicity (Liu et al., 2008; Zhang et al., 2011; Iverson et al., 2013; Mendonça et al., 2015; Calle et al., 2018).

#### Biological Imaging and Nanobiosensors

There is still a great need of methods to create a precise, accurate and space-time resolution detection for investigating the changes in the brain tissue after an ischemic event. The rapid development of nanotechnology has led to promising diagnostic tools for stroke using engineered nanomaterials. In addition to the diagnosis of patients, this technology may aid in understanding the biochemical and pathophysiological mechanisms of stroke using in vitro and in vivo model.

Honjie Dai's group developed a new method to image mouse cerebral vasculature without craniotomy, using throughscalp and through-skull fluorescence imaging in a biological transparent sub-window in the 1.3–1.4 µm wavelength range (the NIR-IIa region). They exploited the intrinsic photoluminescence of single-walled carbon nanotubes (SWNT–IRDye800). This technique allows a fluorescence imaging to a depth of >2 mm, sub-10-µm resolution and imaging rate of ∼5.3 frames per second providing a non-invasive, real-time assessment of blood flow anomaly with high spatial and temporal resolution in a MCAO stroke model (Hong et al., 2014).

Besides the development of new tools for the diagnosis of patients, another equally important approach is the use of biosensors to uncover the biochemical and pathophysiological mechanisms of stroke in animal and cells models. In this scenario, optical CNT-based biosensors show great promise, because of their unique fluorescence properties, allowing high spatio-temporal resolution at biologically relevant concentrations (Soleymani, 2015; Polo and Kruss, 2016).

Strano's group has studied and developed a new class of biosensors: fluorescent sensors for the detection of biomolecules, formed from single-wall semiconductor carbon nanotubes conjugated to polymers, called CoPhMoRe (corona phase molecular recognition) (Kruss et al., 2013; Zhang et al., 2013). These SWCNT-based biosensors operate basically with the fluorescence quenching and/or peak shift in the NIR-II window, as a response to surface adsorption events of certain molecules. As discussed above, the SWCNT fluorescence in the near infrared (nIR) is ideal for biomedical applications because it is in the window of optical transparency of biological tissues (O'connell et al., 2002). The most used polymers attached to the sensor are specific DNA sequences. As demonstrated in the work of Kruss et al. (2014), DNA oligonucleotides bound to PEG macromolecules (to enhance biocompatibility) interact with SWCNT by wrapping the oligonucleotide chain to the carbonic surface, producing a biocompatible hybrid that functions as a strong and selective dopamine sensor. In this work, an increase in the sensor intrinsic fluorescence response was demonstrated when in contact with dopamine. Thus, a detection specificity to certain molecular species, such as adenosine 5<sup>0</sup> -triphosphate (ATP) (Kim et al., 2010), hydrogen peroxide (H2O2) (Jin et al., 2010), neurotransmitters (Kruss et al., 2014, 2017) and nitric oxide (NO) was achieved (Iverson et al., 2013; Ulissi et al., 2014).

These studies have addressed the detection of these biomolecules in vitro, through direct interaction in cell culture. However, as demonstrated by Iverson et al. (2013), it is also possible to apply CoPhMoRe technology in vivo. These authors used SWCNTs wrapped with DNA oligonucleotides functionalized with PEG (to enable systemic injections) as an in vivo nitric oxide sensor. They used a rodent model injected with RcsX tumor cells to cause liver inflammation and in situ generation of NO molecules in the liver. A strong fluorescence quenching effect in the presence of NO free radicals was detected (1 µM detection limit). The tissue autofluorescence and background signals were successfully removed from the characteristic NIR-II signals of the SWCNT reporters, based on their spectral differences. It is also noteworthy that SWCNT-fluorescence exhibits no blinking, no bleaching and a large Stokes shift (Kruss et al., 2013). These promising results support the potential of this technology for investigating the pathophysiological mechanisms of stroke, considering its ability to detect molecules possibly involved in an ischemic event, such as neurotransmitters, NO and H2O2. The detection at a singlemolecule level, long-term sensing and regeneration/reversibility of the sensor response may find applications in new cellular

assays for disease diagnostics, molecular signaling, and detection of inflammatory signals.

## CONCLUDING REMARKS

Stroke, being one of the leading causes of death and disabilities worldwide, still lacks better ways to diagnose and treat the affected patients. Many studies have focused on neuroprotective drugs to provide guard to the damaged brain tissue and they were successful in improving animals' behavior and neurological score but translational approaches did not show the same efficiency yet. As discussed here, the techniques currently available for the diagnosis of stroke also lack sensitivity.

In this context, nanotechnology has emerged, with effective means of improving drug delivery to the brain and, more specifically, to the ischemic region. As shown in the present review, there is a great potential of liposomes and carbon nanomaterials for applications in the diagnosis and therapy of ischemic stroke. Interestingly, both types of nanostructures exhibit very distinct characteristics and ranges of applications.

The major advantages of liposomes are their reduced toxicity, excellent biocompatibility and ability to accommodate a large variety of bioactive and contrast agents, with very different physicochemical characteristics. Therefore, several neuroprotective agents have been evaluated under the encapsulated form in liposomes, in experimental models of ischemic stroke. The most promising results, regarding drug delivery to the brain and neuroprotection, were achieved using PEGylated liposomes bearing targeting ligands for BECs and/or ischemic tissues receptors. The proof of concept of nanotheranostic approaches using MRI, fluorescent QD or NIR probes was also established. However, in order to translate experimental results into successful clinical applications, there are still some important issues to be addressed. Regarding the route of administration, intra-arterial route has shown higher drug delivery efficiency to the brain, in comparison to the intravenous or intranasal routes. Thus, it is felt that more effective targeting strategies are still needed to improve therapeutic efficacy of these nanosystems by intravenous or intranasal routes. Another complementary approach would be to select the most effective and least toxic neuroprotectant, like for instance the Ang-(1-7) peptide hormone. Finally, the high complexity of liposome nanosystems and related stability and cost issues make necessary the development of specific nanoplatforms for industrial production. In that sense, the experience acquired in the development of existing pharmaceutical liposome-based product should be useful.

We have shown here that carbon based nanomaterials, including nanotubes, fullerenes and graphene, display

### REFERENCES

great versatility regarding size, morphology and surface physicochemical properties. CNTs started to be investigated for ischemic stroke and potential applications for therapy, biosensing and imaging have been identified based on their antioxidant action, their intrinsic photoluminescence, their ability to cross the BBB, transitorily decrease the BBB paracellular tightness, carry oligonucleotides and cells and induce cell differentiation. Therefore, their unique physicochemical properties and the improvement of functionalization methods, now offer a wide range of potential applications. However, considering that CNTs are not naturally encountered in the human body, an important step for clinical translation is the full evaluation of biodistribution and toxicological effects of these materials (considering shape, size, functionalization, and possible catalytic particles contaminants) after systemic administration. Thus, CNTs are at an early stage of development and much work is still needed to reach practical applications. It is felt that CNTs have their greatest potential as nanobiosensor for investigating the biochemical and pathophysiological mechanisms of stroke using in vitro and in vivo model, thus contributing to the development of more effective therapies for this pathology.

## AUTHOR CONTRIBUTIONS

FF and AM defined the outline of the manuscript. AM, GB, and LF wrote the first draft of the Section "Introduction." AM and LF wrote the first draft of the Sections "Physiopathology of Ischemic Stroke" and "Current Limitations in the Diagnosis and Therapy of Ischemic Stroke." LF and FF wrote the first draft of the Section "Liposomes as Drug Nanocarriers." GB wrote the first draft of the Section "Carbon-Based Nanomaterials." FF and AM fully revised the first draft of the manuscript. All the authors contributed to Abstract, Concluding remarks, and the final form of the manuscript.

## FUNDING

This work was supported by the Brazilian agencies Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq, Grant Nos. 305659/2017-0, 401216/2014-4, and 431571/2016-3), Coordenação de Aperfeicoamento de Pessoal de Nível Superior (CAPES, Grant No. PNPD 20131163-32001010007P5) and Fundação de Amparo a Pesquisa do Estado de Minas Gerais (FAPEMIG, Grant Nos. RED-00007-14, APQ-03129-16, and APQ-03069-16) for financial support. FF is recipient of fellowship from CNPq (Grant No. 305659/2017-0), GB from CAPES-PNPD postdoctoral scholarship, and LF received a graduate scholarship from CNPq (Grant No. 170690/2017-2).

Adeoye, O., Hornung, R., Khatri, P., and Kleindorfer, D. (2011). Recombinant tissue-type plasminogen activator use for ischemic stroke in the United States: a doubling of treatment rates over the course of 5 years. Stroke 42, 1952–1955. doi: 10.1161/STROKEAHA.110.61 2358

Abbott, N. J., Rönnbäck, L., and Hansson, E. (2006). Astrocyte-endothelial interactions at the blood-brain barrier. Nat. Rev. Neurosci. 7, 41–53. doi: 10.1038/nrn1824


toxicity to systemic brain delivery. J. Control. Release 241, 200–219. doi: 10.1016/j.jconrel.2016.09.033





**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Fernandes, Bruch, Massensini and Frézard. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Nano-Architectural Approaches for Improved Intracortical Interface Technologies

Youjoung Kim1,2, Seth M. Meade1,2† , Keying Chen1,2† , He Feng1,2, Jacob Rayyan1,2 , Allison Hess-Dunning1,2 and Evon S. Ereifej1,2 \*

<sup>1</sup> Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, <sup>2</sup> Advanced Platform Technology Center, Louis Stokes Cleveland Department of Veterans Affairs Medical Center, Cleveland, OH, United States

Intracortical microelectrodes (IME) are neural devices that initially were designed to

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Catalina Vallejo Giraldo, National University of Ireland Galway, Ireland Erin Purcell, Michigan State University, United States

#### \*Correspondence:

Evon S. Ereifej eereifej@aptcenter.org; eereifej@gmail.com

†These authors have contributed equally to this work.

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 20 April 2018 Accepted: 14 June 2018 Published: 17 July 2018

#### Citation:

Kim Y, Meade SM, Chen K, Feng H, Rayyan J, Hess-Dunning A and Ereifej ES (2018) Nano-Architectural Approaches for Improved Intracortical Interface Technologies. Front. Neurosci. 12:456. doi: 10.3389/fnins.2018.00456 function as neuroscience tools to enable researchers to understand the nervous system. Over the years, technology that aids interfacing with the nervous system has allowed the ability to treat patients with a wide range of neurological injuries and diseases. Despite the substantial success that has been demonstrated using IME in neural interface applications, these implants eventually fail due to loss of quality recording signals. Recent strategies to improve interfacing with the nervous system have been inspired by methods that mimic the native tissue. This review focusses on one strategy in particular, nano-architecture, a term we introduce that encompasses the approach of roughening the surface of the implant. Various nano-architecture approaches have been hypothesized to improve the biocompatibility of IMEs, enhance the recording quality, and increase the longevity of the implant. This review will begin by introducing IME technology and discuss the challenges facing the clinical deployment of IME technology. The biological inspiration of nano-architecture approaches will be explained as well as leading fabrication methods used to create nano-architecture and their limitations. A review of the effects of nano-architecture surfaces on neural cells will be examined, depicting the various cellular responses to these modified surfaces in both in vitro and pre-clinical models. The proposed mechanism elucidating the ability of nanoarchitectures to influence cellular phenotype will be considered. Finally, the frontiers of next generation nano-architecture IMEs will be identified, with perspective given on the future impact of this interfacing approach.

Keywords: nano-architecture, topography, intracortical microelectrodes, neuroinflammation, mechanotransduction

## INTRODUCTION

Intracortical microelectrodes (IME) were initially designed for research purposes to enable researchers in the late 1930s an ability to improve the understanding of the nervous system (Hodgkin and Huxley, 1939; Renshaw et al., 1940; Grundfest and Campbell, 1942; Grundfest et al., 1950). The first clinical use of neural electrode technology was in 1985, when the FDA approved the use of cochlear prosthetics (Spelman, 1999). Since then, clinical implementation of IME has been

employed to treat patients with numerous neurological diseases and injuries, such as amyotrophic lateral sclerosis (ALS) and spinal cord injuries (Gilja et al., 2015; Schroeder and Chestek, 2016; Ajiboye et al., 2017). Unfortunately, an impediment preventing the clinical deployment of IME technology is the complex inflammatory response occurring after electrode implantation, leading to decreased recording quality (Chestek et al., 2011; Jorfi et al., 2014; Kozai et al., 2015). The initial insertion of IME produces an injury in the local brain tissue, breaching of the blood brain barrier (BBB), eliciting an influx of chemical and biological markers, resulting in an inflammatory response (Polikov et al., 2005; Lau et al., 2013; Potter et al., 2013; Kozai et al., 2015). The early failure of IME has instigated substantial research in the development of next generation electrodes.

There are numerous types of IME, such as silicon microelectrode arrays, metal micro/nano-wires, carbon nanotubes (CNTs), and conductive polymers, which have been categorized by their backbone material (for full reviews on these IME types, see Voge and Stegemann, 2011; Fernández and Botella, 2017). Nevertheless, biomimetic alterations and advancements have inspired the design of the most recent IME technology. Engineers and scientists reason that biomimetic alterations to create new electrodes that reflect the properties of the brain will allow for better biocompatibility of the implants in vivo, which may lead to improved quality and longevity of recordings (Nguyen et al., 2014; Patel et al., 2016; Chen R. et al., 2017; Ereifej et al., 2017; Wei et al., 2018). For example, Tybrandt et al. (2018) have developed stretchable electrode grids that allow for high density and high-quality chronic recordings, which reflect the modulus of the brain better than either traditional metal or silicon-based electrodes. Liu (2017) have created a flexible electrode grid, neural mesh that, once injected into the brain, unfolds and is able to record freely moving rats with stable recording quality and coherence chronically. This neural mesh is able to follow micro-movements in the brain caused by mechanical movements of the subject or growth, decreasing chances of neural shear damage, and allowing long-term recording (Liu, 2017). The latter innovation has gained significant interest, most notably, Elon Musk's company Neuralink, has also created a similar neural mesh, called neural lace, which will allow human integration with artificial intelligence (Winkler, 2017; Wu and Rao, 2017).

Amidst the variety of biomimetic electrode types, we introduce the concept of nano-architecture in this review as a class of biomimetic surface alteration for IMEs. We use the term nano-architecture to encompass all topographical surface modifications, such as nano-grooves, nano-pillars, nanofibers, and materials with inherent structural components. The inspiration of creating nano-architecture on IME surfaces is based on the architecture of the brain, specifically the extracellular matrix (ECM). The ECM is composed of a 3D and high-aspect ratio architecture (Wu et al., 2006; Millet et al., 2010). The 3D environment allows cells to have topographical cues which will allow them to differentiate and perform their specific functions (Kriparamanan et al., 2006). Several studies have shown that surfaces that can mimic the architecture of the natural in vivo environment will consequently result in an improved biocompatible response (Curtis et al., 2004; Kotov et al., 2009; Ding et al., 2010; Millet et al., 2010; Zervantonakis et al., 2011). Nano-architecture substrates indicate increase in initial protein adsorption, thus leading to subsequent attachment and proliferation of cells (Ereifej et al., 2013b; Nguyen et al., 2016). Alignment of neuronal cells in the brain have also been shown to depend on the roughness and direction of the substrate surface patterns (Khan et al., 2005; Johansson et al., 2006; Ereifej et al., 2013b; Park et al., 2016; Kim et al., 2017). Although the exact mechanism is not completely understood, it is thought that nanoarchitecture is able to indirectly guide the growth and alignment of neurons (Nguyen et al., 2016). Which is beneficial for IME implementation, since enabling neuron growth and proliferation near the implant may allow for improved recording quality. In addition to changes in morphology and protein adhesion, nano-architecture has also been implicated in changes to cell differentiation, phenotype, and gene expression (Kotov et al., 2009; Ereifej et al., 2013a; Yoo et al., 2015; Nguyen et al., 2016; Thompson and Sakiyama-Elbert, 2018).

The goal of the subsequent sections of this review will be to emphasis the role of the architecture with protein and cell interactions, specifically with the central nervous system cells, in order to convey the rationale behind nano-architecture approaches. The biological inspiration of nano-architecture approaches will be explained as well as leading fabrication methods used to create nano-architecture and their limitations. We will then explore the effects of nano-architecture surfaces on neural cells, depicting the various cellular responses to these modified surfaces in both in vitro and pre-clinical models. The proposed mechanism elucidating the ability of nanoarchitectures to influence cellular phenotype will be considered. Finally, the frontiers of next generation nano-architecture IMEs will be identified, with perspective given on the future impact of this interfacing approach.

## THE ROLE OF ARCHITECTURE FOR BRAIN HOMEOSTASIS AND PHYSIOLOGICAL PROCESSES

In order to convey the rationale behind nano-architecture approaches, an understanding of the brain's ECM is crucial. The brain's ECM is made up of components created by the cells within it: neurons, astrocytes, oligodendrocytes, and microglia (Lau et al., 2013). There are three main ECM components, the basement membrane (basal lamina), the perinueonal net, and the neural interstitial matrix (Lau et al., 2013). The basement membrane, which lies around the cerebral vasculature, is composed of laminin, collagen IV, nidogen, and heparin sulfate proteoglycans (also called perlecans). These proteins support the cellular interactions between the brain capillary endothelial cells (BCECs), pericytes, and astrocytes (Thomsen et al., 2017). Collagen IV makes up about 50% of the basement membrane, and plays an essential role for the creation of suprastructures with laminin in the basement membrane (LeBleu et al., 2007). Laminin is the second most common non-collagenous protein

in the basement membrane, and are vital to ensuring proper scaffolding in the ECM (LeBleu et al., 2007). Nidogen is a glycoprotein important in connecting laminins to collagen, and make up 2–3% of the basement membrane (LeBleu et al., 2007). All of the proteins in the basement membrane have been shown to demonstrate a role in the maintenance of the BBB and homeostasis (LeBleu et al., 2007; Farach-Carson et al., 2014; Thomsen et al., 2017). The perineuronal net is a lattice that wraps around neurons and brings dendrites closer to soma of neurons, creating very close synaptic contacts. It is composed of a hyaluronan backbone, chondroitin sulfate proteoglycans, tenascins, and hyaluronan and proteoglycan link proteins (Sorg et al., 2016; van't Spijker and Kwok, 2017). Expressed late in postnatal development, the formation of the perineuronal net signals the maturation and decreased plasticity of the nervous system, while also increasing synaptic stability (McRae and Porter, 2012). Tenascins have been shown to help regulate neuron differentiation and migration, while link proteins stabilize the non-covalent binding of proteoglycans, such as chondroitin sulfate proteoglycan 1, and hyaluronan, allowing for load bearing capabilities and support (Tsai et al., 2014; Oohashi et al., 2015). Finally, the neural interstitial matrix connects neurons and vasculature, composing 15–20% of the total brain (Lei et al., 2017).

The three components of ECM in the brain, including the molecular weight and size of each specific protein are illustrated in **Figure 1**. It is important to note that the architecture of the individual ECM components are on the nanometer scale, a unique characteristic that is crucial to replicate onto neural implants. Nano-architecture etchings onto neural devices is thought to reduce the foreign body response by providing architectural cues for proteins and cells to respond to, that are similar to the natural ECM environment. The ECM structure and components enable cellular interactions by means of protein expression, adhesion, and cellular sensing of the ECM environment. Cells are signaled to release expression factors and non-soluble ECM molecules that are necessary for cell adhesion, proliferation, morphology, and phenotypic changes (Selvakumaran et al., 2008; Jang et al., 2010; Yoo et al., 2015; Schulte et al., 2016a; Yang et al., 2017). Likewise, a nanoarchitecture substrate will allow cells to sense structural cues, which triggers a similar cellular response observed when cells are in their natural environment surrounded by the ECM (Selvakumaran et al., 2008; Yoo et al., 2015; Schulte et al., 2016a; Yang et al., 2017). The nano-architecture surfaces provide cues to initiate the production of ECM molecules necessary for initial cell attachment to surfaces (Cui et al., 2018). Furthermore, the nanoarchitecture surface results in an increased adsorption of those ECM molecules necessary for cell attachment (Salakhutdinov et al., 2008; Ereifej et al., 2013b; Woeppel et al., 2018). These surface-cell topographical interactions initiate the cellular response that can lead to decreased inflammation around implanted IME.

The homeostasis and health of the brain relies on the proper function of all three components of the ECM. Several neurological diseases have been associated with damage to ECM proteins leading to abnormal structural and topographical deviations. Aggregations of misfolded α-synuclein (i.e., lewy bodies) are hallmarks of Parkinson's disease (Lim and Yue, 2015; Yoo et al., 2015). Misfolded aggregates of superoxide dismutase 1 (SOD1) and transactive response DNA binding protein 43 kDa (TDP-43) are characteristic of amyotrophic lateral sclerosis (ALS) (Sahl et al., 2016; Ciryam et al., 2017). Moreover, aggregates of huntingtin (Htt) exon 1 are known hallmarks of Huntington's disease (Sahl et al., 2016; Ciryam et al., 2017). Protein aggregation into inclusion bodies such as Lewy bodies are usually broken down via neuronal autophagy (Heras-Sandoval et al., 2014; Lim and Yue, 2015). However, the neurodegeneration observed following inflammation or trauma due to the implantation of IME may lead to a decrease in the neuronal phagosomes that enable this process to function. The architecture of misfolded proteins leads to aggregation and clumping that would normally not occur in properly folded proteins. Therefore, aberrant 3D structure of the misfolded proteins may lend insight into how proteins and other ECM components can react to a smooth surface compared to a nano-architecture surface.

Several of the hallmark protein aggregations and misfolding involved with the neural diseases described above, have also been thought to play a role in the inflammatory and oxidative stress response following neural electrode implantation. In fact, it has been shown that the implantation of neural electrodes leads to increased expression of proinflammatory and oxidative stress genes (Karumbaiah et al., 2013; Ereifej et al., 2017, 2018; Bennett et al., 2018; Falcone et al., 2018). Not only does increased oxidizing molecules such as nitric oxide (NO) from mitochondria, neurodegeneration, and the other molecules secreted in this pathway play a part in the failure of implanted electrodes, but they may also play a part in the generation of inclusion bodies leading to neuronal death. Thus, it is imperative to control the inflammatory and oxidative stress response, as well as the protein adhesion and conformation around implanted neural electrodes. Perhaps, electrodes with nano-architecture inspired by the healthy ECM neural tissue may play a role in mitigating the inflammatory and oxidative stress response. The hypothesis is that implants with architectures similar to the architectures of the native brain environment, will provide cells an opportunity to maintain their quintessential inactivated phenotypes after electrode implantation. Chronic foreign body response can be controlled, health of neurons can be maintained and high-quality chronic recordings can be enabled. There are limited technologies that can incorporate nano-architecture onto neural electrodes due to the desired size of surface features, the geometry, material, and manufacturing needs of the electrode.

## TECHNOLOGIES TO CREATE NANO-ARCHITECTURE

The addition of nano-architecture onto neural implant surfaces has the ability to improve the biocompatibility as well as induce varying morphology and phenotype of the cells around them. Therefore, fabrication methods that yielded high accuracy and reproducibility to create nano-architectures are imperative. Some fabrication techniques to incorporate nano-architecture

on device surfaces include, focused ion beam (FIB) etching to create nano-grooves, electron beam lithography (EBL) to create nano-lines, and a combination of photolithography and nano-sphere lithography to make nano-structures (Ereifej et al., 2017; Kim et al., 2017; Nissan et al., 2017). Furthermore, there are nano-scale materials that can also be utilized to create nano-architecture, such as CNTs, nano-wires, and bio-inspired materials (Christopherson et al., 2009; Dugan et al., 2010; Fernández and Botella, 2017). The following sections will discuss some of the commonly used fabrication techniques and materials utilized to create nano-architecture.

## Fabrication Methods to Create Controlled Nano-Architecture

Nano-architectural effects on implant interface have been investigated for silicon, titanium, and some polymethyl methacrylate (PMMA)-based electrodes. Many methods can be used to create the nano-architecture on substrates, such as photolithography, optical lithography, soft lithography, nano-sphere lithography, and nano-stencil (Blattler et al., 2006; Kriparamanan et al., 2006; Das et al., 2008; Ding et al., 2010). The benefits of using fabrication methods discussed below, is to create specific patterns, geometries, and sizes that are reproducible. **Table 1** summarizes the characteristics of the fabrication methods discussed below.

#### Electron Beam Lithography

Electron beam lithography is a nano-fabrication technique in which an electron-sensitive resist is selectively exposed to an electrode beam to form a high-resolution pattern. EBL can achieve resolutions less than 10 nm when used in a maskless scanning configuration that uses a highly focused electron beam (Tseng et al., 2003; Dalby et al., 2008; Yang and Leong, 2010). Alternatively, throughput can be increased at the expense of resolution when a diffuse electron source is projected through a thin mask to expose the resist (Tseng et al., 2003). After spincoating the resist to a thickness of 50–500 nm on the substrate surface and exposing to the electron beam, the resist is developed in a solvent to remove unwanted material (**Figure 2A**). The nanoarchitectural pattern can then be transferred to the underlying substrate using standard etching techniques. PMMA is the most commonly used positive resist for EBL, though negative EBL resists are also used (Tseng et al., 2003). Positive resists become more soluble after electron beam exposure, while negative resists form crosslinks after electron beam exposure.

Electron beam lithography is most compatible with patterning silicon or thin metal films that can dissipate the excess charge from the electrode beam (Lee et al., 2007; Dalby et al., 2008; Scholten and Meng, 2016). Insulating films are more challenging to pattern with EBL because electrons that pass through the resist become trapped at the substrate surface, forming variations in resist surface potential. The electrons can then be deflected and result in distorted patterns (Joo et al., 2006). However, researchers have developed methods to enable EBL on insulating glass substrates by controlling the beam energy and on polymer substrates by exposing through a thin conductive metal film (Joo et al., 2006; Scholten and Meng, 2016).

Electron beam lithography requires high-cost instrumentation, and throughput is limited for high-density features due to the serial nature of the exposure and low sensitivity of the resist to the electron beam (Yang and Leong, 2010). An advantage of EBL is its ability to be used to fabricate a single master mold that can be reused several times [e.g., for nano-imprint lithography (NIL)] (Dong et al., 2005a).

#### Nano-Imprint Lithography

Nano-imprint lithography relies on the contrast in a resist thickness produced by mechanical deformation when pressed



with a rigid mold or stamp with a pre-defined relief structure. The mold is typically fabricated by patterning Si or SiO<sup>2</sup> by combining other nano-fabrication tools and techniques, such as EBL, with reactive ion etching techniques. The resist layer is spin-coated onto the sample surface. In contrast to EBL, the mechanical properties of the resist layer are of prime importance as the Young's modulus of the resist must be lower than the Young's modulus of the mold. In hot embossing NIL, thermoplastic resists are imprinted with the mold at temperatures 70–90◦ above T<sup>g</sup> , then cooled before releasing from the mold (**Figure 2B**) (Guo, 2007; Choi et al., 2013). Alternatively, low-viscosity UV-curable resists can be cured by UV light at ambient temperatures while being imprinted with the mold (Schulz et al., 2000; Guo, 2007). After forming the relief pattern in the resist, the underlying substrate can then be etched using standard processes (Al-Abaddi et al., 2012). By offering a variety of resist materials with a range of fabrication conditions, NIL can be used to pattern nano-architectures on a variety of substrate materials, including silicon and polymers (Eom et al., 2015; Baquedano et al., 2017). Resolutions less than 100 nm and even approaching 2 nm have been achieved using NIL (Dong et al., 2005b; Li et al., 2011).

Nano-imprint lithography can be performed on a larger area than EBL or FIB, providing an efficient and cost-effective method of nano-scale patterning at the wafer scale (Chen and Ahmed, 1993; Al-Abaddi et al., 2012). Additionally, the same mold can be used to pattern multiple samples, and therefore is more costeffective than serial nano-patterning methods (Al-Abaddi et al., 2012).

#### Focused Ion Beam Lithography

Focused ion beam nano-machining technology is a direct–write technique for selectively ablating the substrate surface with a finely focused, high-current ion beam (Lehrer et al., 2001; Raffa et al., 2008). FIB can produce a wide variety of features with nano-scale resolution and high-aspect ratio (Watkins et al., 1986; Veerman et al., 1998; Raffa et al., 2008). The magnitude of the ion beam current modulates the ion beam spot size within a range of 3 nm to 2 µm (Lehrer et al., 2001; Raffa et al., 2008). A primary advantage of FIB is that it can be used on a wide variety of materials, including silicon, metals, and polymers (Veerman et al., 1998; Lanyon and Arrigan, 2007; Ziberi et al., 2009; Menard and Ramsey, 2010). Additionally, FIB can be applied on nonplanar surfaces and can be used for post-processing on individual devices (**Figure 2C**) (Reyntjens and Puers, 2001). Users must be cautious, however, as ion implantation and re-deposition of ablated material can lead to damaged nano-structures or structures that do not match the intended geometry (Vermeij et al., 2018). Overall, FIB allows for a high degree of control and flexibility in feature geometries, compatible materials, and surface requirements, at the expense of a slow throughput rate that limits the potential for mass production (Heyderman et al., 2003).

#### Nano-Scale Materials Producing Nano-Architecture

In addition to the pre-defined nano-architecture fabrication techniques described above, nano-scale topography can also be

Frontiers in Neuroscience | www.frontiersin.org

devices post-processing, directly onto the manufactured device (Ereifej et al., 2017).

polymer, heating the polymer to imprint the stamp, and then finally cooling and releasing the stamp from the imprinted polymer. (C) An advantage of focused ion beam lithography (FIB) is that it can produce nano-architecture features on devices of various shapes. FIB can be used to fabricate nano-architecture on neural

applied to biomedical devices by depositing or growing nanoscale materials on a substrate. At the nano-scale, exact geometries cannot be defined to the extent possible using EBL or FIB due to randomness in orientation, size, and positioning. However, at the device scale, feature size, and density can be controlled. Several reviews discuss these materials in great depth, and therefore we will only briefly summarize some of the nanoscale materials (Kotov et al., 2009; Shah, 2016; Wang et al., 2017). CNTs are hollow carbon tubes with a nanometer-scale diameter and appealing electrical, mechanical, and biological characteristics (Wang et al., 2017). Gabay et al. (2005) used ironbased nanoparticles as catalysts for CNT growth by chemical vapor deposition. Neurons demonstrated a strong affinity to grow on the CNT clusters and send out neurites to connect clusters (Gabay et al., 2005). Polycaprolactone was extruded through aluminum oxide membranes to form high-aspect ratio nano-wires for neural tissue engineering applications (Bechara et al., 2010). Silicon nano-wires can be produced by chemical vapor deposition epitaxial growth or by etching, and produce very high-aspect ratio structures oriented perpendicular to the substrate (Kim et al., 2007). Electrospinning can be used to cover a substrate with polymer nano-fibers 200–1500 nm in diameter (Christopherson et al., 2009). Bio-related materials, such as cellulose nano-whiskers of 10–15 nm diameter, have formed nano-topographies by spin-casting onto a substrate (Dugan et al., 2010). Some of the nano-scale materials also offer unique advantages to improve the functional qualities of implanted neural devices, such as reducing neural recording electrode impedance. The nano-scale materials can be used to cover a large area without necessitating specialized serial patterning equipment, thus lending these strategies to larger-scale manufacturing.

## NANO-ARCHITECTURE EFFECT ON NEURAL CELLS (IN VITRO)

Electrode implant performance and success depends heavily on the reaction of the cells within the local and surrounding areas of the brain. While the initial inflammatory response is beneficial for maintaining homeostasis, chronic inflammation may lead to glial scarring, neurodegeneration, and oxidative stress. All of these events may lead to failure of the implant via mechanical breakdown of the implant or reduced to no signal recording (Polikov et al., 2005; Kang et al., 2011; Kozai et al., 2015; Potter-Baker et al., 2015; Bennett et al., 2018; Ereifej et al., 2018). It has been shown that protein adsorption onto the surface of the implant plays a large role in the behavior and response of the cells (Ereifej et al., 2013b; Nguyen et al., 2016). To reiterate, although the exact mechanisms are unknown, cells are able to respond to various architecture geometries within their environment, leading to changes in differentiation, morphology, phenotype, gene expression, signaling molecules, cytokines, and protein production (Frimat et al., 2015; Marcus et al., 2017; Skoog et al., 2017; Eyster and Ma, 2018b). Tanaka and Maeda (1996) showed that substrate surface-cell topographical interactions may influence neural cell response more than the surface chemistry interactions. When microglia cells were seeded on living astrocyte monolayers, fixed astrocyte monolayers, and a glass coverslip, 80% of microglia showed ramification and processes elongation on the fixed and living astrocyte monolayers while almost no microglia showed ramification on the coverslip (Tanaka and Maeda, 1996). Similarly, Chen et al. (2010) showed that 500 nm parallel grooves imprinted onto poly(s-caprolatctone), poly(lactic acid), and poly(dimethylsiloxane) was able to reduce the foreign body response in macrophage behavior, independent of the material's chemistry. The implications of these studies highlight the importance of nano-architecture to enhance positive cellimplant interactions. The ECM of the brain inspires nanoarchitecture approaches, aiming optimize the protein adsorption, and the subsequent response of the cells. Therefore, it is important to explore the effect of nano-architecture on the primary types of cells present in the brain: neurons, astrocytes, and microglia. It is important to note that nano-architecture substrates will elicit varying cellular responses depending on the cell type. This is an important finding for neural electrode implementation, because the goal of optimal neural electrodes is to improve the neural attachment without negatively affecting glial cell activation. The following subsections will discuss the effects of nano-architecture on neural cells.

## Role of Nano-Architecture on Astrocytes

Astrocytes are the most abundant glial cell in the brain, bridging neurons to the vasculature of the CNS. Astrocytes maintain the BBB and homeostasis of the brain via secreted factors that either promote or disrupt barrier development (Herndon et al., 2017). During inflammation, astrocytes regulate the breakdown of the basement membrane to allow infiltrating macrophages and other immune cells access to the effected site (Herndon et al., 2017). In addition to the regulatory and health-sustaining role of astrocytes, it has been shown that neurons grow preferentially according to the track provided by astrocytes, effectively guiding the growth, and alignment of neurons (Sofroniew, 2009; Hsiao et al., 2015). Since the quality of the recorded signal is dependent on the distance of the neuron to the electrode, it may be beneficial to have astrocyte adhesion onto an implant surface, so that neurons will align in favorable positions.

Initial protein adsorption onto the implant surface has been implicated in the adhesion and proliferation patterns of astrocytes. Ereifej et al. (2013b) showed that reduced protein adsorption resulted in a reduction in glial fibrillary acidic protein gene expression from astrocytes cultured onto PMMA nanogrooved surfaces 200 nm deep and either 277 nm wide or 555 nm wide compared to non-patterned surfaces. Additionally, Ereifej et al. (2013b) found an increase in fibronectin and collagen adsorption rate onto the nano-patterned surfaces, suggesting a change of protein conformation, due to the increase of astrocyte adhesion onto the nano-patterned substrates. As protein adsorption is an important initial factor for astrocyte adhesion and proliferation, groups have begun studying the effects of coating proteins onto implant surfaces and studying those effects on astrocytes. Commonly used proteins used in the effort to increase favorable astrocyte-implant interactions are those found naturally in the brain ECM, such as fibronectin

and collagen. These have been used to coat and create nanoarchitecture for increasing astrocytic adhesion and inactivation. Zuidema et al. (2014) used electrospun polylactic acid (PLLA) fibers randomly oriented (2.38 ± 0.46 µm average diameter) and aligned (2.49 ± 0.32 µm average diameter), coated with fibronectin as a substrate to seed astrocytes in order to direct astrocyte migration and extension. This study exhibited the aligned fibers directed astrocytic migration, which is thought to positively modify the neuroprotective properties of glial cells (Zuidema et al., 2014). Chen P. et al. (2017) summarized that astrocytes tend to align in the direction of the aligned fibers and have a rectangular morphology, compared to the circular morphology of astrocytes seeded on randomly oriented fibers. Moreover, Frimat et al. (2015) showed that nanogrooves 108 nm high cut into polydimethylsiloxane (PDMS) are likewise an ideal approach to align and guide astrocytes as they migrate.

An implication of astrocytes cultured on nano-architecture substrates altering their cellular morphology and adhesion is that they may also exhibit altered phenotype to help guide subsequent neuron growth and migration closer to the nanoarchitecture implant. Ereifej et al. (2013a) revealed astrocytes near nano-patterned PDMS substrates (150 ± 2 nm depth, 117 ± 11 nm ridge length, and 170 ± 16 nm groove width) had decreased neuroinflammatory markers and cytokines. Astrocytes in organotypic brain slices, cultured with nanopatterned substrates, aligned along the nano-pattern grooves, thus altering the cell morphology and downstream phenotype (**Figure 3A**). The astrocytes cultured with nano-patterned substrates exhibited decreased expression GFAP, TNFα, and IL-1β, which are important factors in chronic inflammation (Ereifej et al., 2013a). The mechanism behind this phenomenon is discussed in Section "Proposed Mechanism of Cellular Response to Nano-Architecture Surfaces" of this review. As can be seen collectively from these studies, nano-architecture surfaces can be pivotal in reducing the pro-inflammatory cytokines and activated astrocytes occurring after neural electrode implantation.

## Role of Nano-Architecture on Microglia/Macrophages

At the onset of IME insertion, resident microglia activate and the compromised BBB allows for infiltrating macrophages to enter the local damaged tissue (Ravikumar et al., 2014). Microglia and macrophages form the first line of defense against invading pathogens via phagocytosis and release of cytotoxic molecules, cytokines, reactive oxygen intermediates, proteinases, and complement proteins (Thameem Dheen et al., 2007; Li and Barres, 2017). Chronic activation of microglia and macrophages leads to neurodegeneration (Thameem Dheen et al., 2007; Li and Barres, 2017). In response to the detrimental outcomes of chronically activated microglia and macrophages around implanted IME, a consensus desire has grown to reduce the level of activated microglia and macrophages from binding to substrate surfaces. In this venture, nano-architecture has been explored as a way to control activation of microglia and macrophages and decrease inflammatory molecule release.

For example, Luu et al. (2015) found that macrophages cultured on nano-groove titanium surfaces with widths ranging from 400 nm to 5 µm had an elongated morphology aligned with the nano-pattern and trended toward an anti-inflammatory phenotype depicted by significantly higher expression of interleukin 10 (IL-10), and anti-inflammatory cytokine. Notably, Persheyev et al. (2011) observed elongation of BV-2 microglia cells and increased actin-rich microdomains when cultured on substrates with nano-spikes <70 nm, indicating a ramified quintessential phenotype (**Figure 3B**). Additionally, Saino et al. (2011) found that the secretion of proinflammatory molecules from microglia was dependent on the diameter of the PLLA fibers they were cultured on. In fact, PLLA scaffold with nano-fibers (diameters 0.61 ± 0.18 and 0.55 ± 0.16 µm for random and aligned fibers, respectively) reduced the level of proinflammatory molecules compared to the same scaffold with microfibers (diameters 1.53 ± 0.32 and 1.60 ± 0.25 µm for random and aligned fibers, respectively) or a flat PLLA film (Saino et al., 2011). Similarly, Pires et al. (2015) discovered that microglia cultured on electrospun poly(trimethylene carbonateco-ε-caprolactone) fibers with 1.09 ± 0.1 µm diameter exhibited elongated morphology signifying decreased activation, compared to a flat surface of the same chemistry. They additionally showed that when media from the microglia culture was introduced to astrocytes, astrogliosis was not exacerbated, further signifying the microglia were not activated (Pires et al., 2015). Cumulatively, these studies indicate the promise of nanoarchitecture to influence microglia and macrophage phenotype. This observation can be translated to IME implanted into the brain, thus potentially reducing the glial cell activation, chronic inflammation, and neurodegeneration.

### Role of Nano-Architecture on Neurons

Neuron distance from the electrode determines the quality of the recorded signal (Buzsáki, 2004). Chronic inflammation may lead to decreased neuronal density around the implant due to neuronal death (Jorfi et al., 2015). Therefore, it is crucial to guide neurons and keep neuronal density around the implant high. As discussed earlier, initial protein adsorption causes astrocyte adhesion, which neurons grow over according to the tracks laid by astrocytes. Neuron growth cones are able to sense the architecture in the environment due to mechanotransductive components in the cell membrane, which signal to the cytoskeleton (Section Proposed Mechanism of Cellular Response to Nano-Architecture Surfaces explains this mechanism in detail). This mechanotransductive pathway leads to specific growth patterns exhibited when neurons are cultured onto nano-architecture substrates (Nissan et al., 2017; Eyster and Ma, 2018b).

Growth cones are called so because they resemble cones on the tips of extending neurites. Jang et al. (2010) found that when neurons were seeded on a surface with nano-grooves 350 nm wide and 350 nm high, fillopodia at the growth cone tips aligned along the direction of the patterns. Frimat et al. (2015) have explored the response of neurons seeded on a substrate with a nano-patterned surface made of PDMS nano-grooves 108 nm high. They found neurite extensions aligning to the patterns,

nano-grooves. NSC on the nano-architectures exhibited higher ratio of differentiation into dopaminergic and glutamatergic neurons compared to NSC cultured on flat substrates. Scale bar is 50 microns (Yang et al., 2017). Reproduced in part from Yang et al. (2017) with permission of The Royal Society of Chemistry. License number 4332650271766.

as well as alignment of the soma of neurons (Frimat et al., 2015). Additionally, Xie et al. (2016) found that 47% of neurons seeded on a nano-patterned microseive array with nano-grooves 230 nm wide with a period of 600 nm showed alignment along the direction of the grooves. Recently, Woeppel et al. (2018) found that roughening the silicon substrate surface with silica nanoparticles 60 nm in diameter on average, followed by L1 protein adhesion lead to increase neuron outgrowth. Notably, Nissan et al. observed an increase in the number of neurites and branching points toward more complicated structures, when neurons were seeded onto silver nano-lines (180–500 nm wide, 160 nm high, and 700 nm apart) made via EBL. However, they found that the neurons consistently aligned at a ∼45◦ angle to the nano-lines (Nissan et al., 2017). This is thought to be due to how neurons form focal adhesions on the different surfaces. Neural fillopodia extend from the neural cell body and probe around its environment with focal adhesions to see if the environment is suitable for adhesion and growth (Marcus et al., 2017; Eyster and Ma, 2018b). Focal adhesions signal to intracellular cytoskeletal components such as talin and paxillin as a result of surface– protein interactions, which lead to directional growth (Eyster and Ma, 2018b). Due to the relationship between focal adhesions and cytoskeletal components, the organizational structure of focal adhesions may be the reason for neuron cell alignment (Eyster and Ma, 2018b).

Nano-architecture has also been implicated with cell differentiation via changes in gene expression. Yoo et al. (2015) found that seeding fibroblasts on substrates with nano-grooves 300 nm wide and 400 nm apart led to cell alignment and expression of dopamine markers, leading to reprogramming of fibroblasts into functional dopaminergic neurons. The dopamine expression led to gradual acquisition of dopaminergic neuronal characteristics, and then full differentiation into this class of neuron (Yoo et al., 2015). Additionally, Yang et al. (2017) found that seeding neural stem cells onto a conductive nanogroove substrate with groove sizes from 150 to 300 nm, led to alignment and neural guidance via increased focal adhesions and cytoskeletal rearrangements (**Figure 3C**). The neural stem cells then exhibited enhanced differentiation and maturation as elevated levels of neurite extension and neural markers Tuj1 and NeuN were observed (Yang et al., 2017). In comparison, those neural stem cells seeded on a non-patterned or non-conductive coated substrate showed lower levels of neurite extension and neural markers (Yang et al., 2017).

Collectively, the implications of nano-architecture on astrocyte, microglia/macrophage and neuron cells' morphology, phenotype and differentiation may lead to advents of next generation IME design. The studies reviewed in this section suggest nano-architecture has the ability to control glial cell activation by reducing the expression of inflammatory

markers and maintaining cellular morphology to a ramified quintessential phenotype. Remarkably, nano-architechture has been shown to increase neuron adhesion and extension, which may translate to increased neuron density around implanted IME. The combined results from the aforementioned studies advocate the use of nano-architecture to control the cellular response to biomaterials, however, conclusions specifying the exact nano-architecture to elicit a particular cellular response cannot currently be determined. Given that each study utilized a distinctive biomaterial, differing from other studies examining similar geometries of nano-architecture, it is problematic to establish an optimal nano-architecture, unbiased of the biomaterial, for a precise cell function. Thus, the imperativeness for performing studies evaluating the various nano-architectures described within this review, on the same biomaterial, would greatly benefit the field. Identification of nano-architectures to control cellular response is necessary to inform the proper design of next generation IME. The translation of the aforementioned in vitro studies to pre-clinical models has been nominal, but the few studies achieved proved insight for future IME designs.

## NANO-ARCHITECTURE EFFECT ON NEURAL CELLS (IN VIVO)

Originally utilized with tissue engineering applications, nanoarchitecture approaches have been employed with orthopedic and organ replacement technologies to serve as scaffolds that promote cell adhesion and viability (Karazisis et al., 2016, 2017). The potential of nano-architecture to improve biocompatibility and integrate IME in the neural tissue has become increasingly apparent. One of the proposed goals of applying nanoarchitecture onto the surfaces of IME has been to mitigate neuroinflammation. Unfortunately, there has not been substantial translation of the in vitro findings [see Nano-Architecture Effect on Neural Cells (In Vitro)] to pre-clinical models. There is a gap in the literature evaluating the effects nano-architecture with IME on the inflammatory response and recording quality in vivo. **Table 2** highlights selected findings of next generation neural probes utilizing nano-architecture features implanted preclinically.

## Role of Nano-Architecture Mitigating Neuroinflammation in Vivo

He et al. (2006) demonstrated that nano-structured laminin coatings can limit astrocytic encapsulation of implanted neural electrodes at both 1 day and 4 weeks post-implantation in mice. Although these nano-structured probes showed increased microglia/macrophage activation at 1 day post-implantation, significantly decreased activation was seen when compared to the smooth silicon controls at 4 weeks. Another approach was investigated by Ereifej et al. (2017) involving a gallium ion beam etched nano-pattern (200 nm wide, 200 nm deep, and 300 nm apart) on silicon electrodes that significantly increased neuronal viability 100–150 µm from the implant site at 4 weeks post-implantation (**Figure 4A**). Additionally, tissue implanted with the nano-patterned probes had lower expression of proinflammatory genes at 4 weeks compared to the smooth control group (Ereifej et al., 2017). Bérces et al. (2016) demonstrated similar outcomes, observing no chronic differences for astrocytic and microglial activation, but an increase in neuronal viability around silicon electrodes nano-patterned through low-pressure chemical vapor deposition. These electrodes exhibited a nanopillar structure 520–800 nm long with 150–200 nm diameters along the entire surface of the electrode that they found to be more biocompatible than a similar structure on the microscale (**Figure 4B**). Notably, Moxon et al. (2004) demonstrated a 70– 90% nano-porous silicon-based electrode can confer reduced gliosis and increased neuronal viability utilizing an anodic etching method which provides a porous silicon thin film over ceramic electrodes. Testing over chronic recording time points showed the nano-structured coating created no deviation from proper functioning of the electrode nor altering of electrical properties of recording sites. Collectively, these pre-clinical studies indicate that the application of nano-architecture onto the IME surface can reduce glial cell activation and increase neuron density, suggesting potential improvement of recording quality.

## Role of Nano-Architecture for Improving Recording Quality

Nano-architecture on neural electrode surfaces can confer many potential electrophysiological benefits including improved recording quality, greater stimulation efficiency, and the ability to use smaller, less invasive electrodes (Heim et al., 2012). It has been shown inflammation can limit the quality of recordings from neural electrodes due to astrocytic encapsulation and microglial associated oxidative stress that limits the detection of single neurons (Nolta et al., 2015). To combat this inflammatory response, research is aimed at minimizing electrode geometry. Vitale et al. (2015) showed how CNT yarn electrodes utilize their unique microscale properties to improve biocompatibility which allows for microscale recording and stimulating electrode arrays with small contact surface areas with lower impedance than metal electrodes of similar size. They showed 15 times lower impedance resulted from CNT electrode compared to platinum iridium control electrodes. Further, despite not improving signal to noise ratio (SNR), the CNT structure revealed increased biocompatibility, which enabled increased recording quality and longevity. Although minimizing the IME geometry may seem promising, there are challenges that correspond to the decreased contact surface area.

Smaller contact surface area, results in higher noise and increased impedance, as well as lower SNR and recording quality. Therefore, increasing contact site surface area by nanoarchitecture modifications may offer an elegant solution to improve recording signal quality. Examples of nano-architecture modifications include, porous platinum black, golden nanoflakes or -pillars, CNTs, conducting polymers such as polypyrrole and poly(3,4-ethylenedioxythiophene):poly (styrenesulfonate) (PEDOT:PSS) (Venkatraman et al., 2011; Furukawa et al., 2013; Sessolo et al., 2013; Kim et al., 2014; Pas et al., 2018). Abidian et al. (2007) investigated PEDOT coated contacts on

#### TABLE 2 | Highlights of selected studies implanting nano-architecture neural probes.


electrodes implanted into rats that demonstrated significantly lower impedance and higher SNR 7 weeks post-implantation (**Figure 4C**) (Abidian et al., 2007). These benefits of nanopatterned contacts showed by Abidian et al. (2007) met expectations and opened the door for the investigation of other methods to utilize nano-scale alterations and fabrication methods to improve recording contacts. Brüggemann et al. (2011) investigated a gold nano-pillar sturcture (300–400 nm high, 60 nm diameter) to increase contact surface area which lowered impedance and improved extracellular recording performance in vitro. Vallejo-Giraldo et al. (2017) showed that alteration of the nano-pattern on electrodes using anodized indium tin oxide (ITO) can be modulated to find different levels of glial activity, neural cell survival, and promotion of neural network activity in vitro. They concluded that anodized ITO electrodes with nano-structure features can be employed to deposit insulator and charge carrier regions within the same electrode system (**Figure 4D**) (Vallejo-Giraldo et al., 2017). Piret et al. (2015) found that by using a CNT scaffold to grow nano-structured boron-doped diamond coatings on neural electrode contacts, the impedance was lowered and the recording quality in vitro was improved. This structure included another nano-pillar structure similar to other discussed above with 500 nm a diameter of and one micron height (Piret et al., 2015). Cummulatively, the aforementioned studies indicate that the application of nanoarchitecture on electrode contacts shows promise to improve recording quality of neural electrodes.

Overall, these in vivo studies suggest a role for nanoarchitecture to reduce neuroinflammation surrounding implanted neural electrodes and improve rescoring quality. It would be beneficial to see future studies comparing various size and geometries of these nano-architectures to provide insight into the optimal surface for reducing inflammation. Furthermore, it would be remarkable to examine and compare the various fabrication methods to create the desired nanoarchitectures onto neural devices to identify best practices for manufacturing electrodes with nano-architected surfaces.

#### PROPOSED MECHANISM OF CELLULAR RESPONSE TO NANO-ARCHITECTURE SURFACES

Although the exact mechanisms are not well understood, nanoarchitecture is thought to affect the cellular response directly or indirectly via the effects of protein adsorption onto the implant surface (Kriparamanan et al., 2006). It has been shown that nano-architecture effects protein adsorption, with various surface geometries and sizes having different rates, amounts, and conformation of adsorbed protein (Kriparamanan et al., 2006). For example, surfaces with 4 nm height, had lowrandomly oriented protein adsorption, but surfaces with a height ranging from 1 to 2 nm, had very high-protein adsorption (Kriparamanan et al., 2006). Additionally, cell morphology is effected by the nano-architecture of the implant surface, as the actin filaments and focal adhesion structures generally align along the direction of the grooves, depending on the cell type (Schulte et al., 2016a). Following cell seeding onto nanoarchitecture surfaces, actin aggregation was observed, followed by microtubule alignment (Selvakumaran et al., 2008; Yoo et al., 2015; Yang et al., 2017). This showed that one of the first events to occur after seeding is the rearrangement of the cytoskeleton (Selvakumaran et al., 2008; Yoo et al., 2015; Yang et al., 2017). Teixeira et al. (2003) showed that surfaces with

70 nm wide grooves seeded with human corneal epithelial cells aligned themselves according to the direction of the grooves, while the same type of cells seeded in a non-textured surface did not show alignment in any direction. This indicates that nano-architecture effects the alignment of the cells, and thus also effects the cells' morphology, adhesion, and function. It is important to understand the mechanisms of molecular and cellular activities as a response to topographical nanoarchitecture in order to design a successful neural electrodes (Schulte et al., 2016b). The following sections will describe the proposed mechanism of cellular response to nano-architecture surfaces.

## Role of Proteins: Absorption, Conformation, Integrin Signaling, and Sensing

After a neural electrode is implanted into the brain, ECM proteins immediately aggregate and attach to the electrode surface, thus playing an essential role in determining the duration and stability of the implant (Selvakumaran et al., 2008). It has been shown that topographical cues can modify protein absorption and consequently influence cell interactions via modified receptors, which will lead to changes in mechanotransductive signaling (Selvakumaran et al., 2008). These changes may positively affect cell-implant interactions and ultimately improve implant biocompatibility.

Cell adhesion peptides in the ECM are always the first interaction with an implant (Seidlits et al., 2010). Protein absorption initiates cell adhesion, alignment, and outgrowth of neurites. Yang et al. (2013) observed an increase in protein adsorption onto a nano-structured surface (300 nm wide grooves or pillar gap) compared to a smooth surface, explaining that a possible explanation for the increased protein adsorption may be a result of protein unfolding once adsorbed onto the implant surface, exposing more functional groups for subsequent adhesion of cells. This exposure of amino acids was explored by Webster et al. (2001), who found that when vitronectin is adsorbed onto nano-phase alumina, the protein unfolded, leading to more exposed functional groups that could facilitate cell adhesion and growth. It was proposed by Blattler et al. (2006), that uncontrolled, non-specific interactions between biological molecules and the implant are the reason why implants fail. Blattler et al. (2006) further explained that the modes of failure of implants (i.e., immune reactions), are a result of inadequate protein adsorption onto the implant surface.

Middle box: Integrin binding to adhered surface proteins will transmit force across the cell membrane through FAK (blue rectangle) and other adapter proteins (pink, green, brown, and orange). These proteins then bind to actin in the cytoskeleton (dark blue strands) and transmit force to the intermediate filaments (light blue strands) which send the transmitted force to the cell nucleus. Right box: The final morphology of the cell and the cell nucleus are altered, thus effecting the downstream phenotype of the cell.

In addition to protein adsorption on the surface of the implant, integrin proteins residing in a cell's transmembrane are responsible for the cellular response to changes in the ECM via transmitting force to the cell's cytoskeleton, thus playing an essential role in cell-substrate binding (Tate et al., 2004; Ingber, 2010). Filopodial probing is crucial for recognition of topographical surface characteristics. Filopodia are thin and only a few micrometers long; they are protrusive processes made by parallel bundles of filamentous actin (Dalby et al., 2004). Molecular receptors such as integrins and cadherins reside on the tips, which behave as sensors for extracellular environments. The length of filopodium outside of the cell membrane is limited to approximately 5 µm, so the range of topographical nano-patterns sensing is limited (Dalby et al., 2004). Albuschies explained that various contact angles between filopodia and substrate result in different performances in sensing of nano-patterned surfaces, which provide guidance to cell alignment (Albuschies and Vogel, 2013). Integrins behave as a mediator of cell adhesion to regulate cellular activities. Integrin binds to ECM proteins in the process of cell recognition, and transmits force across the cell membrane (Tate et al., 2004; Ingber, 2010). Following which adaptor proteins bind to actin in the cytoskeleton, where forces from actin filaments are transmitted to intermediate filaments (Roca-Cusachs et al., 2012). Intermediate filaments are the only cytoskeletal component that have direct access to the filaments in the nucleus (also known as the nucleoskeleton) (Dalby et al., 2007; Eyster and Ma, 2018a). The thought is that the cytoskeletal mechanical stimulation can lead to the rearrangement of interphase chromosomal DNA through the intermediate filaments, thereby effecting gene expression (Bloom et al., 1996; Dalby et al., 2007). Thus, explaining the interactions between cells and their extracellular environment (i.e., nanoarchitecture implant surface) and the mechanism initiating the signaling pathways for cellular morphology and phenotypic changes (Tate et al., 2004; Roca-Cusachs et al., 2012). **Figure 5** illustrates the aforementioned mechanism of a cell interacting with a nano-architecture surface.

Yang et al. (2013) illustrated that proteins participating in topographical cues can affect signaling of mechanotransduction events by quantitatively analyzing the α5β1 integrin binding to fibronectin-coated nano-patterned substrates with nanoscale shapes, both grooves and pillars. It was found that an increase in focal adhesion correlated with an increased concentration of integrin binding (Yang et al., 2013). Integrin proteins are associated with the strength of the focal adhesions involved in cell-substrate binding, which is directly related to neuron sensitivity to activate signaling pathways of mechanotransduction (Yang et al., 2013; Blumenthal et al., 2014). A later study by Yang et al. (2017) used topography in diverse dimensions to explore focal adhesions and subsequent cell differentiation. Phosphorylated focal adhesion kinase (FAK) gene expression was found to be higher on the smaller nano-scale patterns compared to smooth surfaces (Yang et al., 2017). FAK is a mechanosensitive protein inside the cell and can be activated by integrin binding (Yang et al., 2013). Furthermore, tracking of the α5β1 integrin binding on various substrates showed increased integrin clustering, which was correlated with increased focal adhesion, and contributed to the increase in neuronal density and astrocyte differentiation (Blumenthal et al., 2014). Nano-architecture has also been shown to differentiate

cells through mechansotransductive pathways, specifically the mitogen-activated protein kinase/extracellular signal regulated kinase (MEK-ERK) pathway (Yang et al., 2013, 2017). When the MEK/ERK pathway was blocked, there was an observed reduction of downstream signaling from FAK, resulting in a reduction of cell alignment, focal adhesions, neurite outgrowth, and differentiation (Yang et al., 2013). Furthermore, nanopatterned substrates were found to enhance focal adhesions thus leading to neuronal differentiation into dopaminergic and glutamatergic neurons (Yang et al., 2013, 2017).

Changes in cellular activities have a close relationship with protein expression, which can be controlled through nano-architecture surfaces. For example, proteins involved in the regulation of neuronal cytoskeletal organization were upregulated consistently from cells seeded on a nano-structured substrate (Schulte et al., 2016b). Essential proteins involved with axon and synapse microenvironment, as well as vesicle transport and membrane trafficking were also affected by these nanostructured substrates (Schulte et al., 2016b). Maffioli et al. (2017) described that cell–nano-topography interaction can modify cell function such as calcium signaling and/or homeostasis. The effects of surface nano-architecture on cellular functions can be explained by the morphological changes the cells exhibit while on these surfaces (Maffioli et al., 2017).

### Role of Mechanotransduction

Topographical features are important to neural interfacing in terms of local cells, since mechanotransductive components in cells are able to perceive topographical features in the microenvironment. As a result, cells are able to convert mechanical stimuli information to corresponding physiological signals (Altuntas et al., 2016; Schulte et al., 2016a,b; Shi et al., 2016; Bonisoli et al., 2017; Maffioli et al., 2017). Those physiological signals are capable of eliciting further cellular responses and affecting cell function (Bonisoli et al., 2017). Alterations in protein expression are associated with multiple processes: cell–cell adhesion, glycocalyx and ECM, integrin activation and membrane-F-actin linkage, cell–substrate interaction and integrin adhesion complexes, actomyosin organization/cellular mechanics, and nuclear organization and transcriptional regulation (Yang et al., 2015). All of these processes are closely related to mechanotransductive signaling (Yang et al., 2015). It has been exhibited that nano-patterned surfaces, rather than the material itself can cause dramatic change in protein activities (Schulte et al., 2016a).

Despite the lack of clear mechanisms behind cell–surface interactions, there is a clear relationship between mechanosensitive molecules and the cytoskeleton. Schulte et al. (2016a,b) investigated the mechanism of interfacing of cells and nano-patterned surface by studying supersonic cluster beam deposition (SCBD) of zirconia nanoparticles. Under in vitro conditions, a nanoscopic architecture of the adhesion regions was enforced by cell and nano-patterned interfacing, which had an effect on focal adhesion dynamics and the cytoskeletal organization (Schulte et al., 2016b). This also showed that cell morphology was effected by changes in the cytoskeleton structure (Schulte et al., 2016b; Maffioli et al., 2017). This is thought to influence signaling events and cell behaviors because the shape of the cell's nucleus changes during this process (Ingber, 2010; Maffioli et al., 2017). Furthermore, not only is cell morphology changed, but cell rigidity is also decreased, and mechano-transduction was shown to change transcription factors of neuronal differentiation and protein expression (Schulte et al., 2016b). In fact, it has been shown that activation dynamics of transcription factors was susceptible to mechanical stimuli from topographical features, resulting in cellular protein profile modifications (Schulte et al., 2016a). Several proteins contributing to adhesion and/or cytoskeletal organization lost their functions, which consequently affected neuronal differentiation processes (Schulte et al., 2016a,b).

Living cells usually have a long range of propagation. However, this becomes short-ranged when the actin bundles of the cells are disrupted, or the pre-stress in the actin bundles are inhibited (Wang, 2017). When a force is exerted locally, Src and Rac1 can be directly activated within 300 ms and up to 30–60 µm away (Na et al., 2008; Poh et al., 2009; Wang, 2017). Src (a non-receptor tyrosine kinase protein) and Rac1 (a member of the Rc subfamily within the Rho family of GTP-ases) activation propagates along the plasma membrane along microtubule-dependent mechanisms, causing an elevation of the rigidity of the ECM (Na et al., 2008; Poh et al., 2009; Wang, 2017). The Rho/Rac GTPase signaling pathway is particularly important in cell/nano-feature mechanism studies (Eyster and Ma, 2018a). RhoA has been shown to play a role with controlling cell adhesion (Kaibuchi et al., 1999), cell spreading (Thodeti et al., 2003), cytoskeletal stress fiber formation (Buhl et al., 1995), as well as stem cell differentiation (McBeath et al., 2004). Moreover, mechanotransductive signaling through the cytoplasm is extremely fast, about 40 times faster than chemosignaling, as a result of the pre-stressed fibers which are stiffer than the rest of the cytoplasm, allowing stresses and stimuli to travel along the whole length of the cell quickly (Na et al., 2008; Poh et al., 2009; Wang, 2017). The ability for fast mechanotransductive signaling compared to chemosignaling, even across long distances such as 30–50 µm, may contribute to the importance of surface architecture on the cellular morphology of cells and why architecture may even be more important than surface chemistry.

## Role of Nano-Architecture on Cell Phenotype and Differentiation

The addition of nano-architecture on surfaces has been implicated to alter the phenotype of cells as well as control cell differentiation. Additionally, nano-architecture topographical features can provide guidance to neuronal extension, such as the direction and length of neuron growth, which promote neuronal regeneration (Altuntas et al., 2016; Schulte et al., 2016a; Shi et al., 2016; Bonisoli et al., 2017; Maffioli et al., 2017; Yang et al., 2017). Maffioli et al. (2017) explored mechanosensing/transduction and cell differentiation with multiple surface topographies. By quantitatively analyzing phosphoproteomic data of diverse topographical profiles, it was observed that the dynamic and

complex modulation of the entire signaling network was affected by the cell's interaction with nano-structures, which contributes to distinct cellular behaviors (Maffioli et al., 2017). Schulte et al. (2016a) reported that nano-topographical features that mimic ECM could control the level of maturation of neural networks. By observing neuron morphology with atomic force microscopy, Schulte et al. (2016a) found that neurons grown on nano-patterned surfaces expanded more neurites and accelerate synaptogenesis, which significantly regulates neuronal differentiation and maturation.

Conductive polymers with electrical conductive properties that can give electrical stimulus are commonly used biomaterials involved in nano-patterned topography studies, which is favorable for neuronal differentiation (Ingber, 2010; Altuntas et al., 2016; Shi et al., 2016; Bonisoli et al., 2017). Bonisoli et al. (2017) combined topographical features with a conductive polymer coating to support neuronal growth and differentiation, and axonal guidance. The device conveyed multiple stimuli, both mechanotransductively and electrically, to cells and brain tissues, which effectively promoted neurite growth (Bonisoli et al., 2017). Yang et al. (2017) seeded human neural stem cells (hNSCs) over nano-patterned titanium substrates. They found that cells seeded on the nano-patterned substrate showed alignment and significant focal adhesions, which led to enhanced neuronal differentiation (Yang et al., 2017). Altuntas et al. (2016) extended the application of nanoporous anodized alumina membranes (AAMs) to neural implant coatings. The conductive property of film was achieved by coating AAMs with a thin conducting layer (CAAMs). The conductive AAMs showed that they were favorable for neurite extension and proliferation under electrical stimulation but poor cell adhesion performance (Altuntas et al., 2016). The nano-porous featured AAMs without conducting layer gave topographical cues and thus had excellent neuronal cell adhesion (Altuntas et al., 2016). With nerve growth factor embedded, naked AAMs could achieve similar effect of neurite extension compared to electrically stimulated CAAMs (Altuntas et al., 2016).

While the topographical surface of neural implants provides mechano-transductive cues so that absorbed proteins are modified, the aggregated cells also change their morphology due to local environmental changes (Szarowski et al., 2003). Attachment and clustering of microglia and astrocytes on the implant surface is common in in vivo studies (Szarowski et al., 2003). The additional mechano-tranductive signaling of topography is of great importance in cell–protein interactions, which further changes the cell phenotype and has effects on neuro-inflammation (Szarowski et al., 2003). Neuronal cell differentiation is also known to be affected by topography. McMurray et al. (2011) was able to design a nano-patterned substrate that could control differentiation of mesenchymal stem cells. However, unlike other groups where speeding up or specifying a certain type of cell differentiation was investigated, McMurray et al. (2011) tested the ability of a certain nano-structure to delay differentiation. They were able to identify a nano-structure surface (120 nm pits in a square configuration spaced 300 nm apart, with an offset level near zero) able to delay differentiation of stem cells to remain in their undifferentiated phenotype for over 8 weeks postseeding (McMurray et al., 2011). In this in vitro investigation of adult stem cells differentiation, the nano-structured surface promoted the effect of small RNAs correlated with cell signaling and metabolomics, which manipulated the long-term differentiation of mesenchymal stem cells (McMurray et al., 2011). Christopherson et al. (2009) demonstrated the effect of nano-scale topographical cues on neuronal differentiation and outgrowth processing of human stem cells. The dimensions of nano-patterned geometries enhanced the alignment of neural process outgrowth with the direction of the nano-patterns (Christopherson et al., 2009). The increase in neuronal alignment and outgrowth processing significantly promotes compatibility with implanted devices (Christopherson et al., 2009). Yoo et al. (2015) elucidated that nano-grooved patterns made via UVlithography were able to enhance fibroblast differentiation into dopaminergic neurons, concluding that nano-pattern substrates could serve as an efficient stimuli for cell differentiation. Collectively, the examples provided here utilizing architecture in the nano-scale, give evidence that nano-architecture have an effect on the phenotype, morphology, and differentiation of cells.

## CONCLUSION AND FUTURE PERSPECTIVES

Integration of neural electrodes into the brain tissue lies heavily on reducing the neuroinflammatory response. An understanding of the natural environment and how cells interact and communicate with each other is crucial when designing next generation neural electrodes. A growing body of literature is investigating the effect surface architectures have on controlling cell behavior, differentiation, and phenotype. Here, we reviewed only the nano-scale architectures, as nanoscale surface modifications have shown promise in controlling protein adsorption, reducing glial cell inflammatory markers, guiding axonal direction, and cell differentiation. Nano-scale architectures are inspired by the native in vivo environment, specifically the ECM which cells receive their signals and cues from. Advancements in fabrication techniques and novel biomaterials have allowed the addition of nano-scale features to be added to neural implants within the manufacturing process or even post-processing. For example, altering the architecture of commercially used microelectrodes, such as the Michigan electrode, utilizing FIB lithography, allows seamless translation to research labs and potentially patients. Additionally, the fabrication processes to create nano-architectures can be done on numerous materials. The feature sizes and shapes can be limitless with the multitude of fabrication methods. Nanoarchitectures can include multiple features on various parts of the neural electrodes depending on the desired outcome. Features can be integrated onto the contact sites to reduce impedance and increase recording quality as well as around the electrode insulating layer to reduce the foreign body response. The nanoscale features on the electrode's insulating layer help to reduce

the foreign body response by controlling the protein adsorption, conformation, integrin signaling thereby influencing the cellular morphology, and downstream phenotype. The promise of incorporating nano-architechture on the surface of neural implants has been implicated by the countless aforementioned examples. Nano-architecture can be utilized to control cell phenotype, differentiation, growth, adhesion, migration, and morphology.

This approach can be utilized on various neural implants, including stimulating electrodes, DBS probes, closed loop sensors. Unfortunately, there is a lack of preclinical studies evaluating abovementioned in vitro studies presented throughout this review. Given that the methods to incorporate nanoarchitecture onto neural electrodes are feasible, we predict the gap in the literature evaluating this approach will be filled in the upcoming years. Although a proposed mechanism of how nano-architectures can control cellular response was reviewed, it is crucial to consider diverse features will elicit different cellular responses. Thus, a thorough understanding of cellular responses to specific nano-architectures will inform the design and development of improved neural prosthetic and neuromodulatory devices. The implementation of nanoarchitectures onto these devices is hypothesized to reduce the foreign body response and create seamless device tissue integration. The ability for nano-architectured implants to more closely mimic the brain's architecture is an interesting avenue of discovery and research, because it allows for the future development of the optimal implant that integrates seamlessly with CNS tissue.

#### REFERENCES


## AUTHOR CONTRIBUTIONS

YK and EE contributed substantially to the conception and design of the work, drafting and revising the manuscript for important intellectual content, approved the final version to be published, and agree to be accountable for all aspects of the work. SM, KC, HF, JR, and AH-D drafted corresponding sections of the manuscript. All authors (EE, YK, SM, KC, HF, JR, and A-HD) approved the final version to be published and agree to be accountable for all aspects of the work.

## FUNDING

This study was supported by United States (US) Department of Veterans Affairs Rehabilitation Research and Development Service Career Development Awards #RX001664-01A1 (CDA-1, EE), #RX002628-01A1 (CDA-2, EE), and #RX001841-04 (CDA-2, AH-D) from the United States (US) Department of Veterans Affairs Rehabilitation Research and Development Service. The contents do not represent the views of the U.S. Department of Veterans Affairs or the United States Government.

### ACKNOWLEDGMENTS

The authors acknowledge Erika Woodrum of the Cleveland Functional Electrical Stimulation (FES) Center for exceptional artistry of **Figures 1**, **2C,** and **5**.





**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Kim, Meade, Chen, Feng, Rayyan, Hess-Dunning and Ereifej. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

1001

Funct. Mater. 28:1700905

doi: 10.1088/1361-6463/aa6e18

Eng. 7, 291–301. doi: 10.1089/10763270152044152

fnins-12-00456 July 13, 2018 Time: 16:9 # 20

Tsai, H.-L., Chiu, W.-T., Fang, C.-L., Hwang, S.-M., Renshaw, P. F., and Lai, W.- F. T. (2014). Different forms of tenascin-C with tenascin-R regulate neural differentiation in bone marrow-derived human mesenchymal stem cells. Tissue

Tseng, A. A., Chen, K., Chen, C. D., and Ma, K. J. (2003). Electron beam lithography in nanoscale fabrication: recent development. IEEE Trans. Electr.

Tybrandt, K., Khodagholy, D., Dielacher, B., Stauffer, F., Renz, A. F., Buzsáki, G., et al. (2018). High-Density stretchable electrode grids for chronic neural

van't Spijker, H. M., and Kwok, J. C. (2017). A sweet talk: the molecular systems of perineuronal nets in controlling neuronal communication. Front. Integr.

Veerman, J., Otter, A., Kuipers, L., and Van Hulst, N. (1998). High definition aperture probes for near-field optical microscopy fabricated by focused ion beam milling. Appl. Phys. Lett. 72, 3115–3117. doi: 10.1063/1.121564 Venkatraman, S., Hendricks, J., King, Z. A., Sereno, A. J., Richardson-Burns, S., Martin, D., et al. (2011). In vitro and in vivo evaluation of PEDOT microelectrodes for neural stimulation and recording. IEEE Trans. Neural Syst.

Vermeij, T., Plancher, E., and Tasan, C. (2018). Preventing damage and redeposition during focused ion beam milling: the "umbrella" method.

Wang, M., Mi, G., Shi, D., Bassous, N., Hickey, D., and Webster, T. J. (2017). Nanotechnology and nanomaterials for improving neural interfaces. Adv.

Wang, N. (2017). Review of cellular mechanotransduction. J. Phys. D 50:233002.

Watkins, R., Rockett, P., Thoms, S., Clampitt, R., and Syms, R. (1986). Focused ion beam milling. Vacuum 36, 961–967. doi: 10.1016/0042-207X(86)90148-X Webster, T. J., Schadler, L. S., Siegel, R. W., and Bizios, R. (2001). Mechanisms of enhanced osteoblast adhesion on nanophase alumina involve vitronectin. Tissue

Wei, X., Luan, L., Zhao, Z., Li, X., Zhu, H., Potnis, O., et al. (2018). Nanofabricated ultraflexible electrode arrays for high-density intracortical recording. Adv. Sci.

Eng. Part A 20, 1908–1921. doi: 10.1089/ten.TEA.2013.0188

Pack. Manufact. 26, 141–149. doi: 10.1109/TEPM.2003.817714

recording. Adv. Mater. 30:e1706520 doi: 10.1002/adma.201706520 Vallejo-Giraldo, C., Pampaloni, N. P., Pallipurath, A. R., Mokarian-Tabari, P., O'connell, J., Holmes, J. D., et al. (2017). Preparation of cytocompatible ITO neuroelectrodes with enhanced electrochemical characteristics using a facile

anodic oxidation process. Adv. Funct. Mater. 28:1605035

Rehabil. Eng. 19, 307–316. doi: 10.1109/TNSRE.2011.2109399

Ultramicroscopy 186, 35–41. doi: 10.1016/j.ultramic.2017.12.012 Vitale, F., Summerson, S. R., Aazhang, B., Kemere, C., and Pasquali, M. (2015). Neural stimulation and recording with bidirectional, soft carbon nanotube fiber microelectrodes. ACS Nano 9, 4465–4474. doi: 10.1021/acsnano.5b01060 Voge, C. M., and Stegemann, J. P. (2011). Carbon nanotubes in neural interfacing applications. J. Neural Eng. 8:011001. doi: 10.1088/1741-2560/8/1/01

Neurosci. 11:33. doi: 10.3389/fnint.2017.00033

# Nanoparticle-Based Systems for Delivery of Protein Therapeutics to the Spinal Cord

#### Juan C. Infante\*

Department of Molecular and Cellular Biology, Harvard College, Harvard University, Cambridge, MA, United States

Recent studies have demonstrated that delivery of protein therapeutics to the spinal cord may promote functional axon regeneration, providing a pathway for recovery of certain motor skills. The timeframe for delivery of protein therapeutics, however, must be modulated to prevent bulk release of the therapeutics and minimize the frequency of implantations. This perspective examines both affinity-based and nanoparticle-based strategies for delivery of neurotrophic factors (NFs) to the spinal cord in an effective, safe, and tunable manner.

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Mihai Moldovan, University of Copenhagen, Denmark Hari S. Sharma, Uppsala University, Sweden Ruxandra Vidu, University of California, Davis, United States

#### \*Correspondence:

Juan C. Infante juaninfantesamper@gmail.com

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 07 April 2018 Accepted: 26 June 2018 Published: 19 July 2018

#### Citation:

Infante JC (2018) Nanoparticle-Based Systems for Delivery of Protein Therapeutics to the Spinal Cord. Front. Neurosci. 12:484. doi: 10.3389/fnins.2018.00484 Keywords: protein therapeutics, drug delivery systems, nanoparticles, spinal cord injuries, FGF1, Liraglutide

## INTRODUCTION

Spinal cord injuries (SCIs) often have debilitating effects on patients and can severely impact quality of life. Each year there are about 20,000 new injuries and currently there may be up to 350,000 SCI patients in the United States (Spinal Cord Injury (SCI) Facts and Figures at a Glance, 2016). Finding a treatment for these injuries is therefore of utmost importance and axon regeneration may prove to be one of the most promising alternatives.

Several groups have shown that inducing axon regeneration is in fact possible. A 2010 study found that following peripheral nerve injury, mice that were locally treated with insulin-like growth factor 1 (IGF-1) displayed significant gains in axon number and diameter (Apel et al., 2010). Other studies have also found that IGF-1 can enhance axon growth in cultured corticospinal motor neurons (Ozdinler and Macklis, 2006; He, 2010). In fact, it is thought that IGF-1 plays a crucial role in CNS development since double-knockout IGF-1 mice display problems in the creation of functional nerve connections (Ozdinler and Macklis, 2006). IGF-1 and other protein therapeutics are therefore promising options for the treatment of SCIs.

However, in order for axon regeneration to be clinically meaningful in the context of paralysis, it must be accompanied by at least some functional recovery of motor skills. Most studies up to date have been unable to induce significant elongation in severed or injured CNS motor axons (Axon Regeneration in the Central Nervous System, 2016). Since these axons are usually only able grow a few millimeters, functional recovery is difficult to achieve. A 2015 study found that coexpression of IGF1 and osteopontin (OPN) could promote axon regeneration in retinal ganglion cells (Duan et al., 2015; He and Jin, 2016). It has also been shown that overexpression of fibroblast growth factor 1 (FGF-1) in the spinal cord improves functional recovery after SCI, as measured by the Basso-Beattie-Bresnahan locomotion scale (Li et al., 2018). Liraglutide, a glucagon-like peptide 1 analog that is already an approved treatment for type 2 diabetes, has similarly been implicated in functional recovery (Chen et al., 2017). FGF-2, brain-derived neurotrophic factor (BDNF), and

neurotrophin-3 (NT-3) have also been shown to enhance motor neuron outgrowth following injury (Boyce and Mendell, 2014; Santos et al., 2016). In fact, some research suggests that synergistic administration of NFs may be most beneficial (Logan and Ahmed, 2006). Thus, the question becomes whether such protein therapeutics can be reliably delivered to the spinal cord. Studies have successfully delivered protein therapeutics, and seen associated functional improvements, through implantation of adeno-associated virus (AAV) vectors into the spinal cord (Li et al., 2018). These findings are very promising in terms of their potential applications to patients in the short-term future. However, constant AAV implantations as a form of human treatment seem to be impractical. This perspective thus examines both affinity-based and nanoparticle-based strategies for delivery of neurotrophic factors (NFs) to the spinal cord in an effective, safe, and tunable manner.

## AFFINITY-BASED SYSTEMS

There are two different approaches for delivering FGF-1, Liraglutide, and other NFs in general. The first one involves an affinity-based delivery system (**Figure 1**). This system consists of a polymeric matrix that has a ligand for our therapeutic of interest immobilized on the matrix. The rate of release of the therapeutic can therefore be controlled by modulating the dissociation kinetics of our protein therapeutic and its matrix-immobilized ligand (Vulic and Shoichet, 2014). Several heparin-based delivery systems (in which heparin is the ligand) have been studied to date. FGF-2 release from heparin-gelatin-PEGDA gels occurred over a 42-day timeframe, but this matrix is not biocompatible (Vulic and Shoichet, 2014). Heparin-based biocompatible matrices have achieved delivery of FGF-2 over comparable timeframes (Vulic and Shoichet, 2014). Similarly, heparin-based release of NT-3 and BDNF has been sustained for 14 and 7 days, respectively (Vulic and Shoichet, 2014). However, affinity-based systems exhibit certain limitations: mainly that protein therapeutics may be inadvertently immobilized or that ligand-protein binding may sterically hinder release of the therapeutic (Vulic and Shoichet, 2014). Similarly, the fact that practical ligands, such as heparin, can bind multiple therapeutics, or that some therapeutics such as IGF-1 must first bind binding proteins before binding to heparin, introduces too many layers of control into these systems and adds unnecessary complexity to the modulation of dissociation kinetics. Therefore, it is useful to explore the benefits that nanoparticle-based delivery systems provide.

## NANOPARTICLE-BASED SYSTEMS

The second approach involves a nanoparticle delivery system which essentially modulates release of the protein therapeutics by degradation of the particle matrix. The proteins are encapsulated in the nanoparticles and, after an initial period of rapid release, are slowly released as the particle matrix progressively degrades (Vulic and Shoichet, 2014). However, issues surrounding this approach include the use of organic solvents, such as chloroform, that are not only toxic but may also denature the therapeutic of interest (Vulic and Shoichet, 2014). In addition, usual nanoparticle preparation protocols involve extensive use of sonication which may lead to protein aggregation and thus damaging of a protein's bioactivity (Stathopulos et al., 2004; Vulic and Shoichet, 2014). However, nanoparticle-based systems may be able to sustain protein release for more than two weeks, despite the rapid release period in the first 24 h (Swed et al., 2014). This release-profile appears promising for translational studies since the goal is to reduce the frequency of injections and implantations as much as possible. The following section therefore proposes, in detail, future experimental directions for designing non-toxic nanoparticle delivery systems that may modulate the slow release of NFs without compromising their bioactivity.

### Experimental Strategy

The first challenge when designing nanoparticle systems involves ensuring that the therapeutics of interest are successfully encapsulated. Initial investigations could prepare basic water-inoil micelles to gauge the efficacy of protein encapsulation through

FIGURE 1 | Schematic of principles behind Affinity-Based Release Systems (Vulic and Shoichet, 2014) – Published under ACS Author Choice License, which permits copying and redistribution of the figure for non-commercial purposes.

standard procedures. A fluorescent protein, GFP for instance, could be encapsulated inside a micelle using chloroform, even though this organic solvent could never be used in vivo due to its toxicity. Since GFP is soluble in water, one would expect to observe fluorescence in the aqueous layer only if the protein was successfully encapsulated in the micelle. Successful encapsulation could suggest that hydrophilic proteins may be stably enclosed in hydrogels and other similar systems. We were able to conduct this experiment and will discuss the results in the following sections.

Further investigations could involve the preparation of poly-lactic-co-glycolic acid (PLGA), a biodegradable material, nanoparticles using a non-separation method with nontoxic, and non-denaturing, organic solvents (Swed et al., 2014). Glycofurol (GF) and isorbide dimethyl eher (DMI), presumably non-toxic and non-denaturing solvents, may be independently used to determine whether one of them leads to more effective encapsulation of hydrophilic fluorescent proteins of similar weights to the NF of interest (outline in Appendix) (Tran et al., 2012; Swed et al., 2014). It is crucial to use proteins that have similar weights since this can have a significant effect on the diffusion rate of the therapeutic (Swed et al., 2014). The size differences between FGF-1 and Liraglutide, for instance, could pose a potential problem when attempting to co-deliver these proteins as therapeutics, since the MW of FGF-1 is about 17 kDa while that of Liraglutide is about 4 kDa. It would not be as ideal to have different release rates for FGF-1 and Liraglutide, although this could perhaps be attenuated by using some of the principles in affinity-based systems (perhaps by using heparin or some other ligands) to modulate release rates. The main feature of this preparation method is that sonication is not needed and the risk for protein aggregation can therefore be minimized (Tran et al., 2012).

After encapsulation of the therapeutic of interest, it is necessary to verify its structural integrity to ensure that no major denaturation took place. Protein samples released within the first 24 h can be collected in a phosphate buffer (would not degrade the protein) and compared to wild-type protein through circular dichroism (CD) spectroscopy (Swed et al., 2014). The release profile of the protein therapeutic can then be estimated by measuring fluorescence intensity of protein samples in the phosphate buffer at different time-points (each day until no more apparent changes in intensity can be detected).

## Micelle Results

No fluorescence was detected in the aqueous layer. This result can be explained by multiple factors, ranging from the use of chloroform as the organic solvent to the need for sonication in the protocol. As previously mentioned, certain organic solvents may cause proteins to denature. It is possible that chloroform caused GFP to denature, which could possibly prevent the protein from fluorescing. This does not necessarily mean that encapsulation was unsuccessful, although it would not be very useful to deliver a denatured protein that has lost its bioactivity. Alternatively, it is also possible that the lack of fluorescence was due to a failure in encapsulation. Another possibility is that aggregation of GFP led to quenching of the fluorescence signal, since, as mentioned before, sonication may lead to protein aggregation (Stathopulos et al., 2004). Although our experiments with these basic micelles proved to be unsuccessful, it would be interesting to determine whether the use of non-toxic, non-denaturing procedures would yield positive results.

### PLGA Expected Results CD Spectroscopy

Since the protein is expected to be successfully encapsulated in the PLGA nanoparticle without being denatured, the CD spectra of the released protein (found in the phosphate buffer) and the wild-type protein should be identical (**Figure 2**) (Swed et al., 2014). This would indicate that the structure of the protein did not change during encapsulation and that the protein likely retained its bioactivity.

#### Fluorescence Analysis

When analyzing the protein samples in the phosphate buffer, fluorescence intensity should gradually increase each day until no more changes can be detected. At this point, it is reasonable to assume that the majority, if not all the protein, has already been released from the nanoparticle. The rate of change of the fluorescence intensity can then be used to estimate the rate of release of the protein therapeutic.

## DISCUSSION

Recent literature suggests that nanoparticle delivery systems may be quite effective for patient applications. For our purpose, the blood brain barrier (BBB) poses an extra difficulty. The nanoparticles must not only be non-toxic and able to modulate the slow release of the protein-therapeutic, but must also be small enough to cross the BBB. PLGA nanoparticles may be the solution to all three problems. First, PLGA is particularly useful because it is a biodegradable material, which addresses toxicity concerns. Second, previous studies have been able to prepare PLGA nanoparticles for drug delivery in the 200 nm range, which is required for the particles to cross the BBB (Tamilselvan et al., 2014). Finally, some PLGA-based delivery systems have sustained release of therapeutics for up to 30 days, providing a reasonable timeframe for translation to patients; a timeframe which may further improve once regulation is fine-tuned (Vulic and Shoichet, 2014). Nanoparticle delivery methods prove to be a promising alternative for delivering protein therapeutics that may promote functional axon regeneration. If the discussed nanoparticle systems can be effectively used in a mouse model, clinical trials may be coming soon.

## AUTHOR CONTRIBUTIONS

The author confirms being the sole contributor of this work and approved it for publication.

## ACKNOWLEDGMENTS

fnins-12-00484 July 17, 2018 Time: 16:7 # 4

This work was conducted as part of the Spring 2017, MCB 91r course series at Harvard College, under the supervision of

## REFERENCES


Dr. Zhigang He (Boston Children's Hospital). The author is entirely responsible for the content, which does not necessarily represent the views of Harvard College, Dr. He, or Boston Children's Hospital.


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Infante. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

## APPENDIX

## – PLGA PROTOCOL

fnins-12-00484 July 17, 2018 Time: 16:7 # 5

**Adapted from (Tran et al., 2012; Swed et al., 2014).**

## **Non-Toxic PLGA Protein Encapsulation**

#### **Materials**


#### **Procedure 1 (2 weeks)**

	- A total of 10 µg fluorescent protein dissolved in 2M NaCl solution containing 15% w/v Lutrol F68.
	- Mixture of 120 µl GF and 75 µl DMI were added to 5 µl of Tris–HCL 0.05M? first solution was then added to this.
	- Prepare 12% w/v solution of PLGA in GF.
	- Leave mixture under stirring for 48 h and then leave to stand for 7 days at RT? You should observe a macroscopic evolution of polymer solution.
	- Before any further use, filter solution through a 0.2-µl pore size filter.
	- First, add 100 µl of protein precipitates into 200 µl of polymer solution to obtain a suspension of protein precipitates in polymer solution.
	- The 50–150 µl of ethanol were added into the mixture, right before 0.9–1.8 ml of 1% F-68 solution was added into the mixture to start the phase separation (Solution X).
	- A total of 15 ml of 1% Lutrol solution in glycin buffer 1.25 mM pH (9–11) introduced (Solution Y).
	- After 15 min, 15–25 ml of Solution Y added into the suspension and this suspension was left to stand for 12–24 h? At the end, the pH should be around 7.
	- Successively centrifuge at 1000 × g for 30 min and then 2000 × g for 30 min? Eliminate the supernatant? Final volume should be about 1 ml.
	- Freeze-dry the suspension

#### **Procedure 2 (2 weeks)**

	- Fluorescent protein dissolved in 0.16M NaCl solution to obtain concentration of 20 mg/ml. A total of 25 µl of this solution was then added to 975 µl of glycofurol (GF).

#### **Analyzing Structural Integrity**


#### **Estimating Rate of Release**

• Track fluorescence intensity each day. Fluorescence intensity should gradually increase each day until no more changes can be detected.

# An in Vivo Mouse Model to Investigate the Effect of Local Anesthetic Nanomedicines on Axonal Conduction and Excitability

Mihai Moldovan1,2 \*, Susana Alvarez<sup>1</sup> , Christian Rothe<sup>3</sup> , Thomas L. Andresen<sup>4</sup> , Andrew Urquhart<sup>4</sup> , Kai H. W. Lange<sup>3</sup> and Christian Krarup1,2

<sup>1</sup> Department of Neuroscience, University of Copenhagen, Copenhagen, Denmark, <sup>2</sup> Department of Clinical Neurophysiology, Rigshospitalet, Copenhagen, Denmark, <sup>3</sup> Department of Anesthesia, Nordsjællands Hospital, Hillerød, Denmark, <sup>4</sup> Department for Micro- and Nanotechnology, Technical University of Denmark, Lyngby, Denmark

Peripheral nerve blocks (PNBs) using local anesthetic (LA) are superior to systemic analgesia for management of post-operative pain. An insufficiently short PNB duration following single-shot LA can be optimized by development of extended release formulations among which liposomes have been shown to be the least toxic. In vivo rodent models for PNB have focused primarily on assessing behavioral responses following LA. In a previous study in human volunteers, we found that it is feasible to monitor the effect of LA in vivo by combining conventional conduction studies with nerve excitability studies. Here, we aimed to develop a mouse model where the same neurophysiological techniques can be used to investigate liposomal formulations of LA in vivo. To challenge the validity of the model, we tested the motor PNB following an unilamellar liposomal formulation, filled with the intermediate-duration LA lidocaine. Experiments were carried out in adult transgenic mice with fluorescent axons and with fluorescent tagged liposomes to allow in vivo imaging by probe-based confocal laser endomicroscopy. Recovery of conduction following LA injection at the ankle was monitored by stimulation of the tibial nerve fibers at the sciatic notch and recording of the plantar compound motor action potential (CMAP). We detected a delayed recovery in CMAP amplitude following liposomal lidocaine, without detrimental systemic effects. Furthermore, CMAP threshold-tracking studies of the distal tibial nerve showed that the increased rheobase was associated with a sequence of excitability changes similar to those found following non-encapsulated lidocaine PNB in humans, further supporting the translational value of the model.

Keywords: liposomes, peripheral nerve block, threshold-tracking, in vivo imaging, lidocaine

## INTRODUCTION

Adequate pain management has been shown to improve the rate and quality of patient recovery following surgery (Zaslansky et al., 2015). Peripheral nerve blocks (PNBs) using local anesthetic (LA) are superior to opioid analgesia which have a high rate of undesirable systemic effects (Wu and Raja, 2011; Aguirre et al., 2012). Voltage-gated Na+ channel (VGSC) blockers acting as LA (e.g.,

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Angelika Lampert, Uniklinik RWTH Aachen, Germany Mario Valentino, University of Malta, Malta

> \*Correspondence: Mihai Moldovan moldovan@sund.ku.dk

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

> Received: 20 April 2018 Accepted: 02 July 2018 Published: 26 July 2018

#### Citation:

Moldovan M, Alvarez S, Rothe C, Andresen TL, Urquhart A, Lange KHW and Krarup C (2018) An in Vivo Mouse Model to Investigate the Effect of Local Anesthetic Nanomedicines on Axonal Conduction and Excitability. Front. Neurosci. 12:494. doi: 10.3389/fnins.2018.00494

**193**

bupivacaine, lidocaine, etc.) are the most widely used drugs in PNBs. Single PNB only lasts several hours (Wu and Raja, 2011). Catheter-based PNBs using LA infusions can extend analgesia (Madsen et al., 2018); however, their efficiency is limited by mechanical factors and can expose patients to large quantities of LA (Ilfeld, 2017). Alternatively, several LA carriers with extended release have been developed, among which liposomes have been shown to be the least toxic (Weiniger et al., 2012). A multilamellar liposome formulation of bupivacaine (Exparel) has so far failed to show PNB of major nerves (Hadzic et al., 2016) and remains of limited indication (Vyas et al., 2016). Further experimental work is required to understand the in vivo effects of liposomal LA nanomedicines.

In vivo rodent models for PNB have focused primarily on assessing behavioral responses following LA such as the motor performance on an inverted mesh (Leszczynska and Kau, 1992), the tail flick latency to a thermal stimulus (Grant et al., 1993), or the vocalization responses to an electrical shock (Grant et al., 2000). Such sensory-motor behavioral techniques were successfully used to investigate the protracted effect of liposomal LA (Epstein-Barash et al., 2009), and nevertheless, the effect on axonal function remains poorly investigated. In a previous study in human volunteers, we found that it is feasible to monitor the effect of LA in vivo by combining conventional conduction studies with nerve excitability studies by thresholdtracking (Moldovan et al., 2014). Here, we aimed to develop a translational mouse model where the same neurophysiological techniques can be used to investigate liposomal formulations of LA in vivo. To challenge the validity of the model, we tested an unilamellar liposomal formulation, consisting of a single lipid bilayer (Betageri and Parsons, 1992) which has faster release profile than multilamellar liposomes (Silva et al., 2016). The liposomes were filled with the LA lidocaine which has a shorter duration of action than bupivacaine in extended-release formulations (Huynh et al., 2015).

## MATERIALS AND METHODS

## Mice, Anesthesia, and Experimental Design

Investigations in vivo were carried out in adult mice of C57Bl background (3- to 5-month-old) with axonal expression of the Yellow Fluorescent Protein (YFP) reporter (homozygote Thy1- YFP) obtained from the Jackson Laboratory, Bar Harbor, ME, United States. General anesthesia was ensured using a 1:1 mixture of Hypnorm/Midazolam (5 mg/ml). A volume of 0.1 ml/10 g from the mixture was injected subcutaneously for induction, and then maintained with 50% hourly for up to 4 h as needed.

Using in vivo neurophysiological recordings (**Figure 1**), we monitored the recovery of tibial nerve conduction (**Figure 2**) and excitability (**Figure 3**) following injection of the LA. Slow subcutaneous injection of a volume of 50 µl at the ankle engulfed the tibial nerve in an easily observable bubble (**Figure 1C**) ensuring the reproducibility of the LA exposure without the need of ultrasound visualization as in the case of human PNBs (Moldovan et al., 2014). The monitoring of the systemic toxicity was carried out using a 1-channel electrocardiogram (ECG) recorded between the forepaws (**Figure 1B**) (MP150 with an ECG100C module, Biopac Systems Inc., United States).

At the completion of the electrophysiological experiments, the lateral aspect of the leg was exposed for imaging studies (**Figure 4**) after which the mice were killed by cervical dislocation.

This study was carried out in accordance with the recommendations of directive 2010/63/EU of the European parliament and of the council on the protection of animals used for scientific purposes. The protocol was approved by the Danish Animal Experiments Inspectorate.

## Electrophysiological Setup

For in vivo electrophysiological investigations, the mice were placed on a temperature-controlled pad (HB 101/2, LSI Letica) set to 37◦C (Moldovan and Krarup, 2006). Stimulation and recording were carried out using percutaneously inserted Ptneedle electrodes (Moldovan and Krarup, 2006; Wild et al., 2018). Electrical stimuli generated from a constant current stimulator (DS4, Digitimer Ltd.) were delivered proximally to the sciatic nerve at the sciatic notch (**Figures 1A,B**) and distal to the tibial nerve at the ankle (**Figures 1A,C**). The evoked compound muscle potential (CMAP) of the plantar muscle was recorded using a bandpass filter of 10 Hz– 6 kHz (Neurolog NL820A amplifier with NL844 Pre-Amplifier, Digitimer Ltd., United Kingdom). A ground reference electrode was inserted into the back of the mouse. The near-nerve electrode placement was carried out ensuring the lowest threshold current.

The signals were digitized into a PC-based recording and control system via a multifunction I/O device (PCI-6221, National Instruments Corporation Ltd., United Kingdom). The amplitudes of the CMAP evoked from the proximal (**Figure 1D**) and distal (**Figure 1E**) stimulation sites were measured peak-topeak (ppA). The CMAP latency was measured to the first take-off from baseline (**Figures 1D,E**).

It should be noted that both from the proximal stimulation site (**Figure 1D**) as well as from the distal stimulation site (**Figure 1E**), the evoked CMAP was followed by a secondary motor response ("F-wave") resulting from motoneuronal backfiring following antidromic activation in humans (Magladery and Mcdougal, 1950) and rodents (Meinck, 1976; Robertson et al., 1993; Moldovan et al., 2012). Although the F-waves were not specifically quantified, their presence was considered as an indication of the functional integrity of the motoneurons and the alpha-motor fibers.

## Nerve Conduction Studies

We have previously established a setup for the continuous monitoring of tibial nerve conduction by stimulation at the ankle for up to 4 h (Moldovan and Krarup, 2006; Alvarez et al., 2008). Nevertheless, overcoming the partial voltagegate Na+ channel block by lidocaine could require very large stimulation currents (Moldovan et al., 2014). Given the small distances, using such high currents at ankle could cause the stimulus to" jump" and activate the unanesthetized part of the

corresponding latency (Lat) to the first take-off from baseline. Note that from both stimulation sites, the evoked CMAP was followed by a secondary motor response (F-wave) resulting from antidromic activation of the motor neuron pool.

tibial nerve. To avoid this potential confounder, we tested the changes in conduction across the PNB by proximal stimulation (**Figures 1A,B**). Stimulation and recording were controlled using a custom control software developed in MATLAB (version R2013b, MathWorks Inc., United States). We delivered negative rectangular stimuli of 1 ms duration at 0.5 Hz and recorded 1000 ms sweeps with a sampling frequency of 125 KHz allowing synchronized CMAP (**Figure 2A**) and ECG recordings (**Figure 2C**). The amplitudes at proximal stimulation were averaged over 1-min and expressed, relative to the 1st minute

FIGURE 2 | Development and recovery of conduction block after injection of the local anesthetic (LA). (A) The plantar compound muscle action potential CMAP evoked by proximal stimulation at the sciatic notch after non-encapsulated lidocaine (L. HCL) exposure. After 5 ms, the recordings are presented at 10-fold larger amplitude to facilitate the visualization of F-waves; (B) The relative changes in peak-to-peak CMAP amplitude (ppA) were shown for the non-encapsulated lidocaine (L. HCl) experiment in (A) versus 2 other different experiments using liposomal lidocaine (L. Lip); (C) ECG traces recorded during exposure to liposomal lidocaine. Note that the heart rate, measured from the QRS complexes, remained between 8 and 10 s (corresponding to 480–600 beats per minute). The ppA recovery corresponding to the ECG after 60 min is indicated in blue in panel (B).

prior to injection (**Figure 2B**). The stimulation current was set at twofold the current required to obtain a 100% ppA CMAP and kept constant for the entire duration of the recording.

#### Nerve Excitability Studies

The changes in nerve excitability at the PNB site were monitored by distal stimulation at the ankle (**Figures 1A,C**) and tracking changes in the threshold current (**Figures 3A,B**) required to evoke a 40% ppA CMAP (Bostock et al., 1998) using QtracS stimulation software for recording and QtracP software for analysis (©Institute of Neurology, London, United Kingdom). We used that same TRONDH multiple excitability protocol (Kiernan and Bostock, 2000), that we used in our human PNB study (Moldovan et al., 2014). We have previously adapted this protocol for use in mice (Moldovan et al., 2009, 2011). This allowed recording of the full sequence of excitability measures (**Figure 3C**) as previously described in detail (Bostock et al., 1998; Kiernan and Bostock, 2000; Kiernan et al., 2000): charge–duration relationship to measure the rheobase, threshold

stimulation of the tibial nerve at the ankle (within the nerve area exposed to LA). The thick traces indicate the CMAPs closest to 40% of the maximal CMAP amplitude, referred to as threshold CMAPs; (B) The stimulus-response measurements corresponding to the recordings in (A). The stippled lines indicate the current required to evoke the threshold CMAP (symbols), referred to as threshold; (C) The multiple measures of nerve excitability by "threshold-tracking" protocol, measuring the changes in threshold when varying stimulus duration (charge–duration), when conditioning with 100 ms depolarizing (upward) and hyperpolarizing (downward) currents that were set to 40% of threshold (threshold electrotonus), when conditioning with a range of 200 ms polarizing currents that were stepped from +50% (top) to –100% (bottom) of threshold (current–threshold relationship) and following a supramaximal conditioning stimulus (recovery cycle). The arrows indicate that following liposomal lidocaine there was: an increase in the charge–duration slope, a reduction in threshold changes during both depolarizing and hyperpolarizing threshold electrotonus, a marked increase in the current–threshold slope on depolarization, and an increased in refractoriness of the recovery cycle.

electrotonus to measure accommodation to shifts in depolarizing and hyperpolarizing membrane potential, current–threshold relationship to measure input conductance, and the recovery of excitability following conduction of the action potential to measure refractoriness.

#### Imaging Studies

In vivo imaging (**Figure 4**) was carried out using probe-based confocal laser endomicroscopy (pCLE) using a Cellvizio single band, with a 488-nm laser (Mauna Kea Technologies, France) which allows for the investigation of fluorescent blood vessels and axons (Wong et al., 2009). The frames were captured via a S-1500 objective (**Figure 4A**) and exported using the Cellvizio Viewer for PC (version 1.4.2).

#### Liposomal LA Formulation

We formed an unilamelar liposomes using the established methods of freeze drying lipids followed by rehydration in lidocaine hydrochloride (2% w/v) phosphate-buffered saline. This solution was pressure extruded through 100 nm filters to reduce liposome size polydispersity. The liposomes comprised of hydrogenated soy L-α-phosphatidylcholine (HSPC), cholesterol

(Chol), 1,2-distearoyl-sn-glycero-3-phosphoethanolamine-Nmethoxy (polyethylene glycol)-2000 (DSPE-PEG2000), and 1-palmitoyl-2-(dipyrrometheneboron difluoride)undecanoylsn-glycero-3-phosphocholine (TopFluor PC) in a mol% ratio of 54.8:40:5:0.2, respectively. Polyethylene glycol conjugated lipid was introduced into the liposome bilayer to increase residency time (Eriksen et al., 2017) and TopFluor PC was incorporated to allow visualization. The resulting liposome size (Mean ± SD) was found to be 196.2 ± 2.2 nm with a size polydispersity index of 0.134 ± 0.03 as determined by dynamic light scattering. A non-encapsulated lidocaine salt solution (20 mg/ml, L. HCl) served as control. All chemical components were purchased from commercial suppliers (Sigma–Aldrich, DE, and Avanti Polar Lipids, United States).

pCLE imaging; (C) With further advancement of the imaging probe between the muscle planes (blunt dissection), the tibial nerve could be readily identified due to its

fluorescent fibers (axonal YFP reporter expression). Occasionally, liposomal accumulations (star) could be identified near the fibers.

## Recovery of Conduction

fnins-12-00494 July 25, 2018 Time: 17:15 # 7

Proximal stimulation of the tibial nerve fibers at the sciatic notch (**Figures 1A,B**) evoked a CMAP (Mean ± SD) with a ppA of 12 ± 4 mV and a latency of 1.7 ± 0.1 ms (**Figure 1D**). Distal stimulation of the tibial nerve fibers at the ankle (**Figures 1A,C**) evoked a CMAP (Mean ± SD) with a ppA of 16 ± 5 mV and a latency of 1.2 ± 0.1 ms (**Figure 1E**). Note that the proximal CMAP was not larger than the distal CMAP indicating that the CMAP measurements reflected tibial-nerve innervated muscles with minimal contamination (if any) from common peroneal nerve innervated muscles.

An experiment with non-encapsulated lidocaine is detailed in **Figure 2A**. Within 3 min following LA injection, there was a complete abolishment of the CMAP and F-waves evoked by proximal stimulation which recovered completely within 30 min (**Figure 2A**). The time-course of CMAP amplitude (ppA) recovery, measured relative to the minute before LA, is presented in **Figure 2B**. In contrast, our model was able to detect a slower recovery after liposomal lidocaine, with good reproducibility in repeated studies (**Figure 2B**). Furthermore, we found that following liposomal lidocaine the CMAP evoked by proximal stimulation recovered rapidly to above 50% ppA whereas the subsequent recovery was much slower (**Figure 1B**), so that a full recovery was observed after 3 h (data not shown). Although liposomes were absorbed, as indicated by pCLE (**Figure 4B**), there were no ECG signs of bradycardia (**Figure 2C**) that could indicate a confounding effect of lidocaine toxicity.

#### Recovery of Excitability

During the slow CMAP recovery phase after liposomal lidocaine (**Figure 2B**), liposomal accumulations could be demonstrated by pCLE near the tibial nerve fibers at the ankle (**Figure 4C**) suggesting a persistent release. An experiment with uncaged lidocaine is detailed in **Figure 3**. We found that during the partial CMAP recovery phase, evoking the maximal CMAP by distal stimulation in the nerve exposed to LA required a much larger stimulation current (**Figure 3A**) with a right-shift in the stimulus-response curve (**Figure 3B**). At that time, the current required to evoke the CMAP from the proximal stimulation site remained undistinguishable from the value prior to LA (data not shown).

To explore the mechanisms of impaired tibial nerve excitability at the LA site, we carried out multiple measures of excitability by "threshold-tracking" (**Figure 3C**). Consistent with the larger stimulus current required to evoke the CMAP, we found a marked increase in rheobase (Weiss, 1901; Bostock et al., 1998), as indicated by the slope of the charge–duration relationship (**Figure 3C**). Moreover, the observed increase in rheobase after liposomal lidocaine was associated with (**Figure 3C**) a reduction in threshold changes during both depolarizing and hyperpolarizing threshold electrotonus, a marked increase in the current–threshold slope (input conductance) on depolarization, and an increased in refractoriness of the recovery cycle (**Figure 3C**).

## DISCUSSION

We developed an in vivo mouse model to explore the effect of local anesthetic nanomedicines on axonal function. We tested the validity of the model by comparing the effect of a liposomal lidocaine formulation versus non-encapsulated lidocaine on motor axon function and found a slower conduction recovery along the tibial nerve following liposomal lidocaine. Furthermore, the model opened the possibility to combine conduction studies with "threshold-tracking" excitability studies in the same nerve, allowing a translational mechanistic insight into the liposomal release in vivo.

Assessment of the effect of liposomal local anesthetic nanomedicines on sciatic nerve block in rodents has been previously carried out previously by behavioral measures (Yin et al., 2016). The sensory (thermal nociceptive) blockade was investigated in a hot-plate paradigm (Sagie and Kohane, 2010), whereas the motor blockade was investigated using the extensor postural thrust test (Sagie and Kohane, 2010). Although these behavioral studies provided a valuable approach to distinguish the motor versus sensory selectivity of the nanomedicine, they offered little information on the actual changes in axonal function. It is long known that the susceptibility to an anesthetic conduction block is different between neuronal populations, reflecting differences in membrane properties (Gasser and Erlanger, 1929). Nevertheless, motor population was found to be at least as good an indicator for changes in the duration of PNB as the sensory fibers (Dietz and Jaffe, 1997; Gokin et al., 2001) even though they do not directly reflect the analgesic effect. Furthermore, the fact that the myelinated motor axon population is more functionally homogenous than the sensory fiber population (Gasser and Erlanger, 1929), can convey an advantage in characterizing the excitability changes.

The LA lidocaine is thought to impair the function of the voltage-gated Na<sup>+</sup> channels (VGSC) (Sheets and Hanck, 2007) reducing the axonal "safety factor" for conduction (Tasaki, 1953) in excitable tissues of the nervous system (Nathan and Sears, 1961) as well as the heart (Austen and Moran, 1965). Although the pCLE imaging studies indicated liposome accumulation within veins (**Figure 4B**), the ECG (**Figure 2C**) remained within normal limits (Ho et al., 2011). We therefore do not think that the observed effects on axonal function were confounded by lidocaine cardiotoxicity (Fox and Kenmore, 1967).

We confirmed an increase in the threshold for electrical stimulation during recovery after LA (Gasser and Erlanger, 1929; Gasser and Grundfest, 1939). Similar deviations in excitability were found in our previous study using non-encapsulated lidocaine in humans (Moldovan et al., 2014) which supports the translational value of our mouse model. These excitability deviations differed from those observed with selective VGSC block following accidental tetrodotoxin poisoning in humans (Kiernan et al., 2005). It is likely that the protracted period of reduced nerve excitability following LA (Tabatabai and Booth, 1990) reflected both the partial recovery of the Na<sup>+</sup> conductance as well as an impairment of the axolemmal passive electrical properties (Moldovan et al., 2014) due to a physical effect of lidocaine dissolved within the membrane (Kassahun et al., 2010).

In further studies, excitability changes could allow the detection of early signs of toxicity (Moldovan et al., 2009) in combination with imaging of morphological changes (Beirowski et al., 2005; Alvarez et al., 2008, 2013).

## AUTHOR CONTRIBUTIONS

CK and KL initiated the study. MM made the neurophysiological control software. AU made the liposomal lidocaine formulation.

## REFERENCES


All the authors contributed to the experiments, analyzed the data, discussed the results, and wrote the manuscript.

#### FUNDING

The project was supported by the Danish Medical Research Council grants (4092-00100/2014 and 7016-00165/2017), the Lundbeck Foundation, and the Foundation for Research in Neurology and Jytte and Kaj Dahlboms Foundation.



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Moldovan, Alvarez, Rothe, Andresen, Urquhart, Lange and Krarup. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# A Novel Microbiosensor Microarray for Continuous ex vivo Monitoring of Gamma-Aminobutyric Acid in Real-Time

Imran Hossain<sup>1</sup> , Chao Tan1,2, Phillip T. Doughty<sup>2</sup> , Gaurab Dutta<sup>1</sup> , Teresa A. Murray<sup>2</sup> , Shabnam Siddiqui<sup>2</sup> , Leonidas Iasemidis<sup>2</sup> and Prabhu U. Arumugam1,2 \*

1 Institute for Micromanufacturing, Louisiana Tech University, Ruston, LA, United States, <sup>2</sup> Center for Biomedical Engineering and Rehabilitation Science, Louisiana Tech University, Ruston, LA, United States

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Christine Kranz, Universität Ulm, Germany Nancy Jo Leidenheimer, Louisiana State University Health Sciences Center Shreveport, United States

> \*Correspondence: Prabhu U. Arumugam parumug@latech.edu

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 31 March 2018 Accepted: 03 July 2018 Published: 07 August 2018

#### Citation:

Hossain I, Tan C, Doughty PT, Dutta G, Murray TA, Siddiqui S, Iasemidis L and Arumugam PU (2018) A Novel Microbiosensor Microarray for Continuous ex vivo Monitoring of Gamma-Aminobutyric Acid in Real-Time. Front. Neurosci. 12:500. doi: 10.3389/fnins.2018.00500 Gamma-aminobutyric acid (GABA) is a major inhibitory neurotransmitter that is essential for normal brain function. It is involved in multiple neuronal activities, including plasticity, information processing, and network synchronization. Abnormal GABA levels result in severe brain disorders and therefore GABA has been the target of a wide range of drug therapeutics. GABA being non-electroactive is challenging to detect in real-time. To date, GABA is detected mainly via microdialysis with a high-performance liquid chromatography (HPLC) system that employs electrochemical (EC) and spectroscopic methodology. However, these systems are bulky and unsuitable for real-time continuous monitoring. As opposed to microdialysis, biosensors are easy to miniaturize and are highly suitable for in vivo studies; they selectively oxidize GABA into a secondary electroactive product (usually hydrogen peroxide, H2O2) in the presence of enzymes, which is then detected by amperometry. Unfortunately, this method requires a rather cumbersome process with prereactors and relies on externally applied reagents. Here, we report the design and implementation of a GABA microarray probe that operates on a newly conceived principle. It consists of two microbiosensors, one for glutamate (Glu) and one for GABA detection, modified with glutamate oxidase and GABASE enzymes, respectively. By simultaneously measuring and subtracting the H2O<sup>2</sup> oxidation currents generated from these microbiosensors, GABA and Glu can be detected continuously in real-time in vitro and ex vivo and without the addition of any externally applied reagents. The detection of GABA by this probe is based upon the in-situ generation of α-ketoglutarate from the Glu oxidation that takes place at the Glu microbiosensor. A GABA sensitivity of 36 ± 2.5 pA µM−<sup>1</sup> cm−<sup>2</sup> , which is 26-fold higher than reported in the literature, and a limit of detection of 2 ± 0.12 µM were achieved in an in vitro setting. The GABA probe was successfully tested in an adult rat brain slice preparation. These results demonstrate that the developed GABA probe constitutes a novel and powerful neuroscientific tool that could be employed in the future for in vivo longitudinal studies of the combined role of GABA and Glu (a major excitatory neurotransmitter) signaling in brain disorders, such as epilepsy and traumatic brain injury, as well as in preclinical trials of potential therapeutic agents for the treatment of these disorders.

Keywords: electrochemical, GABA, glutamate, biosensor, microarray, neurochemical, ex vivo

The development of multiplexed neural probes for real-time sensing of neurochemicals is a critical step in the study and effective treatment of brain disorders. Abnormal neurochemical signaling is an underlying signature of brain dynamical disorders such as epilepsy, Parkinson's and Alzheimer's, traumatic brain injury, as well as drug addiction (Cahill et al., 1996; Robinson and Wightman, 2007; Robinson et al., 2008; Sandberg and Garris, 2010; Willuhn et al., 2010). Therefore, it is crucial to be able to monitor and understand the long-term spatiotemporal dynamics of key neurochemicals in the brain. Gammaaminobutyric acid (GABA), a major inhibitory neurotransmitter, is essential for normal neuronal activity, information processing and plasticity, and for neuronal network synchronization (Caudill et al., 1982; Smith and Sharp, 1994; Bhat et al., 2010). GABA's function is impaired in psychiatric and neurological disorders, inflammation and immune diseases, and therefore has been the target in a wide range of drug therapies (Caudill et al., 1982; Smith and Sharp, 1994; Ting Wong et al., 2003; Tian et al., 2004; Bhat et al., 2010; Auteri et al., 2015). GABA is non-electroactive, and it is therefore challenging to detect it in real-time using electrochemical (EC) and spectrophotometrical methods (Cifuentes Castro et al., 2014). To date, in vivo GABA levels are detected mainly via microdialysis on a highperformance liquid chromatography (HPLC) system with EC and spectroscopic detection methods (Kehr and Ungerstedt, 1988; Rowley et al., 1995b; Monge-Acuña and Fornaguera-Trías, 2009; Reinhoud et al., 2013). Since these methods are relatively insensitive to GABA, one must derivatize the solution by making it more conducive to electric signals. Several studies have used HPLC with pre/post derivatized columns using 2,4,6 trinitrobenzenesulfonic acid, o-phthaldialdehyde (OPA)-sulfite and OPA-alkylthiols to separate GABA and then detect it electrochemically at picomolar concentrations by a glassy carbon electrode (GCE) in rat brains (Caudill et al., 1982). The use of OPA-butylthiol was first proposed by Kehr, who infused nipecotic acid and 3-mercaptopro-pionic acid to obtain a faster and more sensitive determination of GABA (Kehr and Ungerstedt, 1988). Rowley et al. (1995b) extended this derivatization technique to separate seven amino acids and concluded that their method could detect, with good sensitivity, the stimulated levels of GABA and glutamate (Glu), a major excitatory neurochemical in a rat hippocampus. Acuna et al. also used this method to separate GABA, Glu and glutamine in rat brain homogenates with higher accuracy and repeatability (Monge-Acuña and Fornaguera-Trías, 2009). Reinhoud et al. (2013) used microbore columned Ultra-HPLC to detect catecholamines such as dopamine (DA) and serotonin (HT-5). The main benefit of this method was that analytes with large differences in retention time could be separated in a single run (Reinhoud et al., 2013). Commercial HPLC-ED systems (Alexys, Dionex) are now available that utilize GCEs at ∼0.8 V for amino acid detection. However, this state-of-the-art technology is bulky and unsuitable for real-time continuous GABA monitoring, which is a key technology gap in the chemical neuroscience field.

The second-best detection method is based upon an amperometry (AM) technique during which GABA is detected indirectly using biosensors (Mazzei et al., 1996; Badalyan et al., 2007, 2014; Sekioka et al., 2008). The main advantage of AMbased biosensors are that they can be easily miniaturized into multiple microarrays and are highly suitable for ex vivo and in vivo studies (Garguilo and Michael, 1994; Hascup et al., 2007). Biosensors selectively oxidize GABA into a secondary electroactive product or reporter molecule in the presence of enzymes, similar to the detection method used for Glu (Hascup et al., 2007) or acetylcholine (Garguilo and Michael, 1994). Electroactive reporter molecules such as β-nicotinamide adenine dinucleotide phosphate (NADPH) or hydrogen peroxide (H2O2) are usually generated through a series of enzymatic reactions by adding nicotinamide adenine dinucleotide phosphate (NADP) as a co-factor, or α-ketoglutarate reagents externally, and then electrochemically detecting them on a modified GCE. The current generated by electrochemically oxidizing them can be used as a quantifying index of GABA's presence. In AM-based GABA biosensors, GABASE, which consists of two enzymes, γ-aminobutyrate aminotransferase (GABA-T) and succinic semialdehyde dehydrogenase (SSDH) converts GABA into Glu (henceforth called GluGABA) and succinic semialdehyde (SSA) in the presence α-ketoglutarate (reaction 1, **Figure 1A**). For reaction 1 to occur, α-ketoglutarate must be present in the sample. The α-ketoglutarate can be added to the sample externally or it can be obtained from oxidizing Glu that is ubiquitously present in the brain microenvironment (henceforth called GluE) using the glutamate oxidase (GOx) enzyme (reaction 2, **Figure 1A**). Subsequently, there are two pathways by which reaction 1 can proceed to electrochemically or optically generate active molecules that indicate the presence of GABA. The first approach is based on SSA reacting with NADP in the presence of SSDH to form NADPH that is then detected optically [UV spectrophotometry or colorimetry (Sethi, 1993) or electrochemically (reaction 3, **Figure 1A**)]. Mazzei et al. (1996) developed a GABA biosensor based on reaction 3 using a horseradish peroxidase-modified GCE. However, the main disadvantage of reaction 3 is that the electrode surface fouls rapidly due to the irreversible nature of NADP<sup>+</sup> adsorption. Sekioka et al. (2008) partially addressed this challenge using an electron cyclotron resonance (ECR) sputtered carbon electrode. Since there is a critical need for ex vivo and in vivo studies to detect GABA long term and NADP to NADPH conversion is irreversible, NADP must be continuously replenished. Badalyan et al. (2007) addressed this problem of continual NADP additions by employing periplasmatic aldehyde oxidoreductase instead of SSDH, and mediators such as ferricyanide, phenoxazines, ferrocene derivatives, quinones, and bipyridinium salts instead of NADP.

Niwa et al. (1998) employed a very different approach that relied on GOx to convert the GluGABA generated in reaction 1 into α-ketoglutarate and H2O<sup>2</sup> (henceforth called H2O2(GABA) , (i.e., H2O<sup>2</sup> generated from the Glu that in turn is generated from GABA; reaction 4, **Figure 1A**) and then detecting it on an osmium-poly(vinylpyridine) gel-horseradish peroxidasemodified GCE. Applying reactions 1 and 4, researchers were

inserted completely within brain slices.

able to detect GABA with adequate sensitivity and selectivity in the presence of DA, HT-5 and ascorbic acid (AA). However, both approaches are incapable of continuously monitoring the changes in GABA levels in real-time since they require additions of reagents such as NADP and α-ketoglutarate. A biosensor technology that can accurately measure GABA in real-time continuously and without any external intervention is technically challenging and yet unrealized. In this work, we report the development and validation of a GABA probe based upon a platinum (Pt) microelectrode array (MEA) (**Figure 1B**) in an in vitro setting and then used the probe for ex vivo measurements in brain slices. The GABA probe uses two types of microbiosensors, namely a Glu microbiosensor (located in Site 1, **Figure 1C**) and a GABA microbiosensor (located in Site 2, **Figure 1C**) that are uniquely modified with GOx only at Site 1 for reaction 2 to occur and with GOx and GABASE at Site 2 for reactions 1, 2, and **4** to occur. Each site in the GABA microarray probe consists of two Pt microelectrodes that are separated by 100 µm. By simultaneously measuring and subtracting the oxidation currents of H2O<sup>2</sup> generated from the two microbiosensors, i.e., IH2O<sup>2</sup> from H2O2(E) at Site 1 (henceforth called IH2O2(Site1)) and IH2O<sup>2</sup> from H2O2(E) and H2O2(GABA) at Site 2 (henceforth called IH2O2(Site2)), GABA (IGABA = 1I = IH2O2(Site2) <sup>−</sup> IH2O2(Site1)) can be detected continuously in real time (Scheme 1, **Figure 1A**) without adding α-ketoglutarate externally (**Figure 2**). This is possible because α-ketoglutarate generated in reaction 2 is used in reaction 1. Scheme 1 can be readily implemented ex vivo and in vivo because the ubiquitous presence of Glu<sup>E</sup> allows in situ generation of α-ketoglutarate, and thus reaction 1 to occur continuously. The SSA generated in reaction 1 is converted to SA when periplasmatic aldehyde reductase is present on the electrode surface (reaction 5, **Figure 1A**) (Badalyan et al., 2014). The other salient features of the GABA probe are: (1) eight individually electrically addressable Pt microelectrodes that can easily be multiplexed to simultaneously measure other important neurochemicals, such as Glu, DA, adenosine and HT-5, through suitable surface modifications, which is not possible with other

commonly available electrodes for chemical sensing, e.g., carbon fiber microelectrodes; (2) GABA and Glu microbiosensors can be placed in close proximity to provide precise measurements of local GABA level changes; (3) an ability to detect GABA realtime without adding reagents (i.e., truly self-contained); (4) the location of MEAs along the long shank allows GABA sensing at multiple depths in the brain; and (5) allows simultaneous sensing of neurochemicals and field potentials for multimodal recordings, which is not possible with the current neurochemical technologies.

## MATERIALS AND METHODS

#### Chemicals

Phosphate buffered saline (PBS), bovine serum albumin (BSA), glutaraldehyde, GABA, GABASE from Pseudomonas fluorescens and α-ketoglutarate disodium salt was purchased from Millipore-Sigma (MO, United States). Glutamate oxidase was purchased from Cosmo Bio United States (CA, United States).

#### GABA Probe Preparation

The platinum (Pt) MEA (8-TRK probe) was purchased from Center for Microelectrode Technology (CenMeT, United States). The MEA consists of eight Pt microelectrodes (50 µm × 100 µm, two microelectrodes per site) and the sites are spaced at 1 mm apart. Each site has two closely spaced (100 µm apart) microelectrodes (**Figure 1C**). Since the in vitro experiments were carried out in a stirred solution in a beaker, we do not expect to see any effect or variability particularly on the Glu signal due to this spatial variation. For ex vivo measurements in brain tissue slices, the two Pt microelectrodes (located in Site 2, as shown in **Supplementary Figure S3A**, and spaced 100 µm apart, see **Supplementary Material**) were coated with GOx and GABASE+GOx, respectively.

#### Enzyme Aliquot Preparation

The GOx enzyme with the BSA and glutaraldehyde was coated in Site 1 as per Burmeister et al. (2013). For Site 1, the GOx enzyme was mixed in DI water to prepare aliquots of 0.5 U/µL and stored in −80◦C. Prior to coating, they were thawed first at 4◦C and then at room temperature. DI water (985 µL) was added to 10 mg BSA in a 1 mL centrifuge tube. After allowing the BSA to dissolve, 5 µL of glutaraldehyde (25% in water) was added to the solution. We kept the solution mixture (1% BSA and 0.125% glutaraldehyde) at room temperature for ∼5 min. A 4 µL of the mixture was added to 1 µL of GOx (0.5 U/µL) and centrifuged to form the final enzyme-matrix mixture of 0.1 U/µL GOx/0.8% BSA/0.1% glutaraldehyde. Similarly, for Site 2, DI water (986.7 µL) was added to 13.33 mg BSA in a 1 mL centrifuge tube. After allowing the BSA to dissolve, 6.67 µL of glutaraldehyde (25% in water) was added to the solution. We kept the solution mixture (1.33% BSA and 0.166% glutaraldehyde) at room temperature for ∼5 min. Next, 3 µL of the mixture was added to 1 µL of GOx (0.5 U/µL) and 1 µL GABASE (0.5 U/µL) and centrifuged to form the final enzyme-matrix mixture of 0.1 U/µL GOx/0.1 U/µL GABASE/0.8% BSA/0.1% glutaraldehyde. For the GABASE-only site, the procedure used for Site 1 was followed except that GABASE instead of GOx was used.

#### Enzyme Coating Procedure

Under a Nikon stereomicroscope (Model, SMZ18), three drops (0.05 µL/drop) of the respective enzyme-matrix mixture was applied manually at each site using a microsyringe (Hamilton <sup>R</sup> , Model 701 N). Then the probe was stored for 48 h in an aluminum foil covered storage container with no exposure to light prior to use. **Figure 2** shows the cross-sectional schematic of the GABA probe with reaction pathways in Sites 1 and 2.

#### Electrochemical Measurements

For amperometry measurements, a multichannel FAST-16mkIII <sup>R</sup> potentiostat (Quanteon, LLC, Nicholasville, KY, United States) in a 2-electrode setup was used with an Ag/AgCl electrode as the reference electrode. The applied potential was set at +0.7 V for H2O<sup>2</sup> detection. Note: This applied potential can be reduced to +0.3 V vs. Ag/AgCl when modified using platinum black as reported in the literature (Ben-Amor et al., 2014). The experiment was carried out in a 40 mL buffer solution. The analytes were introduced into the solution using a syringe pump (KD Scientific, Legato <sup>R</sup> 100 syringe pump) to obtain the desired concentrations (M). The solution was continuously stirred at 200 rpm and maintained at 37◦C. All measurements were repeated 6 times (n = 6). The Fast analysis <sup>R</sup> software provided by Quanteon was used for data analysis. Sensitivity was defined as the change in current for each unit of analyte addition. Sensitivity was calculated from the slope (pA/µM) of the calibration curves. Then the slope was converted into nAµM−<sup>1</sup> cm−<sup>2</sup> by dividing it by the Pt microelectrode area (5 × 10−<sup>5</sup> cm<sup>2</sup> ). The limit of detection (LOD) was calculated by dividing (3 times the standard deviation of 10 points from the baseline) by the least squares slope, which is based on the FAST 2014 software manual provided by Quanteon. The baseline is the signal that was obtained when no electroactive analyte was present in the solution. Two-tailed Students t-test was performed (n = 6) at two different confidence intervals. They are 99.99% (p < 0.0001) and 95% (p < 0.05). The values lie within p < 0.0001 unless otherwise stated. The value which lies within p < 0.05 are represented with (<sup>∗</sup> ) in the bar charts and tables. One-way ANOVA was performed (n = 6) with significance defined as p < 0.05 to verify if sensor-to sensor variation (in the same site) is significant. Error value is shown as mean ± SEM.

## Recording GABA and Glutamate in Brain Tissue

#### Animal Care and Use

Male Sprague Dawley rats were housed on a 12 h on – 12 h off cycle with food and water provided ad libitum, according to a Louisiana Tech University IACUC protocol, the Guide for the Care and Use of Laboratory Animals and the AVMA Guidelines on Euthanasia.

#### Hippocampal Slice Preparation

Hippocampal slices were prepared from an adult Sprague Dawley rat that was anesthetized using 5% isoflurane gas prior to decapitation and rapid removal of the brain. The brain was immediately placed into ice cold artificial cerebral spinal fluid (aCSF) containing (in mM): 135 NaCl, 3 KCl, 16 NaHCO3, 1 MgCl, 1.25 NaH2PO4, 2 CaCl2, and 10 glucose, bubbled with 95% O2/5% CO<sup>2</sup> (carbogen) (Song et al., 2005). The slicing chamber of an OTS-5000 tissue slicer (Electron Microscopy Sciences) was filled with aCSF at 4◦ C and then 500-µm thick coronal sections were cut and transferred to a holding chamber filled with aCSF maintained at 35◦C and bubbled with carbogen. Slices were incubated for at least 60 min prior to recording. Thereafter, one slice was transferred to a liquid-air interface of a BSC1 chamber (Scientific Systems Design, Inc.) with the slice suspended on a nylon net at the liquid-air interface with continuously dripping aCSF (37◦C) bubbled with carbogen. Waste products were removed by continuous suction from the recording chamber (**Supplementary Figure S3B**, see **Supplementary Material**).

#### GABA Recording in Rat Hippocampal Slices

The microbiosensors were coated with a size-exclusion polymer (m-phenylenediamine, mPD) to prevent the interferents reaching the microbiosensor surface and to enhance the probe selectivity (Wilson et al., 2017). **Supplementary Figure S4** (see **Supplementary Material**) demonstrates the ability of the mPD coating to block dopamine and ascorbic acid effectively. The tradeoff here is that, with an mPD coating, a ∼20% decrease in the sensitivity of the probe to GABA was observed (**Supplementary Figure S5**). Similar decrease in sensitivity values was observed at the Glu microbiosensor as well. The mPD layer was electrochemically deposited (cycling between +0.2 V and +0.8 V, 50 mV/s, 20 min in 10 mM mPD solution). A pair of 160-µm diameter tungsten stimulation electrodes was placed in the Schaffer collateral CA1 pathway within 200 µm of the microbiosensor probe sites (Song et al., 2005). An A365 stimulus isolator (World Precision Instruments) was used to deliver 100-µA direct current pulses to the stimulus electrodes; pulse widths were regulated by transistor-transistor logic (TTL) input from an Arduino microcontroller. Current detected at the probe sites was plotted in real time.

#### Data Analysis for ex vivo Recordings

Results from ex vivo, hippocampal recordings were analyzed using OriginPro 2017. Measurements are reported as the mean ± square error of the mean (SEM). ANOVA was performed for comparisons of means and significance was defined as p ≤ 0.05. Rise times (Tr10−90) were defined as the elapsed time between 10 and 90% from the baseline to the peak current of the stimulation response. The Rise Time Gadget tool in OriginPro 2017 was used to calculate the rise time.

### RESULTS AND DISCUSSION

## Calibration of GABA Probe in the Presence of α-Ketoglutarate

Studies have shown dependence of the GABA current response (pA) on concentration of α-ketoglutarate (Niwa et al., 1998),

FIGURE 3 | GABA probe calibration in different concentrations of GABA (5, 10, 20, and 40 µM) and α-ketoglutarate (5, 10, 20, 40, 100, 200, and 500 µM) in 1X PBS. (A) Current response at GABA microbiosensor in Site 2 and Glu microbiosensor in Site 1 in PBS only (background or control – red dashed curve, blue dashed curve, respectively) and in 100 µM α-ketoglutarate in 1X PBS (red and blue solid curves, respectively). (B) Current response at GABA microbiosensor for other concentrations of α-ketoglutarate. (C,D) Current response and linear fitting at GABA microbiosensor for different GABA concentrations and α-ketoglutarate concentrations at ≥ 40 µM. Legends: 40 µM (red), 100 µM (green), 200 µM (blue) and 500 µM (magenta). The microbiosensors were biased at + 0.7 V vs Ag/AgCl reference. The solution was stirred at 200 rpm and maintained at 37◦C. Linear fit parameters obtained: 40 µM α-keto sensitivity = (0.55 ± 0.077 pA/µM), R <sup>2</sup> = 0.99728; 100 µM α-keto sensitivity = (1.74 ± 0.13 pA/µM), R<sup>2</sup> = 0.99582; 200 µM α-keto sensitivity = (1.03 ± 0.13 pA/µM), R<sup>2</sup> = 0.99582 and 500 µM α-keto sensitivity = (1.51 ± 0.13 pA/µM); R<sup>2</sup> = 0.99582 at GABA microbiosensor. Two-tailed Students t-test was performed (n = 6, p < 0.0001, <sup>∗</sup>p < 0.05). One-way ANOVA was performed (n = 6, p < 0.0001) to verify that sensor-to sensor variation (in the same site) is not significant. Error value is shown as mean ± SEM.

TABLE 1 | GABA sensitivity and LOD for different α-ketoglutarate concentration.


Error value is shown as mean ± SEM. Two-tailed Students t-test was performed (n = 6, p < 0.0001, <sup>∗</sup>p < 0.05). One-way ANOVA was performed (n = 6, p < 0.0001) to verify that sensor-to sensor variation (in the same site) is not significant.

which is an important molecule in physiological functions, for example in the Krebs cycle (Tretter and Adam-Vizi, 2005). Therefore, we first studied the electrochemical response of the Glu and GABA microbiosensors (Sites 1 and 2) in the presence of different concentrations of α-ketoglutarate (1–500 µM) in the phosphate buffered saline (PBS) solution. **Figure 3A** shows the typical AM responses at Sites 1 and 2 in 1X PBS supporting electrolyte (background or control, blue dashed, red dashed curves), and to varying concentrations of GABA (5, 10, 20 and 40 µM) in 100 µM α-ketoglutarate solution prepared in 1X PBS (blue solid, red solid curves). These values of concentration in the micromolar range were chosen because of their relevance to the ones encountered in the brain microenvironment where GABA is typically present (Badalyan et al., 2014). For example, GABA levels are in the range of 20–70 µM in rat brain slices, (Grabauskas, 2004), and up to 1.25 µM/cm<sup>3</sup> in the human brain (Ke et al., 2000) as measured by proton magnetic resonance spectroscopy. The AM response was recorded in different concentrations of α-ketoglutarate solution, first by allowing the microbiosensors to stabilize in the solution for up to 240 s, and then injecting GABA at 1 min time intervals to obtain the desirable concentration (**Figures 3B,C**). From **Figure 3A**, as expected, we observe that the Glu microbiosensor at Site 1 did not exhibit a response to GABA because of the absence of the GABASE enzyme. Also, there was no enzymatic activity of GOx in converting GABA into Glu and then into H2O2. This indicates that the GABA conversion is highly selective at Site 2 that has

GABASE and not at Site 1. The GABA microbiosensor at Site 2 responded to GABA when the α-ketoglutarate concentration was at least 40 µM (**Figure 3B**). A transient spike in the signal was observed during the injection of the solution in the beaker. However, the signal was stabilized a few seconds following the injection of the solution. Sometimes the time to stabilization was a bit longer (e.g., in the case of 40 µM and 500 µM α-ketoglutarate experiments). This might be due to a few bubbles in the micro syringe pump that disturb the solution more in certain experiments than others. The other data points for the same α-ketoglutarate concentrations did not show similar spikes. The highest sensitivity was observed at 100 µM. From **Figure 3D**, the sensitivity is 36 ± 2.5 pA µM−<sup>1</sup> cm−<sup>2</sup> and the LOD is 2 ± 0.12 µM (n = 6), which is 10-fold higher than that of similar AM-based microsensors (Niwa et al., 1998). The sensitivities at 40, 200, and 500 µM of α-ketoglutarate were 12 ± 1.7, 20 ± 2.4 and 28 ± 2.5, respectively, and the LOD was 7 ± 0.7, 4.0 ± 0.4, and 3 ± 0.24, respectively (**Table 1**). This GABA response to α-ketoglutarate concentration is in agreement with previously published literature (Niwa et al., 1998). One possible reason for the decrease of GABA sensitivity at highest α-ketoglutarate concentrations could be due to their scavenging of H2O<sup>2</sup> as suggested by previous studies (Nath et al., 1995; Long and Halliwell, 2011). Another study (Badalyan et al., 2014) showed a similar trend where the GABA sensitivity was highest at 1 mM α-ketoglutarate and then decreased at much higher concentrations. The LOD achieved using the GABA microbiosensor is 2–7 µM, which is lower than the clinically-relevant concentrations (Grabauskas, 2004) and similar to the values achieved by alternative methods (Ke et al., 2000) in the human brain. Sensitivities differ slightly

between microelectrodes, which are likely due to variations in the quantity of enzymes that are manually applied to each site. Any potential defects in the surface of the electrodes may also lead to a difference in sensitivity. But this could be remedied by employing an array of GABA and Glu microbiosensors and by applying appropriate statistics (e.g., averaging the current values, etc.) in the future. This sensitivity variation can be further minimized by employing micro spotting techniques that are fully automated and dispense very precise volumes of enzyme solutions. Next, to determine the linear range of the calibration plots, we generated the plots for 5–500 µM GABA concentrations versus different α-ketoglutarate concentrations. We observe that the GABA current values saturate, and saturation depends on the α-ketoglutarate concentration (**Supplementary Figures S1, S2**, see **Supplementary Material**). For example, for 40 µM α-ketoglutarate, the GABA signal saturation is at 50 µM. Whereas in 100, 200 and 500 µM α-ketoglutarate concentrations, the GABA signal saturation occurs at 100 µM. The trend in sensitivity in the linear range is the same as before. For 100 µM α-ketoglutarate, the GABA sensitivity is highest and becomes lower at other concentrations of α-ketoglutarate.

## Calibration of the GABA Probe in the Presence of Glutamate

The GABA probe was calibrated in the presence of a range of concentrations (5–80 µM) of Glu, which mimics the brain microenvironment both in healthy and diseased states. For example, the basal concentration of Glu in the extracellular space is up to 20 µM (Moussawi et al., 2011), while Glu concentrations in cerebrospinal fluid are ∼10 µM. During seizures, Glu levels

FIGURE 4 | GABA probe calibration in different concentrations of Glu (5, 10, 20, 40, and 80 µM). (A) Current response at GABA microbiosensor in Site 2 and Glu microbiosensor in Site 1 (red and blue solid curves, respectively). (B) Inset showing the linear fitting for GABA and Glu microbiosensors (red and blue dotted lines). Linear fit parameters obtained: Site 2, GABA microbiosensor: Glutamate sensitivity = (6.67 ± 0.38 pA/µM), R<sup>2</sup> = 0.99984; Site 1, Glu microbiosensor: Glutamate sensitivity = (4.55 ± 0.11 pA/µM), R<sup>2</sup> = 0.99926. (C) The difference in the current response between the microbiosensors (blue bars). The current response at GABA microbiosensor that was coated with GABASE enzyme only (no GOx enzyme) (red bars). Two-tailed Students t-test was performed (n = 6, p < 0.0001, <sup>∗</sup>p < 0.05). One-way ANOVA was performed (n = 6, p < 0.0001) to verify that sensor-to sensor variation (in the same site) is not significant. Error values are shown as mean ± SEM. Note: the error bars are too small for the blue dotted data for the naked eye to see. The microbiosensors were biased at + 0.7 V with respect to an Ag/AgCl reference electrode. The solution was stirred at 200 rpm and maintained at 37◦C. No α-ketoglutarate was added during any of the experiments.

sensitivity = (8.2 ± 0.17 pA/µM); R<sup>2</sup> = 0.99863 at GABA microbiosensor. Two-tailed Students t-test was performed (n = 6, p < 0.0001). One-way ANOVA was performed (n = 6, p < 0.0001) to verify that sensor-to sensor variation (in the same site) is not significant. Error value is shown as mean ± SEM. The microbiosensor was biased at + 0.7 V vs Ag/AgCl reference. The solution was stirred at 200 rpm and maintained at 37◦C. No α-ketoglutarate added during all the experiments.

increase 4-fold and GABA levels decrease (Rowley et al., 1995a; Kanamori and Ross, 2011; Medina-Ceja et al., 2015). Glu is a major excitatory neurochemical that is ubiquitously present as L-glutamate in its anionic form (glutamic acid) in the brain environment (henceforth called GluE) (Moussawi et al., 2011). One of the objectives of this study was to monitor Glu<sup>E</sup> as an in-situ source for the generation of α-ketoglutarate, which aids in the continuous real-time GABA monitoring at Site 2, and thus does not rely on the addition of α-ketoglutarate externally. Firstly, we calibrated the two microbiosensors by injecting Glu at various concentrations (5, 10, 20, 40, and 80 µM) in 1X PBS buffer solution. **Figures 4A,B** shows the response of the two microbiosensors. The GABA microbiosensor (Site 2) consistently exhibited a slightly higher Glu response than that of the Glu microbiosensor (Site 1). The Glu sensitivity of Site 2 and Site 1 are 132 nA µM−<sup>1</sup> cm−<sup>2</sup> and 90 nAµM−<sup>1</sup> cm−<sup>2</sup> , respectively. The difference in the current response from the two microbiosensors increases for higher Glu concentrations (**Figure 4C**, blue bars). To further understand this, we modified Site 2 with only GABASE and no GOx. Ideally, there should not be any response from the GABA microbiosensor, however, a small response was observed (**Figure 4C**, red bars). This confirms our hypothesis that some non-selective activity of GABASE is due to Glu oxidation. Others have made similar observations where GABASE showed weak enzyme activity toward Glu compared to GOx (Niwa et al., 1998). The large response could also be due to the presence of more enzymes per unit volume (0.2 U/µl) that somehow collectively create more active sites (Arima et al., 2009). To account for this difference in the Glu response, henceforth called the background noise, I<sup>b</sup> [shown in **Figure 4C** (blue bars)], the I<sup>b</sup> was subtracted from the difference in the currents (IGABA) at the two sites in order to obtain the final current response to GABA (details discussed later).

The next calibration step was to test different GABA solutions (0, 5, 10, and 20 µM) in 1X PBS buffer and repeat the TABLE 2 | Sensitivity and LOD in Site 1 (GOx only) and Site 2 (GOx + GABAse).


Error value is shown as mean ± SEM. Two-tailed Students t-test was performed (n = 6, p < 0.0001, <sup>∗</sup>p < 0.05). One-way ANOVA was performed (n = 6, p < 0.0001) to verify that sensor-to sensor variation (in the same site) is not significant.

above Glu calibration (**Figure 5**). These experiments were performed without adding α-ketoglutarate externally. At Site 1, Glu<sup>E</sup> is oxidized to α-ketoglutarate and H2O2(E) (reaction 2). This α-ketoglutarate then reacts with GABA at Site 2 and produces GluGABA (reaction 1) followed by reaction 4, which generates H2O2(GABA) and more α-ketoglutarate. These reactions and pathways were shown in **Figure 2**. At the GABA microbiosensor (Site 2), in the case of no GABA in the solution, the current response (IH2O2(Site2)) is due only to the changing Glu levels in the solution (**Figure 5A**, red curve). When GABA is present in the solution, the IH2O2(Site2) response is from both GABA and Glu oxidation and we expect it to be larger than the response when there was no GABA. Therefore, higher GABA concentrations appear to induce a greater response (**Figure 5A**, blue, green, and magenta curves) at Site 2 and greater IGABA, which is the GABA signal (Scheme 1). **Figure 5B** shows the sensitivity of the GABA microbiosensor at different GABA and Glu concentrations. With increasing GABA and Glu concentrations, the sensitivity of the GABA microbiosensor increases and

this is because of increased availability of α-ketoglutarate for reaction 1. The sensitivity and the LOD of the two microbiosensors is shown in **Table 2**. The GABA sensitivity increased by ∼25% at 20 µM GABA concentrations. The sensitivity reported here is greater than that of the Pt based Glu sensors published in the literature (Tseng et al., 2014)


TABLE 3 | Stimulation pulse parameters and rise time of the stimulated response.

Values are expressed in mean ± SEM. Two-tailed t-test was performed (n = 3, p < 0.05). <sup>∗</sup>Two-tailed t-test was performed (n = 6, p < 0.05).

The LOD is comparable to other Glu sensors (Khan et al., 2011).

## Quantification of GABA Using the IGABA and IH2O2(E) Current Values

Finally, the GABA signal was quantified as IGABA = IH2O2(Site1) <sup>−</sup> IH2O2(Site2) . The IGABA is plotted for varying GABA and Glu concentrations in **Figure 6A** after subtracting the I<sup>b</sup> noise. The positive values for IGABA at all concentrations of GABA and Glu confirms GABA detection at Site 2. As expected, the IGABA increases as GABA concentrations increase. The GABA calibration curves, following linear approximation of IGABA at various Glu concentrations, is shown in **Figure 6B**. A steeper slope is evident at higher GABA concentrations. Values of the slope are 2.7 ± 0.2 pA/µM, 2.9 ± 0.3 pA/µM and 3.5 ± 0.2 pA/µM for 5, 10, and 20 µM GABA, respectively. To better understand the GABA signal dependence on Glu concentrations, IGABA values were plotted in terms of different molarity ratios of GABA:Glu (1:1, 1:2, 1:4, and 1:8) for different GABA concentrations (**Figure 6C**). It is known that GABA and Glu maintains a certain balance in the human brain by means of the glutamate-glutamine (GABA) cycle (Hertz, 2013) And they exist in a certain molarity ratio based upon the state of the brain. For example, in epilepsy, this cycle becomes imbalanced and Glu levels are elevated (Rowley et al., 1995a; Kanamori and Ross, 2011; Medina-Ceja et al., 2015). The data clearly suggest that the IGABA value is greatly dependent on both GABA and Glu concentration, i.e., the IGABA increases as GABA and Glu levels increases. This is evident from **Figure 6C**, which shows that, for a given GABA concentration, the IGABA value is larger for higher GABA:Glu ratios. So, in this approach, for a given IGABA value, the GABA concentration can vary. For example, for an IGABA value of 100 pA, the GABA concentration can be 5, 10, or 20 µM. This is because the GABA signal is dependent on the local availability of α-ketoglutarate, which is dependent on the local Glu concentration. Thus, there is no one IGABA value for a given GABA concentration. This problem can be solved by considering the IGABA value from Site 2 and the IH2O2(E) value from Site 1. From the IH2O2(E) value, the local Glu concentration can be measured. Once the local Glu concentration is known, (x-coordinate in **Figure 6B**), and since the IGABA value is already known (y-coordinate in **Figure 6B**), their intersection yields the local GABA concentration. For example, let us say that the IH2O2(E) value from the Glu microbiosensor is 175 pA and then from **Figure 4B**, the Glu concentration will be 50 µM. And, this 50 µM Glu is the x-coordinate in **Figure 6B**. Next, let us say that the IGABA value

FIGURE 7 | Ex vivo recording of stimulated release of Glu and GABA in rat hippocampal slice preparation. The amperometry method was used to record current with the microbiosensor biased at + 0.7 V with respect to an Ag/AgCl reference electrode. (A) Current responses to unipolar stimulation (tungsten wires, 100 µA) were recorded on a F.A.S.T. 16mkIII system (Quanteon, Kentucky; red traces, Glu microbiosensor; black traces, GABA microbiosensor). Stimulation pulse parameters (Pulse A–E) are listed in Table 3 and range from 1 s to 5 ms. Conversion of peak current measurements to Glu and GABA concentrations are listed in Table 4 for points 1 and 2 (see numbers with arrows). Insets B, C, and D show details of the responses to shorter pulse widths. (E) Processed GABA signal with Glu signal from responses to Pulse C–E stimulations in 7A. The GABA trace (blue trace) is the difference between the signals from the GABA-glutamate and the Glu microbiosensor sites (red trace). (F) Inset shows the rise time, tr10−90, for the GABA signal (blue curve) and the Glu signal (red curve) from the boxed region in E. Arrows indicate the slope of line drawn from tr10−<sup>90</sup> for GABA (blue arrow points to line) and tr10−<sup>90</sup> for Glu (red arrow points to line). The rise times for GABA were faster than for Glu, as the difference in the slopes of the lines illustrate.

parameters: slope = 0.21 ± 0.006 pA/µM, R<sup>2</sup> = 0.99699, the linear fit for 10 µM GABA (green dashed), linear fit parameters: slope = 0.63 ± 0.09 pA/µM, R <sup>2</sup> = 0.9699 and linear fit for 20 µM GABA (blue dashed), linear fit parameters: slope = 1.14 ± 0.2 pA/µM, R<sup>2</sup> = 0.98967. Values are expressed in mean ± SEM. Two-tailed t-test was performed (n = 3, p < 0.05).

is 175 pA, which is the difference between the IH2O<sup>2</sup> values obtained from the two microbiosensors. Again, the IGABA value is the y-coordinate in **Figure 6B**. So, from **Figure 6B**, with (x, y) as (50 µM, 175 pA), the intersection of the lines falls on the blue dashed line that corresponds to a GABA concentration of 20 µM.

Finally, in this work, for the in vitro experiments, the microbiosensors were not coated with selective coatings such as nafion and m-phenylenediamine (mPD) that have shown to completely block potential electroactive interferents such as dopamine and ascorbic acid. For the ex vivo testing, we coated the microbiosensors with mPD to achieve selectivity of the probe (Wilson et al., 2017).

## Real-Time Measurement of GABA and Glutamate in Rat Hippocampal Slice Preparation

Simultaneous and continuous real-time detection of GABA and glutamate was accomplished using electrically stimulated release in a hippocampal slice model. We used a range of 100-µA pulse widths to induce release of the neurotransmitters (see **Table 3**) to determine the responsiveness of the sensor to varying levels of stimulation which included single pulses ranging from 1 s to 25-ms in duration and a pulse train of ten 5-ms pulses. The GABA signal was derived by subtracting the signal from the Glu microbiosensor from the GABA microbiosensor. As expected, the amplitude of GABA and glutamate release scaled with pulse width (**Figure 7**). In some cases, GABA had a shorter peak duration, and in all cases the concentration of GABA rose faster than glutamate concentration (**Figure 7F**). For example, the mean rise time (± SEM) for a 25-ms stimulation was 3.12 ± 0.35 s for GABA and 6.94 ± 0.9 s for glutamate (n = 6, p < 0.05). Both GABA and glutamate leak out of neuronal synapses after neurons release these neurotransmitters. Mechanisms exist to quickly scavenge and recycle these neurotransmitters, but some molecules diffuse

TABLE 4 | Conversion of current to glutamate and GABA concentration in ex vivo recordings.


<sup>1</sup>Number corresponds to signal trace number in Figure 7A. <sup>2</sup>From Figure 7E, the IH2O2(E) value, i.e., the local Glu signal is measured. <sup>3</sup>Then the local Glu concentration is known from Figure 8A. <sup>4</sup>The IGABA value is the difference between the IH2O<sup>2</sup> values obtained from the two microbiosensors. <sup>5</sup>Now, knowing the Glu concentration, which is the x-coordinate in Figure 8B and the IGABA value, which is the y-coordinate in Figure 8B, one can find the GABA concentration for the two points.

through the extracellular space (Danbolt, 2001; Robinson and Jackson, 2016; Boddum et al., 2016). Thus, there is a slight delay from stimulation to response, as well as a long decay period as GABA and glutamate are eventually cleared. Both of these dynamic processes are evident in the traces shown in **Figure 7** with a rapid, but not immediate increase in neurotransmitter concentration, and a slower decline to baseline representing release and uptake, respectively.

A calibration curve was constructed before performing the ex vivo recordings in order to convert current from GABA release to GABA concentration at the probe (**Figure 8A**). This calibration curve is constructed based on the procedure detailed in **Figure 4B**. The data plotted in **Figure 8B** is constructed in the same way as that of **Figure 6B**. Peak current measurements in **Table 4** represent a range of stimulated release of GABA and glutamate. These measurements correspond to curves labeled 1–2 in **Figure 7A**. Peak concentrations ranged from 5 to 35 µM for glutamate and 5–13 µM for GABA. Thus, these probes can measure GABA and glutamate at concentrations that are well below normal levels (Grabauskas, 2004; Moussawi et al., 2011) making them suitable to study impaired release in disease states. Furthermore, numerous cycles of stimulated

release with consistent current amplitude for each level of stimulation and without adding any exogenous substrates, such as a-ketoglutarate, support the premise that endogenous products of the conversion of glutamate provide the substrate for the GABASE reaction. This is an important capability for future in vivo applications.

#### CONCLUSION

In this work, we report a novel GABA microarray probe that can detect GABA without the addition of any external reagents such as α-ketoglutarate and NADPH in vitro. The GABA probe consists of two microbiosensors that were modified with GOx and GOx+GABASE enzymes. By simultaneously measuring and subtracting the oxidation currents of H2O<sup>2</sup> generated from the microbiosensors. GABA was detected with a sensitivity of 36 ± 2.5 pA µM−<sup>1</sup> cm−<sup>2</sup> and LOD of 2 ± 0.12 µM. We demonstrate a new detection method that will assist neuroscientists to better understand the combined role of GABA (a major inhibitory neurochemical) and Glu (a major excitatory neurochemical) in real-time in the brain. The key benefits of the proposed approach are: (1) the probe can be easily multiplexed to simultaneously measure other important neurochemicals, which is not possible with other commonly used electrodes for chemical sensing, (2) ability to detect GABA in real time without adding reagents (i.e., truly selfcontained), (3) it is based on an established, commercially available Pt MEA platform that is suitable for future in vivo recordings, (4) the location of the MEAs along the long shank allows GABA and Glu sensing at multiple depths in the brain, and (5) it can simultaneously sense neurochemicals and field potentials for multimodal (e.g., neurochemical and neuroelectrical) recordings, which is not possible with the current neurochemical technologies. Furthermore, we demonstrated the utility of the microbiosensor microarray to simultaneously record fluctuations in electrically stimulated GABA and glutamate release continually and in real time in a rat hippocampal slice preparation. Moreover, we have shown that GABA release can

## REFERENCES


be detected over repeated stimulations without adding substrate compounds externally. Future work for testing the GABA probe using in vivo animal models is anticipated.

## AUTHOR CONTRIBUTIONS

IH designed the experiments, performed the experiments, analyzed the data, prepared the figures and manuscript. CT performed the experiments. PD performed the ex vivo experiments and analyzed the data. GD analyzed the in vitro data and figures. TM designed the ex vivo experiments, analyzed the data, and prepared the ex vivo section of the manuscript. SS designed the in vitro experiments, analyzed the data, and reviewed the manuscript. LI provided guidance and reviewed the manuscript. PA conceptualized and designed the GABA probe, detection scheme and experiments, provided guidance, prepared the figures, analyzed the data, and prepared the manuscript.

## FUNDING

This work was supported by the National Science Foundation through the OIA/EPSCoR grant 1632891 on Probing and Understanding the Brain: Micro and Macro Dynamics of Seizure and Memory Networks.

## ACKNOWLEDGMENTS

We thank Haocheng Yin at AMRL, Louisiana Tech University for assistance in imaging.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnins. 2018.00500/full#supplementary-material


and brain altered function. Curr. Neuropharmacol. 12, 490–508. doi: 10.2174/ 1570159X13666141223223657


**Conflict of Interest Statement:** A report of invention (ROI 2017-09) is filed with Louisiana Tech University's Office of Intellectual Property and Commercialization and filed a provisional patent.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Hossain, Tan, Doughty, Dutta, Murray, Siddiqui, Iasemidis and Arumugam. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Cortical Network Synchrony Under Applied Electrical Field in vitro

Min D. Tang-Schomer1,2,3 \*, Taylor Jackvony<sup>4</sup> and Sabato Santaniello3,5

<sup>1</sup> Department of Pediatrics, UConn Health, Connecticut Children's Medical Center, Farmington, CT, United States, <sup>2</sup> The Jackson Laboratory for Genomic Medicine, Farmington, CT, United States, <sup>3</sup> CT Institute for the Brain and Cognitive Sciences, University of Connecticut, Storrs, CT, United States, <sup>4</sup> School of Medicine, UConn Health, University of Connecticut, Farmington, CT, United States, <sup>5</sup> Biomedical Engineering Department, University of Connecticut, Storrs, CT, United States

Synchronous network activity plays a crucial role in complex brain functions. Stimulating the nervous system with applied electric field (EF) is a common tool for probing network responses. We used a gold wire-embedded silk protein film-based interface culture to investigate the effects of applied EFs on random cortical networks of in vitro cultures. Two-week-old cultures were exposed to EF of 27 mV/mm for <1 h and monitored by time-lapse calcium imaging. Network activity was represented by calcium signal time series mapped to source neurons and analyzed by using a community detection algorithm. Cortical cultures exhibited large scale, synchronized oscillations under alternating EF of changing frequencies. Field polarity and frequency change were both found to be necessary for network synchrony, as monophasic pulses of similar frequency changes or EF of a constant frequency failed to induce correlated activities of neurons. Group-specific oscillatory patterns were entrained by networklevel synchronous oscillations when the alternating EF frequency was increased from 0.2 Hz to 200 kHz. Binary responses of either activity increase or decrease contributed to the opposite phase patterns of different sub-populations. Conversely, when the EF frequency decreased over the same range span, more complex behavior emerged showing group-specific amplitude and phase patterns. These findings formed the basis of a hypothesized network control mechanism for temporal coordination of distributed neuronal activity, involving coordinated stimulation by alternating polarity, and time delay by change of frequency. These novel EF effects on random neural networks have important implications for brain functional studies and neuromodulation applications.

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Axel Hutt, German Meteorological Service, Germany Jiang Wang, Tianjin University, China

#### \*Correspondence:

Min D. Tang-Schomer Min.Tang-Schomer@jax.org; mints2013@gmail.com

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 12 May 2018 Accepted: 22 August 2018 Published: 21 September 2018

#### Citation:

Tang-Schomer MD, Jackvony T and Santaniello S (2018) Cortical Network Synchrony Under Applied Electrical Field in vitro. Front. Neurosci. 12:630. doi: 10.3389/fnins.2018.00630 Keywords: neural synchronization, neural interface, in vitro culture, neurostimulation, silk biomaterials, network analysis

## INTRODUCTION

Synchronized neural activities underlie many cognitive and behavioral responses during normal brain functioning (Buzsaki and Draguhn, 2004) and neurological disorders such as epilepsy (Burns et al., 2014; Yaffe et al., 2015) and schizophrenia (Uhlhaas and Singer, 2010). Neurons organize into functional networks that generate synchronized activities either spontaneously (Kirkby et al., 2013; Luhmann et al., 2016) or upon exogenous stimulus (Zhang and Poo, 2001; Tagawa et al., 2008). This process involves intrinsic molecular programs at the cellular level (Mathie et al., 2003; Holtmaat and Svoboda, 2009; Rebola et al., 2010; Bagley and Westbrook, 2012) and large scale (ensembles)

**215**

information processing at the network level (Marom and Shahaf, 2002). Stimulating the central nervous system (CNS) with applied electric or magnetic field has become a common tool for probing neural networks in functional studies of the brain (Wagner et al., 2007; Bestmann et al., 2015). The applied electromagnetic fields affect CNS by generating a distributed electric field (EF) around the brain tissue underneath (McIntyre and Grill, 2002; Frohlich, 2014). Despite the wide-ranging neuro-modulatory effects of exogenous EF on the nervous system, the underlying mechanism for induced network changes remains elusive.

Major challenges for functional studies lie in the complexity of neural networks and the highly variable dynamics of neuronal responses. Neuronal response depends on the stimulus as well as the cell's intrinsic properties. Studies (Lafon et al., 2017; Hutt et al., 2018) have recently shown that high-frequency periodic stimulation can lead to waveform-dependent changes in the oscillatory dynamics of neurons while randomly fluctuating stimulation linearizes the neuron response function. Other studies, instead, have shown that the neuron's response change as a function of the intensity, duration, and polarity of the stimulus (McIntyre and Grill, 2002; Wagenaar et al., 2004) and these changes are noticed at multiple spatial scales, from small neuronal clusters to large populations (Yoo et al., 2004; Yu et al., 2017, 2018). Neuronal sensitivity depends on the cell's channel protein and receptor composition, synapse maturate state, and cell morphology (O'Brien et al., 1998; Mathie et al., 2003; Holtmaat and Svoboda, 2009; Rebola et al., 2010; Bagley and Westbrook, 2012; Yi et al., 2017). Models with defined network architecture and cell compositions, such as ex vivo brain slices or specific CNS pathways, are used to determine conditions capable of evoking functionally relevant responses. For example, continuous stimulation at 5 Hz corresponds to the resting state at hippocampal synapses (Bito et al., 1996) and context-dependent stimuli mimics neuronal activities during learning (Larson et al., 1986; O'Keefe and Recce, 1993; Staubli et al., 1999). However, it is unclear how those region- or pathway-specific findings can be applied elsewhere in the CNS.

In vitro cultures of dissociated neurons combined with multi-electrode arrays (MEAs) provide an alternative model for studying neural networks. Such cultures retain many in vivo features, including connectivity and cell type distribution, as well as synaptic and cellular level plasticity, e.g., see (Marom and Shahaf, 2002) for a review. In vitro cortical cultures allow much more detailed observation and manipulation than intact brains. Cultures exhibit spontaneous periodic calcium transients or bursting activities (Robinson et al., 1993; Maeda et al., 1995; Jimbo et al., 2000; Opitz et al., 2002), with increased propensity for synchronized bursting as the culture matures (Kamioka et al., 1996; Tateno et al., 2002; Sun et al., 2010). Interrogated by site-specific stimuli with varying temporal and spatial features, in vitro cortical networks exhibit in vivo-relevant adaptive behavior (Eytan et al., 2003). Studies have shown pathwayspecific (rather than neuron-specific) changes in neuronal responsiveness, including potentiation or depression (Jimbo et al., 1999), and stimuli context-dependent plasticity (Bakkum et al., 2008a). Network-level signal propagation involves intrinsic firing of random neurons, recruitment of other neurons, and repetitive excitation leading to synchronous burst firing (Jimbo et al., 2000; Chao et al., 2007; Bakkum et al., 2008a; Kryukov et al., 2008; Manyakov and Van Hulle, 2008; Sun et al., 2010; Yu et al., 2011a,b, 2012, 2013). These studies have led to attempts to develop a self-learning network with in vitro cortical cultures, for example, with a closed-loop training algorithm to guide the network toward a pre-determined activity state (Wagenaar et al., 2005; Bakkum et al., 2008b), and with patterned stimuli to increase burst firing probability of a selective loci (Shahaf and Marom, 2001). However, there are considerable viabilities regarding the design of MEA systems, the network topology and variable endogenous activities of heterogeneous neuronal populations (Morin et al., 2005). In this study, we sought to identify stimulation conditions that can induce synchronized activities of a random network of in vitro cortical cultures. To avoid some of the model-specific features with point stimulation, we applied a uniform field with substrate-embedded electrode pair spanning the culture.

An in vitro cortical culture has a small world topology (Watts and Strogatz, 1998) characterized by dense local clustering of neighboring nodes (neurons) and a short path length (axon connections) between any pair of nodes. Since smallworld network architecture is widespread in biological neural networks, properties of the living neuronal network will have broad implications to other nervous systems (Bassett and Bullmore, 2006). The effect of stimulations (including electric and magnetic fields) on synchrony dynamics of nervous system has been extensively studied in mathematics and computational neuroscience (Yu et al., 2013). These theoretical studies have produced mechanistic insights regarding network dynamics, for example, chaos phase locking (Yu et al., 2011b), time-delayed feedback for synchrony suppression (Yu et al., 2013), simulated tACS effects (Ali et al., 2013; Frohlich, 2015). However, the insights gained from such modeling strategies can only be fully leveraged when used in conjunction with experimental approaches. Our goal is to develop a cell culture-based model with quantifiable and controllable population-wide synchronous activities to be used for testing network theories.

Population-wide analysis of neuronal activities requires the detection of families of neurons having a similar activity pattern, so that the original neuronal network can be decomposed into distinct clusters. With electrical recordings, algorithms are needed for detecting bursts and defining their attributes (e.g., duration) as unitary events, and for correlation analysis of the time series of bursts (Parodi et al., 1998; Chao et al., 2007; Sun et al., 2010). Alternatively, calcium live imaging can be used to monitor large populations of neurons within a field of view simultaneously. Synchronized calcium transients are direct result of propagation of bursts of action potentials that are generated periodically by in vitro cortical cultures (Robinson et al., 1993). When mapped onto the source neurons, calcium time series allow for direct comparison of the temporal and spatial patterns of neuronal activities. In this study, we developed computational analysis of calcium signals based on graph theory and network community detection to identify functionally correlated neuronal clusters. We tested a local greedy-optimization algorithm (Blondel et al., 2008)

to automatically determine the best partition of the neuronal population (i.e., number of communities and composition of each detected community) with minimal computational cost. Communities returned by the algorithm are entirely based on calcium signals and therefore capture a common behavior across neurons.

The experimental setup of this study built upon a previously developed neural-electric interface, consisting of dissociated cortical neurons growing on a silk fibroin film with embedded gold wires (Tang-Schomer et al., 2014a,b). Silk fibroinbased biomaterial has found extensive applications in neural engineering, as biosensors (Domachuk et al., 2009; Kim et al., 2010), neural probes (Kim et al., 2009; Tien et al., 2013), and for tissue engineering of the nervous system (Tang-Schomer et al., 2014c; Chwalek et al., 2015a,b). The silk film-based neuralelectric interface has shown evoked calcium influx of in vitro cultured rat cortical neurons by applied EF (Tang-Schomer et al., 2014a,b). This study examined cortical network activities under different stimulus patterns, varying in frequencies and directions, and evaluated temporal and spatial associations of neuronal populations. The findings revealed novel EF effects on random neural networks and provided guiding principles for control of network synchrony in vitro.

## MATERIALS AND METHODS

## Silk Film Supported Neural-Electric Interface

Silk films were processed as previously reported in (Tang-Schomer et al., 2014b). Briefly, silk fibroin (1–2%) solution extracted from Bombyx mori silkworm cocoons (Tajima Shoji Co., Yokohama, Japan) was prepared. A pair of gold wires (100 µm diameter, SPM Inc., Armonk, NY, United States) were positioned at 6 mm apart onto a PDMS mold (16 mm diameter) and immersed in silk solution. After drying in air, the silk film (∼5 µm thickness) was peeled off the mold with the gold wires embedded in the film. Films were rendered waterinsoluble by β-sheet formation via water annealing in a waterfilled desiccator for >5 h. To prepare for cell culture, the film was UV sterilized, coated with 20 µg/ml poly-L-lysine (Sigma-Aldrich, St. Louis, MO, United States) overnight, washed and dried prior to introducing cells.

#### Primary Cortical Neuronal Culture

The rat brain tissue dissociation protocol was approved by Tufts University Institutional Animal Care and Use Committee and complies with the NIH Guide for the Care and Use of Laboratory Animals (IACUC # B2011-45). Cortices from embryonic day 18 (E18) Sprague Dawley rats (Charles River, Wilmington, MA, United States) were isolated, dissociated with trypsin (0.3%, Sigma) and DNase (0.2%, Roche Applied Science, Indianapolis, IN, United States) followed with trypsin inhibition with soybean proteins (1 mg/mL, Sigma), centrifuged, and plated at 200,000–625,000 cells/cm<sup>2</sup> in neuro-basal media (Invitrogen, Carlsbad, CA, United States) supplemented with B-27 neural supplement, penicillin/streptomycin (100 U/mL and 100 µg/mL), and GlutaMax (2 mM, Invitrogen). Cultures were maintained in 37◦C, 100% humidity and 5% CO<sup>2</sup> in an incubator (Forma Scientific, Marietta, OH, United States) for up to 16 days in vitro (DIV 1–16). Cultures of DIV 14–16 were used for stimulation.

## Electrical Stimulation

The interface cultures were set up with extensions of the silk protein film-embedded gold wires connected to an electrical stimulator, as previously described in (Tang-Schomer et al., 2014b). The field potential was set at 160 mV between the electrodes and validated with an oscilloscope. A functional generator (Tenma Universal Test Center 72-5085, MCM Electronics, Centerville OH, United States) delivered biphasic, rectangular waves with frequencies ranging from 0.2 Hz to 200 kHz. Monophasic pulses (0.1 ms) were delivered by a Grass S44 stimulator and SIU5 stimulation isolation unit at frequencies ranging from 0.2 Hz to 2 kHz. A total of 12 cultures from six independent batches of cells (i.e., rats) were used. Voltage applied across each silk film was verified prior to stimulation with an oscilloscope. No cellular damage was observed during all our experiments, based on morphological characterization.

## Calcium Imaging and Image Analysis

Calcium dye Fluo-4 AM (Invitrogen) was used to visualize changes in intracellular calcium concentration. Calcium imaging with fluorescent calcium indicators is a reliable method to monitor action potential activities (Burnett et al., 2003), as intracellular calcium concentration rises transiently during electrical activity to levels that are 10–100 times higher (Berridge et al., 2000). Calcium imaging of bulk-loaded fluorescent indicators can be used to record the spiking activity of hundreds of neurons (Kwan, 2008).

Experiments were performed in controlled saline solution (CSS: 120 mM NaCl, 5.4 mM KCl, 0.8 mM MgCl2, 1.8 mM CaCl2, 15 mM glucose, and 25 mM HEPES, pH 7.4). Cultures were loaded with 1 µg/ml dye solution (in PBS containing 0.2% DMSO) at 37◦C for 30 min, washed with PBS, and incubated in fresh media for another 30 min. The cultures were mounted onto a confocal microscope (Leica TCS SP2, Leica Microsystems, Wetzlar, Germany) within an environmental chamber with the temperature maintained at 37◦C.

During stimulation, time-lapse fluorescence images were acquired with the same optical settings (at Ex/Em of 488/525 nm). For field stimulation, we imaged a 30-section z-stack every minute for 45–60 min. Time-series fluorescence images of one focal plane at the middle-point of a z-stack were used for image analysis. For pulse stimulation, images at a fixed focal plane were acquired every 10 s (i.e., 1t = 10 s) for 20–30 min.

NIH Image J software suite was used to quantify the fluorescence intensity. Circular selection was made for each cell body, and the mean fluorescence intensity was measured. A neuron's fluorescence intensity at a specific time point t (Ft) divided by the intensity at time 0 (F0, no stimulation) of the same neuron gave the calcium signal change and reported as Ft/F0.

## Network Analysis and Unsupervised Community Detection

For the analysis of the functional connectivity between neurons, a.k.a. network analysis (Newman, 2010), sample distribution of the fluorescence intensities at time 0 (F0) was estimated and the values of mean (µ0) and standard deviation (σ0) were computed. Each cortical culture was assumed to be representative of all cortical cultures, as a common practice with primary cortical culture-based studies. To further compare different cell cultures, we normalized each fluorescence measurement against the global average. Specifically, each fluorescence intensity time series F<sup>t</sup> was normalized by subtracting µ<sup>0</sup> and dividing by σ0. This normalization procedure aimed at preserving the range of fluorescence intensities observed across each cortical culture while removing time-series-specific biases. The normalized fluorescence intensity time series were then used to run the network partition algorithm described in section 3 and to identify functional clusters.

A local greedy-optimization algorithm was used to automatically determine the best partition of the neuronal population (i.e., the best number of communities and composition of each community) with minimal computational cost. We defined the communities returned by the algorithm as functional clusters as the neurons within the community had fluorescence time series with high degree of temporal correlation. Specifically, we envisioned each neuron in the culture as a node in a fully connected network, i.e., we assumed an edge between nodes i and j, for all i, j = 1, 2, 3, . . ., N, where N is the number of labeled neurons in the culture. For each pair (i, j), a weight wi,<sup>j</sup> was assigned to the edge between i and j, with wi,<sup>j</sup> being the Pearson correlation coefficient between the normalized fluorescence intensity time series estimated for neuron i and neuron j, respectively. The functional network is univocally defined by the weighted adjacency matrix (Newman, 2010).

$$\mathcal{A} = \begin{bmatrix} \mathbf{0} & \boldsymbol{\omega}\_{1,2} & \cdots & \boldsymbol{\omega}\_{1,N} \\ \boldsymbol{\omega}\_{2,1} & \mathbf{0} & \cdots & \boldsymbol{\omega}\_{2,N} \\ \vdots & \vdots & \ddots & \vdots \\ \boldsymbol{\omega}\_{N,1} & \boldsymbol{\omega}\_{N,2} & \cdots & \mathbf{0} \end{bmatrix} \tag{1}$$

which is a N × N symmetric matrix and has zeros on the main diagonal because no node forms edges with itself. We applied a static community detection algorithm (Newman, 2010) on matrix A to identify meaningful group structures in the neuronal network. A community is a set of nodes (i.e., cultured neurons) that are connected among one another more densely than they are to nodes in other communities, and nodes within a community may share similar structural or functional properties (Newman, 2010).

We used the Louvain algorithm (LA) (Blondel et al., 2008) to partition matrix A in communities. Briefly, LA identifies communities in a network by optimizing a quality function known as "modularity index"Q (Newman, 2010), which measures the density of edges inside the communities compared to edges between communities. Communities are estimated by comparison between the assigned network and a null model (Newman-Girvan null model) (Newman, 2010) and high modularity index values indicate large separation between communities. Because LA is a locally greedy optimization algorithm, we ran the community detection procedure for a total of 100 optimizations and used a consensus partition method (Lancichinetti and Fortunato, 2012) to obtain a consistent community partitioning in each network. After the functional clusters were determined, individual neurons were color-coded accordingly onto the original fluorescence image, to compare with their physical partitioning.

## RESULTS

An electrical field was imposed to the cortical neurons by a pair of substrate-embedded gold wires spanning the in vitro culture (**Figure 1A**). **Figure 1A-a** shows a biphasic, rectangular wave. **Figure 1A-b** shows the simulated EF distribution by the COMSOL software, as described previously in (Tang-Schomer et al., 2014b). The biphasic wave introduced EF of alternating polarity during the positive and negative phases of the wave function, at the rate of the wave frequency.

In conventional stimulation experiments, the parameters of stimulus (amplitude, frequency, duration) are determined by paring with intracellular recording of evoked responses of targeted neurons (Jayakar et al., 1992; Bagley and Westbrook, 2012). This study used voltage (160 mV across 6 mm) that showed frequency-dependent calcium responses of cortical neurons in a similar system (Tang-Schomer et al., 2014a,b). Our setup generated a theoretical EF strength of 27 mV/mm, above the threshold extracellular voltage gradient of 5–10 mV/mm for evoked neuronal response (Jefferys, 1995) (**Figure 1B**).

## Network Synchronization Under Alternating EF With Increasing Frequencies

When we monitored cortical cultures without stimulation for 10 min a time, no oscillatory calcium responses were found, and calcium signals fluctuated within 20% of the baseline level. Cortical cultures under alternating EF of a constant frequency (i.e., 2 and 10 Hz) also failed to produce synchronized activities. Only square waveform was tested in the study.

To our surprise, when biphasic, rectangular waves with field polarity alternating from 0.2 Hz to 200 Hz were applied, largescale, synchronized oscillations of cortical neurons were observed (**Figure 2**). **Figure 2A** shows the example of neurons stained with fluo-4, a calcium indicator, adjacent to a silk film-embedded gold wire electrode (field of view, 750 by 750 µm).

**Figure 2B** shows the stimulation protocol of alternating EF with increasing frequencies and the average calcium signal time series of the cortical culture. The experiment was conducted in a temperature controlled (37◦C) environmental chamber and lasted for less than 1 h. Stimulus was introduced at the third minute of live imaging and increased from 0.2 Hz by 10-fold a time to 200 kHz for 6 min per condition. The average calcium signal showed synchronous oscillations of approximately 15 min wave length (70 neurons measured).

cortical culture. (a) A biphasic rectangular wave. (b) COMSOL simulation of the electric field (160 mV peak-to-peak amplitude, 6 mm-apart electrode distance), as previously described (Tang-Schomer et al., 2014b). Field distribution at the positive and negative phases of the wave is shown. (B) Cluster analysis. (a) Labeled neurons n1, n2, . . ., n<sup>N</sup> are envisioned as the nodes of an all-to-all graph (right) and the connection between any two neurons n<sup>i</sup> , n<sup>j</sup> is weighted by the Pearson correlation coefficient wi,<sup>j</sup> between the correspondent fluorescence intensity time series (left). (b) Correlation-based weighted adjacency matrix A before (left) and after (right) sorting the neurons according to the community partition given by the Louvain algorithm (LA). Color-map reports the range of correlation coefficients.

the original fluorescence image of the culture. Neurons in Cluster 1 (non-responders) in white, and those in Cluster 2 (super-responders) in red. Scale bar, 100 µm.

The community detection algorithm was used to sort the neurons based on the statistical significance of the differences of their calcium signals, and two clusters were identified (**Figure 2C**). Neurons in the same cluster were highly correlated (i.e., Pearson's correlation coefficient >0.5) (yellow). Neurons belonging to different clusters were either poorly correlated (i.e., Pearson's correlation coefficient close to 0) or negatively correlated (i.e., Pearson's correlation coefficient close to −1) (blue).

Calcium signal time-series were then color-coded (**Figure 2D**) according to whether they referred to neurons in Cluster 1 (black) or Cluster 2 (red). Cluster 1 contained "non-responders" with calcium signals fluctuating close to the baseline level. Cluster 2 contained "super-responders" with calcium signal increases >5 folds of the baseline level.

When mapped onto the original image, the functional clusters found remarkable match with the neurons' physical partitioning (**Figure 2E**). Neurons belonging to the same functional cluster resided in close proximity to each other and separate from neurons belonging to the other cluster (Cluster 1, nonresponders, white; Cluster 2, super-responders, red).

## Entrainment of Sub-populations' Oscillations by Network Synchronization

By manual examination of calcium signal traces, we further divided the apparent non-responders into two groups, i.e., "modest-responders" with <5-fold signal changes but displaying synchronized activities and the rest as "noisy-responders." When mapped onto the original image, these sub-populations were found to belong to distinctive physical groups (**Figure 3A-a**): the super-responders and modest-responders as two separate neuronal aggregates (in red and white circles, respectively), and the noisy-responders consisting cells dispersed in the surrounding areas (arrows). **Figure 3A-b** displays representative images at specific time-points (in minute), demonstrating different fluorescence changes of neuronal sub-populations.

**Figure 3B** shows the average calcium time series from superresponders (red, n = 14, 20%), modest-responders (blue, n = 17, 24%), and noisy-responders (black, n = 39, 56%). The superresponders had peak signal levels of approximately 10-fold increases, compared to <2-fold changes of the other groups (**Figure 3B-a**). Further close examination of the low-amplitude signal changes of the modest-responders and noisy-responders revealed that they, too, exhibited synchronized oscillations (**Figure 3B-b**). Notably, all three sub-populations' oscillatory patterns were entrained by the network-level synchronous oscillation. The sub-population showed group-specific amplitude and phase patterns. For example, the modest-responders (blue) had opposite phase responses than the super-responders (red), i.e., peaks in one plot correspond to troughs in the other plot and vice versa. The noisy-responders' signal trace (black) had its major peaks in phase with other sub-populations but contained two smaller peaks.

#### Symmetrical Phase Changes and Dependence on Group-Specific Spontaneous Activities

To understand the phase differences between the superresponders and modest-responders, we examined individual calcium time series (**Figure 3C-a**: super-responder; **Figure 3C-b**: modest-responder). The traces showed remarkable synchrony within each sub-population. The synchronous oscillations of the two groups exhibited opposite phase changes.

We then focused our analysis on the initial period of the experiment, i.e., when the culture was switched from being unstimulated for the first 3 min to 0.2 Hz alternating EF for another 6 min. **Figure 3D** shows the average calcium signal time series of the first 6 min (a, super-responder; b, modestresponder). The two sub-populations had opposite activity trends prior to stimulation, with spontaneous calcium signal increases and decreases, respectively. There were significant differences of their signal levels at the third minute compared to the first minute. Upon stimulation, the different calcium responses continued their opposite trajectory that were further enhanced by the 0.2 Hz EF. Significant differences were observed in the signal level at 1-min post-stimulation (i.e., the fourth minute) compared to right before the stimulation (i.e., the third minute).

The above findings showed that a random network of cortical culture contained sub-populations of distinctive physical partitioning and endogenous activity levels. Alternating EF of increasing frequencies induced synchronization within each sub-population as well as across the entire network, while retaining group-specific oscillatory patterns. The binary response of activity-increase or decrease contributed to the opposite phase patterns of different sub-populations.

## Symmetrical Sub-population's Oscillatory Patterns Under Alternating EF With Decreasing Frequencies

Considering that applied EF of a constant frequency failed to induce network synchronization, we suspected that the context of EF frequency change was critical. We therefore conducted a different experiment, in which a different cortical culture was exposed to alternating EF of decreasing frequencies (**Figure 4**). We applied biphasic, rectangular waves (peak-topeak 160 mV) with frequencies starting from 200 kHz at

the third minute and decreased by 10-fold to 0.2 Hz for 6 min per condition. **Figure 4A** shows the fluo-4 stained neurons adjacent to a silk film-embedded gold wire; the wire was right blow the imaged area outside the field of view.

**Figure 4B** shows the average calcium time series (black dots) from 63 neurons measured. The mean activity level appeared to be mostly flat with a down-ward trend, with large variance of each data point. When we plotted the variance spread (**Figure 4B**, crosses), measured as the ratio of standard derivation to the mean, a dependence on EF frequency was found. The baseline variance of 26% decreased to 7% after 5 min of 200 kHz stimulation, maintained at approximately 12% during 20 kHz stimulation, and rose progressively as the frequency decreased, until reaching 102% at the end of the experiment (total time < 1 h). These results suggested that there were mixed responses of different subpopulations, and functional association of these groups depended on EF frequency change.

We used the community detection algorithm to automatically group the 63 neurons into two functional clusters (**Figure 4C**). Neurons within a cluster were highly correlated and poorly or negatively correlated to neurons in the other cluster, thus reflecting a marked functional separation between clusters. Color-coded calcium signal time series in **Figure 4D** revealed cluster-specific signal patterns that were previously obscured in the total average trace (**Figure 4B**). Cluster 1 (black) neurons had increased activity and Cluster 2 (red) neurons decreased activity. Notably, there was symmetry of plots between the two groups with peaks in one group corresponded to troughs of the other group.

## Suppression of Spontaneous Activity by High Frequency Alternating EF

To better understand the differences between Cluster 1 and Cluster 2, we mapped individual neurons onto the original image (**Figure 5A-a**, Cluster 1 in white. Cluster 2 in red). The functional clusters matched neuronal physical groups, as neurons belonging to the same cluster were in proximity to one other and separate from the other cluster. **Figure 5A-b** displays representative images at specific time-points, demonstrating different fluorescence changes of neuronal sub-populations.

**Figure 5B** shows the average calcium time series of the two clusters (Cluster 1, black. Cluster 2, red) and demonstrates groupspecific oscillatory patterns. Both clusters started with opposite spontaneous activity changes (increase versus decrease), had suppressed activities during 200 kHz and 20 kHz stimulation, and continued with opposite activity changes in terms of amplitude and phase patterns as the frequency decreased. The average time series in Cluster 1 and Cluster 2 had significantly different trends (two-way ANOVA test with cluster label and time as factors, P-value P < 10−<sup>10</sup> for the cluster factor). Moreover, for each frequency depicted in **Figure 5B**, we tested whether the responses of Cluster 1 and Cluster 2 were significantly different (Wilcoxon Rank-Sum test, P-value P < 0.05) and we found that the responses during stimulation at 200 Hz, 20 Hz, 2 Hz, and 0.2 Hz

were significantly different. Finally, we looked at the sample distribution of the correlation coefficient of the normalized fluorescence signals in Cluster 1 and Cluster 2 and we found that the distributions were significantly different (0.18 ± 0.60 versus 0.22 ± 0.88, Cluster 1 versus Cluster 2, mean ± S.D, Wilcoxon Rank-Sum test, P-value P < 10−10). The sample distribution of the correlation coefficient was uniformly distributed between −1 (anti-phase) and +1 (in-phase) for Cluster 1 while it was polarized around −1 and +1 (bimodal distribution) for Cluster 2, thus indicating a stronger level of intra-cluster correlation for Cluster 2.

We then examined the differences of the two sub-populations during the initial period of the experiment (**Figure 5C**), when the culture was switched from being unstimulated for the first 3 min to under 200 kHz alternating EF stimulation for another 6 min. Cluster 1 (**Figure 5C-a**) and Cluster 2 (**Figure 5C-b**) neurons had calcium signal increase of 15 ± 6% (n = 33, P-value P < 0.01) and decrease of 36 ± 10% (n = 30, P-value P < 0.01), respectively, at the third minute compared to the first minute. Upon stimulation of 200 kHz alternating EF, the opposite calcium signaling trends were attenuated, and both sub-populations headed toward the baseline level.

## Network Desynchronization Under Alternating EF With Decreasing Frequencies

By closer examination of each neuron's activity, we further manually divided the clusters into four groups based on similarities of their fluorescence time series. This manual process re-grouped the neurons with subjectively determined similarities of the calcium signals, with no exclusion or other assumptions, i.e., clusters 1a (n = 20, 32%), 1b (n = 13, 21%), 2a (n = 16, 25%), 2b (n = 14, 22%) (**Figure 6**). **Figure 6A-a** shows the general distribution of the sub-populations. Cluster 1a and 2a contained two well-separated neuronal aggregates. Cells interspaced in surrounding areas were contained in Cluster 1b and Cluster 2b. **Figure 6A-b** shows each cluster's average calcium signal time series. The variance at each data point remained consistent within each group in contrast to the highly variable total average response in **Figure 4B**, indicating similar intra-group but different inter-group signal patterns. All clusters showed suppressed activities under 200 kHz and 20 kHz stimulation. However, starting from 2 kHz, there was great divergence of activity trends with group-specific oscillatory patterns as the frequency decreased.

**Figure 6B** shows pair-wise comparison between the clusters of calcium signal time series. An opposite trend was observed between the sub-population-specific oscillatory patterns, as highlighted in gray (Wilcoxon Rank-Sum test, P-value P < 0.05). Cluster 1a showed phase symmetry (i.e., peak versus trough) and opposite activity changes (i.e., increase versus decrease) with Cluster 2a (**Figure 6B-a**) and Cluster 2b (**Figure 6B-b**) under all frequencies (2000–0.2 Hz). Cluster 1b showed symmetric phase and activity changes in specific frequency ranges, with Cluster 2b between 2000 and 20 Hz (**Figure 6B-c**) and Cluster 1a at ≤2 Hz (**Figure 6B-d**).

Taken together, these behaviors suggested a network desynchronization process. The initial globally suppressed network diverged into two groups, Cluster 1 and Cluster 2 with opposite activity trends and phase patterns. As the alternating EF frequency decreased, neurons in Cluster 2 split into subgroups of 2a and 2b with oscillations of synchronized phase patterns but different amplitudes. Neurons in Cluster 1 split into subgroups of 1a and 1b that initially had synchronized phase patterns and different amplitudes, but under further decreased EF frequency, exhibited opposite phase patterns.

## Lack of Synchronized Activity Under EF Without Polarity Change or Continuous Frequency Change

To examine the role of EF polarity in network synchrony, we designed a different set of stimulation experiments with monophasic EF of similar frequency changes as the alternating EF (**Figure 7**); different batches of cortical cultures were used. **Figure 7A** shows wave function comparison of biphasic EF and monophasic pulse trains of a fixed 0.1 ms pulse duration. The pulse train captured the initial moment of field potential change upon each stimulus at the same frequency as the corresponding biphasic waves. However, the pulse trains lacked field polarity change of the biphasic waves.

The pulse train was delivered at frequencies ranging from 0.2 Hz to 2 kHz for 3 min for each condition, and calcium fluorescence images were collected every 10 s. **Figures 7B,C** show fluorescence images of neurons (a) and corresponding calcium time series (b) under conditions of increasing frequencies and decreasing frequencies, respectively. In both scenarios, most of the neurons showed activity fluctuation within 20% of the baseline levels, and only selective neurons reported spiking activities as shown in **Figures 7B-b,C-b** (non-spiking activities were omitted). Statistical analysis of fluorescence intensity time series from individual neurons determined that neuronal activities in both scenarios were largely uncorrelated. Pearson's correlation coefficients between spiking neurons were close to 0, indicating that these neurons activated independently from one another.

In another set of experiments, we examined the role of frequency change by introducing a 3-min zeroing period (i.e., no stimulation) in-between frequency changes of alternating EF; frequencies were changed from 0.2 Hz to 200 kHz or vice versa in similar orders as previous experiments (**Figures 3**, **4**). Only a few random neurons showed spiking activities, and no synchronized oscillations were found (data not shown).

## Hypothesis of Coordinated Stimulation by Alternating EF

Based on these findings, we proposed a hypothesis of network synchrony control by applied EF of alternating polarity (**Figure 8**). Applied EF results in the polarization of the membrane of the nearby cells (Jefferys, 1995; Bikson et al., 2004; Radman et al., 2009). In general, neuronal elements are depolarized near cathode and hyperpolarized near anode. However, the spatial distribution of such polarization under a

FIGURE 6 | Network desynchronization under alternating EF with decreasing frequencies. (A, a) Functional clusters 1a, 1b, 2a, and 2b mapped onto the original fluorescence image of the culture under alternating EF with decreasing frequencies. Scale bar, 100 µm. (b) Average calcium signal time series of each sub-population. Legend: Cluster 1a, black; Cluster 1b, blue; Cluster 2a, red; Cluster 2b, yellow. (B) Pair-wise comparison of sub-population's calcium signal time series. (a) Cluster 1a versus Cluster 2a. (b) Cluster 1a versus Cluster 2b. (c) Cluster 1b versus Cluster 2b. (d) Cluster 1a versus Cluster 1b. Gray background highlights the symmetrical areas of the plots.

corresponding calcium time series (b) under pulse trains of decreasing frequencies. Only neurons with significant spiking activities are shown in the calcium time series and marked onto the corresponding images. Scale bar, 100 µm.

uniform EF is highly variable, depending on cell biophysics and morphologies (Bikson et al., 2004; Radman et al., 2009; Yi et al., 2017). By extending these concepts to a neuronal network, we hypothesized that different populations are depolarized under a same uniform EF, and that as the field polarity changes, the populations switch to the other activation state (i.e.,

populations. The initial high frequency (i.e., 200 kHz) stimulation suppresses all activities. As the frequency decreases, the timing between neuronal activation increases. Therefore, the subpopulations are less likely to fire together, resulting in divergent oscillation patterns of different amplitude (d versus e in the early stages, and f versus g) or different phase patterns (d versus e).

hyperpolarization versus depolarization). Therefore, biphasic EF would result in coordinated stimulation of neuronal populations.

As illustrated in **Figure 8B**, two populations (1 and 2) with different EF threshold are located at different distances from a nearby electrode; the other electrode would be too far away to impose direct effect. When the electrode is cathode (left), population 1 is activated (in green, + indicating depolarization) and population 2 non-activated (in white, − indicating no change or hyperpolarization). When the electrode turns to anode (right), population 1 is in-activated and population 2 activated. **Figure 8C** illustrates the resulting calcium transients upon neuronal activation (left for population 1; right for population 2). There would be a time delay of the population activation, as the inverse of two times of the EF frequency. Control of the frequency (**Figure 8D**) would provide a means to temporally associate or dissociate the two neuronal sub-populations' evoked activities.

**Figure 8E** illustrates the hypothesized network synchrony control by EF alternating frequency. In vitro studies of random cortical networks have shown that repetitive, time stimulation of loosely associated neurons can induce synchronized bursts of the neurons and their neighbors (Tateno and Jimbo, 1999; Shahaf and Marom, 2001). Increasing EF frequency would be analogous to repetitive stimulus with increasingly shorter timing. In addition, the wide range of frequencies could activate many sub-populations of different responsiveness. It would result in network synchrony (**Figure 8E**, left). The initial response to the applied EF would depend on neurons' endogenous activities, as shown in **Figure 3D**. Moreover, binary responses to EF (activity increase or decrease) would lead to symmetrical phase pattern, as shown in **Figures 3**, **5**. Therefore, group-specific oscillations with different amplitude or opposite phase patterns would be expected (**Figure 8E**, left, a, b, c).

Conversely, decreasing EF frequency could dissociate the endogenous activities of different neuronal sub-populations (**Figure 8E**, right). High frequency EF is known to suppress neuronal activities (Wagenaar et al., 2004; Chao et al., 2005; Birdno and Grill, 2008), also shown in our studies with the initial 200 kHz stimulation (**Figure 5C**). As the frequency decreases, the timing between neuronal activation increases, and the subpopulations are less likely to fire together, resulting in divergent

oscillatory patterns. Population-specific responsiveness could be the different amplitudes, for example, **Figure 8E-f** versus g, or different phase patterns as d versus e.

## DISCUSSION

We presented results of the behavior of a random cortical network under applied electrical field. Each neuron's activity was captured by calcium live imaging and matched to its physical location in the network. Calcium signal time series were subjected to cluster analysis for unbiased detection of neuronal communities of similar activity patterns. Spatial and temporal associations of neuronal activities revealed large scale, synchronized oscillations of a random network under alternating EF of changing frequencies. EF without polarity change or frequency change failed to produce synchronized activities among neurons. These findings formed the basis of a hypothesized network control mechanism, involving coordinated stimulation of different sub-populations by alternating field polarity. Change of EF frequency was critical for control of the time delay of group-specific activities, by associating or dissociating different sub-populations via frequency increases or decreases, respectively. These novel EF effects on random neural networks provide important understanding of network synchrony underlying brain functions and neuromodulation applications.

## Neural Network Manipulation and System Setup

A thin silk fibroin-based film with embedded gold wires provided the interface system for in vitro cortical cultures. Compared to rigid MEA substrates, the flexible and transparent silk film provides greater ease and superb compatibility with in vitro neuronal cultures (Tang-Schomer et al., 2014b) as well as in vivo brain implants (Kim et al., 2010; Tang-Schomer et al., 2014c). The wire embedding method simplifies interface fabrication compared to the lithographic process for surface electrodes (Tang-Schomer et al., 2014a), with excellent interface stability requiring no additional adhesives or bonding. Regarding signal transmission, the thin silk film (∼5 µm) poses no significant barrier (>90% conductivity) (Hronik-Tupaj et al., 2013). The gold wire provides double layer capacitive charging (Brummer and Turner, 1977) and modifies the ionic composition near the electrode. By applying charge-neural biphasic field, potential pH buildup at the electrode-solution interface would be eliminated and field propagation increased at high frequencies (Wagenaar et al., 2004; Graves et al., 2011). These features support the use of silk film-based neural-electric interface as a suitable system for investigating EF effects on neural networks.

Sorting activities onto source neurons and grouping them based on common behaviors are not trivial tasks (Morin et al., 2005). Recorded electrical signals have superior temporal resolution allowing for temporal correlation analysis, for example, the delay between stimulus and the first evoked pulse. Temporal correlation of these signaling events forms the basis for inferring functional association of distributed neuronal populations. In comparison, the temporal features of calcium signals are less sharp (Robinson et al., 1993), and slow fluorescence imaging further limits the temporal resolution. In this study, we used confocal 3D imaging to maximize captured neurons that took almost 1 min for each z-stack. The slow sampling rate precluded us from examining fast events. Faster imaging in future could allow for detailed temporal analysis. Nevertheless, calcium imaging provides undisputable spatial resolution and allows for the signal trace to be mapped to the source neuron. The individually traceable time series have provided a multi-dimensional picture of the network dynamics for each cell at each time point.

We used community detection in functional networks for the unsupervised identification of neuronal communities that, within a given culture, exhibit homogenous fluorescence-based discharge patterns. Community detection is an established area of network analysis (Newman, 2010) and it has been recently used to unravel structural and dynamical properties of complex neuronal networks such as the epileptogenic brain network in patients with drug-resistant epilepsy (Khambhati et al., 2015), circadian-clockrelated networks of neurons in the suprachiasmatic nucleus (Park et al., 2016), and networks of ganglion cells from retina (Billeh et al., 2014). Community detection algorithms, though, are typically applied to large (i.e., more than 1000 nodes) networks while the LA was used in our study on small-size (i.e., up to 70 nodes) neuronal networks. As the size of the network grows, however, the community detection remains feasible. Locallygreedy, resolution-adaptive algorithms (Bassett et al., 2013) and null models (Newman, 2010) are available to guarantee fast neuron clustering, while avoiding the detection of spurious and statistically nonsignificant communities.

## Point and Distributed Electrical Fields for Network Stimulation

Point-source pulse stimulation is the most commonly used modality in neurophysiology studies. Specific stimulation frequencies have been associated with functional responses, for example, hippocampal resting activity (5 Hz), long-term potentiation (LTP, 100 Hz, 1–3 s), long-term depression (LTD, 0.5–5 Hz for 5–30 min), or homeostatic synaptic depression (3 Hz, 12–24 h) (Larson et al., 1986; Staubli et al., 1999; Malenka and Bear, 2004; Goold and Nicoll, 2010). However, it is unclear how the parameters developed with defined CNS pathways can be applied to a random network of in vitro cortical culture. In fact, the wide range of frequencies tested in our experiments failed to produce correlated activities of among neurons in culture.

Studies of in vitro cortical networks showed that evoked responses depend on the neuron's endogenous activities (Jimbo et al., 1999; Wagenaar et al., 2005) and that time varying stimulus is more effective in inducing bursting spikes of neuronal ensembles (Shahaf and Marom, 2001; Bakkum et al., 2008b). A series of studies by Jimbo and colleagues reported long-lasting (∼30 min) binary responses of stimuli-induced, large-scale (ensemble) changes in connectivity (Maeda et al., 1995; Jimbo et al., 1999; Tateno and Jimbo, 1999). It was shown that for

a given site of tetanic stimulation, all the activated neurons either increase their responsiveness (potentiation) or decrease their responsiveness (depression) to the stimulus (Jimbo et al., 1999). Our results are consistent with these reports as we showed that (i) there were sub-population-specific responses to the same stimulus, (ii) the initial evoked responses were dependent on group-specific endogenous activities prior to the stimulation, and (iii) the evoked response was binary (i.e., activity increase or decrease) upon stimulation. In addition, our study showed that time varying frequency, but not constant frequency, produced synchronized network activities.

However, evoking network synchrony with point stimulation would require pre-selecting a site for stimulation, matching the initiating stimulus with the selected neuron's responsiveness, and tailoring stimulus time series for each affected neuron (or ensembles) in the network. These tasks would be daunting, if not impossible, for a random network. Alternatively, distributed EF stimulation used in our study would allow different subpopulations to be activated simultaneously. Although speculative at this stage, it is worth noticing that if stimuli-induced changes are operated under the pathway-specific principle as suggested by Jimbo et al. (1999), then group-specific responses would be paced by network-level changes. Hence, intrinsic activity fluctuations would be expected to ride along a slower wave of network oscillation. Indeed, in both scenarios of alternating EF stimulation in our study, the different group-specific calcium signal time series showed oscillatory patterns in synchrony with one other at a time scale (tens of minutes) much longer than previously reported neuronal activities (i.e., milliseconds). The oscillatory patterns were not precisely aligned with the temporal changes of stimulus, in part due to the crude temporal resolution (in minutes) used in the study. Nevertheless, it is interesting to note that the network-level oscillation had a wave length of approximately 15 min, about one round of frequency changes of 6 min per frequency. Focal stimulation studies showed that periodic stimulus can be used to phase lock bursting activities of a local network (Maeda et al., 1995; Darbon et al., 2002). Accordingly, our results imply that the network may not only respond to EF frequency and duration, but also to the change of EF frequency over a longer time scale.

## Network Synchrony Under EF of Alternating Polarity at Changing Frequencies

The most interesting finding of this study is control of network synchrony with EF of alternating polarity and changing frequencies. Field polarity change was found to be essential for network synchronization, as monophasic fields of the same frequency changes failed to produce correlated activities among neurons (**Figure 7**). Other systems have shown that temporal coordination of distributed neuronal activities establishes network synchrony (Singer, 1999). We hypothesized that the alternating field polarity could introduce a time delay of half period of the biphasic wave, and therefore, temporally coordinate the stimulation of different neuronal sub-populations in a network (**Figure 8**). As illustrated in **Figure 8B**, this hypothesis assumes that different sub-populations are activated according to the nearby electrode's status as cathode or anode. In vitro cortical cultures consist of many neuronal types with a wide range of sensitivities to EF as low as 5 mV/mm (Jefferys, 1995; Bikson et al., 2004). In general, neuronal elements are depolarized near cathode and hyperpolarized near anode. However, the spatial distribution of such polarization is modified by a neuron's complex morphology, summation of which would lead to either somatic depolarization or hyperpolarization (Yi et al., 2017). Therefore, it is reasonable to assume different activation state of sub-populations under a uniform EF, which depends on neuron-specific features. Neuronal sensitivities to different stimulus shapes have been examined in studies by Wagenaar et al. (2004) by using MEAs in cortical cultures. It was found that the transition between the positive and negative phases is the most effective stimulus compared to other pulse shapes. It is possible that different sub-populations have different sensitives to the phase transition rate and are activated at different time points throughout the wide range of frequency change span.

Another key factor is change of polarity alternating rate, or the time differential of the EF frequency. Alternating EF of a constant frequency did not produce correlated activities among neurons, neither did introducing resting-periods in-between EF frequency changes. These results suggested that change of EF frequency was necessary for inducing large scale, synchronized activities of neurons. The hypothesis of coordinated stimulation by EF frequency-dependent time delay could explain these findings (**Figures 8D,E**). Increasing the frequency of the biphasic wave would increase temporal correlation of the activation of different sub-populations. Conversely, as the EF polarity change rate decreases, different sub-populations are less coordinated, resulting in more divergent activities with group-specific oscillatory patterns. Neuronal sensitivity to stimulation frequency are widely reported to affect neural network activities, including adaptation (Eytan et al., 2003), phase-locking (Leondopulos et al., 2012), conduction block (Kilgore and Bhadra, 2004), and rhythm modulation (Birdno and Grill, 2008). This is the first report of network response to continuous frequency change.

It seems unnecessary to increase the frequency up to 200 kHz, as high frequency stimulation is known to pace networks to refractory state (Chao et al., 2005; Wagenaar et al., 2005). When we designed the study, we chose to test the broadest range of frequencies that our stimulator can provide. We had also reasoned that the higher frequencies beyond the membrane polarization threshold would, essentially, act as a direct constant field. Our results showed that the synchronized activity persisted at higher frequencies. Given the abovementioned literature, it is unlikely that these activities are direct result of high frequency stimulation. According to our hypothesis, these behaviors would depend on the history of frequency change. Further studies focusing on the regime of 2–200 k will be needed to clarify the input/output correlations and determine the upper limit of EF frequency. It is possible that the induced synchronous oscillation could persist after reaching a frequency threshold. It will also be important to test a wider range of parameters, including frequency range, duration, order of frequency change, etc.

The study had focused on neurons adjacent to the electrode (within 750 µm). Less response would be expected of the neurons in distant areas as there would be little voltage gradient in the middle of the culture. At present, it is unclear whether synchronous oscillation near the electrode can propagate to other part of the network. Detailed mapping of neuronal communities in relation to field polarity and strength will provide insights on how the network communicates changes.

## Implications for Functional Modulation of Neural Networks

Synchronous oscillatory activity in the cerebral cortex plays a crucial role in implementing complex brain functions (e.g., memory, cognition) as well as encoding information (Buzsaìki, 2006). Numerous studies, both in vitro and in vivo, have focused on the mechanisms that sustain oscillations and their synchronization as well as on the relationship between neural oscillations and network dynamics, e.g., for a review, see (Buzsaki and Draguhn, 2004). Abnormal increments in synchronization are reported as a key component in chronic neurological disorders, e.g., Parkinson's disease and epilepsy, and in the impairment of decision-making capabilities (Ross et al., 2013; Tan et al., 2013; Broggini et al., 2016; Cao et al., 2016). Our study demonstrates that widespread oscillations can be induced in a neural population in vitro by using a coordinated electrical stimulation paradigm with biphasic rectangular waves. Our solution may be used to recreate oscillatory conditions in vitro with a fine spatial resolution. The system provides an easy-to-use testbed for reproducing pathological oscillatory activities in large neural populations as well as studying the effects of exogenous inputs (e.g., chemical compounds or novel neuromodulation approaches) on neural oscillations.

Furthermore, noninvasive brain stimulation with electrical or electromagnetic waves provide effective neuromodulation interventions for treating a range of neurological and psychiatric disorders, including deep brain stimulation (DBS) for Parkinson's, essential tremor, and dystonia (Gross and Lozano, 2000; Ferrucci et al., 2008), transcranial direct or alternating current stimulation (tDCS or tACS) therapies for Alzheimer's disease (Ferrucci et al., 2008) and stroke (Hummel et al., 2005) and transcranial magnetic stimulation (TMS) for depression. In particular, tACS applies a weak sine-wave electric current to the scalp to identify a cortical oscillation pattern associated with a specific aspect of cognition or brain function and then to apply frequency-matched stimulation with concurrent assessment of changes in the targeted behavior (Frohlich, 2014). The choice of stimulation frequency has relied on the simple assumption that the stimulation frequency applied is the frequency that will induce or enhance in the network, with an implicit assumption of the linearity of the stimulated system. However, there is little reason to assume that the interaction of periodic stimulation with endogenous cortical network dynamics follows the same rules (Frohlich, 2015). Indeed, the cortical network used in the study demonstrated sub-population specific oscillatory patterns that were susceptible to entrainment by applied EF of a wide range of frequencies. These findings suggest that varied EF with controlled polarity and frequency changes could be a more effective means for neuromodulation than a paradigm with fixed polarity and frequencies. For example, applied alternating EF with increasing frequency may induce entrainment of different endogenous oscillators to synchronize activities of different brain areas. Conversely, alternating EF with decreasing frequency starting from refractory high levels may be used to suppress undesirable synchronized activities in some pathological conditions, and "re-tune" the activities back to the level of background oscillations.

However, in order to translate the study's findings into effective neuromodulation application, specific parameters need to be identified, for example, the upper limit of the EF frequency and time variants of frequency change. In particular, the study only examined square waves and the effects of other waveforms such as sinusoidal inputs are unknown; though other studies have suggested different entrainment properties on cortical oscillations by distinctive waveforms (Hutt et al., 2018). Cell culture-based models provided by this study and other systems (Frohlich, 2015), combined with computational and mathematical simulations (Hutt et al., 2018), will be powerful tools to test different neuromodulation paradigms.

## Limitations of the Study

A major limitation is that different stimulation paradigms were tested with different cell cultures, due to technical constraints of calcium dye use and live imaging, and to avoid potential residual effects of serial stimulations; therefore, direct comparison of different stimulation protocols was lacking. We had made assumption that each cortical cell culture is representative of all mixed cortical cultures, as a common practice with primary cortical culture-based studies. For our analysis, we normalized the fluorescence signals with the global average and standard derivation to prevent culture-specific biases. In addition, the hypothesized mechanism does not rely on neuronal cell composition or network topology that are the major variables between different cultures. Another limitation is the crude temporal resolution due to the 1 min imaging interval (due to slow z-stack confocal imaging) that prevented us from making more precise correlations between the stimulation conditions with neuronal responses. Future studies need faster imaging protocols.

### AUTHOR CONTRIBUTIONS

MDT-S designed the study and performed the experiments in the laboratory of David Kaplan at Tufts University. MDT-S analyzed and interpreted the data at UConn Health and wrote the manuscript at the Jackson Laboratory for Genomic Medicine. TJ assisted with image processing. SS performed the computational analysis and contributed to the preparation of the manuscript.

## FUNDING

TJ was supported by the UConn School of Medicine Summer Research Fellowship. This work was partially supported by the Connecticut Institute of Brain and Cognitive Sciences (IBACS) through Seed Grant 024 (to MDT-S and SS) and the Connecticut Children's Medical Center Strategic Fund (to MDT-S).

### REFERENCES


#### ACKNOWLEDGMENTS

We thank Dr. David Kaplan for support of the work and review of the manuscript. We thank Mark Schomer for his thoughtful comments. We also thank Dr. Steve Moss's laboratory at Tufts University for providing embryonic rat brain tissues.




**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Tang-Schomer, Jackvony and Santaniello. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Rectifying Resistive Memory Devices as Dynamic Complementary Artificial Synapses

#### Dan Berco\*

School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore, Singapore

Brain inspired computing is a pioneering computational method gaining momentum in recent years. Within this scheme, artificial neural networks are implemented using two main approaches: software algorithms and designated hardware architectures. However, while software implementations show remarkable results (at high-energy costs), hardware based ones, specifically resistive random access memory (RRAM) arrays that consume little power and hold a potential for enormous densities, are somewhat lagging. One of the reasons may be related to the limited excitatory operation mode of RRAMs in these arrays as adjustable passive elements. An interesting type of RRAM was demonstrated recently for having alternating (dynamic switching) current rectification properties that may be used for complementary operation much like CMOS transistors. Such artificial synaptic devices may be switched dynamically between excitatory and inhibitory modes to allow doubling of the array density and significantly reducing the peripheral circuit complexity.

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Darsen Duane Lu, National Cheng Kung University, Taiwan Tara Julia Hamilton, Macquarie University, Australia

> \*Correspondence: Dan Berco danny.barkan@gmail.com

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 26 July 2018 Accepted: 01 October 2018 Published: 22 October 2018

#### Citation:

Berco D (2018) Rectifying Resistive Memory Devices as Dynamic Complementary Artificial Synapses. Front. Neurosci. 12:755. doi: 10.3389/fnins.2018.00755 Keywords: artificial neural networks, brain inspired computing, dynamic artificial synapses, memristors, rectifying synapses

#### INTRODUCTION

Ever since the scientific community's revival of interest in memristors (Chua, 1971; Chua and Kang, 1976) was triggered by publications in the last decade such as Strukov et al. (2008), these devices have been extensively used for the implementation of artificial neural networks (ANN) in braininspired computational platforms. Within this domain, crossbar array architectures are promising candidates for achieving high-densities (∼10<sup>15</sup> bits/cm<sup>2</sup> ) similar to the human cerebral cortex (∼10<sup>14</sup> synapses), when configured in 3D stacking (Kügelera et al., 2009), due to the nanoscale device dimensions (Aratani et al., 2007; Jo et al., 2010). In addition, crossbar arrays are very efficient, in terms of calculation time and energy expenditure, when performing matrix-vector dot product operations, that form the basis for machine learning algorithms (Hu et al., 2012). Several demonstrations of resistive random access memory (RRAM) array implementations proved to be very successful in tasks such as image classification (Prezioso et al., 2015). The sizes of these networks range from small scales of only few neurons (Kim et al., 2012; Prezioso et al., 2015), medium levels (Park et al., 2015; Hu et al., 2016) and up to larger scale that incorporate hundreds of neurons (Yao et al., 2017) and even up to 10<sup>5</sup> synaptic connections (Burr et al., 2015).

Resistive random access memory arrays are designed to imitate the functionality of biologic synaptic networks. A chemical synapse is a gapped connection between two neurons through which communication takes place (Pereda and Faber, 1996; Kandel et al., 2000;

Nicholls et al., 2001). A typical neuron can have several thousands of synapses which mostly connect axons in a presynaptic neuron to dendrites in postsynaptic neuron. Inter-neural signaling occurs by the release of neurotransmitters from the presynaptic neuron into the gap (i.e., synaptic cleft) that in turn is collected by receptors in the postsynaptic neuron. The molecular neurotransmitters are kept in sacs called synaptic vesicles. During signaling, these vesicles are released into the synaptic cleft and bind to receptors on the postsynaptic neuron. Once the signal is delivered, the transmitters are evacuated from the receptors through potential mechanisms such as enzymatic degradation, or absorbed back into the presynaptic neuron by specific transporters. The postsynaptic potential response is classified as being either excitatory or inhibitory and determined by the type of neurotransmitter (Glutamate or γ-aminobutyric acid) (Nakanishi, 1992). Two key characteristics resulting from this behavior are the so called long-term potentiation (LTP) and long-term depression (LTD) and the synaptic weight (connection strength) is modulated by this neural activity. A recent study showed that actually both types of neurotransmitter could be released simultaneously during synaptic activity (Root et al., 2014). Moreover, neurotransmitters have been shown to be able to actually exchange roles during early stages of brain development (Ben-Ari, 2002).

Biologic neural networks have evolved over hundreds of millions of years to easily and efficiently perform tasks that state of the art computers find difficult. The operation of a man-made ANN should thus be true to the source as much as possible with respect to a building block artificial device. In order to achieve this target, artificial synaptic devices (ASDs) should imitate as much as possible the traits that biological synapses have. Although current understanding of neural networks and synapses is nowhere near complete, one may assume that such ASDs would be the best option for future ANN implementations. Some of these basic features include having a large dynamic range and multilevel operation to match the analog nature of the biological synaptic weight changes. These traits translate to higher accuracies, more degrees of freedom for weight adjustment and robustness during network training. A positive correlation exists between device dimensions and the number of states it can support (e.g., multilevel resistance) (Kuzum et al., 2013). However, sizing up the device will increases the overall current and power consumption as well as reduce the potential density.

The potential ability to implement symmetric weight changes may play a role in the simplification of an ANN peripheral control system. Biologic neural networks are very adaptive and can easily compensate for asymmetric weight changes especially when hundreds of neurons are involved in determining the weight of a synaptic junction. State machine based control systems on the other hand, are best suited to operate with well known and predictable parameters. In order to deal with asymmetric or random parametric distributions, an elaborate feedback system must be implemented and incorporated into the controller. In this sense, an ASD having a potential for incrementally small changes in both the up and down directions may simplify the controller design. A desirable corresponding weight parameter (e.g., conductance in RRAM) should thus be both symmetric with regards to the direction of change (increase or decrease) and differentially linear in magnitude of step change. Nonetheless, RRAMs (being a promising candidate in terms of low-power and high-density) (Wong et al., 2012) and other types of memristors show non-linear conductance as well as asymmetric conductivity changes (Lee et al., 2010) in response to successive set and reset pulses (Alibart et al., 2012; Chen et al., 2015; Wang et al., 2016) that complicate the task of designing a control state machine and sensing circuitry. This in turn affects both the potentially achievable network accuracy and overall performance.

An even more critical attribute expected from an ideal ASD would be the ability to reconfigure dynamically during real time operation between the excitatory and inhibitory response modes in a similar manner to a biologic synapse. Current RRAM devices are not able to reproduce this feature since they are passive devices by nature (once formed) and intentionally operated to comply with the linearity requirements as much as possible. It is virtually impossible for an RRAM to display this dynamic attribute without adding an additional control terminal to modulate the material properties by the field effect. In other words, it should have a bipolar conductance that is distinct characteristic of active devices. Adding a control terminal would damage the linearity and seriously downgrade the highintegration capabilities.

A common architectural solution is used to work around the dynamic reconfiguration issue by employing a differential approach. Instead of representing a synaptic gap by a single RRAM, two devices are used in a differential manner (Bichler et al., 2012; Prezioso et al., 2015). In this way, the synaptic weight-current is evaluated through a differential amplifier to determine whether the synaptic gap represents an excitatory or an inhibitory state. Needless to say, the prospective array density is reduced to half in addition to the added complexity required from the control and sensory circuitry. Recent publications have demonstrated ASDs with dynamic capabilities (mimicking either LTP or LTD behaviors as a function of a modulation bias) through the use of FET structures (Kim et al., 2013; Tian et al., 2015, 2016, 2017; Yang et al., 2015). As mentioned previously, these devices rely on at least one additional modulation terminal and an associated bias voltage to control the conductance polarity through the field effect. However, they are operated using very large biases (tens of volts) that seriously compromise their integration possibility with modern CMOS architectures. Moreover, the physical dimensions of these devices are very large (tens of micrometers), in addition to having a lateral structure (as opposed to vertical stacking) that is not optimal for high-density 3D integration. State of the art ASDs are thus still far from being able to truly reproduce the behavior of biologic synapses.

Another type of biologic synapses is based on fast conductive links between neurons capable of transmitting and receiving electrical signals (Pereda and Faber, 1996; Kandel et al., 2000). These links contain numerous ion channels (i.e., connexons) scattered along plasma membranes to form the connection between the pre- and post-synaptic neurons (Connors and Long, 2004; Michael et al., 2004). Some of these junctions demonstrate a rectifying behavior while being stimulated by electric pulses

resulting in a preferred direction for ion flow (Hormuzdi et al., 2004; Landisman and Connors, 2005; Haas et al., 2011). In this light, another subclass of micrometer sized RRAMs was shown to have a unidirectional current rectification property along with non-volatile, multilevel resistive states (Yoon et al., 2015; Kim et al., 2016). In addition, Kim et al. (2018) demonstrated a rectifying micrometer ASD with transient (volatile) currentvoltage dependence. Unfortunately, all these devices fall short in terms of both size and high-density integrability that play a key role in the implementation of ideal ANN.

Recently, Berco et al. (2018) presented a proof of concept for a nanoscale current rectifying dynamic RRAM. This ASD, of merely a few nanometers in size, dissipates only several picowatts of power during operation while having the ability to dynamically flip its current rectification direction thus effectively implementing both excitatory and inhibitory synaptic functionalities without the need for a modulation terminal. This operation mode allows for doubling the array size when compared to the common differential approach discussed previously (Bichler et al., 2012; Prezioso et al., 2015). Conductance-weight changes may be implemented by using a digital methodology (as opposed to the ubiquitous analog model). In this manner, a group of nanoscale ASDs are grouped together to represent a single artificial synapse where each member of the group plays the role of a rectifying connexon at the expense of layout resources. Setting the number of current-rectifying devices to a specific direction, being either positive rectification (PR) or negative rectification (NR), in relation to the others (being in the opposite direction), effectively determines the total conductance for each rectification direction. These artificial rectifying connexons (ARCs) may be individually toggled and the overall synaptic weight digitally manipulated in a similar manner to the LTP and LTD in biologic synapses.

## ARTIFICIAL SYNAPTIC DEVICE IMPLEMENTATION

Memristor arrays for hardware implementations of ANN are extremely efficient in performing matrix vector dot products as weighted-sum operations. A widely used RRAM-based differential array architecture is depicted in **Figure 1A** (Bichler et al., 2012; Prezioso et al., 2015). In this approach, two passive devices are used to determine a single synaptic weight in a differential manner thus allowing for both positive and negative parametric values. Programming the analog conductance levels may be done either during network training or on the fly to emulate the LTP and LTD synaptic behaviors. The sensing circuitry is based on a differential amplifier driving an activation function module (marked as f). However, analog RRAM operation (conductivity adjustment) (Lee et al., 2010; Alibart et al., 2012; Chen et al., 2015; Wang et al., 2016) usually requires complex pulsing schemes to account for its non-linear nature which complicates the design of a digital controller. An ARCbased implementation is given in **Figure 1B** (Berco et al., 2018). The operation principal allows for dynamically switching of

FIGURE 1 | (A) A common approach for emulating LTP and LTD using a memristors crossbar array as an ANN weight matrix (Prezioso et al., 2015; Bichler et al., 2012). Two memristors (W1a,1 and W1b,1) represent a single artificial synapse and their induced current, based on a pre-programmed conductance, is summed in a differential manner to determine the synaptic potentiation or depression. (B) An ARC-based implementation can both double the array density and simplify the peripheral circuitry by allowing dynamic switching of the rectification direction thus implementing either an excitatory or inhibitory weight parameter at each junction (much like a CMOS gate shown in the inset). (C) Sample response of an ARC-based ANN to a generic input (1,1,0). (D) A different input (1,0,0) would produce a different response from the same network setting.

the rectification direction resulting in either an excitatory or inhibitory weight parameter at each junction. In this manner, a single active device can be set to either push or pull current (much like the CMOS couple shown in the inset) effectively doubling the prospective density and simplifying the peripheral circuitry considerably.

A sample response of an ARC-based ANN to a generic input (1,1,0) is demonstrated in **Figure 1C**. In this example, some ARCs are configured as a PR-ARC (current direction marked by a red arrow) while the others are configured as an NR-ARC (current direction marked by a blue arrow). A high-voltage input "1" will result in current flowing into the output neuron through the PR-ARC. In the same manner, a low-voltage input "0" will result in current flowing out of the output neuron through the NR-ARC. ARCs that are configured in an opposite direction (in relation to the input value) will produce a zero current response. **Figure 1D** depicts the response of the exact same network configuration to a different input vector (1,0,0). In this manner, both network training and learning may be implemented.

**Figure 2A** depicts a schematic diagram of a lateral structured RRAM device configured as an ARC for illustrative purposes. The device is comprised of a metal-oxide-based resistive switching layer (RSL) placed between two conductive electrodes. The forming of a conductive filament (CF, an aligned path of current conducting defects) is done using a specific current compliance limit that yields an uneven distribution of oxygen vacancies (OV) and oxygen species (O) (Berco and Tseng, 2016; Berco et al., 2018). The figure depicts positively charged OV being pushed away from the anode and accumulated near the cathode while the negatively charged O ions are drawn to the anode. The forming process is arrested by the current compliance setting before a continuous CF from anode to cathode is able to form. The resulting gap (indicated by a blue arrow) yields a current rectifying behavior (Berco et al., 2018).

The dynamic nature of ARCs may be utilized for real time modulation of the ASD junction plasticity, using a plurality of devices, by changing the ratio of the number of PR to NR-ARCs. A digital approach (contrary to the common analog treatment)

FIGURE 3 | Implementation of logic gates with ARCs. The directionality of the ARC determines either a push or pull functionality and as a result the logic output value. (A) OR gate. (B) AND gate.

(Lee et al., 2010; Alibart et al., 2012; Chen et al., 2015; Wang et al., 2016) using ARCs for implementing LTP and LTD synaptic weight adjustment is summarized in **Figure 2B**. Using this concept, a group of ARCs are treated as a single ASD (consuming more area). This implementation was verified with Spice simulations using a behavioral model based on the experimental data published by Berco et al. (2018). The circuit under simulation is composed of 10 ARCs connected in parallel to a single DC voltage source (**Figure 2B** inset) representing the input value (either "0" or "1"). The LTP simulation progresses by consecutive flipping an NR-ARC to a PR-ARC starting from n = 1 to n = 10 and calculating the overall conductance under positive bias of 0.3 V. The LTD simulation is done by consecutive flipping a PR-ARC to an NR-ARC in the same manner under a negative bias of −1.5 V. The conductance results in **Figure 2B** show a good linear behavior when depicted as a function of the ratio of PR-ARCs (n) to NR-ARCs (10−n) and vice versa. The positive slope for PR may thus be used to implement an excitatory synaptic weight change (increased current flow to the postsynaptic circuitry for a positive input vector) while the negative slope for NR may be used for an inhibitory synaptic weight (increased current flow from the postsynaptic circuitry for a negative input vector).

#### ABSTRACT MODEL

**Figure 2C** depicts a block diagram of an abstract model for synaptic operation based on n ARCs connected in parallel as proposed by Berco et al. (2018). The number of PR-ARCs in the group is marked as nPR and of NR-ARCs as nNR. GPR is the combined conductance of the PR-ARCs and GNR of the NR-ARCs. The conductivity of a single PR-ARC is G<sup>u</sup> and G<sup>d</sup> of a single NR-ARC. Both parameters may be modeled after experimental data by using a behavioral lookup table. The synaptic model transitions from an excitatory state to an inhibitory one once the positive conductivity surpasses the negative conductivity and vice versa. The synaptic weight change 1G corresponds to spiking timing alignment in biologic synapses and is determined by network training and operation. In this manner, the number of ARCs which are flipped from a NR state to a PR one is determined by the ratio 1G/G<sup>u</sup> for LTP. In a similar way, the number of ARCs which are flipped from a PR state to a NR one is determined by the ratio 1G/G<sup>d</sup> for LTD.

## LOGIC GATE IMPLEMENTATION

ARCs may also be used for the implementation of logic gates as depicted in **Figure 3**. Programming entire crossbar arrays could

### REFERENCES

Alibart, F., Gao, L., Hoskins, B. D., and Strukov, D. B. (2012). High precision tuning of state for memristive devices by adaptable variation-tolerant algorithm. Nanotechnology 23, 75201–75220. doi: 10.1088/0957-4484/23/7/075201

be utilized in this manner to implement in-memory computing schemes. **Figure 3A** shows the implementation of an OR gate. Both ARCs are programmed to rectify current from the inputs toward the output. A high logic level setting of any of the inputs would result in current flow and charging of the output parasitic capacitance to "1". When both inputs are set low the output will either retain a low logic level or discharge through leakage to "0". **Figure 3B** gives the implementation of an AND gate. In this case, both ARCs are set to rectify in an opposite direction (from output to inputs). Only when both inputs are set to a high level the output will charge through leakage to "1". If any of the inputs is set to a low level, a discharge path will occur which will force the output to "0".

## CONCLUSION

In summary, a nanoscale RRAM with dynamic current rectification properties may be used as an ASD in neural networks to effectively double the array density for some applications and significantly reduce the required complexity from the peripheral circuitry (both sensing and control). This device is analogous to a biologic connexon (gap connection between synapses) that, when aggregated in a group, define the overall synaptic directionality and weight with respect to ion motion. An ARC may be dynamically toggled between positive and negative rectifications states thus allowing for a complementary operation (much like CMOS devices) of artificial synapses (as opposed to the linear analog scheme common to passive RRAM-based networks). In addition, the synaptic weight may be controlled in a digital manner by using a plurality of devices grouped together by changing the ratio of the number of positiverectifying to negative-rectifying ones. Furthermore, the LTP and LTD behaviors of biologic synapses may be emulated as well.

### AUTHOR CONTRIBUTIONS

The author confirms being the sole contributor of this manuscript and has approved it for publication.

#### FUNDING

The author acknowledges the partial funding support by Singapore Ministry of Education under grants MOE2016-T2-1- 102 and MOE2016-T2-2-102 of the group led by Prof. Diing Shenp Ang.

Aratani, K., Ohba, K., Mizuguchi, T., Yasuda, S., Shiimoto, T., Tsushima, T., et al. (2007). A novel resistance memory with high scalability and nanosecond switching. IEDM 2, 783–786.

Ben-Ari, Y. (2002). Excitatory actions of gaba during development: the nature of the nurture. Nat. Rev. Neurosci. 3:728.


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Berco. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

fnins-12-00755 October 17, 2018 Time: 13:56 # 6

# Multicellular Crosstalk Between Exosomes and the Neurovascular Unit After Cerebral Ischemia. Therapeutic Implications

Ana-Maria Zagrean<sup>1</sup> \* † , Dirk M. Hermann2,3† , Ioan Opris<sup>4</sup> , Leon Zagrean<sup>1</sup> and Aurel Popa-Wagner2,3,5 \*

#### Edited by:

Hari S. Sharma, Uppsala University, Sweden

#### Reviewed by:

Alexander A. Mongin, Albany Medical College, United States Ayman ElAli, CHU de Québec Research Center, Canada

#### \*Correspondence:

Ana-Maria Zagrean ana-maria.zagrean@umfcd.ro Aurel Popa-Wagner aurel.popa-wagner@ geriatricshealthyageing.com †These authors have contributed equally to this work

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 02 July 2018 Accepted: 17 October 2018 Published: 06 November 2018

#### Citation:

Zagrean A-M, Hermann DM, Opris I, Zagrean L and Popa-Wagner A (2018) Multicellular Crosstalk Between Exosomes and the Neurovascular Unit After Cerebral Ischemia. Therapeutic Implications. Front. Neurosci. 12:811. doi: 10.3389/fnins.2018.00811 <sup>1</sup> Division of Physiology and Neuroscience, Carol Davila University of Medicine and Pharmacy, Bucharest, Romania, <sup>2</sup> Department of Neurology, Chair of Vascular Neurology, Dementia and Ageing Research, University Hospital Essen, Essen, Germany, <sup>3</sup> Center of Clinical and Experimental Medicine, University of Medicine and Pharmacy of Craiova, Craiova, Romania, <sup>4</sup> Department of Neurological Surgery, University of Miami, Miami, FL, United States, <sup>5</sup> School of Medicine, Griffith University, Gold Coast, QLD, Australia

Restorative strategies after stroke are focused on the remodeling of cerebral endothelial cells and brain parenchymal cells. The latter, i.e., neurons, neural precursor cells and glial cells, synergistically interact with endothelial cells in the ischemic brain, providing a neurovascular unit (NVU) remodeling that can be used as target for stroke therapies. Intercellular communication and signaling within the NVU, the multicellular brain-vesselblood interface, including its highly selective blood-brain barrier, are fundamental to the central nervous system homeostasis and function. Emerging research designates cellderived extracellular vesicles and especially the nano-sized exosomes, as a complex mean of cell-to-cell communication, with potential use for clinical applications. Through their richness in active molecules and biological information (e.g., proteins, lipids, genetic material), exosomes contribute to intercellular signaling, a condition particularly required in the central nervous system. Cerebral endothelial cells, perivascular astrocytes, pericytes, microglia and neurons, all part of the NVU, have been shown to release and uptake exosomes. Also, exosomes cross the blood-brain and blood-cerebrospinal fluid barriers, allowing communication between periphery and brain, in normal and disease conditions. As such exosomes might be a powerful diagnostic tool and a promising therapeutic shuttle of natural nanoparticles, but also a means of disease spreading (e.g., immune system modulation, pro-inflammatory action, propagation of neurodegenerative factors). This review highlights the importance of exosomes in mediating the intercellular crosstalk within the NVU and reveals the restorative therapeutic potential of exosomes harvested from multipotent mesenchymal stem cells in ischemic stroke, a frequent neurologic condition lacking an efficient therapy.

Keywords: exosome, nanovesicles, neurovascular unit, blood-brain barrier, miRNA, stroke, mesenchymal stem cells

## INTRODUCTION

fnins-12-00811 November 3, 2018 Time: 18:56 # 2

At the interface with the bloodstream, neurovascular units (NVUs) are structural and functional multicellular modules consisting of neurons, perivascular astrocytes, microglia, pericytes, extracellular matrix and the endothelial cells of the brain microcirculation. They provide a coordinated neurovascular coupling and maintain a highly selective bloodbrain barrier (BBB) (Abbott, 2002). The dynamic multicellular crosstalk within the NVUs in physiological and pathological conditions could reveal novel cell-targeted therapeutic strategies with impact on the BBB, cerebral homeostasis and brain functions (Attwell et al., 2010; Abbott and Friedman, 2012).

The endothelial cells of the BBB are interconnected by tight and adherens junctions and form a continuous layer. This layer selectively buffers the impact of fluctuations in blood composition on brain interstitial fluid, regulating the brain microenvironment and neuronal signaling (Abbott, 2013). Various transcellular transport systems across the BBB have been described, as carrier mediated transport, receptor-mediated transport, ion transfer, efflux carriage, adsorptive-mediated passage, and fluid-phase endocytosis (Zlokovic, 2008).

Apart from the classical modes of intercellular communication, such as ligand-receptor interactions, direct cell-cell contacts (e.g., gap junctions) or paracrine signaling (Goodenough et al., 1996), a significant experimental evidence has confirmed that several physiological and pathophysiological processes are controlled by the extracellular membrane vesicles, such as exosomes and microvesicles, secreted from various cellular sources into the body fluids and interconnecting cells without direct cell-to-cell contact (Valadi et al., 2007). This type of signaling occurs mainly through exosomes, which are nano-sized vesicles that easily transfer biological information from cell to cell. This is achieved by means of exosomal molecules that would usually not cross membrane barriers. This shows the capability of inducing functional changes in target cells and modulating local and systemic crosstalk (Krämer-Albers and Hill, 2016). In the brain, exosomes are released from all types of cells (Frühbeis et al., 2013) and are bidirectionally transported through the blood-brain communication interfaces, blood-brain and blood-cerebrospinal fluid barriers (Balusu et al., 2016). These blood-brain interfaces are potential pathways for therapeutically administered exosomes.

Given their capacity to easily reach body compartments and connect origin cells with target cells, exosomes have a promising potential to be used in clinical applications. Indeed, exosomes have shown the capacity to serve both as biomarkers and novel therapeutic tools in the nervous system pathologies lacking efficient therapies, such as stroke (Barile and Vassalli, 2017). The cellular interactions within NVU might contribute to (i) the restoration of a well-organized cerebral microvasculature by providing trophic support and a stimulating brain microenvironment (Hermann and ElAli, 2012), and (ii) the remodeling of parenchymal tissue, including axonal sprouting, dendritic growth and synaptic reorganization (Hermann and Chopp, 2012). However, there is still more to explore about the diagnostic benefits and therapeutic roles of exosomes, their production, release, transport, uptake, signaling potential, change of their cargo proteins profile and miRNAs (Zhang and Chopp, 2016). Here, we review the roles of exosomes in mediating the intercellular crosstalk within the NVU and the therapeutic potential of exosomes derived from multipotent mesenchymal stem cells (MSCs) in stroke.

## EXOSOMES' AS A BIOLOGICAL COMMUNICATION TOOL

Exosomes are defined as 30–100 nm sized membrane vesicles derivatives of the endosomal compartment and correspond to the intraluminal vesicles of multivesicular bodies (MVBs) that upon fusion of the MVBs with the plasma membrane are released as exosomes into the extracellular environment (Lener et al., 2015), where they act as signaling organelles for intercellular communication. From the extracellular milieu, exosomes may contact target cells by (i) receptor-mediated adhesion to the cellular plasma membrane, followed by endocytic uptake and internalization, (ii) direct fusion of the exosome membrane with the target cell membrane and subsequent exosomal content release into the recipient cell (Bang and Thum, 2012).

Exosomes' vesicles are homogenous in shape, surrounded by a phospholipid membrane displaying membrane proteins, such as cell-specific receptors, and containing cell-type specific combinations of lipids, metabolites, coding and non-coding RNAs (miRNA, sRNA), single- and double stranded DNA, cytosolic and membrane proteins including enzymes, growth factors, receptors and cytokines (Théry et al., 2001; Lener et al., 2015). Exosomal lipids (e.g., phosphoglycerides, sphingomyelin, cholesterol, ceramide) are important for providing structural stability. Proteins of the exosomes are characteristic for their endosomal origin, and include membrane transport and fusion proteins (annexins, flotillin), proteins involved in cell targeting (tetraspanins, mostly CD9 and CD63) or other proteins correlated with their biogenesis from MVBs, as the tumor susceptibility gene 101 (TSG101) (András and Toborek, 2016). Exosomes also contain heat-shock proteins (Hsp60, Hsp70, Hsp90), known for their neuroprotective potential. Also, they expose low levels of phosphatidylserine and celltype-specific proteins. One of the most important function of the exosomes is targeting cellular pathways in the recipient cells through their RNAs and miRNAs cargo (Ling et al., 2013).

Novel research supports exosomes as a fundamental mechanism of communication in the nervous system, with roles in brain homeostasis and plasticity (Holm et al., 2018), acting as bidirectional cargo in brain-periphery communication and within the brain, in between neurons, glia, vascular and perivascular cells (**Figure 1**). Exosome secretion has been described from (i) depolarized/stimulated cortical neurons, mainly from the somato-dendritic compartments (Faure et al., 2006; Lachenal et al., 2011; Von Bartheld and Altick, 2011), (ii) oligodendrocytes (Frühbeis et al., 2013), (iii) microglia (Potolicchio et al., 2005), (iv) astrocytes when activated by oxidative and heat stress (Taylor et al., 2007), (v) endothelial cells (Dozio and Sanchez, 2017), and (vi) pericytes (Mayo and Bearden, 2015), known to generate MSCs in the

perivascular area of the lesioned or inflamed vessels (Caplan, 2008; Caplan, 2016).

The complex and versatile exosomal signaling was shown to impact the synaptic activity (e.g., neuronal origin exosomes exhibiting neurotransmitter receptors bind neurotransmitters within the synapse to stop signaling), transsynaptic communication, synaptic plasticity, maintenance of myelination, angiogenesis, neurovascular integrity, but also on neuroregeneration and neuroprotection in response to disease conditions (Holm et al., 2018). For example, angiogenesis could be stimulated both by activation of signaling pathways PI3K, ERK1/2, Wnt4/ß-catenin or NF-kB and transfer of the transcription factors STAT3, STAT5, transfer of lipids like S1P, transfer of proteins including VEGF, FGF-2, PDGF, metalloproteases, but also by the transfer of micro-RNA-126, miR-214, miR-296, and miR-150 (for a review, see Todorova et al., 2017). Likewise, after experimental stroke, treatment with exosomes isolated from miR-133b-overexpressing MSCs, significantly increased functional improvement and neurite remodeling/brain plasticity in the ischemic boundary area compared with control animals (Xin et al., 2017b). Recently, it was also reported that miR-26a is a physiological regulator of mammalian axon regeneration by targeting glycogen synthase kinase 3β (GSK3β) in adult mouse sensory neurons in vitro and in vivo (Tsenkina et al., 2015).

Exosomes can also propagate inflammation across the BBB and within the brain, as brain endothelial cells activated by systemic inflammation further activate the neighboring cells in the NVU via secreted exosomes (Balusu et al., 2016; Holm et al., 2018). The pathogenic role of microglia-derived exosomes in the inflammatory response was demonstrated in a model of traumatic brain injury (TBI) i.e., in vitro activated microglia-derived exosomes induced neuroinflammation at the site of injection and around the lesion. Furthermore, circulating enriched exosomes or CD11b-isolated microglia from the TBI brain ex vivo, initiated neuroinflammation following intracortical injection in naïve animals (Verderio et al., 2012; Kumar et al., 2017). The pathogenic effects of microglia-derived exosomes could be mediated by pro-inflammatory mediators TNF-alpha, IL-1β and miR-155 (Kumar et al., 2017).

Also, exosomes contribute to disease spreading by acting like Trojan horses for neurodegenerative agents (e.g., toxins, such as tetanus toxin, protein aggregates, such as phosphorylated Tau, amyloid Aβ or synuclein) (Bellingham et al., 2012; Holm et al., 2018). Through their non-coding RNA cargo and miRNA transfer, exosomes are involved in epigenetic regulation of neuroglial communication within the nervous system, but also in brainbody epigenetic interconnection (Lai and Breakefield, 2012).

#### NEUROVASCULAR UNIT REMODELING IN RESPONSE TO STROKE

Following the failure of acute neuroprotection therapies, major efforts are currently made worldwide to promote neurological recovery and brain plasticity in the subacute and post-acute

phases of stroke. For over more than two decades, therapeutic efforts in the stroke field have focused on the promotion of neuronal survival, which failed to succeed in clinical trials in humans until now (Savitz and Fisher, 2007; Ginsberg, 2008). From failure to translate successful neuroprotection therapies from animal models to humans, it may be concluded that the stimulation of survival alone is without prospect, as long as no successful remodeling of brain tissue stimulated by a permissive microenvironment takes place. Indeed, studies done recently have shown that extensive remodeling occurs in the brain following an ischemic event (Hermann and Zechariah, 2009; Hermann and Chopp, 2012). Currently, there is hope that stroke recovery might be promoted through pharmacological or cell-based therapies. Indeed, promising results from experimental studies have led to clinical trials, the results of which are currently awaited (Lener et al., 2015).

Remodeling of ischemic brain tissue involves interactions between neurons, glial and microvascular cells that create a microenvironment in which neurological recovery may ensue. Neurons and brain capillaries sprout. Neuronal outgrowth enables the formation of functional axons and synapses in the brain both over long [e.g., along pyramidal tract (Andres et al., 2011; Reitmeir et al., 2011)] and short (e.g., within motor cortex (Clarkson et al., 2010; Hermann and ElAli, 2012) distances, thus allowing for the restitution of neuronal networks that were damaged by the stroke event. The remodeling of ischemic brain tissue also includes responses of immature cells, namely of endothelial progenitor cells (EPC), neural progenitor cells (NPC), and inflammatory cells. New blood vessels are formed, and EPC and NPC are attracted to the stroke lesion. Glial cells contribute to the remodeling of the extracellular matrix, enabling neuronal plasticity.

In the process of brain remodeling, proliferating microvascular cells play a supportive role, enabling the migration of neural precursor cells and promoting the remodeling of neurons and glial cells via secretion of growth factors (Hermann and Zechariah, 2009). This rearrangement of cell-cell interactions is followed by the recovery of the BBB, leading to the restoration of brain homeostasis (Hermann and ElAli, 2012).

The remodeling potential of the NVU serves as an important therapeutic target in stroke and other acute neurologic conditions. After stroke, damaged and inflamed endothelium release pro-inflammatory factors and extracellular vesicles (EVs) that pass through the leaky BBB and activate astrocytes and microglia to release pro-inflammatory cytokines (TNFα, IL1β) (Norden et al., 2014). Microglia also release the antiinflammatory cytokine IL-10 that acts on reactive astrocytes to modify their cytokine secretion from a pro-inflammatory profile toward a pro-recovery one, represented mainly by TGFβ. During the post-stroke BBB repair and parenchymal remodeling process, NVU cells cooperate and release pro-recovery factors (e.g., IL-4, IL-10, TGFβ) that switch microglia into a pro-remodeling phenotype that release growth factors (Norden et al., 2014). Moreover, IL-10 acts on the endothelial and vascular cells to modulate vascular repair and remodeling, diminishes leukocyte– endothelial interactions, decreases expression and activation of cytokine receptors, promotes NO-induced vasodilatation and diminish ROS production and oxidative stress by inhibiting a NADPH oxidase subunit (Nox1) with impact on degenerative vascular remodeling (Dammanahalli et al., 2011; Garcia et al., 2017). Overall, IL-10 secreted from the NVU cells, but also from MSCs and their exosomes (Nakajima et al., 2017), is a prosurvival factor for neurons and glial cells that diminishes the post-lesional inflammatory response and limits the secondary damage during the resolution phase (Mosser and Zhang, 2008).

Pericytes behavior in different phases of ischemic stroke were recently described (Yang et al., 2017). Briefly, during the stroke hyperacute phase, pericytes constriction causes capillary occlusion (no-reflow phenomenon). Then, during the acute phase, pericytes have a pro-inflammatory and immune-modulatory action, with consecutive increase in BBB permeability and brain edema. By protecting the endothelium on its abluminal side and through release of neurotrophins, pericytes stabilize the BBB and protect brain parenchyma. Further, during post-stroke recovery phase, pericytes have a neuroprotective activity, promoting angiogenesis, neurogenesis, and brain recovery. The complex multifaceted, multistage pericytes intervention in ischemic-reperfusion injury and repair processes, recommend them for new targeted therapeutic strategies (Cai et al., 2017).

## NEUROVASCULAR UNIT-DERIVED EXOSOMES IN RESPONSE TO STROKE

Pericytes are important players in post-stroke NVU remodeling. Thus, they were shown to become activated and gain multipotent stem cell phenotype after brain ischemia and express the neuroepithelial stem cell marker nestin, with a potential to differentiate into neural and vascular precursor lineages (Nakagomi et al., 2015). Cooperation between endothelial cells and pericytes occurs both through paracrine interaction, but also through an exosomal bidirectional communication and is essential for preserving the microvascular functionality and stability. For example, endothelium or pericyte-derived hypoxic exosomes were shown to induce an angiogenic program (Fan, 2014; Mayo and Bearden, 2015). Secondary to local injury and perivascular inflammation, MSCs are released from their perivascular location and secrete bioactive molecules and exosomes with immunomodulatory and trophic effects, supporting the regenerative microenvironment needed for the post-injury recovery (Caplan and Correa, 2011). A recent work even suggests that perivascular MSCs are adventitial cells, acting as precursors of pericytes and other stromal cells during tissue homeostasis (de Souza et al., 2016). Not all pericytes can generate MSCs. It has been recently shown that from the various subpopulation of existing pericytes, not all of them can act like stem cells, and some act like fibroblasts.

Neuronal exosomes are present at synaptic level, both within pre- and postsynaptic compartments, and transport synaptic receptors (e.g., AMPA receptors, GPCRs) (Koniusz et al., 2016) contributing to synaptic plasticity, both locally and within broader neuronal networks (Chen and Chopp, 2018). The activity within glutamatergic synapses, which is increased in

post-stroke excitotoxic conditions, stimulates neuronal release of exosomes that preferentially bind to adjacent neurons, impacting on interneuronal communication (Chivet et al., 2014). The exosomes released secondary to neuronal depolarization are rich in miRNAs, potentially promoting synaptic plasticity by enabling the rapid translation of associated proteins (Goldie et al., 2014).

The interactions between neurons, glial cells and microvascular cells are finely tuned. They involve mutual cell to cell communication via release of growth factors as well as physical cell-cell interactions across the extracellular matrix that is itself subjected to remodeling processes after stroke (Rosell and Lo, 2008). Considering the complexity of these systems and considering both the structural and functional heterogeneities of brain structures and the heterogeneities of ischemic strokes with regard to their size, etiology, and localization (Hermann and Chopp, 2012), the development of neurorestorative therapies is a true challenge (Hermann et al., 2015).

## THE POTENTIAL USE OF MESENCHYMAL STEM CELLS AND THEIR EXOSOMES FOR STROKE THERAPY

Ischemic stroke is a leading cause of death and long-term disability in industrialized countries, with thrombolysis and interventional vascular recanalization being the only treatments available. Due to severe side effects and a narrow therapeutic time window, only a small proportion of stroke patients receive this therapy. Thus, additional therapeutic concepts are mandatory (Hermann and Chopp, 2012). Strategies that promote neuronal survival in the acute stroke phase have successfully been studied in experimental stroke models, but were not successful in clinical trials. Therefore, the research focus has recently shifted from the acute to post-acute stroke phase (Hermann and Chopp, 2012). After acknowledging that transplanted cells integrate poorly into existing neural networks and that they induce brain remodeling in a paracrine way by secreting a heterogeneous group of nanovesicles, these EVs have been identified as key players that mediate restorative effects of stem and progenitor cells in ischemic brain tissue. Neuroprotection as observed after EV infusion in experimental stroke models is related to stem cell application in stroke. As a matter of fact, stem cell-induced neurological recovery after stroke is not a consequence of cell regeneration but due to paracrine mechanisms of grafted cells, among which stem cell-derived EVs are key mediators (Doeppner et al., 2018).

Blood-brain barrier may block or diminish the access of therapeutic agents within the central nervous system and therefore many nervous system diseases lack an efficient treatment because of a deficient drug delivery vehicle. Considering this important issue, research is nowadays developing nanocarriers for brain targeted drug delivery and exploit the potential use of stem cells to secrete exosomes, as natural nanovesicles rich in biological active molecules. As lipid-bound nanoparticles, exosomes easily interconnect cells and cross selective-permeable membranes such as BBB, thus emerging as versatile tools for new therapeutic strategies (e.g., regenerative, immune-modulatory or anti-tumor therapies), either acting through their biochemically active constituents (e.g., proteins, lipids, genetic material), or serving as natural nonimmunogenic vehicles for drug delivery (Lener et al., 2015).

## Non-exosomal Effects of Mesenchymal Stem Cells

The human brain contains reservoirs of neural stem and precursor cells in the subventricular zone (SVZ) surrounding the lateral ventricles (Bacigaluppi et al., 2009). Although cerebral ischemia triggers the activation of these cells and promotes their migration toward ischemic lesion sites, their siblings hardly survive and differentiate within the ischemic milieu (Doeppner et al., 2012, 2014b). To improve brain remodeling and plasticity, and to bypass limitations of endogenous neurogenesis following ischemic stroke, a variety of approaches started to focus on the transplantation of NPCs or somatic stem cell entities, such as MSCs (Popa-Wagner et al., 2006, 2007, 2011). MSCs secretome comprises growth factors, cytokines, chemokines, extracellular matrix components, genetic material, but also EVs (exosomes and microvesicles), recommending them as versatile tools in clinical applications (Gaceb et al., 2018).

Mesenchymal stem cells, known as "sentinel and safe-guards of injury" (Caplan, 2016), were shown to produce neurotrophic factors such as nerve growth factor, brain-derived neurotrophic factor, or glial-derived neurotrophic factor, explaining their therapeutic potential (Lopatina et al., 2011).

Initially, it was assumed that transplanted NPCs and MSCs home to affected sites and, upon expansion and differentiation, directly replace the lost brain cells to restore tissue functions. In this context, our lab has comprehensively characterized the therapeutic effects of SVZ-derived adult NPCs in a mouse model of ischemic stroke, i.e., transient proximal (i.e., intraluminal) middle cerebral artery occlusion (MCAO). We observed that systemic NPC delivery induces profound brain tissue remodeling, reflected by reduced secondary neurodegeneration, reduced neuroinflammation, reduced astrogliosis and reduced microglial activation, that was associated with functional neurological recovery (Bacigaluppi et al., 2009; Doeppner et al., 2012, 2014a,b). Remarkably, it turned out that systemic intravenous administration of adult NPCs was more effective than intracerebral transplantation. Indeed, systemic administration effectively resulted in the stabilization of BBB integrity. However, just 0.1–0.3% of intravenously transplanted NPCs were detected in the brain and most of them were in an undifferentiated state (Bacigaluppi et al., 2009). These findings imply that NPCs act in a paracrine rather than a cellular mode.

## Exosomal Effects of Mesenchymal Stem Cells

The exosomes' database ExoCarta<sup>1</sup> reports more than 900 species of proteins associated with MSCs-derived exosomes, but

<sup>1</sup>http://www.exocarta.org/

recent data from proteomic analysis, identified more than 2000 proteins in MSC-exosome, many of them being involved in brain repair (Otero-Ortega et al., 2017). These were shown to increase glial production of anti-inflammatory and immunoregulatory mediators, TGFβ1 and IL-10 (Burrello et al., 2016), with significant roles in NVUs' recovery and remodeling. Also, it was recently shown that IL-10 is one of the neuroprotective factors through which transplanted MSCs act after an ischemic stroke. Thus, MSCs overexpressing IL-10 improved neuronal survival in the ischemic hemisphere (Nakajima et al., 2017). Interestingly, MSCs-derived exosomes were shown to exhibit post-stroke changes in their miRNA profile, mostly in the miRNAs actively involved in the repair process by altering gene expression and promoting brain recovery (Liu et al., 2013).

Classically, paracrine effects were thought to be mediated by soluble molecules such as growth factors, cytokines, chemokines and hormones. More recent data, however, demonstrate that several physiological and pathophysiological processes are controlled by exosomes (Cramer et al., 2017). In experimental stroke models, evidence was provided that exosomes exert neuroprotective, proangiogenic and neuronal plasticity-promoting functions (Xin et al., 2013). Thus, systemic administration of MSC-derived exosomes in a rat model of stroke improved functional recovery and enhanced neurite remodeling, neurogenesis, and angiogenesis (Xin et al., 2013). Furthermore, administration of combined xenogenic (from mini-pig) adipose-derived mesenchymal stem cell (ADMSC) and ADMSC-derived exosome therapy has been shown to reduce brain-infarct zone (BIZ) and enhance neurological recovery in rat after acute ischemic stroke (Chen et al., 2016). At molecular level the beneficial effects of MSC-derived exosomes could be mediated by the miR-17-92 cluster. Thus, rats subjected MCAO and treated with miR-17-92 cluster-enriched exosomes, performed significantly better than the control rats treated with MSC exosome alone (Xin et al., 2017a). Similarly, administration of exosomes isolated from miR-133b-overexpressing MSCs lead to increased neural plasticity and improvement of functional recovery after stroke in rats (Xin et al., 2017b).

Based on these observations, we performed a direct head-by-head comparison of the therapeutic effects of MSCs and their exosomes in a murine model of transient intraluminal MCAO, which predominantly affects the striatum and most lateral parts of the overlying cerebral cortex, showing that systemic MSCs and MSC-derived exosomes are equally effective in enhancing stroke-related motor and coordination recovery thereby confirming the beneficial effects of the exosome therapy reported by Chopp and colleagues who observed a significant reduction in neurological impairment that improved gradually over 4 weeks after systemic delivery of MSC-derived EVs (MSC-EVs) in a model of transient MCAO in rats (Xin et al., 2013). Both therapies promoted post-ischemic endogenous neurogenesis and angiogenesis and reversed

the stroke-associated immunodepression (Doeppner et al., 2015).

## CONCLUSION AND PERSPECTIVES

Ischemic stroke is a leading cause of death and longterm disability for which no restorative therapy is available. After stroke, the NVU is compromised and has become a major target for restorative therapies in the central nervous system. Emerging research has revealed that the nano-sized exosomes could be used for the NVU remodeling after stroke, due to their ability to mediate cell-to-cell communication. Considering the side effects typically attributed to cellbased therapies, in particular, malignant transformation of the transplanted cells, MSC-derived exosomes are attractive candidates for stroke therapy, as emphasized by a recent position paper (Lener et al., 2015). Indeed, systemic administration of MSC-derived exosomes is effective in enhancing stroke-related motor and coordination recovery in experimental stroke models, fueling the hope for clinical studies. Nevertheless, for clinical applications we need further studies to shed light on (i) mechanisms of the interaction between exosomes and target cells, (ii) circulation kinetics and biodistribution; (iii) biogenesis mechanism; (iv) potential side effects. For example, tumor-secreted exosomes may act as mediators in cancer metastasis by maintenance and enhancement of tumor microenvironment (Salido-Guadarrama et al., 2014; Cheng et al., 2017; Li et al., 2018). Likewise, several studies have reported high levels of cholesteryl ester (CE), triacylglycerol (TAG) and cardiolipin in exosomal preparations fueling concerns about increasing the risk of stroke instead of having a beneficial effect (Llorente et al., 2007; Van Meer et al., 2008; Strauss et al., 2010; Record et al., 2014; Skotland et al., 2017; Popa-Wagner et al., 2018).

Furthermore, considering that ischemic stroke mainly affects elderly patients, experimental data in aged rodents are urgently required before a clinical proof-of-concept study in human patients should be envisaged (Popa-Wagner et al., 2006, 2007, 2018; Balseanu et al., 2014).

## AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

## FUNDING

This work was supported by grants of the Romanian National Authority for Science Research and Innovation, CNCS – UEFISCDI, project numbers PN-III-P4-ID-PCE-2016-0340; PN-III-P2-2.1-PED-2016-1013 andPN-III-P4-ID-PCE-2016- 0215.

## REFERENCES

fnins-12-00811 November 3, 2018 Time: 18:56 # 7



recovery after stroke in rats. Stroke 48, 747–753. doi: 10.1161/STROKEAHA. 116.015204


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Zagrean, Hermann, Opris, Zagrean and Popa-Wagner. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# A New Frontier: The Convergence of Nanotechnology, Brain Machine Interfaces, and Artificial Intelligence

#### Gabriel A. Silva\*

Departments of Bioengineering and Neurosciences, Center for Engineered Natural Intelligence, University of California San Diego, La Jolla, CA, United States

A confluence of technological capabilities is creating an opportunity for machine learning and artificial intelligence (AI) to enable "smart" nanoengineered brain machine interfaces (BMI). This new generation of technologies will be able to communicate with the brain in ways that support contextual learning and adaptation to changing functional requirements. This applies to both invasive technologies aimed at restoring neurological function, as in the case of neural prosthesis, as well as non-invasive technologies enabled by signals such as electroencephalograph (EEG). Advances in computation, hardware, and algorithms that learn and adapt in a contextually dependent way will be able to leverage the capabilities that nanoengineering offers the design and functionality of BMI. We explore the enabling capabilities that these devices may exhibit, why they matter, and the state of the technologies necessary to build them. We also discuss a number of open technical challenges and problems that will need to be solved in order to achieve this.

Keywords: nanotechnology, neuroscience, machine learning, artificial intelligence (AI), brain machine interface (BMI), brain computer interface, computational neuroscience

## INTRODUCTION

A confluence of technological capabilities is creating an opportunity for machine learning and artificial intelligence (AI) to enable "smart" nanoengineered brain machine interfaces (BMI). The goal is for this new generation of technologies to be able to communicate with the brain in ways that support contextual learning and adaptation to changing functional requirements. This applies to both invasive technologies aimed at restoring neurological function, as in the case of neural prosthesis, as well as non-invasive technologies enabled by signals such as electroencephalograph (EEG). Advances in computation, hardware, and algorithms that learn and adapt in a contextually dependent way will be able to leverage the capabilities that nanoengineering offers the design and functionality of BMI. Eventually, these technologies will be able to carry out learning and adaptation in (near) real time, as external shifting demands from the environment and physiology require them. Ultimately, the goal is to produce personalized individual user experiences for applications such as gaming, and to allow the device to learn and adapt to changing disease requirements in clinical scenarios. In this commentary we explore the enabling capabilities that these devices may

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Hari S. Sharma, Uppsala University, Sweden Michela Chiappalone, Fondazione Istituto Italiano di Technologia, Italy Liang Guo, The Ohio State University, United States

#### \*Correspondence:

Gabriel A. Silva gsilva@ucsd.edu

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 14 June 2018 Accepted: 29 October 2018 Published: 16 November 2018

#### Citation:

Silva GA (2018) A New Frontier: The Convergence of Nanotechnology, Brain Machine Interfaces, and Artificial Intelligence. Front. Neurosci. 12:843. doi: 10.3389/fnins.2018.00843 exhibit, why they matter, and the state of the technologies necessary to build them. We also discuss a number of open technical challenges and problems that will need to be solved in order to achieve this.

## THE OPPORTUNITY FOR "SMART" BRAIN MACHINE AND BRAIN COMPUTER INTERFACES

fnins-12-00843 November 14, 2018 Time: 16:39 # 2

Brain machine and brain computer interfaces (we use these terms interchangeably here) represent technologies designed to communicate with the central nervous system: the brain, spinal cord, and neural sensory retina. Clinically, depending on the design and intent of the technology, the goal can be to record and interpret neural signals in order to execute an intended neural command through an external device, or to achieve neural stimulation, often to restore neural function following disease or trauma, or both (Adewole et al., 2016; Choi et al., 2017; Slutzky and Flint, 2017; Rezeika et al., 2018). Some devices make use of feedback in an attempt to optimize performance, whether physiological or via patient specific intent and instructions (Widge et al., 2018). There is also a growing list of non-invasive brain machine interface technologies not meant for clinical use, primarily driven by innovative startup companies. These technologies are intended to augment the user experience and control interface for gaming and augmented (AR) and virtual reality (VR) applications. Although of course very different than technologies aimed at treating and restoring clinical function and quality of life to patients, this is a market that should not be ignored. Not the least of which because it could provide leveraging resources to the benefit of clinically related research. For example, advances in our understanding of the relevant neurophysiology, cognitive neuroscience, mathematical and engineering aspects of signal processing, and hardware, can significantly impact both the gaming industry as well as clinical devices and neural prosthesis. The brain machine interface market is projected to reach \$1.46B by 2020, with a compound annual growth rate (CAGR) of 11.5% between 2014 to 2020 by one estimate (Allied Market Research, 2015), and a comparable \$1.72B in 2022, with a predicted CAGR of 11.5% between 2012 and 2022 by another estimate (Grand View Research, 2018). Much of this projected growth will be due to non-invasive technologies, with the gaming industry as a market driver roughly on par with healthcare applications.

As significant as these numbers are, these projections primarily reflect enabling technologies for interfacing between neural control and sensory experiences with machines. They do not reflect opportunities that go beyond what is currently possible with the existing state of the art. BMI that can learn and adapt reflect the cutting edge of what is technologically possible, due to a confluence of BMI technologies, in particular nanotechnologies, machine learning and AI, alongside a continued increasing understanding of the relevant neuroscience. AI can provide opportunities to create "smart" BMI that contextually learn and adapt to changing functional requirements and demands. This has the potential to produce personalized individual experiences in gaming and AR/VR, and allow for the changing requirements associated with patient specific disease progression and evolution in clinical applications. This latter point cannot be overestimated, because not only would it accommodate the differing clinical demands of different neurological disorders, it would allow for patient specific adaptation of BMI functionality to the needs of different patients. And it would allow the technology to continue to adapt as disease progression evolves in individuals over time. One of the significant limitations of current state of the art BMI and neural prosthesis is the assumption of one size fits all. In in other words, the assumption that a technology operating under a specific set or range of functionality will properly treat all patients. While we are not aware (yet) of a device or technology that reflects the actualized integration of machine learning and nanotechnology applied to BMI, we argue that the potential and impact of doing make the subject worth exploring. Each on their own, machine learning and nanotechnology, are already being used in the design and function of BMI and neural prosthesis in a number of ways that align with the vision we propose here.

### INTEGRATION OF BMI WITH MACHINE LEARNING

What advantages does machine learning and artificial Intelligence (AI) offer BMI? What exactly are machine learning algorithms learning, and how can they use that information to adapt in a meaningful way? What these algorithms can learn is information provided by feedback and telemetry from the hardware. This could be information about the current state of the output settings of the device, or any kind of external information measured by sensors in the BMI. For example, physiological measurements in response to stimulation, feedback from other algorithms external to the BMI-machine learning system, such as haptic or computer vision feedback, or the internal parameter settings of current stimulation or recording protocols to the device. In the case of internal parameters the algorithms have constant access to variables such as pulse durations and amplitudes, stimulation frequencies, energy consumption by the device, stimulation or recording densities, electrical properties of the neural tissues it is interfacing with (resistances, impedances), and continuous or near continuous levels of biochemical factors such as neurotransmitters or other metabolites. Of course, none of these are mutually exclusive, with multiple types and streams of information possibly being provided to the algorithms in parallel, albeit at likely different sampling resolutions. With this information, machine learning algorithms could then identify subtle and non-trivial patterns and phenomena in the data, ideally in (near) real time, in order to produce desired functional outcomes from the BMI that change dynamically as external (e.g., clinical or functional) requirements demand them. This would necessitate the development and training of machine learning models and algorithms offline as part of the design of the BMI system. The algorithms would need to learn a wide

enough range of the parameter spaces in order to appropriately identify patterns in the data they encounter when online. Subsequent algorithms can then autonomously make decisions about how to use that data. This step does not necessarily have to be part of the brain machine interface system itself, and could be executed with algorithms computing in the cloud, if sufficient bandwidth was available. Or even offline following periodic data downloads, for example. Clearly though, on board decision algorithms that operate in real time with the machine learning algorithms that are identifying patterns in the data would be ideal. This would alleviate issues of data transfer delays, and bandwidth insufficiency. This could also allow for the need to store less data on the device, which could be limited due to physical constrains. Data would only have to be stored long enough for the system to make an autonomous decision, essentially as a moving window that matches the processing capabilities of the algorithms. Of course, it may still be valuable or necessary to store some data or types of data for offline analyses even though they may not be needed for the BMI system to make a decision. For example, in order to understand offline after the fact why the algorithms made the decisions they made and the clinical outcomes of those decisions. With this process complete we close the loop: information is provided to the machine learning algorithms, followed by learning, pattern identification, and subsequent executable autonomous decisions that in turn dynamically change the output of the brain machine interface and how it interacts with the external environment it is interfacing with. This could be the brain itself in the case of a neural prosthesis intended to restore clinical function, or software in a noninvasive BMI that is interfacing that is part of an AR or VR system.

While still early days, a number of research groups have recognized this potential, and are beginning to explore how machine learning could inform and integrate neural stimulation and feedback. Nurse and colleagues (Nurse et al., 2015) have developed a generalized approach that takes advantage of a stochastic machine learning method to classify motor related signals specifically for BMI applications. Importantly, their classifier does not need to rely on the use of extensive a priori data to train the BMI. Their algorithms outperformed other methods on the Berlin BMI IV 2008 dataset, and demonstrated high levels of classification accuracy when tested on datasets derived from EEG signals. In another recent study, Ortega et al. (2018) explored different data pre-processing strategies and convolution neural network architectures for classification tasks derived from EEG signals. Interestingly, they found that a rather straightforward network architecture, when combined with a pre-processing step that analyzed spectral power preserving features of the electrode arrangements, was sufficient to handle the analysis of the data. Their network consisted of a single convolution layer, one connection layer, and single linear regression classifier layer. Their approach allowed them to carry out co-adaptive training on the data to achieve on-line classification. A different study had explored a similar approach. Lawhern et al. (2016) carried have also explored a similar approach.

As discussed above, one of the biggest advantages machine learning may confer on BMI is the ability to achieve real-time or near-real-time modulation of output or stimulation parameters in response to active real-time feedback from physiological signals, the environment, or other internal cues from the system itself, such as possibly the output from other internal algorithms that have processed some amount of data. Most BMI's have a decoder component whose job it is to decode and make sense of neural signals in order to produce executable or actionable outputs. This typically necessitates extensive supervised training in order to optimize the interpretation of recorded neural signals before the decoder can properly correlate observed signals with desired outputs and commands. This training has traditionally required supervised feedback with a human in the loop, typically a technician or clinician often with input from the patient her/himself, thus making the process highly inefficient and intermittent. Training and subsequent adjustments can only occur periodically and are typically time consuming. Furthermore, the mapping from neural signals to actionable outputs is limited to the training data the system is exposed to during the training. This then highly limits the ability of the BMI to respond to variable real world scenarios it may encounter when in use, thus severely limiting its functionality to the patient when such conditions arise. Early work relied on feedback from external sensory references to compute an error between the output of the system and desired supervised target. These included visual and auditory signals (Wessberg et al., 2000; Lebedev et al., 2005), mechanotransduction (Nicolelis and Chapin, 2002; O'Doherty et al., 2011), and direct cortical sensory stimulation (6 Bach-y-Rita and Kercel, 2003). But these approaches are severely limited due to their need for continuous information from an external reference target to adjust the mapping to the output of the BMI. More recent work has addressed some of these limitations by adapting output parameters to unsupervised learning methods such as Bayesian statistical methods and reinforcement learning that do not rely on an external reference (Vidaurre et al., 2011; Orsborn et al., 2012, 2014; Bryan et al., 2013; Huang and Rao, 2013; Bauer and Gharabaghi, 2015). Although in most cases they still require significant training periods. More recent studies have begun to investigate the use of endogenous neural signals directly as the training source in iterative closed feedback loops with the BMI that can respond and adapt in a much more direct way (Suminski et al., 2010; Carmena, 2013). For example, Prassad and colleagues are developing an approach they call Actor-Critic reinforcement learning that does not need to rely on a supervised error signal (Pohlmeyer et al., 2014; Prins et al., 2014, 2017).

In general, the BMI field in general and neural prosthesis field in particular are still exploring machine learning. One of the challenges is that key state of the art methods, such as deep learning, that have had huge successes in other applications may not be the best approach for the constraints imposed by the needs of BMI (Vidaurre et al., 2015). In a recent paper, Panuccio et al. (2018) do an excellent job summarizing the current state and challenges of neural engineering aimed at restoring neural function, including proposing a number of similar requirements discussed in the current paper, that emerging algorithms and

machine learning will need to address in order to build a true adaptive BMI.

An important consideration that such machine learning approaches offer that other methods cannot is the opportunity to develop BMI that adapt to the scaling requirements, both spatial and temporal, necessitated to achieve targeted functional outcomes. In the context of neural stimulation, the optimal density of stimulation required to produce a target response in neuronal populations being stimulated is a complex consideration, and may not always be the highest stimulation density achievable by the device (Shepherd et al., 2013; Patil and Thakor, 2016). What the right stimulation density should be can be a complex question to answer, and often depends on specific physiological and pathophysiological considerations. In many situations we still do not fully understand what the right stimulation density should be and why. Furthermore, the optimal stimulation density is likely to vary from individual to individual even in the same disorder, and within an individual patient the disease can greatly evolve over time as the physiology changes and the body responds and adapts to altered conditions. This could be a function of age or exogenous perturbations such as a response to other treatments, diet, and the psychological state of the patient. Another consideration is that hardware or other algorithms that need to make use of recorded or measured neural data in order to interact with the brain could have different scaling requirements in the data. This would be dependent on what the external query is and how the neural data needs to be used. It reflects the technical capabilities and limitations of the external technologies requesting the data. Under sampling could lead to poor user interactions, for example, a frustrating or confusing AR/VR experience, or the inability of a disabled patient to communicate in a timely or accurate manner. Oversampling would waste computational resources and time. In a research setting, data scaling issues could affect the empirically determined accuracy of a computational model, or how a hypothesis is tested and interpreted. Clinically, it could impact treatment or other clinical decisions. Changing temporal and spatial scaling requirements demanded by exogenous considerations to the BMI present situationally unique challenges that the existing state of the art is not yet able to address in a substantive way. The integration of machine learning and AI with nanoengineered BMI offers the opportunity for these technologies to learn, adapt, and respond to their environments in order to address functionally challenging considerations such as dynamic scaling demands.

### BEYOND THE CURRENT STATE OF THE ART: MACHINE LEARNING ENABLED NANOENGINEERED BMI

In recent years there has been an explosion of work focused on the development and use of nanotechnologies aimed at interacting and interfacing with the brain and central nervous system generally (Silva, 2006, 2007a,b, 2008, 2010; Kotov et al., 2009; De Vittorio et al., 2014; Saxena et al., 2015; Badry and Mattar, 2017; Scaini and Ballerini, 2017; Rosenthal, 2018), and in the context of BMI and neural prosthesis in particular (Webster et al., 2003; Lovat et al., 2005; Fabbro et al., 2012; Nicolas-Alonso and Gomez-Gil, 2012; Seo et al., 2013; Avants et al., 2016; Ha et al., 2016; Scaini and Ballerini, 2017). Considerable recent effort has focused on nanoscale neurotechnologies aimed at recording from and stimulating from the brain at high densities. This has to a significant degree been motivated by federal research efforts in both the United States and the European Union through the Brain Initiative<sup>1</sup> and Human Brain Project,<sup>2</sup> respectively. We do not attempt to review this extensive literature here, but refer the reader to the references and published literature more broadly.

While the confluence of machine learning and nanoengineered BMI and neural prosthesis has not yet occurred, machine learning is playing an increasing role in other aspects of nanotechnology and related molecular-scale research (for example, see the review by Sacha and Varona, 2013). In one example, Albrecht et al. (2017) have written a tutorial for using deep learning convolution neural networks for analyzing and mining single molecule data from DNA sequencing experiments. Ju et al. (2017) recently showed they could use an atomic version of Green's function and Bayesian optimization to optimize the interfacial thermal conductance of Si-Si and Si-Ge nanostructures (Ju et al., 2017). Their method was able to identify optimal structures within a library of over 60,000 candidate structures. And in another striking recent study, Lin et al. (2018) were able to implement a deep learning architecture on an all-optical 3d printed Diffractive Deep Neural Network (D2NN) that were designed and optimized by deep learning. These researchers were able to carry out classification and other imaging tasks without the need or use of any power, except for the input light into their system. The opportunity for BMI and neural prosthesis lies in the ability of machine learning to "learn" (i.e., identify and classify patterns) in highly complex physical and chemical data derived from devices that have been engineered at the nanoscale in order to inform and optimize the design and functional outputs of the devices.

As already alluded to, however, an important consideration in this quest is the realization that existing machine learning and AI algorithms may not be optimal for such needs. Thus, there is a possible opportunity and need for the development of purposely developed machine learning algorithms specifically designed to take advantage of and control nanoengineered BMI devices. Current machine learning methods, in particular deep artificial neural networks (ANN's), are incredibly powerful and continue to show some spectacular progress. What is probably the most surprising is that at its most fundamental, the underlying learning rules responsible for the existing state of the art and success of ANN's are essentially all variants of gradient descent statistical learning methods. But like any method, there are theoretical and practical limitations. The data they operate on must therefore be able to accommodate these constraints. In particular, existing algorithms are dependent on exposure to enormous data sets to train them properly so they can learn (a form of model bias). They can only find associations and patterns in the data that

<sup>1</sup>http://www.braininitiative.org/

<sup>2</sup>https://www.humanbrainproject.eu/en/

already exist (model bias again). There is always a danger of over generalizing from a limited training set (model over fitting). A such, they display an almost complete lack of robustness and ability to adapt beyond the training sets they are exposed to. New data may not achieve further learning (model saturation). And because of these considerations these methods will miss outliers (data sparseness problem). Finally, they require large computational resources and the consumption of huge amounts of energy to properly identify learned patterns. These methods are limited by a set of fundamental engineering challenges inherent to statistical learning. Yet, even with these constraints in mind, and ignoring all the hype currently surrounding machine learning and AI, it is difficult not to be impressed by the accomplishments these methods are achieving. If the data and resources are appropriate to the task being presented to the algorithms, these methods can work remarkably well and it is likely impractical (and even unnecessary) to attempt to develop new methods to supersede them; at least for the foreseeable future. In some cases it is certainly plausible that the machine learning needs of nanoengineered BMI's could be amenable to the current state of the art (see references and discussion above). BMI can generate significant amounts of data, and the range of operating conditions of physiological signals, stimulation parameters, and recording densities specific to given functional tasks are sufficiently well understood, at least from the perspective of defining the extremes of those conditions. Thus, sufficient data over known and practical physiological operating ranges could allow existing machine learning to learn sufficiently in order to guide decision algorithms for adapting the interactions of the BMI with their targets. This is particularly true of nanoengineered brain machine interface's, whereby the degree of synthesis control over the material or device, and spatial and temporal stimulation resolutions and recording densities, can be engineered at the nanoscale. Conceivably, the quality and amount of information nanoengineered BMI could produce, along with the degree of functional control nanoscale engineering provides, are particularly well suited to take advantage of the state of the art in machine learning and AI in order to achieve smart integrated BMI.

At the same time, however, it is worth asking if machine learning and AI architectures designed to learn differently than existing algorithms could provide a degree of functionality and integration with BMI's that does not yet exist. In particular, machine learning methods that mathematically model and abstract specific neurobiological properties of interest. Empirical (i.e., data driven) statistical learning AI works well on problems where bias, sparseness, and saturation are not (or not yet) an issue that limit its learning. But it is precisely learning beyond these constraints that the biological brain excels at. In particular, the ability of the brain to adapt and extrapolate beyond data presented to it, and it is incredible computational and physical robustness to perturbations. These properties go beyond the current stage of the art in machine learning, but could be critical to the sophisticated integration of BMI with the brain. The biological brain represents, learns, and manipulates information very differently than the way existing artificial neural networks, machine learning, and statistical methods "learn" to find patterns in data. The brain primarily learns by analogy and by abstracting beyond the immediate training sets presented to it. It is capable of robustly adapting to different situations and contexts it may not have previously encountered with an incredible degree of plasticity. The computational flexibility, adaptation, and robustness of the brain exceed any existing machine. One extreme example of the human brain's incredible robustness and ability to adapt is evident in a neurological condition called Rasmussen's encephalitis, a rare pediatric chronic inflammatory neurological disorder that typically affects one hemisphere. It is typically characterized by severe and frequent seizures that result in loss of motor function, loss of speech, hemiparesis, encephalitis, and cognitive decline (Freeman, 2005; Varadkar et al., 2014; Venkatesan and Benavides, 2015). Most patients become refractory (stop responding) to medical treatment. In many cases the only effective treatment for seizures is hemispherectomy, whereby portions or the entire affected cortical hemisphere are surgically removed and the corpus callosum cut from the unaffected hemisphere. The corpus callosum is the high speed "ribbon cable" that connects our two sides of the brain. Yet, to varying degrees, the remaining side of the cortex in these patients is able to take up the functions of the excised cortical tissue to a remarkable extent. In many cases these patients are able to function cognitively and physically almost normally considering how much of their brains are removed. (Contrast that with what would happen if you remove even a handful of the transistors or circuits in a computer.) All of this is even more impressive given the computational and energy efficiency with which the brain achieves this - using about 20 watts of power, barely enough to power a dim light bulb, in about 3 lbs of "wetware" that occupies a volume equivalent to a 2 liter bottle of soda.

One final comment worth emphasizing is that although the biological brain exhibits computational properties and an ability to learn that we want to understand and leverage, this does not necessarily mean that we have to reverse engineer the brain to the point that we are modeling or emulating every aspect of how the biology itself implements the brain's internal algorithms. One approach is to abstract away the biological details and capture the core algorithms, i.e., rules, that underlie the property or system being studied in the brain that we want to build into the BMI. The end result are mathematical models that are independent of the underlying biological details, but which capture the functional mechanisms at an appropriate scale of abstraction in order to arrive at algorithmic descriptions that emulate those properties. Admittedly, where that line of abstraction is drawn can be more an art than a science.

#### CHALLENGES AND OPEN PROBLEMS

In this final section we briefly introduce some of the challenges and open problems associated with actually executing the vision discussed above. We do not elaborate in this paper, but leave them open for further discussion and dialog.

First, most (all) of the recent efforts in the development of neurotechnologies aimed at high density recording or stimulation have focused on the physics, chemistry, and engineering of the core nanotechnologies themselves. This is understandable because the fundamental technologies necessary to enable stimulation or recordings at the actual interface with the brain have to come first. They need to precede any methods or technologies intended to modify or make use of data and information such technologies provide. Beyond the actual interface itself, mechanical and operational stability and long term reliability of the devices is critical in order to ensure accurate recordings or stimulation. For example, if the electrodes move or there is excessive reactive gliosis it will severely affect the efficacy and accuracy of the devices, rendering any control or adaptation by machine learning algorithms irrelevant. These reflect fundamental engineering challenges that have attracted significant amounts of work. And while significant progress has been made, it very much remains highly active areas of research. We do not discuss these issues further in this paper (see for example Lega et al., 2011; Gilja et al., 2011; Lu et al., 2012).

Beyond these well-known issues surrounding the fabrication and functionality of BMI devices, there are open problems that have received comparatively less attention. Of particular relevance are questions surrounding how data from these devices can be accessed and used, which are of importance to any discussion about integrating machine learning as part of the overall system. At the nanoscale, the density of recording or stimulation can be so large that the telemetry problem of how do you keep track of all those signals becomes an issue. In other words, how do you keep track of where and when signals are coming from (in the case of recordings) or going to (in the case of stimulation). With high density recordings, e.g., many thousands of signals, it becomes physically impossible to follow the standard micro-scale strategy of having individual leads "read out" signals. Most of the nanotechnologies currently being developed for recording at such extreme densities are being engineered as individual standalone nanoscale devices that can then be deployed in large numbers. But even if each individual device is indeed able to faithfully record local signals, how does one extract that information globally across the entire population of sensors and how does one make sense of the resultant data? In the case of applications that necessitate spatial "corticotopic" information, this question is critical. We do not yet have a clear answer, but the impact of the problem cannot be overstated. Whatever the solutions end up being, they will almost certainly necessitate a combination of developments in nanotechnology, algorithms, and data analyses methods. The analogous problem with neural stimulation at nanoscales is how do you selectively target, i.e., turn on and off, nanoscale electrodes in controlled and coordinated spatial and temporal combinations according to defined optimized protocols to produce the most efficient clinically meaningful stimulation paradigms? As discussed above, these would likely differ from patient to patient and evolve over time in the same patient. Being able to accommodate such changes is at the core of the learning and adaptation machine learning methods applied to BMI and neural prosthesis could provide. The design of BMI devices from a materials and engineering standpoint should be aligned with the implementation and integration of machine learning intended to be deployed as part of the overall system.

Other important considerations are broader topics and go beyond just nanoengineered BMI integration with AI, but no less relevant or important. For example, we do not completely understand the neurophysiology, neural code, and intent of neural signals in the context of information processing. This makes it difficult or not yet possible to develop meaningful machine learning algorithms for controlling BMI. So even if the neural stimulation or neural recording interface technologies were perfected, and even if we could develop efficient and accurate machine learning for closed loop feedback control, it still is not clear what we should be optimizing for. We just do not understand how the brain works well enough to do this. With existing neural prosthesis technologies in particular, it is the brain that adapts to the engineered technology, and not the other way around. Other open problems reflect open engineering challenges, beyond the neuroscience. For example, how can machine learning and AI be efficiently implemented on board the device itself given limited form factors and local computational resources? If access to more significant computational resources on the cloud are required, the usual questions of insuring appropriate bandwidth access becomes important, in particular if such reliance was needed under clinically sensitive situations. Is edge computing a possible emerging alternative?

Finally, it is important to acknowledge and consider ethical challenges that arise from the development and use of these technologies. Neurotechnologies and AI on their own each have important ethical considerations. And in at least one recent commentary the ethical considerations of neuroscience, neurotechnologies, and AI were simultaneously discussed (Yuste et al., 2017). Those authors identified four principles that these technologies must adhere to and respect for each individual: privacy, identity, agency and equality. This needs to be an ongoing and evolving conversation that tracks with the progress of the technology. The potential risks are too high to ignore or defer.

#### CONCLUDING COMMENTS

The integration of machine learning and AI with nanoengineered brain machine and brain computer interfaces offers the potential for significant advances in neurotechnology. BMI's that have the ability to learn and adapt from the environment and situational demands of external requirements offer tremendous possibilities to radically change the treatment and quality of life of patients. It also offers opportunities for non-invasive interactions and collaborations between humans and machines that at the moment are still in the realm of science fiction. It is conceivable that we are approaching an era of personalized individual experiences that will impact both clinical and nonclinical applications. Of course, as with any truly disruptive and paradigm changing progress, there remain many technical challenges that must be overcome, many in no way trivial or easy, and serious ethical questions that have to be thoughtfully considered and navigated. But it is hard not to be excited about

fnins-12-00843 November 14, 2018 Time: 16:39 # 6

the prospects, what it could mean for how we interact with and use technology and computers for everyone, and the life changing effects it could have on the quality of life and well-being of patients who stand to benefit the most.

We end with one last parting consideration. We have argued the position that the machine learning and AI algorithms that will be required to arrive at "smart" nanoengineered brain machine interface systems may include the use of existing state of the art algorithms, but also possibly new neural derived algorithms and machine learning architectures that more directly model computational and systems neuroscience. What we have not argued for, and what is in no way obvious, is a need for artificial general intelligence (AGI) as necessary to achieve this. Advanced applications such as smart adaptive BMI will almost certainly benefit from advanced algorithms that depend on new mathematical models and theory grounded in empirical neurobiological data. But such algorithms in isolation and out of context do not constitute AGI (although they

#### REFERENCES


could conceivably contribute to it). These algorithms need to be able to execute very sophisticated data analyses, pattern recognition, learning, and decision making, but only within the context and embodiment of the neurotechnologies they are supporting. The concept of a self-aware or conscious machine is not required, and should not be confused with the technical considerations that actually are needed, i.e., the discussion in this paper. This distinction is important, because the serious societal and ethical concerns and on-going conversations surrounding AGI are very different than the societal and ethical questions that we need to discuss involving neurotechnologies.

### AUTHOR CONTRIBUTIONS

The author confirms being the sole contributor of this work and has approved it for publication.


fnins-12-00843 November 14, 2018 Time: 16:39 # 7


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Silva. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

fnins-12-00843 November 14, 2018 Time: 16:39 # 8

# Ion Beam Assisted E-Beam Deposited TiN Microelectrodes—Applied to Neuronal Cell Culture Medium Evaluation

Tomi Ryynänen<sup>1</sup> \*, Maria Toivanen<sup>2</sup> , Turkka Salminen<sup>3</sup> , Laura Ylä-Outinen<sup>2</sup> , Susanna Narkilahti <sup>2</sup> and Jukka Lekkala<sup>1</sup>

<sup>1</sup> BioMediTech Institute and Faculty of Biomedical Sciences and Engineering, Tampere University of Technology, Tampere, Finland, <sup>2</sup> NeuroGroup, BioMediTech Institute and Faculty of Medicine and Life Sciences, University of Tampere, Tampere, Finland, <sup>3</sup> Laboratory of Photonics, Tampere University of Technology, Tampere, Finland

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Ikuro Suzuki, Tohoku Institute of Technology, Japan Volker Bucher, Furtwangen University, Germany

> \*Correspondence: Tomi Ryynänen tomi.ryynanen@tut.fi

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 15 December 2017 Accepted: 12 November 2018 Published: 04 December 2018

#### Citation:

Ryynänen T, Toivanen M, Salminen T, Ylä-Outinen L, Narkilahti S and Lekkala J (2018) Ion Beam Assisted E-Beam Deposited TiN Microelectrodes—Applied to Neuronal Cell Culture Medium Evaluation. Front. Neurosci. 12:882. doi: 10.3389/fnins.2018.00882 Microelectrode material and cell culture medium have significant roles in the signal-to-noise ratio and cell well-being in in vitro electrophysiological studies. Here, we report an ion beam assisted e-beam deposition (IBAD) based process as an alternative titanium nitride (TiN) deposition method for sputtering in the fabrication of state-of-the-art TiN microelectrode arrays (MEAs). The effects of evaporation and nitrogen flow rates were evaluated while developing the IBAD TiN deposition process. Moreover, the produced IBAD TiN microelectrodes were characterized by impedance, charge transfer capacity (CTC) and noise measurements for electrical properties, AFM and SEM for topological imaging, and EDS for material composition. The impedance (at 1 kHz) of brand new 30µm IBAD TiN microelectrodes was found to be double but still below 100 k compared with commercial reference MEAs with sputtered TiN microelectrodes of the same size. On the contrary, the noise level of IBAD TiN MEAs was lower compared with that of commercial sputtered TiN MEAs in equal conditions. In CTC IBAD TiN electrodes (3.3 mC/cm<sup>2</sup> ) also outperformed the sputtered counterparts (2.0 mC/cm<sup>2</sup> ). To verify the suitability of IBAD TiN microelectrodes for cell measurements, human pluripotent stem cell (hPSC)-derived neuronal networks were cultured on IBAD TiN MEAs and commercial sputtered TiN MEAs in two different media: neural differentiation medium (NDM) and BrainPhys (BPH). The effect of cell culture media to hPSC derived neuronal networks was evaluated to gain more stable and more active networks. Higher spontaneous activity levels were measured from the neuronal networks cultured in BPH compared with those in NDM in both MEA types. However, BPH caused more problems in cell survival in long-term cultures by inducing neuronal network retraction and clump formation after 1–2 weeks. In addition, BPH was found to corrode the Si3N<sup>4</sup> insulator layer more than NDM medium. The developed IBAD TiN process gives MEA manufacturers more choices to choose which method to use to deposit TiN electrodes and the medium evaluation results remind that not only electrode material but also insulator layer and cell culturing medium have crucial role in successful long term MEA measurements.

Keywords: titanium nitride, microelectrode array, MEA, IBAD, cell culture medium

## INTRODUCTION

A microelectrode array (MEA) is a common tool to measure the electrical activity of various cell types in vitro and to provide an electrical stimulus to the objects under study. The applications of MEAs vary from basic biological research to drug screening and toxicity testing. In neuroscience, it has been found to be applicable for in vitro drug screening and toxicity testing (Johnstone et al., 2010; Ylä-Outinen et al., 2010). Recently, the rise of human pluripotent stem cell (hPSC)-based technologies for human cell-based modeling, including disease modeling, has benefitted from MEA technology (Falk et al., 2016; Odawara et al., 2016).

In its simplest form, MEA consists of a glass substrate, a metal layer containing electrodes, tracks, and contact pads, and an insulator layer with openings on the electrodes and the contact pads. Even though metal electrodes such as Pt, Au, or Ti can be used, they have limitations in their performance. For this reason, metallic microelectrodes are usually coated with a porous material that increases the effective surface area ratio (SAR) and decreases the impedance, leading to a higher signal-to-noise ratio of the electrodes (Bauerdick et al., 2003). Since the early days of MEA (Thomas et al., 1972), platinum black (Pt black) has been one of the most commonly used coating materials for low impedance electrodes. It has excellent electrical characteristics, but in addition to obvious cost issues, a major drawback is that Pt black has been reported to have problems with mechanical stability during long-term use (Heim et al., 2012). Iridium oxide (IrOx), even as a rather common in vivo electrode material (Cogan, 2008), has not reached notable popularity for in vitro microelectrodes. This is likely to be at least partly due to its tendency to lose the low impedance state rather rapidly in a liquid environment (Gawad et al., 2009). Carbon nanotube-based solutions do exist (Gabay et al., 2007; Samba et al., 2014), and even though excellent performance has been reported, they are still more a topic of academic interest than a real choice for active use. The only commonly used substitute for Pt black has been titanium nitride (TiN) (Janders et al., 1996), especially in commercial solutions. Depending on the deposition parameters and methods, the morphology of a TiN thin film may vary a lot from plain to highly columnar. The latter is seen as increased SAR and decreased impedance. Although, some doubts about the performance of TiN exist (Weiland et al., 2002), it can generally be considered as the least problematic high-performance microelectrode coating developed to date. In addition to in vitro electrodes, TiN can be used also in in vivo applications (Stelzle et al., 2001).

There exists a wide range of methods for the fabrication of TiN coatings. Because TiN is applied as the last layer on MEAs in the fabrication process, finding an etching process that is not harmful for the MEA insulator layer, typically Si3N4, and underlying track material, commonly titanium, might be challenging and may require additional process steps for preparing the etch mask. Thus, a lift-off process is favored with TiN. Because photoresist is needed for lift-off and the melting temperature of the glass substrate set limits for the maximum allowed temperature during the TiN deposition process, certain common TiN deposition processes such as atomic layer deposition (ALD) (Xie et al., 2014), thermal chemical vapor deposition (CVD) (Wagner et al., 2008), and physical vapor deposition (PVD) (Gahlin et al., 1995; Peng et al., 2015) techniques must be ruled out when selecting the TiN deposition method for the MEAs. For this reason, reactive sputtering has been the only method used to deposit TiN on MEA electrodes (Egert et al., 1998; Cyster et al., 2002; Bauerdick et al., 2003; Gabay et al., 2007). However, there is also an alternative method for the low temperature deposition of TiN: ion beam assisted deposition (IBAD) in which the e-beam evaporated titanium is bombarded by a flux of low energy nitrogen and argon ions from the ion source to form TiN (**Figure 1**). The dominating mechanism in TiN formation is adsorption of ambient gaseous atoms on the growth surface (Hubler et al., 1988). Over the last three decades, several groups have reported their IBAD TiN experiments for hard coatings (Guzman et al., 1998; López et al., 2001) and more general materials science (Hubler et al., 1988; Huang et al., 2000; Yokota et al., 2004) applicable not only on hard coatings but also, for example, on decoration coatings and microelectronics diffusion barriers. By contrast, as far as we know, IBAD TiN has not been applied on MEAs previously.

In this paper, we evaluated different deposition parameters, including evaporation and nitrogen flow rates, for optimal IBAD TiN microelectrode coating. The coatings are characterized by impedance, charge transfer capacity (CTC) and noise measurements for electrical properties, AFM and SEM for topological imaging, and EDS for material composition. Comparison to sputtered TiN electrodes of commercial MEAs (Multi Channel Systems MCS GmbH) is also reported. To verify the biocompatibility and performance of these novel IBAD TiN microelectrodes, we cultured and measured hPSC-derived neuronal networks for 3 weeks. The neuronal networks were grown in two different cell culture media: neural differentiation medium (NDM) (Heikkilä et al., 2009) and BrainPhys (BPH) supplemented medium recently introduced by Bardy et al. (2015)

to evaluate possible medium derived effects on MEA grown cultures.

#### MATERIALS AND METHODS

#### IBAD TiN Deposition Process Development

Microscope slides (76 mm × 26 mm × 1 mm; Gerhard Menzel GmbH) were used as substrates while optimizing the IBAD TiN deposition parameters. The slides were cleaned with acetone and isopropanol in an ultrasound bath, rinsed with DI water and dried with a nitrogen blow. Cleaned slides were placed in an Orion BC-3000 series box coater (System Control Technologies) equipped with a Telemark 246 e-beam source, Saintech Series III ST55 gridless ion source, Saintech ion current density monitor, and a Meissner trap for 100 nm IBAD TiN depositions. **Table 1** includes different parameter values tested during process development. In all depositions, a filament current of ∼20 A was used, and the vacuum during deposition was in the 10−<sup>5</sup> Torr range (∼10−<sup>3</sup> Pa). The substrate holder was rotated at 5 rpm during the deposition. The 99.995% purity Ti pellets used in both IBAD TiN depositions and later in Ti track depositions in the MEA fabrication were purchased from g-materials. Right after the deposition, the color of the thin films was evaluated visually.

#### AFM Measurements

While optimizing the IBAD TiN deposition process, an atomic force microscope (XE-100 AFM, Park Systems) equipped with an ACTA probe (AppNano; radius of curvature, 6 nm) was used for measuring the effective surface area ratio (SAR). An area of 1µm × 1µm was measured in intermittent mode for each sample. XEI analysis software (Park Systems) was used for calculating the SAR as per the formula

$$SAR = \frac{A\_S - A\_G}{A\_G} \ast 100\tag{1}$$

where A<sup>G</sup> is the plain geometric area and A<sup>S</sup> is the total surface area of the corresponding region. The result was finally given as a mean of two areas measured from the same sample. In addition, the software was used to calculate the root-mean-square roughness Rq.

#### IBAD TiN MEA Fabrication

Microscope slide grade glass plates (49 mm × 49 mm × 1 mm; Gerhard Menzel GmbH) were used as substrates for the MEAs. The slides were cleaned with acetone and isopropanol in an ultrasound bath and oxygen plasma before 400 nm of titanium was e-beam deposited at 5 Å/s on the slides. Electrode sites (30µm in diameter), tracks and contact pads were patterned to the titanium layer in a wet etching [120 H2O: 4 H2O<sup>2</sup> (30%): 3 HF (50%)] process in which PR1-2000A positive photoresist (Futurrex, Inc.) was used as an etching mask. Next, 500 nm of silicon nitride was PECVD deposited as an insulator layer at 300◦C. PR1-2000A was again used as an etching mask when reactive ion etching with SF<sup>6</sup> and O<sup>2</sup> gases was performed with Vision 320 RIE (Advanced Vacuum) to etch the openings on electrode sites and contact pads. The etching mask was not removed after the etching but was reused as a lift-off mask in the IBAD deposition of 400 nm of TiN. For comparison purposes, also MEA versions with 200 nm layer of IBAD TiN as well as MEAs without TiN were fabricated. Just prior to TiN deposition, a 10 min Ar sputter etch was run with the ion source to remove the native oxide layer on titanium electrode sites. IBAD TiN deposition parameters were as follows: anode voltage, 225 V; filament current, 20 A; N<sup>2</sup> flow, 13.2 sccm; Ar flow, 3.3 sccm; ion current density, 14 µA/cm<sup>2</sup> ; ion current density monitor bias, 35.0 V; deposition rate, 2 Å/s; and vacuum, 10−<sup>5</sup> Torr range (∼10−<sup>3</sup> Pa). The substrate holder was rotated at 5 rpm during the deposition. Finally, lift-off was performed in an acetone ultrasound bath. Either an in-house made PDMS ring or SpikeboosterTM 6-well culture chamber (BioMediTech) (Kreutzer et al., 2012) was attached on MEAs to form a pool for Dulbecco's phosphate buffered saline (DPBS) or cell culture media. All of the photolithography masks used in this work were in-house fabricated with a µPG501 direct writing system (Heidelberg Instruments Mikrotechnik GmbH) on chrome mask blanks from Clean Surface Technology Co.

#### Impedance and Charge Transfer Capacity Characterization

The pools on the MEAs were filled with DPBS (PBS Dulbecco w/o Ca++, Mg2+, Biochrom GmbH), and the MEAs were placed in a temperature chamber at 37◦C inside petri dishes for at least 20 h. Subsequently, the MEAs were decreased to room temperature


Parameters of sample 2 (bolded) were chosen to be used in the IBAD TiN MEA fabrication due to the highest surface area ratio.

for at least 1 h before the impedance measurement. MEA-IT60 from MCS, a dedicated device for measuring the impedances of all the microelectrodes of a MEA, was used as the measurement device. The measurement was performed at 1 kHz frequency with the sinusoidal test signal being 100 mV and an external Ag/AgCl pellet acting as a grounding electrode. Faulty electrodes existing both in commercial and in-house made MEAs were excluded before calculating mean values for each MEA. For a few randomly selected electrodes of in-house made IBAD TiN MEAs and, for comparison, also of pure Ti MEAs without TiN electrode coating an additional electrochemical analysis was performed. Frequency dependency of the impedance and CTC were measured with Iviumstat potentiostat (Ivium Technologies B.V.). The frequency range was from 1 to 100 kHz and a Pt wire (ALS-Japan) was used as the counter electrode in the impedance measurement. CTC was integrated from the third CV curve when the voltage was ramped between −0.9 and 0.9 V. The same Pt wire acted as the counter electrode as in the impedance measurements, and the reference electrode was DRIREF-2 (World Precision Instruments). Scan rate was 100 mV/s.

#### Noise Characterization

Noise characterization was performed as part of the cell culture experiments, where the cell culture medium acted as an electrically conducting solvent. After taking the MEAs from the incubator, they were at first left to stabilize in headstage for 3 min without recording the data. Then, the MEAs were measured for 10 min with the MEA2100 MEA system, MC\_Rack software, and temperature controllers TC02 set at 37◦C (all from MCS). The voltage signal was filtered (200–3,000 Hz bandpass), and the noise for each electrode of each MEA was calculated as an estimate of the standard deviation of background noise previously described in Quiroga et al. (2004). In calculating noise values for each MEA-medium combination, electrodes with a noise value above 7.0 µV were excluded as they were considered faulty electrodes. Mann-Whitney U-test was performed to indicate statistical significances: for each MEA type between the media at days 6 and 18, and for each medium type between the MEAs at days 6 and 18. P < 0.05 were considered significant.

### SEM Imaging and EDS

Zeiss Crossbeam 540 FIB-SEM (Carl Zeiss Microscopy GmbH) with a Gemini II SEM column and Oxford Instruments X-Max<sup>n</sup> 80 EDS detector was used in SEM imaging and EDS measurements. In the imaging, the acceleration voltage was 1 kV, and magnifications in **Figure 2b** were 1.16 and 15.35 kX. In the EDS measurements, acceleration voltages from 7 to 15 kV were used.

### Neural Differentiation

The human embryonic stem cell (hESC) line Regea 08/023 was used in the experiments. BioMediTech has approval from the Finnish Medicines Agency (FIMEA) to perform research with human embryos (Dnro 1426/32/300/05). There are also supportive statements from the regional ethical committee of Pirkanmaa Hospital District for the derivation, culturing, and differentiation of hESCs (R05116). Neurons were differentiated from hESCs as previously described (Lappalainen et al., 2010). Neuronal differentiation medium (NDM) consisted of 1:1 DMEM/F12 and Neurobasal medium supplemented with 1 × B27, 1 × N2, 2 mM GlutaMax (all from Gibco Invitrogen), and 25 µ/mL penicillin/streptomycin (Lonza Group Ltd) and, during the differentiation stage, 20 ng/ml of basic fibroblast growth factor (bFGF) (R&D Systems) as previously described (Lappalainen et al., 2010) with or without low-dose naltrexone LDN193189 (100 nM; Stemcell Technologies, Inc.).

#### MEA Preparation and Adherent Culture

MEA preparations were performed as previously described (Heikkilä et al., 2009) with some modifications. MEAs (10 BMT MEAs and 9 60-6wellMEA200/30iR-Ti from MCS) were combined with SpikeBoostersTM (BioMediTech) (Kreutzer et al., 2012) and coated with 0.05% (w/v) polyethylenimine (PEI) incubated overnight, washed with sterile H2O, and coated with 20µg/ml of mouse laminin (both from Sigma-Aldrich) and incubated overnight. A 48-well plate (Thermo Scientific) was coated with 20µg/ml or 10µg/ml of mouse laminin in wells with or without coverslips (Ø 9 mm, VWR).

The 8 week pre-differentiated neurospheres were dissected into small cell aggregates (ø∼50–200µm), and 7–10 of them were plated onto the coated MEA wells and the control 48-well plate, which were filled with NDM. During the 1st week, the medium was gradually switched to BrainPhys medium (BPH) consisting of BPH Neuronal Medium supplemented with 1 × NeuroCult SM1 Neuronal Supplement, 1 × N2 Supplement-A (all from Stemcell Technologies), GlutaMax to 2 mM final concentration, and 25 µ/ml penicillin/streptomycin for half of the cells. Additionally, from 2 days in adherent culture 1 mM cyclic adenosine monophosphate (cAMP) and 200 nM ascorbic acid (AA, both from Sigma-Aldrich) were added to the media and from 7 days after plating 8 ng/ml of bFGF and 10 ng/ml of brainderived neurotrophic factor (BDNF, Gibco Invitrogen) were added to the media. The cells were maintained in a humidified incubator at 37◦C and 5% CO2, and half of the medium was refreshed 3 times per week. The cells were imaged weekly using a phase contrast microscope (Eclipse Ts2R, Nikon). In addition, the control plate was maintained in Cell-IQ (Chip-Man Technologies) 10 days after plating for 26 h with a 1 h imaging interval (**Supplementary Videos 1–3**). Spontaneous activities of neuronal networks were measured for 10 min twice per week.

#### Immunocytochemistry

The control plate cells were fixed after 12 days in adherent culture, and immunocytochemical staining was performed as previously described (Lappalainen et al., 2010). Primary antibodies, rabbit polyclonal anti-Microtubule-Associated Protein 2 (MAP2) (1:400; Millipore), mouse anti-beta-III Tubulin (β-tub) (1:1000; Sigma-Aldrich), chicken anti-Glial Fibrillary Acidic Protein (GFAP) (1:4000; Abcam), mouse anti-Synaptophysin (1:500; Sigma-Aldrich), chicken MAP2 (1:4000; Novus), and chicken β-tub (1:4000; Abcam) were used together with secondary antibodies Alexa 488 donkey anti-rabbit, Alexa 568 donkey anti-mouse and Alexa 647 goat anti-chicken (all 1:400; Invitrogen). In addition, the nuclei of the cells were

stained with 4',6-diamidino-2 phenylindole (DAPI), which was included in the mounting medium (Prolong Gold, Molecular Probes). The cells were imaged with a fluorescence microscope (Olympus IX51, Olympus Corporation) and a laser scanning confocal microscope (LSM 780, Carl Zeiss).

#### MEA Signal Analysis and Statistics

Signal analysis from the MEA data was performed using MATLAB (The MathWorks, Inc.) with a custom-made analysis program based on work by Quiroga et al. (2004) in which the spike detection threshold was set to 5, and spikes larger than 500 times the standard deviation of noise were excluded as artifacts. An electrode was regarded as an active electrode if the spike frequency was more than 0.04 Hz. The threshold was determined by measuring spike rates from MEAs without cells and MEAs with TTX-silenced neuronal cultures (data not shown). For spike waveform analysis, 0.8 ms of voltage signal before and 1.76 ms after the largest absolute value of the spike from the filtered data were clipped. The detector dead time between two waveforms was 1.48 ms. Bursts were detected using a MATLAB code based on work by Kapucu et al. (2012) with additional conditions: burst detection was only applied to channels where the total spike frequency was at least 0.167 Hz (10/min). Thereafter, burst analysis criteria included a median of more than two spikes per burst and more than 1 burst per electrode. For each MEA type, Mann-Whitney U-test was performed to indicate statistical significances between the media at each measurement time point. P < 0.05 were considered significant.

### RESULTS

#### IBAD TiN Process Development

With the assumption that the highest SAR would lead to the lowest microelectrode impedance, we focused on finding the IBAD deposition parameters that would give the highest SAR for the TiN thin film. Briefly, the purpose was to find deposition parameters for a thin film that we expect to give the lowest impedance, not necessarily the purest TiN from material science point of view. The deposition parameters tested while optimizing the IBAD TiN deposition process are presented in **Table 1**. The effect of changing deposition rates from 1 to 5 Å/s was tested for the same anode voltage and gas flow rate. In addition, some experiments with different anode voltages (sample 7), gas flow rates (samples 5–8) and pulsing of the ion beam (sample 8) were tested. Too high evaporation rate (sample 4) or a too low nitrogen flow rate (sample 7) led to gray thin films resembling pure Ti, which indicates that the conditions did not support the formation of TiN. Lower deposition rate (samples 1 and 2) and higher nitrogen content to argon (sample 5) were seen as brownish thin films, which more closely resembled the almost black thin film in MCS's sputtered TiN MEAs. The remainder (samples 3, 6, and 8) were goldish, which is considered to be the color of TiN hard coatings (Jiang et al., 2004).

In the AFM measurements, purple-bronze-colored sample 2 clearly had the highest SAR of 13.1%. The root-mean-square roughness (Rq) was 3.1 nm. The AFM image of sample 2 is shown in **Figure 2a**. According to the assumption of the highest SAR giving the lowest impedance and noise level, we chose the deposition parameters of sample 2 to be used as the deposition parameters in MEA fabrication. **Figure 2b** shows an SEM image of the IBAD TiN microelectrode. The slight pillar-like structure of TiN can be seen in the image. The EDS measurements showed <1% variation for the N/Ti ratio, indicating excellent homogeneity of the coating. However, as N is a light element and produces only the Kα peak that partly overlaps with the Ti Lα peak, EDS is better suited for comparative analysis than exact quantification of the N/Ti ratio.

#### MEA Performance Characterization

The experimental details are shown in **Table 2**. Once we had determined the optimal IBAD TiN deposition parameters, we fabricated a batch of IBAD TiN MEAs (hereafter referred to as BMT MEAs) in a 6-well layout mimicking MCS's 60- 6wellMEA200/30iR-Ti-w/o array design. In this design, the

TABLE 2 | Experimental setup for MCS MEAs and BMT MEAs.


The impedance of each MEA was measured before and after the cell experiments. All the cell experiments (E1–E3) included both NMD and BPH medium tests.

\*Results excluded due to incubator malfunction.

Included measurements are bolded: impedance measurement results for MCS MEAs from before cell experiment E1 and after experiment E2 and for BMT MEA before and after cell experiment E3. Cell experiment results are presented from cell experiment E3.

microelectrodes are grouped in six 3 × 3 electrodes areas for a total of 6 areas with 9 electrodes / MEA. Both in IBAD TiN MEAs and MCS's MEAs the diameter of the electrodes is 30 µm. **Table 3** presents the impedance values for the included results, which are grouped by both MEA type and medium used in the cell experiments. Before the cell experiments, the BMT's IBAD TiN electrodes had ∼2× higher impedance compared with that of MCS's sputtered TiN electrodes, approximately 90 k vs. 45 k, respectively. However, as the impedances of Au, Pt or ITO MEAs, i.e., MEAs without a porous electrode coating are typically ∼10× higher (∼1 M), the impedance of IBAD TiN is still comparable to sputtered TiN electrodes. After the cell experiments, the impedance of both IBAD and sputtered TiN electrodes increased >100 k; thus, in this sense as well, the behaviors of the two electrode types were comparable. For IBAD TiN MEAs in BPH medium the after cell experiments impedance was significantly lower, only 35 k, but this is because of severe insulator layer corrosion (**Figure 7c**). Thus, the result is not reliable, as the impedance in this case is not impedance of the electrodes only but rather impedance of both electrodes and tracks. IBAD TiN MEAs in NDM medium, on the contrary, suffered only minor corrosion (**Figure 7d**), indicating that we can consider their impedance values reliable enough for comparison. The insulator layer of MCS MEAs survived the cell experiments without visible corrosion in both media.

The measurement of impedance as a function of frequency show (**Figures 3A,B**) that the thickness of the IBAD TiN strongly affects to the impedance. Decreasing the thickness from 400 to 200 nm about doubles the impedance (at 1 kHz). As expected, compared with the Ti electrodes without the TiN coating the IBAD TiN coating significantly decreases the impedance and also improves the stability at low frequencies. Charge transfer capacity (CTC) integrated from the third CV curves (**Figure 3C**) was 3.3 ± 0.2 mC/cm<sup>2</sup> for IBAD TiN microelectrodes and about one tenth of that for Ti electrodes without TiN coating. For MCS MEAs CTC of 2.0 ± 0.2 mC/cm<sup>2</sup> was measured.

The noise level of each MEA type and medium combination was evaluated by calculating the estimate for standard deviation of background noise from 10 min cell measurement data (Quiroga et al., 2004). The results are summarized in **Figure 4A**. Briefly, the noise level of BMT MEAs was significantly lower from that of MCS's sputtered TiN MEAs under the same time point and condition (p < 0.001 for all). Moreover, BPH medium decreased the noise significantly compared with that of the NDM medium (at day 6 in BMT MEAs, p ≤ 0.001, and at day 18 in both MEA plate versions, p ≤ 0.001). However, typical examples of raw measurement data plotted in **Figure 4B** show well that, despite the differences in both numerical impedance and noise results, in practice there is no notable difference in the base noise levels and the signal peaks can be separated from the noise equally well with each MEA type and medium combination.

#### Effect of Cell Culture Medium

#### Neuronal Network Formation in NDM and BPH Media and BMT and MCS MEAs

Human pluripotent stem cell (hPSC)-derived neurons were cultured in neural differentiation medium (NDM) and BrainPhys medium (BPH) in control cell culture plastic wells. Both medium types supported the formation of MAP2 and β-tub-positive neuronal networks with expression of synaptophysin (**Figures 5a–f**) during a 12 days follow-up period. However, GFAP-positive astrocytes were only found in cultures supplemented with BPH medium (**Figure 5e**). Even though the network formation was good in both media, the organization of the networks differed between them. The neuronal networks were denser in BPH than those in NDM medium (**Figures 5c,f**). Neuronal cells migrated out of the cell aggregates in both media but more extensively in BPH. In NDM, neuronal cells remaining in the aggregates extended long neurites, which were less common in BPH.

Even though the networks grew well in both media in both MEA types at the beginning of the experiment (**Figures 5g,j,m,p**), after 1–2 weeks the neuronal networks started to retract and form clumps in BPH (**Figures 5k,q**). Network retraction also occurred in some NDM wells but typically later than in BPH (**Figures 5h,n,i,o** vs **Figures 5k,q,l,r**). The results were the same for both MEA types. The cultures were kept for 19–20 days on MEAs until the network retraction was too extensive, especially in BHP medium, for further measurements. Example videos of network growth on control plates after 10 days in adherent culture (**Supplementary videos 1–3**) show the typical behavior of the neuronal networks in both media over 26 h. At this point, the networks were not yet retracting in NDM or BPH on cell culture plastic (**Supplementary videos 1, 2**). However, network retraction in BHP on coverslips was substantial (**Supplementary video 3**). The cell culture experiments were repeated, and similar results were obtained. Network retraction, clump formation, and cell detachment occurred first and were more prominent in BPH than in NDM.

#### Development of Electrophysiological Activity in NDM and BPH Media on BMT and MCS MEAs

The percentage of active electrodes (spike frequency >0.04 Hz) per network was higher in BPH than in NDM at all measurement


\*Unreliable result as the Si3N<sup>4</sup> insulator layer was almost completely corroded.

\*\*Between impedance measurements.

time points (M1 = 6 days, M2 = 11 days, M3 = 13 days, M4 = 18 days, and M5 = 19–20 days) on both BMT and MCS MEAs (**Figure 6A**). The results were statistically significant at most of the time points (BMT M1 p = 0.005, M2 p = 0.035, M4 p < 0.001; MCS M1 p = 0.030, M3 p = 0.045, M4 p < 0.001, M5 p < 0.001). Even though the BPH increased the amount of active electrodes, the median spike frequency in active electrodes was not clearly increased in BPH medium (**Figure 6B**). Depending on the measurement time point, more spikes were recorded in either BPH or NDM medium in both MEA types. Furthermore, the median burst count during the 10 min recording was not higher in BPH than that in NDM (**Figure 6C**). However, more electrodes recorded bursts in BPH medium. The spikes per burst medians were rather similar between the media in both MEA types (**Figure 6D**). Additional MEA analysis results are presented in **Supplementary Tables 1, 2**. Overall, BPH medium increased the amount of active electrodes but did not enhance the spike frequency or network maturation based on the burst parameters.

#### Insulation Layer Corrosion

Corrosion of the insulator layer in BMT MEAs not only affect the reliability of the impedance readings after the cell experiments but the vanishing of the insulation also caused a decrease in the MEA signal amplitudes. Examples of spike waveforms recorded using MEA with a badly corroded insulation layer are presented in **Figure 7a**. In comparison to spike waveforms recorded with MEA still with proper insulation left (**Figure 7b**), the signal amplitude from the badly corroded MEAs is substantially lower. In addition, the

combination at two time points. Data are expressed as the median (band inside the box) with interquartile range (IQR; box) and minimum and maximum values (whiskers) at two time points: 6 and 18 days on MEA. BMT, BioMediTech MEA; MCS, Multi Channel Systems MEA; NDM, neuronal differentiation medium; BPH, BrainPhys medium. \*\*\*p ≤ 0.001. In addition, the noise level of BMT MEAs was significantly lower from that of MCS's sputtered TiN MEAs under the same time point and condition (p < 0.001 for all). (B) Examples of raw measurement data of each MEA type and medium combination at two time points. Curves have been shifted vertically for clarity.

amplitude of the noise was lower in the badly corroded MEAs.

#### DISCUSSION

The aim of this study was to demonstrate that sputtering is not the only method for fabricating TiN microelectrodes, as an alternative method exists. For the sputtered commercial TiN microelectrodes (MCS), we measured an impedance range of 30–50 k, which is in line with the range provided in the manufacturer's brochure. This range is <80 k as reported by Egert et al. (1998) in an early paper describing the sputtered TiN process for MEAs. Thus, it is very likely that by continuing process parameter optimization we could cut some tens of k from the impedance of IBAD TiN microelectrodes. Our results showed that with IBAD TiN we reached impedance levels <100 k, similar to those of current sputtered TiN microelectrodes. The signal-to-noise ratio was also very similar between these types of MEA electrodes. The CTC values reported in literature for sputtered TiN microelectrodes vary a lot. The original record by Janders et al. (1996) was as high as 42 mC/cm<sup>2</sup> , which was later questioned by Weiland et al. (2002) who reported ∼2.4 mC/cm<sup>2</sup> . That is in line with our result for sputtered TiN, 2.0 mC/cm<sup>2</sup> . On the contrary, Gerwig et al. (2012) and Li et al. (2011) have reported both only 0.45 mC/cm<sup>2</sup> . So, if Janders' record is ignored, IBAD TiN seems to perform well against sputtered TiN with its CTC of 3.3 mC/cm<sup>2</sup> also in this aspect. Thus, IBAD TiN microelectrodes can be expected to be competitive also in stimulation use. The produced IBAD TiN MEAs were compatible for cell measurements, especially when TiN is concerned. Overall, for in-house production, the availability of deposition equipment determines which TiN production method is used. For operators with an e-beam but no sputtering system, upgrading the e-beam coater with an ion source could be the most economical choice to obtain tools for in-house TiN deposition.

Here, the IBAD deposition parameters used in MEA fabrication were chosen based on the SAR; higher SAR was expected to result in lower impedance. Mumtaz and Class (1982) linked brownish color to more porous TiN structure, which agrees with our SAR results and color observations. Further, the coating colors of the produced samples were in line with the observations from previous studies (Mumtaz and Class, 1982; Roquiny et al., 1999). Additionally, it might be interesting to also fabricate MEAs with IBAD deposition parameters other than the ones chosen here as optimal, just to confirm whether the highest SAR is the true defining factor of the lowest impedance. The SAR results here are based on AFM sampling of only two 1µm × 1µm areas per sample, which may also leave room for error. However, the measured surface roughness value of 3.1 nm is in good agreement with the value of 3.0 nm by Cyster et al. (2002) for their DC magnetron-sputtered TiN. They also observed a rather strong dependency between the TiN layer thickness and roughness, which is in agreement with our observation of higher thickness meaning lower impedance. In order to keep the MEA surface rather planar and to avoid difficulties in certain process steps there is, however, not much room to play with the TiN thickness. Other parameters commonly connected to IBAD but not evaluated in this study are the substrate temperature and the ion beam incident angle which both may affect on the thin film properties. In our system adjusting those two parameters just was not possible. However, we did observe some temperature increase inside the deposition chamber after the IBAD process, but according to the ion source manufacturer, Saintech, with their ion sources the increase in the substrate temperature should be only very modest 20–30◦C compared to ambient temperature, even if the ion source were operated on full power.

When comparing BMT-and MCS-fabricated MEAs, one should also note that there are some minor differences in the design. Only the electrode area of both BMT and MCS MEAs is equal in layout. Wider parts of the tracks and contact pads, on the contrary, have some "artistic" differences as we did not

FIGURE 6 | MEA activity from NDM and BPH media on BMT and MCS MEAs. (A) Percentage of active electrodes per network (spike frequency >0.04 Hz). Higher percentages of electrodes were active in BPH on both MEA types. (B) Median spike frequency in active electrodes. (C) Median burst count over 10 min. (D) Spikes per burst median. Measurement 1 = 6 days, 2 = 11 days, 3 = 13 days, 4 = 18 days, and 5 = 19–20 days in adherent culture. Median and interquartile ranges are shown.

have MCS's mask layout CAD files available. As MCS brochures reveal only the insulator layer thickness, it is possible that Ti and TiN thicknesses are not equal in both MEA types. These differences can probably generally be ignored, but there was a notable difference in the corrosion resistance of the insulator layer even if both fabricators use 500 nm PECVD Si3N4. As we also observed similar corrosion in a control MEA with no IBAD TiN layer on titanium electrodes, we are confident that the IBAD TiN deposition process itself, despite potential thermal expansion-related issues, or the related rather long ultrasound bath during lift-off, is not the reason for the more corrosionprone insulator layer in BMT MEAs. Our PECVD process only produces lower quality Si3N<sup>4</sup> compared with MCS's process and requires further optimization. One should note that the IBAD TiN process introduced here is by no means connected to our PECVD Si3N<sup>4</sup> process and whoever adapts IBAD TiN to their MEAs is free to use whatever insulator layer they find the most suitable for their application. If one does not want to move to polymer insulators like polyimide or SU-8, one common solution is to replace Si3N<sup>4</sup> with a more stable but harder to etch sandwich structure of SiO2/Si3N4/SiO<sup>2</sup> (Buitenweg et al., 1998; Yeung et al., 2007). Even if we did not observe as strong insulator layer corrosion in MCS's MEAs, these MEAs do still present some corrosion, and we have seen outside this study that, in the long run, the MCS MEA insulator layer does eventually totally wear out as well. Also Wagenaar et al. (2004) have reported long term insulator failure in MCS MEAs. In fact, there are more common studies (Schmitt et al., 2000; Herrera Morales, 2015), where not only Si3N<sup>4</sup> but also many other commonly used insulator materials have been found to have a poor corrosion resistance in a biological environment. It is evident that the MEA community should stop focusing only on developing new electrode materials and put effort on studying the insulator materials as well.

Another interesting finding related to TiN was that the impedance of both sputtered and IBAD-deposited microelectrodes increased greatly during the experiments. As the impedance also increased for the control group of IBAD TiN MEAs not subject to any cell experiments, it is not necessarily only the cells or cell culture medium that either harms the electrode material or leaves some type of impedance-increasing residual on the surface of the MEA. It seems that storing the MEAs in normal room atmosphere between experiments may

insulator layer may have thinned all around, but it is fully corroded only from a small, barely visible area near the electrode as indicated by the red arrow.

be the primary reason for electrode degradation, likely due to the partial oxidation of Ti(N). In our preliminary tests of storing IBAD TiN MEAs in deionized water in a refrigerator, the electrodes retained constant impedance for a 1 week test period. An open question is whether the increasing impedance saturates at some point and the performance of the MEA at this point. As the condition of MEA is often controlled by checking the impedance, the combination of insulator layer corrosion and increasing electrode impedance due to oxidation makes the evaluation of the condition of MEA somewhat tricky; the impedance may stay within certain limits not because of MEA being consistent but because those two factors have compensated for each other.

In the cell culture experiments, we tested 2 media with both type of MEAs. In BPH, a significant increase in active electrode percentages was found compared to NDM. Earlier, similar results have been shown for BPH compared to standard cell culture medium (Bardy et al., 2015). Overall, our results may indicate that BPH enhanced the activity of the neurons directly or enabled denser neuronal network organization compared to NDM (see all the details in **Supplementary Table 1**). Interestingly, BPH did not increase the spike frequencies in this study in contrast to reported by Bardy et al. (2015). The burst count was neither increased in BPH compared to NDM, thus showing in both media the typical developmental increase during 3 weeks follow up as previously reported (Heikkilä et al., 2009). The median spikes per a burst (**Figure 6D**) as well as other burst parameters (**Supplementary Table 2**) showed no statistical differences between BPH and NDM. These results indicate that BPH did not enhance the maturation of the neuronal networks.

Importantly, our results revealed that BPH did not support long-term cell culturing in either of MEA types as neuronal networks started to retract and form cell clumps after 1–2 weeks and resulted in experiment termination in 3 weeks timepoint. Network retraction and cell clumping has been mentioned also in the study of Bardy et al. (2015) as a minor problem while here in our longer study it became a major problem. The long term stability of the network is most important for hPSC-derived neurons that require several weeks or even longer to develop mature neuronal activity (Heikkilä et al., 2009; Odawara et al., 2016).

In summary, we verified that IBAD is a valid method for producing TiN electrodes for MEA systems. Thus, it can be considered as an alternative TiN deposition method for sputtering. We also stated that BPH medium supported the development of neuronal activity on MEAs, although it caused problems in cell behavior and MEA insulator layer stability in long-term cultures. Thus, as insulator material, electrode material, and even cell culture medium can have detrimental effects on recording quality of MEAs especially with long-term cultures, all of these aspects should be carefully evaluated.

#### AUTHOR CONTRIBUTIONS

TR is responsible for the IBAD TiN process development, BMT MEA design and fabrication, AFM, impedance, and CV measurements, and technical data analysis excluding noise analysis performed together with MT and LY-O. MT is responsible for the cell experiments and cell data analysis. TS operated SEM and EDS. TR and MT wrote the manuscript. LY-O, SN, and JL participated in the project design with TR and MT and provided additional support to analysis and writing of the manuscript.

#### FUNDING

This work was funded by Business Finland (formerly known as the Finnish Funding Agency for Technology and Innovation [TEKES]), the Academy of Finland (MEMO [311017, 311021 311022] and grant 286990 for LY-O), the Council of Tampere Region, and the Finnish Culture Foundation (grant for TR).

#### REFERENCES


#### ACKNOWLEDGMENTS

Preliminary results of this study have been published previously as a conference abstract in Ryynänen et al. (2016).

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnins. 2018.00882/full#supplementary-material

extracellular recording of neural networks. J. Neurophysiol. 108, 1793–1803. doi: 10.1152/jn.00711.2011


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Ryynänen, Toivanen, Salminen, Ylä-Outinen, Narkilahti and Lekkala. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Biocompatibility of a Magnetic Tunnel Junction Sensor Array for the Detection of Neuronal Signals in Culture

Daniela Moretti <sup>1</sup> \*, Mattia Lorenzo DiFrancesco1,2, Parikshit Pratim Sharma<sup>3</sup> , Silvia Dante<sup>4</sup> , Edoardo Albisetti <sup>3</sup> , Marco Monticelli <sup>3</sup> , Riccardo Bertacco3,5, Daniela Petti <sup>3</sup> , Pietro Baldelli 1,6 and Fabio Benfenati 1,2

<sup>1</sup> Center of Synaptic Neuroscience and Technology, Istituto Italiano di Tecnologia, Genova, Italy, <sup>2</sup> IRCCS Ospedale Policlinico San Martino, Genova, Italy, <sup>3</sup> Department of Physics, Politecnico di Milano, Milan, Italy, <sup>4</sup> Department of Nanoscopy & NIC@IIT, Istituto Italiano di Tecnologia, Genova, Italy, <sup>5</sup> IFN-CNR, Politecnico di Milano, Milan, Italy, <sup>6</sup> Department of Experimental Medicine, University of Genova, Genova, Italy

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Davide Lovisolo, Università degli Studi di Torino, Italy Rodrigo Lozano, Karolinska Institutet (KI), Sweden Anja Kunze, Montana State University, United States

> \*Correspondence: Daniela Moretti danyfan85@hotmail.com

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 22 August 2018 Accepted: 19 November 2018 Published: 12 December 2018

#### Citation:

Moretti D, DiFrancesco ML, Sharma PP, Dante S, Albisetti E, Monticelli M, Bertacco R, Petti D, Baldelli P and Benfenati F (2018) Biocompatibility of a Magnetic Tunnel Junction Sensor Array for the Detection of Neuronal Signals in Culture. Front. Neurosci. 12:909. doi: 10.3389/fnins.2018.00909 Magnetoencephalography has been established nowadays as a crucial in vivo technique for clinical and diagnostic applications due to its unprecedented spatial and temporal resolution and its non-invasive methods. However, the innate nature of the biomagnetic signals derived from active biological tissue is still largely unknown. One alternative possibility for in vitro analysis is the use of magnetic sensor arrays based on Magnetoresistance. However, these sensors have never been used to perform longterm in vitro studies mainly due to critical biocompatibility issues with neurons in culture. In this study, we present the first biomagnetic chip based on magnetic tunnel junction (MTJ) technology for cell culture studies and show the biocompatibility of these sensors. We obtained a full biocompatibility of the system through the planarization of the sensors and the use of a three-layer capping of SiO2/Si3N4/SiO2. We grew primary neurons up to 20 days on the top of our devices and obtained proper functionality and viability of the overlying neuronal networks. At the same time, MTJ sensors kept their performances unchanged for several weeks in contact with neurons and neuronal medium. These results pave the way to the development of high performing biomagnetic sensing technology for the electrophysiology of in vitro systems, in analogy with Multi Electrode Arrays.

Keywords: magnetic tunnel junction (MTJ), biocompatibility, neuron culture, bio-magnetic field, in vitro, sensor

### INTRODUCTION

In the last 20 years, the study of the magnetic field generated by the electrical activity of the brain has revolutionized neuroscience. In particular, magnetoencephalography (MEG), due to its unprecedented spatial and temporal resolution (Sander et al., 2012; Borna et al., 2017), gained wide clinical applications in detecting and localizing pathological cortical activity in patients with brain tumors or intractable epilepsy (Stufflebeam, 2011). Despite the wide use of this new technique, the nature of MEG signals at local level is still not completely understood (Haueisen and Knösche, 2012). Furthermore, during the last decade, there has been an increasing need to extend biomagnetic signal detection to microscale for higher spatial resolution and system integration in real-time and robust processes (Shen et al., 2018). The big challenges to overcome are (i) preserving the weak biomagnetic signals that can be easily polluted by environmental noise and (ii) detecting magnetic signals, which are heavily damped even at short distances. As described in (Hall et al., 2012), the magnetic field decays one order of magnitude faster than the electric field. In this context, there is evident crucial interest in developing devices for local magnetic recording in in-vitro systems.

Growing primary neuronal cultures directly on top of magnetic sensors offers the possibility to minimize the distance between the source of the biological signal and the detector to maximize sensitivity. However, this brings up potential issues related to the biocompatibility of magnetic sensors and to the viability of neurons in direct contact with them. In addition, in order to maximize the sensitivity to the magnetic signal, one has to take into account that the magnetic field associated with the propagation of the action potential along the axon arises mainly from the axial intracellular currents and is directed perpendicular to the axon (Roth and Wikswo, 1985; Hall et al., 2012).

In the field of magnetic sensors for in-vitro applications, attempts (Barry et al., 2016) have been made by using the nitrogen-vacancy quantum defects in diamond to detect the magnetic field produced by action potentials in the squid and worm giant axons. Using a different technology, promising candidates for the detection of the magnetic field in-vitro are magnetoresistive sensors (Graham et al., 2004) based on Giant Magnetoresistance (GMR) (Martins et al., 2009; Gaster et al., 2011) or Tunneling Magnetoresistance (TMR) (Albisetti et al., 2013; Sharma et al., 2017a) due to their high sensitivity, electrical readout, capability to work at room temperature (RT) and potential compatibility with a cell culture.

Due to the long-term biocompatibility requirements and the high sensitivity of cultured neurons to toxic elements (in particular metals) that can be released by the devices immersed in a saline solution, no attempts have been carried out so far for primary neuronal cultures. Recent studies assessing the capability of these magnetoresistive sensors in detecting a magnetic field of biological origin at the local scale (Barbieri et al., 2016; Caruso et al., 2017) concern macroscopic in vitro and in vivo systems, such as muscle or visual cortex, except for Sharma et al. (2017b), in which very preliminary results on cell viability were shown for neurons cultured for 2 weeks on magnetic tunneling junctions (MTJ) protected by 170 nm of capping layers.

However, a detailed study of the biocompatibility of the system aimed at the optimization of the sensitivity of the platform is still missing. As previously mentioned, low thicknesses of the capping layers are required to overcome the fast decay of the magnetic field. In addition, one has also to take into account that magnetoresistive sensors are sensitive to external magnetic field along only one axis, depending on the magnetic anisotropies of their reference magnetic layers (Sharma et al., 2016).

In the present work we investigate in detail the biocompatibility of MTJ sensors with murine embryonic hippocampal neurons cultured on the top of the device by viability assays, immunocytochemistry and patch-clamp recordings. We studied the dendritic and axonal growth, the formation of synaptic connections and the maturation of the firing properties. Moreover, engineered cultures are grown on top of the sensors in order to maximize the sensitivity of the neuronal-sensor interface. We show that our devices are fully biocompatible up to 3 weeks and preserve their physical integrity and performance.

## MATERIALS AND METHODS

Magnetic tunneling junction-based sensors are composed of a sensor stack that contains several materials, as explained in detail in the following paragraph. A potentially issue in terms of biocompatibility is the presence of neurotoxic materials in the sensor stack, such as cobalt or manganese. Even if these materials are present in low quantity and are isolated by capping layers, we first studied their potential noxious effects on neuronal growth and survival by using ad hoc samples (CoFe films) that mimic the worst case with respect to Magnetic tunneling junction-based sensors.

#### Fabrication of MTJs and CoFe Films

Magnetic tunneling junction-based sensors were grown on Si/SiO<sup>2</sup> substrates by magnetron sputtering with a base pressure of 2·10−<sup>9</sup> Torr and an applied magnetic field of 300 Oe (Albisetti et al., 2013, 2014). The multilayer is composed by (thickness in nm): Ta(5)/ Ru(18)/ Ta(3)/ Ir20Mn80(20)/ Co60Fe40(1.8)/ Ru(0.9)/Co40Fe40B20(2.7)/ MgO(2.5)/ Co40Fe40B20(1.3)/ Ru(5)/Ta(20). Co60Fe<sup>40</sup> and MgO layers were deposited in RF mode while the remaining layers were grown in DC mode. After deposition, the samples were processed with optical lithography and ion beam etching to obtain sensor chips featuring 12 MTJ sensors arrays (**Figure 1**). Each MTJ sensor has a 3 × 40 µm<sup>2</sup> junction area (**Figure 1B**). A 110 nm thick SiO<sup>2</sup> layer was deposited for insulating purposes by magnetron sputtering. Afterwards, a Ti(7)/ Au(100) bilayer was deposited by magnetron sputtering for the contacts. Each sensor is provided with two contacts, the top one to address independently every device, and the bottom one connected to the common ground. An additional step of lithography was performed to planarize the devices, since the two ion milling processes used for the tunnel junction definition and the subsequent contacts deposition leave the sidewalls of the top layers of the magnetic junctions exposed. If the step height between the device top contact and the SiO<sup>2</sup> surrounding is higher than, or comparable with, the thickness of the capping layers, the sidewalls of the junctions cannot be completely sealed. This could imply problems with the biocompatibility and the endurance of the sensors. The planarization of the device was obtained through liftoff, by depositing 180 nm SiO<sup>2</sup> by magnetron sputtering around the devices, giving rise to a step of about 30 nm between the top layer and the substrate. The sensors arrays were then annealed at 310◦C at a pressure of 10−<sup>6</sup> Torr for 1 h; this step was performed to enable the crystallization of the ferromagnetic layers and the tunnel barrier, with the aim of increasing the performance of the sensor (Sharma et al., 2016). In this case, sensitivities in the nT range can be easily achieved (Almeida et al., 2008; Sharma et al., 2017b), while pT sensitivities can be obtained with different strategies. Structures with top-pinned sensing layer and bottom-pinned reference layer present a linear response arising from the cross anisotropies of the two electrodes and up to 60 pT/Hz∧0.5 of sensitivity (Ferreira et al., 2012; Sharma et al., 2016). In addition, a combination of magnetorestrictive and magnetoresistive effects can theoretically enhance the sensitivity of magnetic tunneling junctions up to fT (Pertsev, 2016). Finally, another approach is the use of flux guides, which can concentrate and enhance the magnetic signal in correspondence of the sensor (Chaves et al., 2008).

As last step of the fabrication of our sensors, a capping layer composed by a three-layer of (thickness in nm): SiO2(50)/Si3N4(25)/SiO2(50) was deposited. In this case, the SiO<sup>2</sup> layer was deposited by e-beam evaporation (E-beam Evaporator Evatec BAK 640), while Si3N<sup>4</sup> was grown by magnetron sputtering. The first SiO<sup>2</sup> layer is used as adhesive layer, while Si3N<sup>4</sup> film protects the sensor from the liquid environment, due to its low porosity and high chemical resistance (Vanhove et al., 2013). The final layer of SiO<sup>2</sup> is used to promote the adhesion of the extracellular matrix proteins used for the cell growth (Monticelli et al., 2016).

Cobalt Iron (Co60Fe40) films, 20 nm thick, were grown on Si/SiO<sup>2</sup> substrates by magnetron sputtering in an AJA Orion8 system with a base pressure of 2·10−<sup>9</sup> Torr. Since Cobalt Iron is the most toxic material among those used for MTJs, these samples represent the hardest condition to achieve a neuronal growth. After the deposition, the same capping layer used for MTJs, made of SiO<sup>2</sup> (50 nm)/Si3N<sup>4</sup> (25 nm)/SiO2(50 nm), was deposited as discussed above.

## Preparation of MTJ Chip and CoFe Films for Cell Culture

Samples were first sterilized by immersion in ethanol washed twice in sterile water, dried in a laminar flow hood and further sterilized by UV irradiation for 1 h. The day before dissection substrates were coated with 0.05–0.1% polyethyleneimine (PEI, MW 25.000, Sigma-Aldrich) solution for 45 min. Then, the PEI solution was removed and substrates were washed 4 times with sterile distilled water and stored at 4◦C. To perform random neuronal growth on the samples, the day of dissection, substrates were coated with 100 µl laminin (Sigma-Aldrich) [20µg/ml] and incubated at 37◦C for 2 h. Laminin was removed by aspiration prior to seeding dissociated neurons. All biosecurity and safety procedures were followed, as specifically required by the Health & Safety Office of the Istituto Italiano di Tecnologia (Genova, Italy).

## Engineered Cultures on MTJ

Since the weak biomagnetic signals can be easily polluted by environmental noise, it is essential to reduce the distance between the sensor and biological source. To overcome this issue, we propose a platform to detect neuronal activity in cell cultures, where neurons are grown on top of the magnetic sensor and the distance between the sensor and biological source is dramatically reduced. Moreover, since the magnetic sensors have one defined sensing direction (**Figure 1B**), a critical aspect is to maximize the sensitivity of the system by properly aligning the network to the devices. Considering that magnetic fields produced by the neuronal currents are forcing perpendicular to the current direction, in the most favorable configuration the axons have to be perpendicular to the sensing direction of the probe (Hall et al., 2012). For this reason, beside a condition of random neuronal growth, we also carried out controlled-topology neuronal networks, by neuronal processes to grow along the sensor's major axis.

To this aim, we added the following additional process to sample preparation. After the sterilization and PEI coating, MTJ chips were prepared for patterned coating deposition. Agarose (0.15% w/w in water; Sigma-Aldrich) was prepared by dissolving the agarose powder in MilliQ-water brought to its boiling point in a microwave oven for 2 min. Then, agarose (0.1 µl) was deposited on the sensor area and immediately aspirated to have a thin layer to inhibit random cellular adhesion. To obtain a neuronal network along the sensors, lines of PEI (80 × 20µm) were deposited on the top of the agarose-coated 12-MTJ sensors by a nano-drop ink jet print (NanoEnabler, BioForce Nanoscience Inc.), equipped with SPT-C30 S cantilevers (NanoandMore), (**Video S1**). Relative humidity in the chamber was kept at 70 ± 5%. A small volume (0.3 µl) of the PEI solution was loaded into the cantilever reservoir by a micropipette. Glycerol (5% w/w in water; Sigma-Aldrich) was added to the solution to avoid evaporation. Contact force and withdrawal distance were fine controlled (0.002 nN and 40µm, respectively), in order to obtain the desired size of the PEI lines. The PEI lines were deposited exactly on top of the MTJ sensors, after micrometric alignment of the chip in correspondence with the cantilever. MTJ chips were stored at 4◦C. To improve the neuronal adhesion in the area surrounding sensor active region, laminin was applied. No laminin was deposited on the top of the 12-MTJ sensor where PEI lines were made. On the day of dissection 80 µl laminin [20µg/ml] were deposited outside the sensor active region by 10 µl drops and incubated at 37◦C for 2 h. Laminin was removed by aspiration prior to seeding dissociated neurons.

### Primary Neuronal Cultures

All procedures were carried out in accordance with the guidelines established by the European Communities Council (Directive of November 24th, 1986) and approved by the National Council on Health and Animal Care (authorization ID 227, prot. 4127, 25th March 2008). Primary hippocampal cultures were obtained from Sprague-Dawley rats at embryonic day 18 (E18) (Charles River). Pregnant females were deeply anesthetized with CO<sup>2</sup> and decapitated. Embryos were removed and brains were placed in cold Hank's balanced salt solution (HBSS). After removal of the meninges, the hippocampus was carefully dissected, incubated with 0.125% trypsin for 15 min at 37◦C and mechanically dissociated. Eighty thousands neurons were then plated on each coated devices. Cell culture were maintained in 2 ml of cell culture medium composed by: Neurobasal medium, 2% B-27, 1% penicillin– streptomycin, and 1% Glutamax and maintained at 37◦C in 5% CO2.

## Viability Assay

Cells were incubated for 3 min at RT in extracellular medium (EM; NaCl 135 mM, KCl 5.4 mM, MgCl<sup>2</sup> 1 mM, CaCl<sup>2</sup> 1.8 mM, glucose 10 mM, Hepes 5 mM, pH 7.4), 5 mg/ml propidium iodide (PI), containing 15µg/ml fluorescein diacetate (FDA) and 3.3µg/ml Hoechst-33342. After incubation, cells were washed once in EM and immediately imaged. PI is a redfluorescent nuclear and chromosome counterstain able to permeate exclusively through the membrane of dead cells, FDA a non-fluorescent molecule that is hydrolyzed to fluorescent fluorescein only within live cells and Hoechst a nuclear counterstain binding the DNA of both live and dead cells.

The hardware configuration for the imaging experiments was based on a Nikon Eclipse Ni-U upright microscope equipped with an epifluorescence attachment and a Camera DS-Qi2 (Nikon Instruments). Cells were magnified, with a 20x objective (0.75 NA). For each sample, at least 5 distinct fields of view were acquired. Considering the total number of nuclei identified by Hoechst fluorescence and the apoptotic nuclei, identified by PI fluorescence, the percentages of living cells were calculated for each field as: (Hoechst-positive nuclei – PI-positive nuclei) / (Hoechst-positive nuclei). FDA staining was used as a further marker of cell-membrane integrity and culture viability. Images were analyzed by using the Image J software. Statistical analysis was performed using a commercial package [Sigmastat, Systat Software Inc.].

#### Immunocytochemical Analysis

Cells were washed twice in phosphate buffered saline (PBS) solution, fixed with 4% (w/v) paraformaldehyde in PBS at RT for 20 min and washed two times in PBS. Fixed samples were permeabilized with 0.1% (v/v) Triton X-100 in PBS for 5 min. Blocking solution (PBS, 1% BSA, 5% FBS) was added for 30 min at RT to block nonspecific reactions. The following primary antibodies, diluted in blocking solution, were used: monoclonal anti-MAP2 antibody (Synaptic Systems; mouse #188011, dilution 1:500), polyclonal anti-Tau (rabbit # 314002, dilution 1:1,000) and polyclonal anti-NeuN antibody (guinea pig, # 266004, dilution 1:500). MAP2 is the major microtubule associated protein of brain tissue. Tau is a microtubule-associated protein of the neuronal axons. NeuN is a neuron-specific DNA-binding protein present in most neuronal cell types. After incubation for 2 h at RT, samples were washed 3 times with PBS and incubated with fluorophore-conjugated secondary antibodies for 1 h at RT. Secondary antibodies were: anti-mouse Alexa Fluor 488 (#A11029; Thermo Fisher Scientific, dilution 1:1,000), antirabbit Alexa Fluor 647 (#A21245; dilution 1:1000) and antiguinea pig Alexa Fluor 546 (#A11075; dilution 1:1,000). Samples were mounted using Mowiol 4-88 (81381, Sigma-Aldrich) and stored at 4◦C. In the case of engineered cultures, samples were mounted using Prolong anti-fade reagent containing DAPI (a blue-fluorescent DNA stain, Invtrogen), without using anti-NeuN antibody. Confocal microscopy was performed using an SP8 microscope (Leica Microsystems GmbH) using 40x (1.3 NA) and 63x (1.4 NA) objectives. Confocal images were analyzed with the Leica LAS AF software (Leica Application Suite Advance Fluorescence, version 3.3, Leica Microsystems). Fluorescence microscopy was performed as described above.

## Electrophysiology

Whole-cell patch-clamp recordings of primary cortical rat neurons [14 and 21 days in vitro (DIV)] were performed using borosilicate glass patch pipettes (Kimble) pulled to a final resistance of 3–5 M and under G patch seal. Data were sampled at 20 kHz and low-pass filtered at 4 kHz with an EPC-10 Plus amplifier (HEKA Electronic). Recordings with leak currents >100 pA or series resistance >20M were discarded. Data acquisition was performed using PatchMaster v2.73 software (HEKA Elektronic). Series resistance (Rs) was compensated 80% (2 µs response time) and the compensation was readjusted before

stimulation. Potentials were not corrected for the measured liquid junction potential of 9 mV. All recordings were performed at 22–24◦C. The extracellular "Tyrode" solution contained (in mM): 135 NaCl, 5.4 KCl, 1 MgCl2, 1.8 CaCl2, 5 HEPES, 10 glucose adjusted to pH 7.4 with NaOH, and the intracellular pipette solution contained (in mM): 126 K-Gluconate, 4 NaCl, 1 MgSO4, 0.02 CaCl2, 0.1 EGTA, 10 Glucose, 5 Hepes, 3 ATP-Na2, and 0.1 GTP-Na. Cell membrane capacitance was obtained from the slow time constant component used for capacitance compensation after reaching the whole-cell configuration, and Input resistance was calculated as the Current vs. Voltage (IV) curve slope measured at sub-threshold voltages. Measurements of the firing activity were performed in currentclamp configuration. Resting membrane potential (Vrest) was determined immediately after breakthrough in the whole-cell mode. Spontaneous firing activity was considered for analysis only from those cells with a Vrest between −70 and −50 mV, and the mean firing frequency was calculated as the average reciprocal of the interspike interval. Evoked firing activity was induced by injection of 5 pA current steps lasting 100 ms in neurons maintained at a holding potential (VH) of −70 mV through the injection of a negative current ( IH−70mV). For each patched-neuron we calculated: the minimal current able to evoke the first Action Potential (current threshold), the maximum voltage reached at the Action Potential peak (AP peak), and the maximal rising slope (dV/dt) of the upstroke phase. Whole-cell currents were elicited in voltage-clamp configuration. A protocol consisting of a 200 ms voltage step from the holding potential of −70 to −100 mV, followed by 100-ms linear ramp up to 120 mV was used to evoke voltage-gated conductances. Sodium current peak was calculated as the minimum inward current value measured at the beginning of the ramp phase. Voltagegated macroscopic currents were also evoked by stepping V<sup>H</sup> from −60 to 20 mV for 30 ms with 10 mV increments with 2 s interpulse in order to quantify inward sodium currents (negative peak in the first 10 ms) and steady-state outward potassium currents in the last 5 ms. In all the protocols used, cells were clamped at a Vh of −70 mV before stimulation.

#### RESULTS

#### MTJ Endurance

We first checked whether the magnetoresistive sensors were compatible with prolonged immersion in cell medium (basically an electrolytic solution of NaCl) at 37◦C without deteriorating and loosing their sensor performances. Tunneling magnetoresistance measurements were performed using a Keithley 2611 source meter and an electromagnet driven by a Kepco power supplier. All the devices used here showed a resistance between 1 and 4 k, which depends exponentially on the tunneling barrier thickness (Hayakawa et al., 2005).

The tunneling magnetoresistance (TMR), defined as the normalized difference between the resistance for a given field and that in the parallel state of the magnetization in the two electrodes (i.e., for high negative fields in our case), is plotted in **Figure 2**. In the sensors used in this work, aiming at

14 DIV. Confocal images of the indicated immunoreactivities: MAP-2 in mature neuronal dendritic arborizations, NeuN in neuronal nuclei and Tau in axonal processes. (A) Neurons grown on the top of a MTJ sensor (white arrows indicate the border of the sensor area). (B) Neurons grown outside the sensor active region.

demonstrating the biocompatibility, maximum TMR variations between 20 and 50% were obtained, mainly depending on the CoFeB thickness used in the stack (Wi´sniowski et al., 2008). A low-field sensitivity between 2 and 10 % mT was achieved, together with a linear response and some residual hysteresis (see **Figure 2**, left, inset).

As reported in **Figure 2**, after 31 days in culture the devices presented TMR values comparable to the nominal ones, i.e., to those measured before the experiments. Since the tunneling is highly sensitive to any change in the device structure, we can therefore conclude that the capping layer is impermeable enough to prevent any interaction of the device with neurons and culture medium.

## Cell Viability on CoFe Film

A crucial aspect for ensuring an efficient growth of a neuronal culture onto the sensor surface is to find a proper capping that enables cell viability in spite of the toxic materials of the sensor stack (i.e., cobalt) and, at the same time, provides protection to the magnetic device. To this aim, we studied cell viability on a substrate mimicking the situation of MTJs, but representing a worst case for cell viability in terms of thickness of toxic materials: a CoFe film 20 nm thick deposited on Si/SiO<sup>2</sup> wafers and capped with a SiO2(50)/Si3N4(25)/SiO2(50) three-layer (thickness in nm). Cell growth displayed no qualitative differences between rat hippocampal neurons plated on the CoFe films or glass coverslips (**Figure 3A**). Quantitative analysis of cell viability was performed using a triple staining with PI, FDA and Hoechst as described in the Methods section

The microphotographs of panel (a) show that all the PIpositive cells were negative to FDA (dead cells), while all the PInegative cell were also FDA-positive (live cells). The estimation of the percentage of live cells over the total cell nuclei showed that comparable cell viabilities, higher than 70%, were achieved for both neuronal populations grown on top of CoFe films or on top of control glass coverslips (**Figure 3B**).

We also achieved neuronal growth (data not shown) with CoFe films where the three-layer were grown by plasma enhanced chemical vapor deposition, instead of magnetron sputtering as described above. However, we dropped this technique because this three-layer did not sufficiently protect the CoFe layer due to infiltration issue. These findings validated the three-layer [SiO<sup>2</sup> (50)/Si3N<sup>4</sup> (25)/SiO2(50)] grown by magnetron sputtering that we chose for the CoFe films as suitable capping layers to isolate cells from the toxic materials of the CoFe films. For these reasons, we employed these capping layers to protect the MTJ sensor stack and to guarantee long-term biocompatibility.

### Proper Neuronal Growth on MTJ Chips

We monitored the developing cultures up to 19 DIV by studying the expression of biomarkers of neuronal maturation including: (i) MAP-2 that is expressed only in neuronal cells and labels soma and dendrites, (ii) Tau that stains the axonal processes and (iii) NeuN that is a specific marker for neuronal nuclei. **Figures 4A**, **5** show a proper morphological differentiation of a rich neuronal network that homogenously developed on the top of the MTJ sensor. A similar picture was observed also in the periphery of the device (**Figure 4B**), where the immunostaining shows neurons with a correct growth of MAP-2 positive dendrites and tau-positive axons. Even in culture with random topology, the neuronal cell bodies adhered and developed correctly on the top of the MTJ junction. The device surface underneath the neuronal network is shown in the Tau microphotograph by the autofluorescence signal generated in this wavelength by the MTJ track materials (in particular metals). Moreover, in **Figure 5**, a couple of neurons (indicated with white arrows) spontaneously developed axons parallel to the MTJ. These observations lead us to conclude that no constraints were found on the whole chip surface for a correct neuronal culture growth. This result is important in order to show a proper adhesion and growth of delicate primary neurons on the top of the device and to rule out the potential noxious effect of the neurotoxic materials of the sensor stack (i.e., cobalt or manganese) and possibly of some residual traces of chemicals arising from the photoresist and solvents used in the lithographic process.

Similar results in terms of proper growth and development of the neuronal network were achieved in engineered cultures (**Figure 6**), where neurons were forced to grow with a controlled topology along the MTJ sensors to achieve the best system sensitivity. A reliable confinement of neuronal processes along the MTJ sensors was obtained over time (19 DIV). The PEIpatterned coating deposition enhanced cell adhesion along the MTJ sensors, while the background agarose layer successfully inhibited cell adhesion (Petrelli et al., 2013) outside the MTJ sensor tracks. Several cell nuclei (**Figure 6C**, blue) were located exactly on the top of the sensor and generated rich MAP-2 positive neurite bundles (**Figures 6B,C**, green) that were perfectly aligned along the sensor major axis, i.e., perpendicular to the MTJ sensing direction. This neuronal network architecture could achieve the best system sensitivity conditions: the distance between the source of the biological signal and the detector is

minimized and the biological magnetic field is completely parallel to the sensing direction.

## Proper Spontaneous Electrical Activity on MTJ Chips

To test the functionality of neurons grown on the top of MTJ sensors, we analyzed their electrophysiological properties by means of patch-clamp recordings. We plated primary cortical rat neurons on top of MTJ devices and recorded them from 14 to 21 days after plating. At this age, primary in vitro cortical neurons develop a sustained and persistent spontaneous AP firing activity that we measured in current-clamp configuration (**Figure 7A**).

The passive membrane properties, namely: the neuronal capacitance, which furnishes a rough estimation of the cell size (61.7 ± 13.3 pF), the Vrest that was −53.3 ± 1.3 mV and the input resistance (Rin) that was 259.8 ± 43.9 M (**Figure 7B**) were in full agreement with the corresponding values previously observed under standard culture conditions (Yang et al., 1996; Bean, 2007; Staiger et al., 2016). The spontaneous firing activity that neurons plated on the MTJ devices were able to generate at the resting potential showed a frequency of 7.07 ± 1.99 Hz (**Figure 7B**).

We also studied evoked AP firing (**Figure 7C**) in neurons maintained at a Vh of −70 mV through the injection of a negative current of −31.4 ± 5.7 pA. The mean current threshold necessary to elicit the first AP from a Vh of −70 mV (rheobase) was 134.6 ± 18.0 pA. APs evoked during depolarizing steps reached a maximum membrane voltage (AP peak) of 33.8 ± 3.8 mV with

(AP) firing recorded in current clamp configuration from a cortical neuron. Inset: an enlarged detail of the APs firing activity. (B) From left to right: Box plots of cell capacitance (N = 8), input resistance (N = 10), resting membrane potential (N = 7) and AP firing frequency (N = 5). (C) Evoked AP firing activity elicited by +5 pA increasing current steps from a Holding Potential (VH) of −70 mV. (D) From left to right: Current injected for maintaining the HP at −70 mV (N = 7), minimum current injected for eliciting AP activity from a HP of −70 mV (N = 7), AP peak (N = 7) and maximal AP rising slope (N = 6). (E) Whole-cell currents recorded in voltage-clamp configuration with a ramp protocol depolarizing the cell from −100 to 120 mV in 100 ms (left) and sodium current peak (right) (N = 8). (F) Inward sodium and outward potassium currents evoked by the depolarizing voltage steps from −70 to 20 mV in 10 mV increments. (G) Current vs. Voltage (IV) relationships relative to the mean peak sodium current (left) (N = 6) and the mean steady state potassium current (right) (N = 9). Data are from 3 independent preparations with neurons plated on 3 MTJ sensor chips.

an average maximal rising slope of the upstroke phase of 107.01 ± 16.91 mV/ms (**Figure 7D**).

Lastly, we analyzed voltage-gated sodium and potassium currents underlying the generation of the firing activity under voltage-clamp conditions. A linear ramp stimulation protocol, from −100 to +120 mV, elicited a current response characterized by (i) a fast and transient inward (negative) peak generated by voltage-gated sodium channels, followed by (ii) a slower outward (positive) current activated at more depolarized potentials, generated by voltage-gated potassium channels (**Figure 7E**). The current peak amplitude of the initial negative peak (sodium peak) was −2.4 ± 0.4 nA (**Figure 7E**). Alternatively, voltage-gated sodium and potassium currents were elicited by depolarizing the neuron with increasing voltage steps, from −60 to +20 mV, (**Figure 7F**). Even in this case, faster and inactivating inward sodium currents appeared at the beginning of each voltage step, while slower outward potassium currents were recorded at steady state. Current vs. voltage (I/V) relationships were plotted with the mean sodium and potassium currents as a function of the clamped voltage steps (**Figure 7G**). The resulting voltage-dependence was consistent with data obtained from similar cortical neuronal preparations (Yang et al., 1996; Bean, 2007; Staiger et al., 2016). Indeed, neurons grown in contact with MTJ sensors displayed normal electrophysiological activity, and their spontaneous firing activity generated by the expression of sodium and potassium channels, testifies a correct physiological maturation of the intrinsic excitability properties. These results obtained with the sensitive electrophysiological approach confirm a correct development and maturation of neurons grown in contact with the MTJ sensors.

#### DISCUSSION

In this work, we show the biocompatibility of MTJs for in vitro neuron culture studies. We first demonstrated the preservation of the magnetic proprieties of the MTJ sensors after 31 days of permanence in culture medium. Moreover, we achieved fully viable on-chip neuronal networks for periods longer than 20 DIV and monitored the proper neuronal network growth and maturation under these conditions with immunofluorescence and patch-clamp studies. No constraints were found for primary neuronal culture growth on the MTJ chip surface. We also successfully grew neuronal networks with controlled topology through a micropatterning technique that likely promotes the best conditions to detect neuronal magnetic signals. In conclusion, this work establishes the biocompatibility of a MTJ chip for neuronal studies in vitro.

These results validate the possibility of using a magnetoresistive platform to detect biosignals originating from the spontaneous and evoked electrical activity of primary neurons in culture. Indeed, three main ingredients contribute to the sensitivity of a magnetic platform: the intensity of the magnetic signal in correspondence with the sensor surface, the sensitivity of the MTJ device and the sensitivity of the electronic acquisition platform. Regarding the first point, engineered cultures allow neurons to grow on top of the sensors with neurites aligned in a controlled topology. In this way, the magnetic field generated by the action potentials propagating in the axons can be aligned to the sensitive direction of the sensors. In addition, the thin capping layers optimized in this work allow reducing the distance between the neurons and the sensor surface. In this configuration, highly sensitive magnetic tunneling junctions with sensitivities in the range of low-nT/ pT can be employed to record the weak and rapidly decaying magnetic signals arising from the neuronal activity. Finally, to maintain the low noise levels required by the small signals expected, ad hoc acquisition platforms can be developed. Some of the authors designed and built a platform based on a generation channel to drive all the sensors with a sinusoidal voltage and four low-noise parallel acquisition channels (Sharma et al., 2017b). The front-end acquisition channels in combination

#### REFERENCES


with a Field Programmable Gate Array can process four channels simultaneously with input voltage noise of only 3 nV/<sup>√</sup> Hz.

Our results pave the way for a new generation of biomagnetic chips to study the neuronal magnetophysiology in vitro with high spatio-temporal resolution.

#### AUTHOR CONTRIBUTIONS

DP and FB conceived the study. DM, DP, and FB designed the experiments. PS, EA, MM, and DP fabricated and characterized the sensors. DM carried out neuronal culture experiments and analysis. SD supported the microprinting coating deposition experiments. MD contributed with electrophysiology experiments and analysis. DP, RB, PB, and FB gave conceptual advice and reviewed the manuscript. DM and DP prepared the manuscript. All co-authors agreed to the submission of the final manuscript.

#### ACKNOWLEDGMENTS

This research was supported by Fondazione Cariplo via the project UMANA (Project number 2013.0735). This work was partially performed at Polifab, the micro- and nanotechnology center of the Politecnico di Milano. Biopatterning by nanodropping was performed at Nanophysics department of the Istituto Italiano di Tecnologia.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnins. 2018.00909/full#supplementary-material

Video S1 | The video shows a patterned coating deposition on a top of a MTJ sensor by a nano-drop ink jet print (NanoEnabler, BioForce Nanoscience Inc.), equipped with a cantilever (NanoandMore). The cantilever reservoir contains 0.3µl of the coating solution (PEI). After micrometric alignment of the chip in correspondence of the cantilever, the coating deposition starts in order to obtain PEI line exactly on top of the MJT sensor. This process allow to perform a controlled topology neuronal networks. Scale bar 20µm. Video was edited using VSDC Free Video Editor, version 5.8.9.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Moretti, DiFrancesco, Sharma, Dante, Albisetti, Monticelli, Bertacco, Petti, Baldelli and Benfenati. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Utility of Zebrafish as a Model for Screening Developmental Neurotoxicity

#### Marta d'Amora<sup>1</sup> and Silvia Giordani1,2 \*

<sup>1</sup> Nano Carbon Materials, Center for Sustainable Future Technologies, Istituto Italiano di Tecnologia, Turin, Italy, <sup>2</sup> School of Chemical Sciences, Dublin City University, Dublin, Ireland

The developing central nervous system and the blood brain barrier are especially vulnerable and sensitive to different chemicals, including environmental contaminants and drugs. Developmental exposure to these compounds has been involved in several neurological disorders, such as autism spectrum disorders as well as Alzheimer's and Parkinson's diseases. Zebrafish (Danio Rerio) have emerged as powerful toxicological model systems that can speed up chemical hazard assessment and can be used to extrapolate neurotoxic effects that chemicals have on humans. Zebrafish embryos and larvae are convenient for high-throughput screening of chemicals, due to their small size, low-cost, easy husbandry, and transparency. Additionally, zebrafish are homologous to other higher order vertebrates in terms of molecular signaling processes, genetic compositions, and tissue/organ structures as well as neurodevelopment. This mini review underlines the potential of the zebrafish as complementary models for developmental neurotoxicity screening of chemicals and describes the different endpoints utilized for such screening with some studies illustrating their use.

#### Edited by:

Marius Enachescu, Politehnica University of Bucharest, Romania

#### Reviewed by:

Resham Chhabra, Johns Hopkins University, United States Suresh Jesuthasan, Nanyang Technological University, Singapore

> \*Correspondence: Silvia Giordani silvia.giordani@dcu.ie

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 30 August 2018 Accepted: 06 December 2018 Published: 18 December 2018

#### Citation:

d'Amora M and Giordani S (2018) The Utility of Zebrafish as a Model for Screening Developmental Neurotoxicity. Front. Neurosci. 12:976. doi: 10.3389/fnins.2018.00976 Keywords: zebrafish, model, neurotoxicity, development, chemicals

## INTRODUCTION

Exposure to different chemicals during development induces a significant risk to human health and can cause the onset of different neurological and neuropsychiatric impairments, ranging from attention deficit hyperactivity disorder (Braun et al., 2006; Bellinger, 2013) to autism spectrum disorders (Harrington et al., 2014; Lyall et al., 2017), and Parkinson's disease (Barlow et al., 2007). Moreover, various studies indicate that the brain of early-life organisms is more sensitive to chemicals during a critical period in development, including prenatal and postnatal stages (Giussani, 2011; Perera and Herbstman, 2011). The little progress in acknowledging different chemicals to induce neurotoxic effects, is in part due to the disadvantages and restrictions of the different in vitro and in vivo systems employed to identify adverse effects of chemicals. Different studies have reported the effects induced by different chemicals on the neuronal activity of different cell lines, including mouse and rat primary neuronal cells (Chen et al., 2017; Sethi et al., 2017) and neural precursor cells derived from human induced pluripotent stem cells (Druwe et al., 2015; Ryan et al., 2016). As shown in **Figure 1**, in vitro toxicity studies are cheap, quick and easy, but cultured cells poorly correlate with in vivo mechanisms and therefore the observations have limited translational value. Preliminary in vitro tests confirm zebrafish as promising candidates for intermediate models. The different advantages of zebrafish are illustrated in **Figure 1**. Zebrafish

are significantly more complex than cultured cells and other model systems, such as Drosophila melanogaster (Giacomotto and Segalat, 2010). Moreover, toxicity experiments performed in zebrafish are less expensive and time-consuming than those conducted in rodents (Crofton et al., 2012). Here, we present a brief overview of the different advantages of using zebrafish to assess developmental neurotoxicity and the different endpoints utilized for this screening including some examples illustrating the utiliy of zebrafish studies.

framework, zebrafish represent excellent comparative vertebrate systems.

### ZEBRAFISH AS NEUROTOXICITY MODELS

#### Advantages and Limitations

Zebrafish present several advantages for assessing developmental neurotoxicity, making them excellent in vivo models in this field (Garcia et al., 2016; Kalueff et al., 2016; Wiley et al., 2017). Zebrafish are small-sized animals and, therefore, can be handled easily. They undergo external fertilization with a high fecundity rate, generating large numbers of embryos. Because of their small size, neurotoxicity tests are generally performed by placing embryos in 96 multi-well plates which reduces the amount of waste and chemicals used, as well as cost. Zebrafish are simply soaked in chemical solutions and the compounds penetrate the transparent embryo's external membrane by passive diffusion (d'Amora et al., 2016, 2017, 2018). Hence, zebrafish embryos are ideal for high-throughput screening (Horzmann and Freeman, 2018). Another advantage of using zebrafish is that brain development occurs within 3 days post-fertilization, together with the central nervous system. Zebrafish possess a high degree of genetic, morphological and physiological homology with humans (Howe et al., 2013; Kalueff et al., 2014). In particular, development processes and mechanisms of the central nervous system of zebrafish and other vertebrates are well-conserved (Belousov, 2011). The similarity between these species also includes the development of the blood brain barrier (BBB). This is very important, as the BBB plays a crucial role in protecting the brain against chemical substances (Eliceiri et al., 2011). For instance, the counterparts of many brain subdivisions found in the developing mammalian brain are morphologically identifiable in the developing zebrafish (Wullimann, 2009). Thanks to all these features, particularly to the fast brain development, zebrafish are increasingly utilized as complementary models for in vivo neurotoxicity screening (Fan et al., 2010; Cowden et al., 2012).

However, there are several peculiarities that may limit their use. The most obvious drawback of zebrafish, specifically in comparison with humans, is that they are not mammals. It is not possible to fully control the chemical dose absorbed since zebrafish embryos are not developing inside a placenta and are exposed to chemicals in medium and absorb them directly (Rubinstein, 2006). Furthermore, chemicals can be metabolized in a different manner compared to mammals. In early life stages, zebrafish are surrounded by a protective membrane which may limit the diffusion of some chemicals (Cudd, 2005). In addition, non-water soluble chemicals cannot be easily dispersed in the embryo medium and thus a small amount of solvent has to be used (Maes et al., 2012).

### Neurotoxicity Endpoints

Considering the positive features of zebrafish described above, several approaches have been developed to utilize zebrafish in neurotoxicity screening during the last decade. Effects of different chemicals on brain development can be evaluated by different neurotoxicity endpoints including gene expression patterns, neural morphogenesis and neurobehavioral profiling (Kalueff et al., 2013; Truong et al., 2014; Chueh et al., 2017).

#### Gene Expression Patterns

fnins-12-00976 December 15, 2018 Time: 15:9 # 3

A quick and sensitive method to detect changes in gene expression patterns in zebrafish treated with chemicals, is to quantify markers related to developmental toxicity. Fan et al. (2010) used various nervous system genes as potential markers of neurotoxicity, characterzing their expression profiles in embryos exposed to ethanol by means of quantitative real-time polymerase chain reactions. These markers include the transcripts of genes expressed in neuronal stem cells and/or in developing neurons. Their results showed a decrease or increase of these transcripts during development, and in particular highlighted a significant overexpression of a specific astrocytes marker. This study clearly demonstrates that analyzing the brain gene expression profile is a useful tool to rapidly test the neurotoxicity of chemicals during development (Fan et al., 2010). Studies performed in mammals, including mice (Alfonso-Loeches et al., 2013), rats (Gonzalez et al., 2007), and humans (Jung et al., 2010), reported a similar increase in astrocyte marker expression after treatment with ethanol. This approach to assess the chemical profiling expression of a high number of genes, provide knowledge on how different chemicals affect the developing nervous system.

Yang et al. (2007) treated zebrafish embryos with various concentrations of environmental toxins and analyzed the changes in the profiling expression of hundreds of genes by microarray hybridization. The obtained expression profiles were highly specific for each tested compound, allowing to identify several chemicals from the expression profiles with high probability. This study demonstrated that organ and cell-specific changes in gene expression could be detected by in situ hybridization (Yang et al., 2007).

Following the work of Yang, the group of Ho et al. (2013) focused on the effects of methyl mercury in the nervous system. A genome profiling analysis of treated zebrafish was carried out in conjunction with whole-mount in situ studies of affected genes. An altered expression of various genes involved in different biological functions was found in different neuronal subregions of the brain (Ho et al., 2013).

The gene expression of myelin basic proteins (MBP) was evaluated in zebrafish after treatment with different concentrations of propofol, an anesthetic. The results indicated propofol to be toxic, causing a high decrease in MBP expression levels in the larval central nervous system. In addition, the effects of ibuprofen, diclofenac and paracetamol were assessed by adifferent neuron related expression genes. Ibuprofen and diclofenac exposure down-regulated the neurog1 expression, while ibuprofen up-regulated it (Xia et al., 2017). Li et al. (2018) assessed the expression of neurodevelopmental genes (mbp, syn2a, and α1-tubulin) in larvae treated with Tris (1,3-dichloro-2 propyl) phosphate and chlorpyrifos (CPF), finding the expression to be down-regulated. The neurotoxicity of triphenyl phosphate was investigated by analyzing the expression of genes which are related to neurodevelopment. Exposure caused down-regulation of 1-tubulin, mbp, syn2a, shha, and elavl3, demonstrating the neurotoxic effects of this organophosphate ester (Shi et al., 2018). Embryos and larvae treated with perfluorododecanoic acid resulted in several down-regulated genes, including gap43, α1-tubulin, gfap, mbp, and elavl3.

The use of gene profiling patterns represents a useful neurotoxicity endpoint to assess the potential developmental neurotoxicity of different compounds. However, it is important to assure any modifications in gene patterns are caused by the treatment itself and are not a possible stress response (Spurgeon et al., 2010).

#### Neural Morphogenesis

Different research groups have employed zebrafish to address the effect of chemicals on the central nervous system during the development by morphometric endpoints (Scalzo and Levin, 2004; Parng et al., 2006, 2007). Parng et al. (2006) investigated the biological consequences of different compounds, demonstrating their significant neuroprotective effects in zebrafish. They proposed a new in vivo approach based on evaluation of oxidation-induced apoptosis. The same group tested the neurotoxicity of different drugs, investigating neuronal apoptosis and other parameters by in situ hybridization and immunostaining techniques (Parng et al., 2007). In both studies, obtained data were correlated with previous ones performed in mammals, validating this comparative in vivo system for screening. Evaluation of neuronal apoptosis by acridine orange staining was used as an endpoint to evaluate the neurotoxicity of seven compounds. Three of these chemicals caused specific neurotoxicity in catecholaminergic neurons (Ton et al., 2006). In this case, results were comparable with the mammalian studies. In other works, zebrafish eggs were treated with cypermethrin and its toxic effects on the developing nervous system were evaluated (Shi et al., 2011). Notable signs of apoptosis were observed in the nervous system. Acridine orange staining was also employed to explore the neurotoxicity of fenvalerate (Gu et al., 2010), which caused apoptosis in the brain of embryos and larvae and an alteration in neurodevelopmental genes, leading to brain impairment.

In this framework, another approach analyzes axonal morphology and growth during neuronal development (Yang et al., 2011). Zebrafish embryos were treated with the organophosphorus pesticide chlorpyrifos (CPF) or its oxon metabolite (CPFO) and the in vivo profiling of axonal growth was evaluated. The results showed an inhibition of the axonal growth in primary motoneurons (PMNs) and secondary motoneurons (SMNs), with consequent anomalies in swimming ability. Muth-Köhne et al. (2012) proposed to determine alterations in zebrafish treated with chemicals as a valid method to investigate their neurotoxic effects. To this end, they developed a novel assay based on whole-mount immunostaining of motorneurons using specific antibodies for PMNs and SMNs. The neurotoxic effects induced by thiocyclam, cartap and disulfiram were analyzed. From the three neurotoxins, disulfiram resulted to be the most toxic and thiocyclam the least.

Another morphometric endpoint, commonly used in developmental toxicity assessment, is the in vivo observation of morphological defects in the developing brain. The transparency of zebrafish is one of their peculiarities, helping to observe all

brain cells beginning at early stages. Moreover, it is possible to label and visualize in vivo specific neurons and subsets of axonal tracts by dye microinjection (d'Amora et al., 2016).

Panzica-Kelly et al. (2010) proposed a morphological score system to distinguish the defects induced by different chemical exposures. They analyzed over 30 chemicals and found changes of morphology or size in one or more brain regions of treated zebrafish.

The potential neurotoxicity of triclosan (TCS) on zebrafish, was evaluated by analyzing morphological changes and expression of genes involved in neurodevelopment. Embryos treated with TCS were affected in their CNS structure, with a decrease in synaptic density and axon length. Moreover, expression of α1-Tubulin and Gap43, involved in axon extension, were up-regulated, while expression of Gfap and Mbp, involved in axon myelination, were decreased (Kim et al., 2018).

#### Neurobehavioral Profile

Neurobehavioral changes are the most common neurotoxic endpoints investigated and addressed in zebrafish exposed to chemicals. In particular, the number of movements, spontaneous or induced by stimulation (response to touch), and swimming activity are analyzed. Due to all zebrafish peculiarities, it is easily possible to track in vivo behaviors, using video recording tools.

As in mammals, treating zebrafish with ethanol led to altered swimming activity; in particular, ethanol concentrations of 0.5–1% resulted in hyperactivity, while higher doses caused sedation.

Different studies evaluating a possible neurotoxicity of organophosphorus pesticides reported neurobehavioral changes in zebrafish (Eddins et al., 2010; Yen et al., 2011). In particular, larvae exposure to chlorpyrifos, diazinon, and parathion reduced acetylcholine esterase activity and larval motility. Other pesticides also induced various neurobehavior changes (DeMicco et al., 2010; Liu et al., 2018). Zebrafish larvae treated with different pyrethroids presented neurotoxicity characterized by increased motility (DeMicco et al., 2010; Liu et al., 2018). Triphenyl phosphate, an environmental toxicant, was found to significantly reduce larval locomotor activity (Shi et al., 2018). Weichert et al. investigated the consequences of four different chemicals by quantifying spontaneous locomotion. Their results demonstrated the advantages of using behavioral parameters in detecting neurotoxic effects, in particular when exposed to a chemical with a specific mode of action (Weichert et al., 2017).

Xiao et al. (2018) evaluated the effects of 17 typical fluoroquinolones on zebrafish and reported four different types of neurobehaviors with no influence on locomotor activity, suppression of activity or intermediate responses.

Moreover, different approaches were proposed to test locomotor activity by evaluating tail contractions, touchresponse, and swimming activity in response to chemicals in the microplate format (Kokel et al., 2010; Selderslaghs et al., 2010). The effects of endosulfan I and endosulfan sulfate were characterized in zebrafish by touch response. Larvae treated with acute doses of both compounds presented a reduced response to touch and in some cases, paralysis (Stanley et al., 2009). Irons et al. (2010) developed another drug challenge paradigm for larvae in a microplate format, using alternating light and dark periods, in order to monitor the neurobehavior much quicker. In the same year, Selderslagh et al. developed new methods to evaluate locomotor activity in zebrafish. Spontaneous tail coilings and swimming of embryos treated with chlorpyrifos, a common pesticide, were evaluated using video recording tools (Selderslaghs et al., 2010). Subsequently, they evaluated this method at several developmental stages, investigating the neurotoxic effects of well-known compounds (Selderslaghs et al., 2013). A classification of these chemicals as being neurotoxic or non-neurotoxic obtained in zebrafish showed a 90% similarity with previous data found in mammals (Selderslaghs et al., 2013).

The behavioral effects of benzo[a]pyren were assessed by means of a larval photomotor response assay. This assay allowed tracking the movements over alternating light and dark periods (Knecht et al., 2017). The highest dose of benzo[a]pyrene (4 µM) caused significant hyperactivity. On the other hand, zebrafish exposed to mercury chloride presented a decrease in the number of tail coilings (Abu Bakar et al., 2017).

#### Connections Zebrafish/Humans

Zebrafish assays represent intermediate model systems, that enable high-throughput screening of different chemicals. The use of zebrafish in neurotoxicity research is increasing and different studies underline how these animals can be employed to detect risks for human health, avoiding the ethical constraints of mouse and rat experiments. In this review, we provided multiple examples, from different research groups, using zebrafish as promising models to predict the neurotoxicity of chemicals in mammals, including humans.

However, our understanding of the potential neurotoxicity of chemicals during development has not progressed much. One of the reasons for this is the lack of a common protocol used by researchers; in fact the concentrations of chemicals, the temporal window of chemical exposure, and the method of statistical analyses are different. So far, standard criteria for neurotoxicity are missing.

A systematic comparison of chemical neurotoxicity in zebrafish and mammals is necessary to validate zebrafish as alternative model for human toxicology. Such data will convince chemical companies of the potency and benefits of zebrafish as predictors of neurotoxic effects in humans. We believe that zebrafish will gain more attention and they will become highly popular organisms for testing chemicals.

### CONCLUSION

This mini review gives a brief overview of the potential use of zebrafish to evaluate neurotoxicity during development. Zebrafish possess significant advantages as model organisms and can overcome the limitations of other systems, making them potentially suitable as models in neurotoxicology. Thanks to their peculiarities, zebrafish can be employed as outstanding platforms to efficiently and rapidly evaluate the impact of chemicals on the developing brain. They offer the possibility to

screen several toxicity endpoints by combining different assays, allowing to generate quantitative assessments of a large numbers of chemicals. We believe, the increasing employment of zebrafish in testing chemicals will speed up this process and facilitate the understating of neurotoxicity mechanisms.

#### REFERENCES


#### AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

a multiplexed assay suitable for high-throughput screening. Toxicology 333, 14–24. doi: 10.1016/j.tox.2015.03.011



and dendritic growth in primary rat neurons. Toxicol. Sci. 158, 401–411. doi: 10.1093/toxsci/kfx100


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 d'Amora and Giordani. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

fnins-12-01019 January 7, 2019 Time: 19:45 # 1

# Transcytosis to Cross the Blood Brain Barrier, New Advancements and Challenges

#### Victor M. Pulgar1,2 \*

<sup>1</sup> Department of Pharmaceutical Sciences, Campbell University, Buies Creek, NC, United States, <sup>2</sup> Department of Obstetrics and Gynecology, Wake Forest School of Medicine, Winston-Salem, NC, United States

The blood brain barrier (BBB) presents a formidable challenge to the delivery of drugs into the brain. Several strategies aim to overcome this obstacle and promote efficient and specific crossing through BBB of therapeutically relevant agents. One of those strategies uses the physiological process of receptor-mediated transcytosis (RMT) to transport cargo through the brain endothelial cells toward brain parenchyma. Recent developments in our understanding of intracellular trafficking and receptor binding as well as in protein engineering and nanotechnology have potentiated the opportunities for treatment of CNS diseases using RMT. In this mini-review, the current understanding of BBB structure is discussed, and recent findings exemplifying critical advances in RMT-mediated brain drug delivery are briefly presented.

#### Edited by:

Hari S. Sharma, Uppsala University, Sweden

#### Reviewed by:

Gunnar P. H. Dietz, Georg-August-Universität Göttingen, Germany Marta Bolós, Centro de Investigación Biomédica en Red sobre Enfermedades Neurodegenerativas (CIBERNED), Spain

#### \*Correspondence:

Victor M. Pulgar pulgar@campbell.edu; vpulgar@wakehealth.edu

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 10 August 2018 Accepted: 18 December 2018 Published: 11 January 2019

#### Citation:

Pulgar VM (2019) Transcytosis to Cross the Blood Brain Barrier, New Advancements and Challenges. Front. Neurosci. 12:1019. doi: 10.3389/fnins.2018.01019 Keywords: brain endothelium, transcellular, receptor-mediated transcytosis, drug delivery, CNS diseases

## INTRODUCTION

Brain diseases are among the less understood and poorly treated conditions. In spite of the rapid growth in recent years in drug development, there is still a low success rate of effective therapies focused in diseases of the central nervous system. A main issue hindering therapeutic success is the tightly regulated extracellular environment of the brain tissue which makes reaching macromolecular targets into the brain a great challenge (Pardridge, 2005; Abbott, 2013; Engelhardt et al., 2016). The isolation of the brain tissue from the peripheral circulation is thought to arise from the existence of multi-level "barriers," established in different compartments in the central nervous system of most vertebrates (Cserr and Bundgaard, 1984; Engelhardt et al., 2017) providing protection to the neural tissue. Key to those protective mechanisms is the regulation of the entry of macromolecules from the blood to the brain across the blood-brain barrier (BBB) (Abbott et al., 2006). The BBB regulates an extended surface of interaction between blood and brain. It is calculated that the brain capillary network in humans is approximately 600 km long with a surface of 15–25 m<sup>2</sup> (Wong et al., 2013).

The intimate association between neurons, glial cells, and brain microvessels in the neurovascular unit is being recognized as the functional point for regulation of cerebral blood flow. Among those cell types, the brain endothelial cells are the building blocks of the BBB impeding the entry of most molecules from blood to brain, with the exception of those small and lipophilic in nature. Several recent studies have focused on the functional interactions between endothelial, neuronal and glial cell types and their role on regulating BBB function (Persidsky et al., 2006; Chow and Gu, 2015; Liebner et al., 2018). Since neurons rarely occur at long distance from a brain capillary (Schlageter et al., 1999; Tsai et al., 2009), the BBB also plays a major role in controlling fast delivery of substances to the brain and the local neuronal environment. Due to its extended contact and exchange surface area, most research has focused on the brain endothelium as the therapeutic target to increase brain drug delivery.

## STRUCTURE OF THE BBB

fnins-12-01019 January 7, 2019 Time: 19:45 # 2

Structurally, capillary networks can be divided into continuous non-fenestrated capillaries, continuous fenestrated and discontinuous capillaries. This division is based on their ability to regulate crossing of solutes from blood to tissues; thus continuous fenestrated capillaries are least permeable whereas discontinuous are the most permeable (Aird, 2007a,b). In the BBB, continuous non-fenestrated capillaries, where tight junctions (TJs) connect endothelial cells, form a high-resistance para-cellular barrier limiting the crossing of molecules and ions. Transmembrane proteins are an important part of TJs, they bind the cytoskeleton and link adjacent endothelial cells in a close configuration, eliminating intercellular spaces. Some of the proteins important for TJs structure and function include integral membrane proteins such as members of the claudin family i.e., claudin 3, 5, and 12, ocludins, and junctional adhesion molecules (Anderson and Van Itallie, 2009; Furuse, 2010). Evidences indicate that claudins are essential for the formation of the para-cellular barrier and the structure is stabilized by zona occludens ZO −1, −2, and −3 and additional proteins that link the TJs with the cytoskeleton (Abbott et al., 2006; Furuse, 2010). This structure is further reinforced by the basal lamina, a ∼40 nm thick matrix formed predominantly of collagen type 1V, laminin, and heparan sulfate proteoglycan (Perlmutter and Chui, 1990). Metalloproteinases are other components that contribute to regulation of BBB function in health and disease (Yong, 2005).

Additionally, glial cells such as astrocytes play an important role in development and maintenance of the BBB. Up to 99% of the basal capillary membrane is covered by astrocytes "end feet" and glial-derived factors such as GDNF, angiopoietin-1 and angiotensin II all contribute to BBB integrity (Hori et al., 2004; Abbott et al., 2006; Wosik et al., 2007). Along with astrocytes "end feet," pericytes are also lining the cerebral vasculature, surrounding brain endothelial cells and contributing to the barrier properties of the BBB. Recent advances on pericytes research indicate that this cell type is rather complex with more than one functional definition depending on their location along the arterio-venous capillaries (Attwell et al., 2016). The fact that brain microvessels are enriched in pericytes, and pericyte-deficient mouse mutants showed increased BBB permeability (Armulik et al., 2010) exemplifies the importance of pericytes for BBB control. Pericytes seem to contribute in two ways to BBB integrity: downregulating trans-endothelial permeability and promoting astrocyte-endothelial cells contacts (Armulik et al., 2010). Moreover, growing evidences point now to the importance of the interactions between pericytes and other cell types within the neurovascular unit in health and disease (ElAli et al., 2014).

The multicellular organization occurring at the neurovascular unit involving endothelial cells and astrocytes among others cell types (Willis, 2012) forms the framework where the highly regulated crossing of macromolecules from blood to brain occurs.

## CROSSING THE BBB

The existence of efflux transport systems in brain capillary endothelial cells reinforce the barrier properties of the BBB by removing undesirable substances from the brain to the systemic circulation. Multidrug resistance transporters, monocarboxylate transporters and organic anion transporters/organic anion transporting polypeptide have been implicated in the efflux of drugs from the brain. Consequently, the activity of these efflux transporters limits the effectiveness of CNS targeted drugs (Loscher and Potschka, 2005; **Figure 1C**).

Most of the drug transporters belong to two major classes; adenosine triphosphate binding cassette (ABC) and solute carrier (SLC) transporters. ABC transporters are active transporters coupling efflux against concentration gradients to ATP hydrolysis with P-glycoprotein (P-gp) being the most extensively studied BBB transporter of the ABC family (Mahringer and Fricker, 2016). P-gp is encoded by the multidrug resistance gene 1 (MDR1) and its function is regulated by intracellular factors and environmental toxins (Dauchy et al., 2009).

In order to facilitate the efficient delivery of drugs to the brain, the functional and structural tightness of the BBB needs to be overcome. Strategies used to cross BBB involve para-cellular as well as trans-cellular mechanisms.

## TRANSPORT ACROSS THE BBB

As part of its normal function, the endothelial cells allow the influx of nutrients and regulatory molecules into the brain via passive and active mechanisms. In normal conditions, some passive movement of solutes exists through small intercellular pores located in the TJs (**Figure 1E**). The molecular entities responsible for this transport are largely unknown, although recent evidences point to claudins as pore-forming structures in BBB TJs (Irudayanathan et al., 2017). Since early stage CNS diseases do not show evident BBB alterations, this pathway offers fewer opportunities than trans-cellular transport for drug delivery.

Transport of small molecules trough cells is common in polarized cells. Thus, in brain vascular endothelial cells, hydrophobic molecules with molecular weight lower than 500 Da once they escape the P-gp-type multidrug resistance efflux pumps may diffuse transcellularly from systemic circulation to brain parenchyma (**Figure 1F**). The transport of nutrients, however, requires specialized transporters (**Figure 1D**). Thus, large neutral aminoacid transporters (LAT1) transport aminoacids, nucleosides, and some drugs, while glucose uses the glucose transporter (GLUT1) (Ohtsuki and Terasaki, 2007; Barar et al., 2016).

## TRANSCYTOSIS

fnins-12-01019 January 7, 2019 Time: 19:45 # 3

Transcytosis is a phenomenon present in many different cell types, from neurons to intestinal cells, osteoclasts and endothelial cells. In polarized cells, unidirectional transcytosis refers to the transport of macromolecules from apical to basolateral plasma membranes. Steps along this pathway include endocytosis, intracellular vesicular trafficking and exocytosis. The first of these steps may involve adsorptive (charge dependent) or receptor-mediated internalization (**Figures 1A,B**). Positively charged molecules such as polymers, cationic lipids, albumin and nanoparticles may interact with the negatively charged cell membrane and internalize through adsorptive endocytosis (Lu, 2012). Although initially thought to be attenuated in brain endothelial cells, virtually all endothelial cells display receptor-mediated transcytosis (RMT) (Stewart, 2000). Recent applications of imaging techniques allowed for detailed analyses of transcytosis in brain endothelial cells (Villasenor and Collin, 2017). Several receptors capable of inducing RMT are present in the BBB, such as the insulin receptor, transferrin receptor, and receptors responsible for lipoprotein transport, while others such as albumin receptors are not expressed (Pardridge et al., 1985).

Transport; (E) Paracellular Transport; (F) Diffusion. See text for details.

The intracellular transport of macromolecules is mediated by the vesicular system (Parkar et al., 2009). In brain endothelial cells three types of endocytic vesicles have been identified: clathrin-coated pits involved in most of the RMT, caveolae that participate in adsorptive-mediated endocytosis of extracellular molecules and receptor trafficking, and macropinocytotic vesicles (Mayor and Pagano, 2007). Of these, clathrin-coated vesicles are involved in most of the internalization processes mediated by approximately 20 different receptors in brain endothelial cells.

Once a vesicle is internalized, the common intracellular pathway begins with the initial sorting station, the early endosome (Rodriguez-Boulan et al., 2005; Brooks, 2009; **Figure 2**). In BBB endothelial cells endocytosis occurs at the apical and basolateral membranes with both processes generating its own early endosomes. In polarized cells, routing back to the plasma membrane can occur directly from EE or from recycling endosomes (Thompson et al., 2007). Alternatively, vesicle components can be delivered to late endosomes and targeted for lysosomal degradation. This endosomal trafficking plays an important role in the efficiency of RMT in BBB (Haqqani et al., 2018).

## RMT FOR DRUG DELIVERY TO THE BRAIN

In general, strategies using RMT for drug delivery to the brain involve the generation of a complex between the drug

fnins-12-01019 January 7, 2019 Time: 19:45 # 4

TABLE 1 | Main receptor systems identified mediating receptor-mediated transcytosis (RMT) cargo delivery through the BBB.


(Continued)

#### Table 1 | Continued

fnins-12-01019 January 7, 2019 Time: 19:45 # 5


Evidences of efficient RMT utilizing the Transferrin Receptor, Insulin Receptor, Low Density Lipoprotein Receptor, and single domain llama antibodies are summarized.

of interest and a receptor-targeting entity. This entity may be the endogenous receptor ligand, an antibody targeting the receptor or a mimetic peptide ligand. These two components can be chemically linked or the drug can be incorporated in liposomes or nanoparticles decorated with the RMT-targeting ligand (Jones and Shusta, 2007). Among the most studied targets for RMT in brain endothelial cells are the transferrin receptor, low-density lipoprotein (LDL) receptor and insulin receptor, for reviews see (Lajoie and Shusta, 2015; Paterson and Webster, 2016). In the following section, some examples of the use of these systems are presented with focus in recent advances.

#### Transferrin Receptor

Iron delivery to the brain is accomplished via binding and intracellular trafficking of the iron binding protein transferrin (Tf). The Tf receptor (TfR) has been the target of numerous in vitro and in vivo studies aiming to deliver drugs to the brain (see **Table 1**). Approaches used include liposomes decorated with Tf used for delivery of imaging agents and DNA (Sharma et al., 2013) or the use of an iron-mimetic peptide as ligand (Staquicini et al., 2011). Since the presence of high blood levels of Tf requires competition with the endogenous ligand, alternative methods involving anti-TfR antibodies have been developed (Qian et al., 2002). Challenges using anti-TfR to deliver drugs to the brain via RMT include specificity to the brain tissue, potential lysosomal degradation and significant transport into the brain parenchyma. With the use of protein engineering it has been shown that reducing antibody's affinity for Tf improves release of the antigen-antibody complex in the basolateral side of the BBB endothelial cells (Yu et al., 2011). A correlation has also been suggested between increased antibody's affinity and lysosomal degradation (Bien-Ly et al., 2014) supporting the idea that lower antibody's affinity would help avoid intracellular degradation of the complexes being transported. Studies comparing the brain penetration of monovalent versus divalent antibodies indicate lower lysosomal colocalization of the monovalent form (Niewoehner et al., 2014) and better transcytosis (Johnsen et al., 2018). It appears that in addition to antibody's affinity in physiological conditions, a lower affinity at pH5.5 (lysosomal) also promotes effective transcytosis as suggested by in vitro studies using immortalized human brain endothelial cells (Sade et al., 2014).

The recent successes using TfR in RMT strategies has prompted novel developments aiming to potentiate drug delivery to the brain (Yemisci et al., 2018). Thus, recent reports showed efficient BBB crossing of particles functionalized with anti-TfR antibodies and containing non-permeant drugs of interest for treating brain diseases. Some examples include liposomes containing the MYBE/4C1 antihuman TfR antibody and loaded with the anticancer drug doxorubicin displaying enhanced uptake in human brain endothelial cells (Gregori et al., 2016), and liposomes containing Tf and docetaxel showing greater brain uptake after i.v., injection in rats compared to the drug alone (Sonali et al., 2016). The use of nanoparticles formulated using the Tf system has shown that functionalization with anti-TfR antibodies enhances the delivery of particles carrying relevant drugs such as drugs able to inhibit beta amyloid aggregates (Loureiro et al., 2016). Nanoparticles carrying the chemotherapeutic agent temozolomide have also facilitated enhanced drug uptake by glioblastoma cells (Ramalho et al., 2018). This strategy also exemplifies some of the challenges remaining in the field since gold nanoparticles (AuNPs) coated with the 8D3 anti-TfR antibody injected in mouse are transported through the BBB with low efficiency and most of the particles remain sequestered intracellularly in the endothelial cells (Cabezon fnins-12-01019 January 7, 2019 Time: 19:45 # 6

et al., 2015). Successful uptake by the BBB but low delivery to the brain parenchyma was also reported with quantum dots (Paris-Robidas et al., 2016). The dual functionalization of particles with peptides targeting the TfR to cross the BBB and additional therapeutic agents opens opportunities to specifically modulate gene expression in brain cells as shown by studies of co-delivery of doxorubicin and RNAi targeting the VEGF (Kuang et al., 2016), or siRNA targeting the EGFR (Wei et al., 2016) to glioma cells. The significant reduction in expression of the pro vascularization factors VEGF and EGFR observed in these two studies supports this use of co-delivery systems.

TfR has been used extensively as a model for brain transcytosis, although initial reports came from just one laboratory, later reports supported reproducibility of its use in different settings. Outstanding issues remaining such as brain specificity and low drug uptake will promote further research of this important RMT system.

## Insulin Receptor

Insulin is transported into the brain by the insulin receptor (IR). Similarly to the TfR, anti-IR antibodies have been developed and used in strategies to drug delivery into the brain (see **Table 1**). Following the development of humanized anti-IR antibodies (HIRMAb) that showed good internalization and transport to the brain after intravenous administration in monkeys (Boado et al., 2007), fusion proteins were developed to deliver relevant enzymes as therapies for genetic disorders. One of those examples is a fusion protein between the HIRMAb and α-L-idorunidase (IDUA) an enzyme missing in Hurler's Syndrome, Mucopolysaccharidosis Type I (MSPI), a disorder of brain lysosomal storage (Boado et al., 2008). In pre-clinical studies, HIRMAb-IDUA showed good safety, adequate plasma glucose control, and limited antidrug antibody production (Boado et al., 2009, 2012). Of great interest are recent reports describing clinical studies with HIRMAb-IDUA. In MSPI pediatric and adult patients intravenous infusion of HIRMAb-IDUA describes the first clinical use of RMT to drug delivery into the brain (Pardridge et al., 2018). Although some adverse events reported include reaction at the infusion site, and transient hypoglycemia, the positive neurocognitive and somatic effects observed in pediatric patients (Giugliani et al., 2018) represents a significant advancement on the translational aspects of RMT.

## LDL Receptor

Low-density lipoprotein receptor (LDLR), a single transmembrane glycoprotein able to recognize LDL particles and promote their endocytosis, as well as LDLR-related proteins (LRPs), are present in the BBB and mediate transport of lipoproteins and other ligands through RMT (Hussain et al., 1999; Candela et al., 2008; **Table 1**). Recent in vitro studies showed that LDLR is preferentially located in apical rather than basolateral membranes in brain endothelial cells (Molino et al., 2017) supporting a role for ligand uptake from the circulation. To date no antibodies have been developed targeting the LDLR system, however, LDLR and LRP ligands have been used for drug delivery into the brain. One of those ligands is melanotransferrin, which displays a greater rate of brain transport compared to Tf. In spite of structural homology to Tf, melanotransferrin uses the LDLR and not the TfR to cross the BBB (Demeule et al., 2002). Interestingly, recent reports showed melanotrasferrin delivery and in vivo effectiveness of a fusion protein with an interleukin-1 receptor antagonist in a model of neuropathic pain (Thom et al., 2018). Lipoproteins have also been used to target LDLR for effective brain delivery (Wagner et al., 2012), as described in glioblastoma cells (Nikanjam et al., 2007). Recent developments include functionalization of solid nanoparticles with ApoE, these 160 nm nanoparticles showed efficient clathrin-dependent endocytosis and transcellular transport in human brain endothelial cells (Neves et al., 2017). Other targeting members of the LDLR family include "angiopeps." For example, Angiopep-2 was identified by studying a series of 19 amino acid peptides with the ability to bind the LPR-1 receptor (Demeule et al., 2008). Angiopep-2 was shown to mediate efficient delivery of a conjugate Angiopep-2 placlitaxel to gliomas (Thomas et al., 2009), and more recently antinociceptive properties were demonstrated for an Angiopep-2-neurotensin fusion protein (Demeule et al., 2014). These studies provide evidence of successful delivery of therapeutically relevant agents to the brain via RMT targeting the LDLR family.

## Single Domain Llama Antibodies

Single domain antibodies (sdAbs) are naturally occurring fragments of the antibody's heavy chain that lack the light chain. Among them, sdAbs from camelids specifically FC5 and FC44 have been studied for brain transcytosis of cargo in animal models and their potential warrants further developments (see **Table 1**). FC5 and FC44 recognize α(2,3)-sialoglycoprotein expressed in the luminal side of brain endothelial cells and display advantages over other antibodies such as small size, greater specificity and stability, and low immunogenicity (Arbabi-Ghahroudi, 2017).

## CONCLUSION

Recent advances using RMT are providing alternatives to overcome the barrier properties of the BBB and develop more efficient drug delivery to the brain. Future developments based the TfR, IR, and LDLR and other RMT systems will offer new opportunities in this growing field. However, in spite of clear therapeutic advances shown in animal studies, outstanding challenges remain for the development of efficient and specific RMT-based drug delivery. Although the mechanisms mediating efficient transcytosis through the brain endothelium are still incompletely understood, details about the specific targeting to brain endothelial cells are being revealed. Similarly, the limited brain specific versus systemic drug uptake may explain the lack of success of some potential therapies in non-human primates models of brain diseases.

In addition to increasing knowledge about the factors modulating intracellular trafficking, the generation of fusion proteins with RMT-targeting antibodies as well as functionalization of Nano carriers, an improved understanding of BBB transport, pharmacokinetics, and protein engineering will be needed to potentiate the clinical applicability of RMT.

### REFERENCES

fnins-12-01019 January 7, 2019 Time: 19:45 # 7


## AUTHOR CONTRIBUTIONS

VP was responsible for the design of this review, the literature searches, writing and interpretations presented.


an IgG-GDNF fusion protein that penetrates the blood-brain barrier. Brain Res. 1352, 208–213. doi: 10.1016/j.brainres.2010.06.059


fnins-12-01019 January 7, 2019 Time: 19:45 # 8

multiforme treatment. Int. J. Pharm. 545, 84–92. doi: 10.1016/j.ijpharm.2018. 04.062


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Pulgar. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

fnins-12-01019 January 7, 2019 Time: 19:45 # 9

# Multi-Modal Nano Particle Labeling of Neurons

Lilac Amirav<sup>1</sup> \*, Shai Berlin<sup>2</sup> \*, Shunit Olszakier1,2, Sandip K. Pahari<sup>1</sup> and Itamar Kahn<sup>2</sup> \*

<sup>1</sup> Schulich Faculty of Chemistry, Technion – Israel Institute of Technology, Haifa, Israel, <sup>2</sup> Department of Neuroscience, Ruth and Bruce Rappaport Faculty of Medicine, Technion – Israel Institute of Technology, Haifa, Israel

#### Edited by:

Ioan Opris, University of Miami, United States

#### Reviewed by:

Preston Snee, University of Illinois at Chicago, United States Natalie Julie Serkova, University of Colorado Denver, United States Hari S. Sharma, Uppsala University, Sweden

#### \*Correspondence:

Lilac Amirav lilac@technion.ac.il Shai Berlin shai.berlin@technion.ac.il Itamar Kahn kahn@technion.ac.il

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 28 July 2018 Accepted: 08 January 2019 Published: 01 February 2019

#### Citation:

Amirav L, Berlin S, Olszakier S, Pahari SK and Kahn I (2019) Multi-Modal Nano Particle Labeling of Neurons. Front. Neurosci. 13:12. doi: 10.3389/fnins.2019.00012 The development of imaging methodologies for single cell measurements over extended timescales of up to weeks, in the intact animal, will depend on signal strength, stability, validity and specificity of labeling. Whereas light-microscopy can achieve these with genetically-encoded probes or dyes, this modality does not allow mesoscale imaging of entire intact tissues. Non-invasive imaging techniques, such as magnetic resonance imaging (MRI), outperform light microscopy in field of view and depth of imaging, but do not offer cellular resolution and specificity, suffer from low signal-to-noise ratio and, in some instances, low temporal resolution. In addition, the origins of the signals measured by MRI are either indirect to the process of interest or hard to validate. It is therefore highly warranted to find means to enhance MRI signals to allow increases in resolution and cellular-specificity. To this end, cell-selective bi-functional magnetofluorescent contrast agents can provide an elegant solution. Fluorescence provides means for identification of labeled cells and particles location after MRI acquisition, and it can be used to facilitate the design of cell-selective labeling of defined targets. Here we briefly review recent available designs of magneto-fluorescent markers and elaborate on key differences between them with respect to durability and relevant cellular highlighting approaches. We further focus on the potential of intracellular labeling and basic functional sensing MRI, with assays that enable imaging cells at microscopic and mesoscopic scales. Finally, we illustrate the qualities and limitations of the available imaging markers and discuss prospects for in vivo neural imaging and large-scale brain mapping.

Keywords: MRI, iron oxide nanoparticles, light microscopy, contrast agents, magneto fluorescence nanoparticle labeling

## INTRODUCTION

A central goal in neuroscience research is the development of imaging methodologies for longitudinal single neuron interrogation in the intact animal. High-resolution light-microscopy (LM) imaging of neurons in vivo enables structural and functional mapping, as well as remote optical control with exquisite spatiotemporal resolution; shedding light on some of the most

fundamental questions related to neural morphology and function in health and disease. However, LM offers but a glimpse of the brain in vivo, typically providing access to very small fieldsof-view and depths of the mammalian brain (Silva, 2017; Zong et al., 2017). Thus, non-invasive imaging modalities that can reveal the structure and function of neurons in the entire brain, i.e., at the mesoscale, could provide the complementary means needed for this ambitious endeavor. Magnetic resonance imaging (MRI) is particularly attractive to meet this goal, owing to its soft-tissue imaging capabilities, and the ability to image the entire organ non-invasively and repeatedly over a long period of time, through the intact skull and at any depth, with no known adverse impact on the tissue.

As commonly used, MRI provides an indirect measure of the brain's structure and function, owing to the manner by which signals are acquired — by perturbing the alignment of the highly abundant hydrogen atoms nuclear spins found in this tissue. In structural MRI, this signal is indistinguishable between the intra- versus extra-cellular environments and across cell types. Similarly, the contrast mechanism commonly measured by functional magnetic resonance imaging (fMRI) is also indirect (Logothetis, 2008; Kim and Ogawa, 2012). Intrinsic signals for fMRI arise from the ratio between oxygenated and deoxygenated hemoglobin, a mesoscopic metabolic measure, rather than measures of electrical activity or intracellular physiological events, such as Ca2<sup>+</sup> concentration changes or neurotransmitter release. In recent years, several technological advances, including magnetic fields significantly higher than those commonly used in clinical settings, novel radio-frequency coils with a large number of densely-spaced small coil elements, and the use of unique dielectric materials, allow to indirectly track the activity of smaller and smaller populations of neurons in the brain (Ugurbil, 2016). Despite these innovations, the signals are still minute, of relatively poor signal-to-noise ratio (SNR) and indirect. To overcome these limitations, means to increase contrast of defined targets are under development, in particular MRI-compatible contrastagents that are designed to detect specific molecular targets in the brain (Mukherjee et al., 2017; Ghosh et al., 2018). As contrastagents are used routinely in clinical MRI scans, novel classes of contrast agents suitable for use in humans will certainly enhance and expand the capabilities of the technique. We propose that bi-functional materials that serve as hybrid contrast-agents for multiple imaging modalities at once, notably LM and MRI, will open new fronts for structural and functional whole-brain imaging at higher resolution and with target specificity.

The design and synthesis of bi-functional materials is an active research area with a significant impact on a wide range of technological applications (Corr et al., 2008; Gao et al., 2009; Suh et al., 2009; Qin and Bischof, 2012; Cheng et al., 2014). From an imaging point-of-view, combining several different properties onto a single agent can functionalize it towards many types of imaging and detection modalities; greatly extending its diagnostic, and potentially therapeutic, value. Indeed, emerging multifunctional contrast agents have been shown to label cancerous cells (Sailor and Park, 2012; Dawidczyk et al., 2014; Sanna et al., 2014) or genetically-modified cells in vivo (Kim et al., 2008; Muthu et al., 2014; Ortgies et al., 2016) for detection by MRI and validation by LM. Target selective imaging is a prerequisite for image-guided interventions (including surgery and ablation therapy) (Li, 2014). Beyond imaging, multifunctional contrast agents may also include pharmacological agents to concomitantly visualize and treat diseases in an all-in-one therapeutic and diagnostic (theranostic) approach (Yoo et al., 2011; Lim et al., 2015), for which there is a growing number of examples (Gao, 2018). Specifically, iron oxide nanoparticles are utilized for magneto-responsive therapy, where the responsiveness of the nanoparticles to an external magnetic field is used in order to increase the accumulation of the particles in a target tissue (magnetic targeting), or for exogenous physical stimuli release of cargo gene or drug molecules (Lee et al., 2015). However, this field is still in its infancy, with significant limitations in bona fide labeling of defined cellular targets and at obtaining sufficient particle accumulation at desired locations for gaining higher resolutions; notably to the single-cell level.

Cell-targeted contrast agents could provide the means to increase target-specificity and resolution by adhering and accumulating around or within cells. Extracellular contrast agents are typically aimed at reversibly binding proteins exposed to the extracellular-milieu. However, owing to proteinturn-over, limited expression of the target-protein, expression of some proteins in a large variety of cell types and limited membrane surface-area (Saka et al., 2014), such agents do not typically provide sufficient contrast of defined-cells; especially not for prolonged durations (Mukherjee et al., 2017). Intracellular contrast agents, on the other hand, may bypass several of these limitations and, thereby, potentially provide an extended imaging time-window. For instance, membrane expression levels and protein turn-over are less likely to impact the latter. In addition, the intracellular space (i.e., cytoplasm) greatly exceeds that of the membrane surface, allowing for the accumulation of larger amounts of contrast agents and, consequently, to provide higher signals. Intracellular accumulation also slows washout of the agent, thus extending the imaging periods, and it may also localize the activity of a therapeutic agent, or enable a weak drug (e.g., with a high median effective dose) to become efficient exclusively in the desired cellular population where the drug has been concentrated.

To meet their full potential, intracellular contrast-agents and their further development should benefit from better understanding of their cellular uptake mechanisms. Of particular interest is information on the modes of cellular uptake, their efficiency and kinetics, subcellular distribution of contrast agents following uptake, saturation concentrations, clearance and, notably, toxicity. Gaining control over these parameters will open the door towards novel basic and clinical applications.

A key requirement towards meeting this goal is the ability to track and validate the intracellular accumulation of contrast agents with high spatiotemporal resolution, significantly higher than what is currently afforded by MRI, and over an extended period of time. Here, multifunctional contrast agents that can be detected by both LM and MRI are particularly useful, lending themselves to achieve this task. Fluorescence imaging provides exquisite high-resolution means to explore and validate the

different features of cellular uptake. It can be used to assess the accumulation of specific contrast agents at the cellular, subcellular, protein or even single molecule level, and at very high temporal resolution. To this end, several magneto-fluorescent hybrid systems are currently under development.

### MAGNETO FLUORESCENCE HYBRID NANOPARTICLES

In order to combine LM and MRI, appropriate agents for both imaging modalities first need to be considered based on their signal strength, toxicity, stability and size. Several classes of MRI-compatible contrast agents are available (Geraldes and Laurent, 2009); with iron oxide nanoparticles meeting most of these requirements, namely provide a strong and stable MRI-signature, with little effects on cellular physiology (e.g., Uchiyama et al., 2015; and see review Shen et al., 2017). Consequently, iron oxide nanoparticles are commonly employed in the field. In particular, these have been rendered bi-modal by their conjugation to organic fluorescent dyes, such as fluorescein isothiocyanate, rhodamine B, and Cy5.5; all commonly employed bright markers for LM (Wysocki and Lavis, 2011). However, organic dyes may undergo rapid photobleaching and/or photochemical degradation (shortening the imaging time-window), or produce cytotoxic byproduct (e.g., reactive oxygen species; Zheng et al., 2014) and damage to the biological system under investigation. To address the instability of organic dyes, several reports describe encapsulation of dyes within silica to provide protection (Ow et al., 2005; Piao et al., 2008), but this provides only a moderate protection and comes at the cost of larger particles size. Quantum dots (QDs), inorganic fluorescent semiconductor nanoparticles, are an attractive alternative (Kim et al., 2004; Medintz et al., 2005). QDs have a high molar extinction coefficient and fluorescence quantum yield, broad absorption, and narrow tunable emission spectra, in addition to excellent temporal stability and resistance to photobleaching. These make QDs particularly advantageous over other fluorescent agents. However, coupling of QDs to MR contrast agents, specifically iron oxide nanoparticles, is not straightforward, and likely the reason why this has not been commonly achieved. This is mainly because a direct contact between the semiconductor and magnetic domain can lead to strong electronic coupling and strong attenuation of the QDs' fluorescence. Indeed, traditional heterodimer structures, namely placing the semiconductor QD directly on the surface of an iron oxide nanoparticle, are not sufficiently bright for optical imaging (Selvan et al., 2007). Fluorescence quenching can be minimized, or prevented entirely, if effective separation is achieved (Boldt et al., 2011;Bigall et al., 2012; Feld et al., 2015; Harris et al., 2016). Hence, alternative synthetic strategies for the fabrication of magneto-fluorescent materials include conjugation of separate nano-constructs, or co-encapsulation into organic structures or inorganic materials (Kim et al., 2006; Kim and Taton, 2007; Insin et al., 2008; Park et al., 2008; Roullier et al., 2008; Erogbogbo et al., 2010; Fan et al., 2010; Kas et al., 2010; Di Corato et al., 2011; Shibu et al., 2013; Cho et al., 2014; Lee et al., 2015). These conceptual designs are expected to prevent undesirable interactions within the hybrid that could abrogate the respective properties.

Two prominent examples of such designs are presented in **Figures 1a,b**. Bawendi and co-workers (Chen et al., 2014) developed colloidal superstructures comprised of close-packed magnetic nanoparticle cores that are fully surrounded by a shell of fluorescent QDs, and the core-shell superparticle is coated with a protective silica shell. These super-nanoparticles exhibit high magnetic content (with T<sup>2</sup> relaxivity of 402.7 mM−<sup>1</sup> S −1 ) along unperturbed fluorophore loading, with an overall diameter of ∼100 nm. Weller and co-workers (Feld et al., 2015) used polystyrene to co-encapsulate iron oxide nanoparticles with quantum rods, thereby preserving the fluorescence and magnetism of the separate components, with particle diameters ranging from 74 to 150 nm, and relaxivity of 164 mM−<sup>1</sup> S −1 .

Recently, we presented a novel design strategy for the fabrication of ultra-small (∼15 nm hydrodynamic size) magnetofluorescent nanoparticles (Pahari et al., 2018). In lieu of conjugation or co-encapsulation of the separate magneto and fluorescent components into an insulating matrix, both were combined into a single entity with a unique morphology. The optically active semiconductor QD was encapsulated directly into a hollow paramagnetic iron oxide shell that serves as the MRI contrast agent (**Figure 1c**). Despite their small size, these nanoparticles provide contrast enhancement with a relaxivity of 304 mM−<sup>1</sup> S −1 , making them comparable to the much bigger aforementioned systems, and superior to commercial contrast agents such as Fe3O<sup>4</sup> nanoparticles (26.8 mM−<sup>1</sup> S −1 ). This so-called "yolk-shell" morphology prevents the undesired interactions within the hybrid that would diminish the properties of the semiconductor and magnetic domains. Furthermore, this architecture offers a high level of tunability with respect to size, composition and surface functionalization. The photoluminescence quantum yield of the yolk-shell nanoparticles was typically ∼16% (at 450 nm excitation light), comparable to the ∼12% photoluminescence quantum yield (at 405 nm) of the silica coated core-shell superparticles. This yield was sufficient for single particle tracking (Chen et al., 2014), and for optical microscopy imaging of nanoparticle clusters within single cells, or of 3D cell cultures (as seen in **Figure 2**; Pahari et al., 2018). The decrease in photoluminescence quantum yield compared with that of free QDs, that could be as high as 94%, is attributed to the overlapping absorption with that of the iron oxide.

Noteworthy is the fact that for the morphologies presented in **Figure 1**, the fluorescent QDs may be altered with no effect on overall size, surface characteristics and magnetic properties of the particle. Hence, we anticipate that future hybrids will replace the Cd-based semiconductor with nontoxic and biocompatible alternatives and incorporate materials with efficient near-infrared (NIR) optical activity in both absorption and fluorescence; wavelengths more suitable for in vivo applications. This adjustment will also minimize the spectral overlap between the absorption of the core and that of the

hollow shell, and is expected to improve the photoluminescence quantum yield.

An interesting alternative for multimodal imaging is provided by incorporation of manganese ions into the QD, hence offering the semiconductor itself paramagnetic characteristics (Sitbon et al., 2014). These structures may serve as T1-contrast agents, influencing the longitudinal relaxation time, whereas superparamagnetic iron oxide nanoparticles are employed as T2-contrast agents, influencing the transverse relaxation time.

#### INTRACELLULAR ACCUMULATION OF NANO-PARTICLES

Particle-size is a critical criterion that will determine the mechanism by which it will be internalized into the cell, with major influences on cellular uptake efficiency and kinetics, and subcellular distribution (He et al., 2010; Huang et al., 2010; Verma and Stellacci, 2010; Shang et al., 2014; Zhang et al., 2015). Note that this refers to the overall size of the structure, including the inorganic and organic shells. The modes of internalization can vary quite extensively, ranging from processes such as opsonization and phagocytosis (Gustafson et al., 2015), through clathrin/caveolar-mediated endocytosis (Hachani et al., 2017) and receptor-mediated endocytosis (Qiao et al., 2012), to macroand pinocytosis. The size of a contrast agent will also impact the means by which it escapes from the endosomal/lysosomal compartments and enters the cytoplasm (for a review see, Oh and Park, 2014).

A significant factor for in vivo imaging is the effect of the hydrodynamic diameter of the contrast agent on its resident circulation time (Kievit and Zhang, 2011). Nanoparticles with a hydrodynamic diameter of less than 10 nm, i.e., smaller than the pore size of the glomerulus, are subjected to rapid excretion and thus have relatively short dwell-time in the blood stream (Chapman et al., 1999; Olmsted et al., 2001; Choi et al., 2007; Liu et al., 2013; Ehlerding et al., 2016). On the other hand, nanoparticles with an overall size larger than 100 nm may be recognized and removed quickly from the blood stream by the reticuloendothelial system (opsonization by white blood cells, namely macrophages), which will also result in short circulation time (Reimer and Tombach, 1998). Ultrasmall superparamagnetic iron oxide nanoparticles that are smaller than 50 nm can escape phagocytosis to some extent with a prolonged circulation time (Shen et al., 1993; Wang et al., 2001; Li et al., 2005). An overall size of 15 nm was found to

of the cells along with the inner volume of the cells containing NPs (magenta). (f) Higher power image showing a single cell filled with at least few dozens of NPs.

result with relatively longer circulation time, affording prolonged MRI signal enhancement effects for over an hour, and collection of high-resolution MR images of blood vessels (Kim et al., 2011). Therefore, longitudinal measurements in the intact animal that depend upon longer circulation time will require unique adjustment and optimization of the overall size of the agent. Together, nanoparticles should not be too small, to prevent rapid central nervous system clearance, and should not be too large, to avoid uptake by macrophages. Thus, the synthetic strategy of choice is critical for appropriate functionality of the magnetofluorescent hybrid structure. In this regard, the yolk-shell strategy may present a convenient design for experiments that require smaller dimensions.

In addition to size, the shape, surface chemistry and aggregation or agglomeration habits of the nanoparticles will also influence the type of internalization process (Albanese and Chan, 2011). In particular, surface coating is found to play an important role in successful delivery of nanoparticles into the intracellular environment (e.g., see **Figure 2**), and it may additionally offer opportunities for selective labeling of relevant targets. The surface of our yolk-shell magneto-fluorescent nanoparticles is that of the iron oxide shell, rather than silica or polystyrene. This allows implementation of the vast accumulated knowledge that is available in the literature for surface coating and functionalization of iron oxide for the purpose of biological compatibility. Hence, insights on cell internalization and accumulation of the yolk shell magneto fluorescent nanoparticles, which are acquired via utilization of the fluorescent marker, could be directly implemented for similar iron oxide therapeutic nanoparticles. A prominent example relies on distinctive biochemical characteristics of target diseases such as tumor or inflammatory tissue, with surface-functionalization that is designed for improved cell and tissue specific distribution of nanoparticles for localized therapeutic effect, and minimal

whole-body toxicity (Lee et al., 2015). In addition, along side functionality of the iron oxide nanoparticles as therapeutic hyperthermic agents (Petryk et al., 2013; Hervault and Thanh, 2014), or for controlled drug release through the application of an external magnetic field (Yoo et al., 2011), the QD core could easily be replaced with Au (Shevchenko et al., 2008; Jain et al., 2012) or FePt (Gao et al., 2008) for cancer therapy. Note that such functionalities are hindered or lost entirely if the iron oxide nanoparticles are coated within thick silica or polystyrene shell.

In the course of our work, the CdSe@CdS@hollow-Fe2O<sup>3</sup> yolk-shell magneto-fluorescent nanoparticles (MFNP) were functionalized first with Tiron (disodium 4,5-dihydroxy-1,3 benzenedisulfonate) for their incorporation to cells (Korpany et al., 2013). This resulted with stable and well suspended waterdispersed nanoparticles of 14–16 nm hydrodynamic particle size (as determined by dynamic laser scattering analysis). Alternative ligand coatings such as polymer or PEG resulted with drastic increase of the hydrodynamic size, with no added stability. Yet, the intracellular accumulation of Tiron-coated nanoparticles was found to be variable and, at times, limited. Hence, dopamine was examined as an alternative coating (Sherwood et al., 2017). We, and others, envisioned the dopamine-coat to play several critical roles. First, it was expected to increase the solubility of the nanoparticle. Second, it could be recognized by membrane receptors or transporters to promote cell-specific receptormediated endocytosis and internalization of the particles into cells of interest, neurons in particular. Lastly, this coat was also envisioned to play a role in reducing the ability of extracellular proteins to adhere to the nanoparticles, thereby negating their increase in size and clearance by engulfing cells. Though successful to some extent, we have obtained weak LM signals from cells suggesting only moderate cytoplasmic fill. To further enhance endocytosis, we turned to established chemical reagents that are commonly employed for the introduction of foreign material into cells such as DNA. The MFNP were incubated with the transfection reagent lipofectamine, glycerol; a water soluble and hygroscopic lipid, calcium phosphate precipitates and spermidine; a naturally occurring polyamine commonly used in biolistic transfection of neurons. Strikingly, and despite the common features of the reagents, preliminary observations showed that spermidine yielded the best results; with cells displaying detectable intracellular particles with clear evidence of aggregation (∼1 µm in diameter; **Figure 2**). Notably, external modifications and coatings did not affect the functionality of the encapsulated QDs, as evident by their emission fluorescence spectrum outside or within cells (**Figures 2a,b**).

The observation of aggregates was surprising, as the nanoparticles consistently remain monomeric in solution in vitro. This therefore suggested to us that the particles are undergoing agglomeration, despite the dopamine coat, and this process is likely mediated by the physiological and/or intracellular conditions. Though unsettling at first, we quickly realized that this process could be beneficial for intracellular trapping of the nanoparticles because as small particles may easily enter the cell, so can they exit it. Deliberate and controlled aggregation may serve for entrapping and accumulating nanometer particles within the cell more readily, in turn providing stronger MRI and fluorescent signals, not to mention increasing the stability of the particles over time due to lower exposure to degradation. This observation also suggests that with further design, nanoparticles could be engineered to undergo activity-dependent aggregation. A relevant and interesting example to explore is aggregation that is induced by the binding of a certain biological moiety that surges in an activity-dependent manner, such as Ca2+-ions (Okada et al., 2018). In this regard, we envision that the coat of the nanoparticles could include small molecules or protein fragments (e.g., Li et al., 1999; Atanasijevic and Jasanoff, 2007; Henig et al., 2011) that bind Ca2+, e.g., calcium-specific aminopolycarboxylic acid such as 1,2-bis(2 aminophenoxy)ethane-N,N,N',N'-tetraacetic acid (BAPTA) or Calmodulin, respectively. A plausible scenario would be that following neural activity and sharp rise in cytoplasmic Ca2+ concentration, intracellular nanoparticles would bind Ca2+, thus altering the reactivity of the coating. Ca2<sup>+</sup> binding may influence the effective columbic repulsion of the coating, or enhance the affinity and binding to a secondary moiety on the surface of surrounding nanoparticles. Regardless the exact mechanism, it is expected to change the particles' dispersivity in a manner that would result with strong aggregation. This aggregation, once it exceeds a certain size, is expected to result in a detectable MR signal localized to the cell. As Ca2+-binding is reversible, this phenomenon would also be reversible, providing MRI-readout for neuronal activity.

## CONCLUSION

Intracellular magneto-fluorescent contrast agents are of great significance for the development of imaging methodologies for single cell longitudinal measurements in the intact brain. These bi-functional markers enable coupling between high resolution LM with the non-invasive mesoscale imaging capabilities of MRI. Yet the roadmap towards enhancement of MRI signals from select specific cells requires improved understanding of cellular internalization and accumulation mechanisms. In particular, modes of cellular uptake, their efficiency and kinetics, subcellular distribution of contrast agents following uptake, clearance and, notably, toxicity need to be better understood and accurately characterized. These are all influenced, and at times dictated, by the size, shape and surface chemistry of the contrast agent nanoparticles, and should be properly considered when choosing the synthetic strategy. We envision that insights on the correlation between the agents' characteristics and their intracellular distribution would enable manipulation of the latter for improved specificity, and ultimately for functional readout of neuronal activity.

### ETHICS STATEMENT

All of the experiments involving animals were conducted in accordance with the United States Public Health Service's Policy on Humane Care and Use of Laboratory Animals and study protocol was approved by the Institutional Animal Care and Use Committee of Technion – Israel Institute of Technology.

## AUTHOR CONTRIBUTIONS

fnins-13-00012 January 30, 2019 Time: 17:58 # 7

LA, SB, and IK developed the concepts described in this work and wrote the manuscript. SO, SP, and SB carried out the experiments. All authors analyzed the results described in the manuscript and approved the final version of the manuscript. This work is in partial fulfilment for the Degree of Doctor of Philosophy (Ph.D.) for SO.

#### REFERENCES


#### FUNDING

This study was supported by the Israel Science Foundation (770/17 and 1096/17), National Institutes of Health (1R01NS091037), the Adelis Foundation and the Allen and Jewel Prince Center for Neurodegenerative Disorders of the Brain. The research was carried out in the framework of the Russell Berrie Nanotechnology Institute (RBNI) and the Lorry I. Lokey Interdisciplinary Center for Life Sciences and Engineering at the Technion. Dr. Pahari expresses his gratitude to Israel Council for Higher Education for the PBC postdoctoral fellowship.



cervical mucus. Biophys. J. 81, 1930–1937. doi: 10.1016/S0006-3495(01)75 844-4



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Amirav, Berlin, Olszakier, Pahari and Kahn. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Human Brain/Cloud Interface

Nuno R. B. Martins1,2 \*, Amara Angelica<sup>3</sup> , Krishnan Chakravarthy4,5, Yuriy Svidinenko<sup>6</sup> , Frank J. Boehm<sup>7</sup> , Ioan Opris8,9, Mikhail A. Lebedev10,11,12, Melanie Swan<sup>13</sup> , Steven A. Garan1,2, Jeffrey V. Rosenfeld14,15,16,17, Tad Hogg<sup>18</sup> and Robert A. Freitas Jr.<sup>18</sup>

<sup>1</sup> Lawrence Berkeley National Laboratory, Berkeley, CA, United States, <sup>2</sup> Center for Research and Education on Aging (CREA), University of California, Berkeley, & LBNL, Berkeley, CA, United States, <sup>3</sup> Kurzweil Technologies, Newton, MA, United States, <sup>4</sup> UC San Diego Health Science, San Diego, CA, United States, <sup>5</sup> VA San Diego Healthcare System, San Diego, CA, United States, <sup>6</sup> Nanobot Medical Animation Studio, San Diego, CA, United States, <sup>7</sup> NanoApps Medical, Inc., Vancouver, BC, Canada, <sup>8</sup> Miami Project to Cure Paralysis, University of Miami, Miami, FL, United States, <sup>9</sup> Department of Biomedical Engineering, University of Miami, Coral Gables, FL, United States, <sup>10</sup> Center for Neuroengineering, Duke University, Durham, NC, United States, <sup>11</sup> Center for Bioelectric Interfaces of the Institute for Cognitive Neuroscience of the National Research University Higher School of Economics, Moscow, Russia, <sup>12</sup> Department of Information and Internet Technologies of Digital Health Institute, I.M. Sechenov First Moscow State Medical University, Moscow, Russia, <sup>13</sup> Department of Philosophy, Purdue University, West Lafayette, IN, United States, <sup>14</sup> Monash Institute of Medical Engineering, Monash University, Clayton, VIC, Australia, <sup>15</sup> Department of Neurosurgery, Alfred Hospital, Melbourne, VIC, Australia, <sup>16</sup> Department of Surgery, Monash University, Clayton, VIC, Australia, <sup>17</sup> Department of Surgery, F. Edward Hébert School of Medicine, Uniformed Services University of the Health Sciences, Bethesda, MD, United States, <sup>18</sup> Institute for Molecular Manufacturing, Palo Alto, CA, United States

#### Edited by:

Hari S. Sharma, Uppsala University, Sweden

#### Reviewed by:

Vassiliy Tsytsarev, University of Maryland, College Park, United States Brent Winslow, Design Interactive, United States

> \*Correspondence: Nuno R. B. Martins nunomartins@lbl.gov; nunorbmartins@gmail.com

#### Specialty section:

This article was submitted to Neural Technology, a section of the journal Frontiers in Neuroscience

Received: 10 September 2018 Accepted: 30 January 2019 Published: 29 March 2019

#### Citation:

Martins NRB, Angelica A, Chakravarthy K, Svidinenko Y, Boehm FJ, Opris I, Lebedev MA, Swan M, Garan SA, Rosenfeld JV, Hogg T and Freitas RA Jr (2019) Human Brain/Cloud Interface. Front. Neurosci. 13:112. doi: 10.3389/fnins.2019.00112 The Internet comprises a decentralized global system that serves humanity's collective effort to generate, process, and store data, most of which is handled by the rapidly expanding cloud. A stable, secure, real-time system may allow for interfacing the cloud with the human brain. One promising strategy for enabling such a system, denoted here as a "human brain/cloud interface" ("B/CI"), would be based on technologies referred to here as "neuralnanorobotics." Future neuralnanorobotics technologies are anticipated to facilitate accurate diagnoses and eventual cures for the ∼400 conditions that affect the human brain. Neuralnanorobotics may also enable a B/CI with controlled connectivity between neural activity and external data storage and processing, via the direct monitoring of the brain's ∼86 × 10<sup>9</sup> neurons and ∼2 × 10<sup>14</sup> synapses. Subsequent to navigating the human vasculature, three species of neuralnanorobots (endoneurobots, gliabots, and synaptobots) could traverse the blood–brain barrier (BBB), enter the brain parenchyma, ingress into individual human brain cells, and autoposition themselves at the axon initial segments of neurons (endoneurobots), within glial cells (gliabots), and in intimate proximity to synapses (synaptobots). They would then wirelessly transmit up to ∼6 × 10<sup>16</sup> bits per second of synaptically processed and encoded human–brain electrical information via auxiliary nanorobotic fiber optics (30 cm<sup>3</sup> ) with the capacity to handle up to 10<sup>18</sup> bits/sec and provide rapid data transfer to a cloud based supercomputer for real-time brain-state monitoring and data extraction. A neuralnanorobotically enabled human B/CI might serve as a personalized conduit, allowing persons to obtain direct, instantaneous access to virtually any facet of cumulative human knowledge. Other anticipated applications include myriad opportunities to improve education, intelligence, entertainment, traveling, and other interactive experiences. A specialized application might be the capacity to engage in fully immersive experiential/sensory experiences, including what is referred to here

as "transparent shadowing" (TS). Through TS, individuals might experience episodic segments of the lives of other willing participants (locally or remote) to, hopefully, encourage and inspire improved understanding and tolerance among all members of the human family.

Keywords: brain/cloud interface, brain-computer interface, brain-to-brain interface, brain-machine interface, transparent shadowing, neuralnanorobots, neuralnanorobotics, nanomedicine

#### INTRODUCTION

"We'll have nanobots that... connect our neocortex to a synthetic neocortex in the cloud... Our thinking will be a.... biological and non-biological hybrid." — Ray Kurzweil, TED 2014

There is an incessant drive in medicine toward the development of smaller, more capable, efficacious, and costeffective devices and systems. The primary driver of this quest relates to the cellular and sub-cellular genesis of human disease, at which scale, nanodevices can directly interact and potentially positively influence disease outcomes or prevent them altogether, particularly in regard to brain disorders (Kandel et al., 2000, Kandel, 2001; Zigmond et al., 2014; Chaudhury et al., 2015; Fornito et al., 2015; Falk et al., 2016). The pursuit of ever smaller tools to treat patients is approaching a pivotal juncture in medical history as advanced nanomedicine — specifically, medical nanorobotics — is expected to serve as a dynamic tool toward addressing most human brain disorders. The goal is to finally empower medical professionals to treat diseases at individual cellular and sub-cellular resolution (Freitas, 1998, 1999b, 2003, 2005a,c, 2007, 2016; Morris, 2001; Astier et al., 2005; Patel et al., 2006; Park et al., 2007; Popov et al., 2007; Mallouk and Sen, 2009; Martel et al., 2009; Kostarelos, 2010; Mavroides and Ferreira, 2011; Boehm, 2013).

The application of nanorobots to the human brain is denoted here as "neuralnanorobotics." This technology may allow for the monitoring, recording, and even manipulation of many types of brain-related information at cellular and organellar levels (Martins et al., 2012, 2015, 2016). Medical neuralnanorobots are expected to have the capacity for realtime, non-destructive monitoring of single-neuron and singlesynapse neuroelectric activity, local neuropeptide traffic, and other relevant functional data, while also allowing the acquisition of fundamental structural information from neuron surfaces, to enhance the connectome map of a living human brain (Sporns et al., 2005; Lu et al., 2009; Anderson et al., 2011; Kleinfeld et al., 2011; Seung, 2011; Martins et al., 2012, 2015, 2016). Non-destructive neuralnanorobotically mediated whole-brain monitoring coupled with single-cell repair capabilities (Freitas, 2007) is anticipated to provide a powerful medical capability to effectively treat most, or all of the ∼400 known brain disorders, including, most notably: Parkinson's and Alzheimer's (Freitas, 2016), addiction, dementia, epilepsy, and spinal cord disorders (NINDS, 2017).

Neuralnanorobots are also expected to empower many nonmedical paradigm-shifting applications, including significant human cognitive enhancement, by providing a platform for direct access to supercomputing storage and processing capabilities and interfacing with artificial intelligence systems. Since informationbased technologies are consistently improving their priceperformance ratios and functional design at an exponential rate, it is likely that once they enter clinical practice or non-medical applications, neuralnanorobotic technologies may work in parallel with powerful artificial intelligence systems, supercomputing, and advanced molecular manufacturing.

Furthermore, autonomous nanomedical devices are expected to be biocompatible, primarily due to their structural materials, which would enable extended residency within the human body (Freitas, 1999a, 2002, 2003). Medical neuralnanorobots might also be fabricated in sufficient therapeutic quantities to treat individual patients, using diamondoid materials, as these materials may provide the greatest strength, resilience, and reliability in vivo (Freitas, 2010). An ongoing international "Nanofactory Collaboration" headed by Robert Freitas and Ralph Merkle has the primary objective of constructing the world's first nanofactory, which will permit the mass manufacture of advanced autonomous diamondoid neuralnanorobots for both medical and non-medical applications (Freitas and Merkle, 2004, 2006; Freitas, 2009, 2010).

It is conceivable that within the next 20–30 years, neuralnanorobotics may be developed to enable a safe, secure, instantaneous, real-time interface between the human brain and biological and non-biological computing systems, empowering brain-to-brain interfaces (BTBI), braincomputer interfaces (BCI), and, in particular, sophisticated brain/cloud interfaces (B/CI). Such human B/CI systems may dramatically alter human/machine communications, carrying the promise of significant human cognitive enhancement (Kurzweil, 2014; Swan, 2016).

Historically, a fundamental breakthrough toward the possibility of a B/CI was the initial measurement and recording of the electrical activity of the brain via EEG in 1924 (Stone and Hughes, 2013). At the time, EEG marked a historical advance in neurologic and psychiatric diagnostic tools, as this technology allowed for the measurement of a variety of cerebral diseases, the quantification of deviations induced by different mental states, and detection of oscillatory alpha waves (8–13 Hz), the so-called "Berger's wave." The first EEG measurements required

**Abbreviations:** AIS, Axon initial segment; B/CI, brain/cloud interface; BCI, brain–computer interface; BMI, brain–machine interface; BTBI, brain-tobrain interface; EEG, electroencephalography; fMRI, functional magnetic resonance imaging; FNIRS, functional near-infrared spectroscopy; TS, transparent shadowing.

the insertion of silver wires into the scalps of patients, which later evolved to silver foils that were adhered to the head. These rudimentary sensors were initially linked to a Lippmann capillary electrometer. However, significantly improved results were achieved through the use of a Siemens double-coil recording galvanometer, which had an electronic resolution of 0.1 mv (Jung and Berger, 1979).

The first reported scientific instance of the term "brain– computer interface" dates to 1973, ∼50 years following the first EEG recording, when it was envisioned that EEG-reported brain electrical signals might be employed as data carriers in human–computer communications. This suggestion assumed that mental decisions and reactions might be probed by electroencephalographic potential fluctuations measured on the human scalp, and that meaningful EEG phenomena should be viewed as a complex structure of elementary wavelets that reflected individual cortical events (Vidal, 1973).

Currently, invasive<sup>1</sup> and non-invasive brain–computer interfaces and non-invasive brain-to-brain communication systems have already been experimentally demonstrated and are the subject of serious research worldwide. Once these existing technologies have matured, they might provide treatments for completely paralyzed patients, eventually permitting the restoration of movement in paralyzed limbs through the transmission of brain signals to muscles or external prosthetic devices (Birbaumer, 2006). The first reported direct transmission of information between two human brains without intervention of motor or peripheral sensory systems occurred in 2014, using a brain-to-brain communication technique referred to as "hyperinteraction" (Grau et al., 2014).

The most promising long-term future technology for nondestructive, real-time human–brain–computer interfaces and brain-to-brain communications may be neuralnanorobotics (Martins et al., 2016). Neuralnanorobotics, which is the application of medical nanorobots to the human brain, was first envisaged by Freitas, who proposed the use of nanorobots for direct real-time monitoring of neural traffic from in vivo neurons, as well as the translation of messages to neurons (Freitas, 1999b, 2003). Other authors have also envisioned B/CI and predicted that in the future, humans will have access to a synthetic non-biological neocortex, which might permit a direct B/CI. Within the next few decades, neuralnanorobotics may enable a non-destructive, real-time, ultrahigh-resolution interface between the human brain and external computing platforms such as the "cloud."

The term "cloud" refers to cloud computing, an information technology (IT) paradigm and a model for enabling ubiquitous access to shared pools of configurable resources (such as computer networks, servers, storage, applications, and services), that can be rapidly provisioned with minimal management effort, often over the Internet. For both personal or business applications, the cloud facilitates rapid data access, provides redundancy, and optimizes the global usage of processing and storage resources while enabling access from virtually any location on the planet. However, the primary challenge for worldwide global cloud-based information processing technologies is the speed of access to the system, or latency. For example, the current round-trip latency rate for transatlantic loops between New York and London is ∼90 ms (Verizon, 2014). Since there are now more than 4 billion Internet users worldwide, its economic impact on the global economy is increasingly significant. The economic impact of IoT (Internet of Things) applications alone has been estimated by the McKinsey Global Institute to range from \$3.9 to \$11.1 trillion per year by 2025. The global economic impact of cloud-based information processing over the next few decades may be at least an order of magnitude higher once cloud services are combined in previously unimagined ways, disrupting entire industries (Miraz et al., 2015). A neuralnanorobotics-mediated human B/CI, potentially available within 20–30 years, will require broadband Internet access with extremely high upload and download speeds, compared to today's rates.

Humankind has at its core a potent and ceaseless drive to explore and to challenge itself, to improve its collective condition by relentlessly probing and pushing boundaries while constantly attempting to breach those barriers that tenuously separate the possible from the impossible. The notions of human augmentation and cognitive enhancement are borne of these tenets.

This drive includes an incessant quest for exploration and a constant desire for social interaction and communication both of which are catalysts for rapidly increasing globalization. Consequently, the development of a non-destructive, real-time human B/CI technology may serve as an intimate, personalized conduit through which individuals would have instantaneous access to virtually any facet of cumulative human knowledge and also the optional specialized capacity to engage in myriad real-time fully immersive experiential and sensory worlds.

## THE HUMAN BRAIN

## The Quantitative Human Brain

The human brain comprises a remarkable information storage and processing system that possesses an extraordinary computation-per-volume efficiency, with an average weight of 1400 g and a volume of ∼1350 cm<sup>3</sup> , contained within an "average" intracranial volume of ∼1,700 cm<sup>3</sup> . A brief quantification of the brain's constituents and operational parameters includes ∼1,350 cm<sup>3</sup> (∼75%) brain cells, ∼200 cm<sup>3</sup> (15%) blood, and up to ∼150 cm<sup>3</sup> (10%) of cerebrospinal fluid (Rengachary and Ellenbogen, 2005). The raw computational power of the human brain has been estimated to range from 10<sup>13</sup> to 10<sup>16</sup> operations/sec (Merkle, 1989). The human brain's functional action potential based information is estimated as 5.52 × 10<sup>16</sup> bits/sec (Martins et al., 2012), with a brain power output estimated at 15–25 W and a power density of

<sup>1</sup>For the purposes of this paper, the term "invasive" is defined as a medical procedure or device that imparts quantifiable physiological damage (at any level) to a patient. In the case of the envisaged nanomedically enabled B/CI, the assumption is that millions of micron-scale nanorobots will be non-invasive — i.e., they may ingress into a patient and subsequently auto-situate themselves at various sites within the human brain, with no disruptive functional physiological or experiential effects. Or in some cases, they may be minimally invasive.

1.1–1.8 × 10<sup>4</sup> W/m<sup>3</sup> at an operating temperature of 37.3◦C (Freitas, 1999b).

When considering the human brain at the regional level, an exceptional component is the neocortex (**Tables 1**, **2**), which has a highly organized neural architecture that encompasses sensorimotor, cognitive, and emotional domains (Alexander et al., 1986; Fuster and Bressler, 2012). This cortical structure consists of mini-columnar and laminar arrangements of neurons that are linked via afferent and efferent connections distributed across multiple brain regions (Lorento de Nó, 1938; Mountcastle, 1997; Shepherd and Grillner, 2010; Opris, 2013; Opris et al., 2011, 2013, 2014, 2015). Cortical minicolumns consist of chains of pyramidal neurons that are surrounded by a "curtain of inhibition" formed by interneurons (Szentágothai and Arbib, 1975).

At the cellular level, the average human brain is estimated to contain (86.06 ± 8.2) × 10<sup>9</sup> neurons, with ∼80.2% (69.03 ± 6.65 × 10<sup>9</sup> neurons) located in the cerebellum, ∼19% (16.34 ± 2.17 × 10<sup>9</sup> neurons) located in the cerebral cortex, and only ∼0.8% (0.69 ± 0.12 × 10<sup>9</sup> neurons) located throughout the rest of the brain (Azevedo et al., 2009). The human cerebellum and cerebral cortex together hold the vast majority (99.2%) of brain neurons (Azevedo et al., 2009). Another approximation, based on combining estimates for the different brain regions, produced a similar value of 94.2 ± 11.3 × 10<sup>9</sup> neurons for the whole human brain (Martins et al., 2012).

Glial cells comprise another brain-cell type (**Figure 1**). The average number of glial cells in the human brain is estimated to be 84.61 ± 9.83 × 10<sup>9</sup> (Herculano-Houzel, 2009), with the population of glial cells in the neocortex estimated at from 18.2 to 38.6 × 10<sup>9</sup> (Karlsen and Pakkenberg, 2011). The ratio of glia to neurons likely has functional relevance (Nedergaard et al., 2003) and varies between different brain regions. While the whole-brain glia/neuron ratio is ∼1:1, there are significant differences between

TABLE 1 | Neocortical measures (Pakkenberg and Gundersen, 1997; Stark et al., 2007a,b).


TABLE 2 | Enumeration of neurons and synapses in the human neocortex (Tang et al., 2001; Sandberg and Bostrom, 2008; Karlsen and Pakkenberg, 2011).


FIGURE 1 | Artistic representation of neurons (with blue processes) and glial (white) cells. [Image credit: Yuriy Svidinenko, Nanobotmodels Company].

brain domains. For example, the glia/neuron ratio of the cerebral cortex is 3.72:1 (60.84 billion glia; 16.34 billion neurons) but only 0.23:1 (16.04 billion glia; 69.03 billion neurons) in the cerebellum; the basal ganglia, diencephalon, and brainstem have a combined ratio of 11.35:1 (Azevedo et al., 2009).

In addition, synapses, numbering (2.42 ± 0.29) × 10<sup>14</sup> in the average human brain, are collectively estimated to process information at spiking rates of (4.31 ± 0.86) × 10<sup>15</sup> spikes/sec, empowering the human brain to process data at (5.52 ± 1.13) × 10<sup>16</sup> bits/sec (Martins et al., 2012). Synapses are elements of the neural network that play a critical role in processing information in the brain, being involved in learning, long-term and short-term memory storage and deletion, and temporal information processing (Black et al., 1990; Bliss and Collingridge, 1993; Kandel, 2001; Fuhrmann et al., 2002; Lee et al., 2008; Holtmaat and Svoboda, 2009; Liu et al., 2012). Synapses are also key effectors for signal transduction and plasticity in the brain. Proper synapse formation during childhood provides a substrate for cognition, whereas improper formation or functionality leads to neuro-developmental disorders including mental retardation and autism (Rollenhagen and Lübke, 2006; Mcallister, 2007; Rollenhagen et al., 2007). Synapse loss, as occurs in Alzheimer's patients, is intimately associated with cognitive decline (Dekosky and Scheff, 1990; Terry et al., 1991; Scheff and Price, 2006).

#### Processing Units

Structural cellular or sub-cellular elements of the human brain are considered as information processing units if they are involved in significant functional input/output changes in electrochemically based brain-data storage and/or processing systems.

There is some disagreement in the current scientific literature regarding the quantification of this "significance" metric. This incongruity has led various authors to consider different cellular and subcellular structures as fundamental elements of human brain storage and its computation system, encompassing (aside from neurons and synapses): dendritic trees, axons, proteins, and even neural microtubules (Koch et al., 1983; Bialek, 1993;

Juusola et al., 1996; Zador, 1998; Manwani and Koch, 2001; London and Häusser, 2005; Ford, 2010).

Estimates for whole-brain electrical data processing rates range from 1.48 × 10<sup>11</sup> bits/sec. to a high of 3.2 × 10<sup>29</sup> bits/sec (Sandberg and Bostrom, 2008; Martins et al., 2012). The human brain might even have more than 100 times higher computational capacity than previously thought, based on the discovery that dendrites may generate nearly 10 times as many electrochemical spikes as do neuron soma, and are hybrids that process both analog and digital signals (Moore et al., 2017). This finding may challenge the long-held belief that spikes in the soma (body of the neuron) are the primary means through which perception, learning, and memory formation occur. Dendrites comprise more than 90% of neural tissue, so knowing that they are much more active than the soma would fundamentally alter our understanding of how the brain processes information. As dendrites are ∼100 times larger by volume than neuronal bodies, the immense number of firing dendritic spikes would suggest that the brain may indeed possess significantly higher computational power than earlier estimated.

However, there is currently a consensus that neurons and synapses constitute the fundamental electrochemical processing units of the human brain (Gkoupidenis et al., 2017; Jackman and Regehr, 2017).

The roles of neurons in electrical information processing include receiving, integrating, generating, and transmitting action-potential-based information (Koch, 1997; Koch and Segev, 2000; Zhang, 2008). However, several neuronal noise sources influence the reliability and precision of neuronal signaling, so stimulus-response functions are sometimes unreliable and are dissociated from what is being encoded via spike activity (Bialek and Rieke, 1992).

The other fundamental consensual processing units of electrochemical information are synapses. Synapses are a core component of the neuron network that process information and are involved in learning and memory, with synapse dimensions and morphologies reported as playing a fundamental role in long- and short-term memory storage and deletion. Synapses are also engaged in signal transduction and plasticity, ensuring one-way transmission of signals, and are involved in temporal information processing to allow complex system behaviors, along with acting to decelerate electrical signals (Puro et al., 1977; Black et al., 1990; Bliss and Collingridge, 1993; Kandel, 2001; Rollenhagen and Lübke, 2006; Rollenhagen et al., 2007; IBM, 2008; Lee et al., 2008; Holtmaat and Svoboda, 2009). The role of synapses as processing units of the human brain is reinforced by the results of computational simulation, which indicate that the computational power of a network is increased using dynamic synapses. This suggests that emulation of biological synapses is a prerequisite for the development of brain-like computational systems (Maass and Zador, 1999; Fuhrmann et al., 2002; Kuzum et al., 2012). A recently developed ultra-low-power artificial synapse for neural computing has demonstrated the capacity to provide 500 distinct states (Van de Burgt et al., 2017).

Real-time monitoring of the whole human brain (by placing neuralnanorobots within each neuron and nearby synaptic connections to record/transmit data from localized neuron and synapse spiking) may provide redundant data that might be employed in the development of validation protocols.

## THE CLOUD

Due to the immense volume of data involved, data transfer to and from living human brains and the cloud may likely require the use of supercomputers with artificial intelligence algorithms. Current von Neumann-based-architecture supercomputers with massive numbers of processors are either centralized (composed of large numbers of dedicated processors) or distributed (based on a large number of discrete computers distributed across a network, such as the Internet).

One estimate of maximum computational speed required to handle the electrical data in the human brain is 5.52 × 10<sup>16</sup> bits/sec (Martins et al., 2012). Several centralized and distributed supercomputers have processing speeds that are significantly higher than this estimate (Martins et al., 2012). As of November 2018, the fastest supercomputer worldwide was Summit, developed at the United States Oak Ridge National Laboratory (Tennessee), with 122.3 petaflops on the High Performance Linpack (HPL) benchmark. This computational model may be questionable, however, as computers are based on von Neumann architecture, whereas brain circuits are not; and brains operate in a massively parallel manner, whereas computers do not (Nagarajan and Stevens, 2008; Whitworth and Ryu, 2008).

The Internet consists of a decentralized global system, based on von-Neumann-architecture-based computers and supercomputers, used for data transfer across processing and storage units. The global storage capacity of Internet data centers in 2018 was 1450 exabytes (Statistica, 2018). Van den Bosch et al. (2016) estimate that the storage capacity of the World Wide Web doubles every 3 years, with its computational capacity doubling every 1.5 years.

However, once brain data is interfaced with supercomputers in near real-time, the connection to supercomputers in the cloud will be the ultimate bottleneck between the cloud and the human brain (Knapp, 2013). This challenge includes, in particular, the bottleneck of the bandwidth required to transmit data worldwide. According to one study, "Global Internet traffic in 2021 will be equivalent to 127 times the volume of the entire global Internet in 2005. Globally, Internet traffic will reach 30 GB per capita by 2021, up from 10 GB per capita in 2016" (Cisco, 2017). This speed is forcing innovation to deal with bandwidth constraints. Conventional fiber-optic cables transfer trillions of bits/sec between massive data centers. As of October 2018, the average Internet peak connection speed was 189.33 Mbps in Singapore and 100.07 Mbps in the United States (Kemp, 2018). Several commercial efforts to increase Internet speeds are presently underway, including the recently built \$300 million fiber-optic cable between Oregon, Japan, and Taiwan. In 2016, much of the world's Internet traffic

was transmitted via undersea fiber-optic cables; the 6,600 kmlong MAREA Facebook/Microsoft-owned cable was estimated to carry 160 Tb/sec of data across the Atlantic Ocean (Hecht, 2016). Current commercial 4G networks provide broadband speeds of up to 100 Mbits/sec. However, United States carriers have stated that they plan to deploy 5G technology in 2020 that will eventually "bring speeds of around 10 gigabits per second to your phone. That's more than 600 times faster than typical 4G speeds of today's mobile phones, and 10 times faster than Google Fiber's standard home broadband service" (Finley, 2018).

## POTENTIAL OF CURRENT TECHNOLOGIES TOWARD A BRAIN/CLOUD INTERFACE

#### Nanoparticles, Nanotubes, and Nanodots

One promising near-term technology that may enable an interface with brain-based neural networks is magnetoelectric nanoparticles, which may be employed to enhance coupling between external magnetic fields and localized electric fields that emanate from neural networks (Yue et al., 2012; Guduru et al., 2015). Magnetoelectric nanoparticles might also induce nanoparticles to traverse the blood–brain barrier (BBB) by applying a direct-current magnetic field gradient to the cranial vault. Magnetoelectric nanoparticles have already been utilized to control intrinsic fields deep within the mouse brain and have permitted the coupling of external magnetic fields to neuronal electric fields. A strategy developed for the delivery of nanoparticles to the perineuronal environment is expected to provide a means to access and eventually stimulate selected populations of neurons (Freitas, 1999b).

The delivery of nanoparticles into the human brain will indeed pose a formidable challenge. For intravenous injection, at least 90% of nanoparticles have been observed to be sequestered within tissues and organs prior to reaching the brain (Calvo et al., 2001), so intra-arterial injections might be more reliable. Steering nanoparticles to selected brain regions may also be achieved using external magnetic fields (Li et al., 2018). Since it has been shown that certain customized nanoparticles may damage dopaminergic and serotoninergic systems, a further detailed analysis of the biodistribution and metabolism of nanoparticles will be required. Further, the risk of infection, inflammatory reactions, potential immunogenicity, cytotoxicity, and tumorigenicity must be effectively addressed prior to the in vivo application of nanoparticles in humans (Cupaioli et al., 2014).

The use of carbon-nanotube-based electrical stimulation of targets deep within the brain has been proposed as a novel treatment modality for patients with Parkinson's disease and other CNS disorders (Srikanth and Kessler, 2012). This strategy utilizes unidirectional electrical stimulation, which is more precise and avoids the surgical risks associated with deep macroelectrode insertion, used with current methods of deep brain stimulation (Mayberg et al., 2005; Taghva et al., 2013) that employ long stereotactically placed quadripolar macroelectrodes through the skull. When intended for use as a component of a B/CI system, carbon-nanotube-based electrical stimulation would also require a two-way information pathway at single-neuron resolution for neuronal electrochemical information recording.

Fluorescing carbon nanodots (synthesized using D-glucose and L-aspartic acid) with uniform diameters of 2.28 ± 0.42 nm have been employed to target and image C6 glioma cells in mouse brains. Excellent biocompatibility, tunable fullcolor emission, and the capacity to freely penetrate the BBB might make fluorescing carbon nanodots viable candidates as tagging agents to facilitate the implementation of nanomedical B/CI technologies (Zheng et al., 2015). However, fluorescing carbon nanodots might be problematic, since crossing the BBB is a challenging process for ∼98% of all small molecules (Pardridge, 2005; Grabrucker et al., 2016). This is primarily due to the BBB forming a dynamic, blood-and-brain-regulated, strict physical, transport, metabolic, and immunologic barrier while it is permeable to O<sup>2</sup> and CO<sup>2</sup> and other gaseous molecules, as well as water and other lipid soluble substances (Serlin et al., 2015), the barrier is very restrictive to large molecules. However, small peptides may cross the BBB by either non-specific fluid-phase endocytosis or receptor-mediated transcytosis (RMT) mechanisms.

Optically based nanotechnologies, including optical imaging methods, have demonstrated valuable applications at the cellular level. For example, quantum dot fullerenes have been employed for in vitro and in vivo cellular membrane potential measurements (Nag et al., 2017).

### Injectable "Neural Lace"

A recently proposed technology for the potential integration of brain neural networks and computing systems at the microscale is referred to as "neural lace." This would introduce minimally invasive three-dimensional mesh nanoelectronics, via syringe-injection, into living brain tissue to allow for continuous monitoring and stimulation of individual neurons and neuronal networks. This concept is based on ultraflexible mesh nanoelectronics that permit interfaces with non-planar topographies. Experimental results have been reported using the injection and unfolding of sub-micrometer-thick, centimeterscale macroporous mesh nanoelectronics through needles with diameters as small as 100 µm, which were injected into cavities with a >90% device yield (Liu et al., 2015). One of the other potential applications of syringe-injectable mesh nanoelectronics is in vivo multiplexed neural network recording.

Plug-and-play input/output neural interfacing has also been achieved using platinum electrodes and silicon nanowire fieldeffect transistors, which exhibited a low interface contact resistance of ∼3 (Schuhmann et al., 2017). Dai et al. (2018) also demonstrated "stable integration of mesh nanoelectronics within brain tissue on at least 1 year scales without evidence of chronic immune response or the glial scarring characteristic of conventional implants." This group also showed that the activities of individual neurons and localized neural circuits could be monitored and stimulated over timelines of eight months or more, for applications such as recording of alterations in the activities of specific neurons as the brain ages (Dai et al., 2018).

## Neural Dust

fnins-13-00112 March 29, 2019 Time: 17:42 # 7

Future human B/CI technologies may preferably require longterm, self-implanting in vivo neural interface systems, a characteristic that is absent from most current BMI technologies. This means that the system design should balance the size, power, and bandwidth parameters of neural recording systems. A recent proposal capable of bidirectional communication explored the use of low-power CMOS circuitry coupled with ultrasonic delivery of power and backscatter communications to monitor localized groups of neurons (Seo et al., 2013). The goal was to enable scalability in the number of neural recordings from the brain, while providing a path toward a longer-duration BMI. This technology currently employs thousands of independent free-floating 10–100 µm scale sensor nodes referred to as "neural dust." These nodes detect and report local extracellular electrophysiological data, while using a subcranial interrogator that establishes power and communications links with each of the neural dust elements. Power transmission is accomplished ultrasonically to enable low-efficiency (7%, 11.6 dB) links, yielding ∼500 µW of received power (>10<sup>7</sup> higher than the ∼40 pW EM transmission available at a similar-size scale) with a 1 mm<sup>2</sup> interrogator, which may eventually provide ∼10 µm sensing nodes.

### Brain–Machine Interface (BMI)

Brain–machine interface technology is currently being pursued via invasive neural interfaces composed of neural microchip sensor arrays that contain a plurality of electrodes that can detect multicellular signals. These are available for several brain areas (e.g., visual cortex, motor cortex neuroprosthetics, hippocampus, and others) (Berger et al., 2005; BrainGate, 2009).

There are currently two different types of BMI systems. One type samples the neural activity of a single brain and unidirectionally controls an external device (Lebedev, 2014), while the other type (sensory BMI) includes sensory feedback from the device to the brain (O'Doherty et al., 2011). Noninvasive neural BMI interface strategies include the use of EEG, magnetoencephalography (MEG), fMRI (Miyawaki et al., 2008) and optical strategies, including fNIRS (Naseer and Hong, 2015). One 8-channel EEG signal-capture platform, built around Texas Instruments' ADS1299 analog front-end integrated circuit, may soon be printable at home, thus democratizing low-resolution brain-data-extraction technologies (OpenBCI, 2019).

Neurophotonics integrated with prosthetics, which links artificial limbs and peripheral nerves using two-way fiber-optic communications to enable the ability to feel pressure or temperature, is expected to permit high-speed communications between the brain and artificial limbs. Neuralnanorobots are anticipated to optimize interfaces using advanced touch-sensitive limbs that convey real-time sensory information to amputees, via a direct interface with the brain (Tabot et al., 2013).

At the cellular level, attempts to achieve a direct junction between individual nerve cells and silicon microstructures are being pursued. Neuron-silicon junctions were spontaneously formed using the nerve cells of a mammalian brain, which permitted direct stimulation of nerve cells (Fromherz and Stett, 1995; Offenhausser, 1996; Vassanelli and Fromherz, 1997; Schätzthauer and Fromherz, 1998). Currently, nanoelectronics devices utilizing carbon nanotubes and silicon nanowires can detect and identify neuronal biomolecular chemical secretions and their bioelectrical activities (Veliev, 2016). An array of nanowire transistors can detect, stimulate, or inhibit nerve impulses and their propagation along individual neurites (Freitas, 1999b; Zeck and Fromherz, 2001; Patolsky et al., 2006). To demonstrate experimental minimally invasive neuron cytosolic recording of action potentials, a nanotransistor device was placed at the tip of a bent silicon nanowire to intracellularly record action potentials (Tian et al., 2010; Duan et al., 2011). Vertically arranged gold nanowire arrays have been used to stimulate and detect electrical activity at the nanoscale from simultaneous locations within neurons (Saha et al., 2008). High-density arrays of nanowire FETs enabled mapping signals at the subcellular level – a functionality that is not possible with conventional microfabricated devices (Timko et al., 2010).

In principle, neuralnanorobotics may empower a nearoptimal BCI with long-term biocompatibility by incorporating silicon, platinum, iridium, polyesterimide-insulated gold wires, peptide-coated glassy carbon pins, carbon nanotubes, polymerbased electrodes, silicon nitride, silicon dioxide, stainless steel, or nichrome (Niparko et al., 1989a,b; Edell et al., 1992; Yuen and Agnew, 1995; Huber et al., 1998; Malmstrom et al., 1998; Decharms et al., 1999; Normann et al., 1999;Mattson et al., 2000; Kristensen et al., 2001; Parak et al., 2001; Freitas, 2003). Neural electrodes can be implanted without producing any detectable damage beyond the initial trauma and brief phagocytosis, which are typically limited to the edges of the electrode insertion pathway (Babb and Kupfer, 1984) (Freitas, 2003). Several types of neural electrodes are presently employed to interface with the brain via cochlear implants at scala tympani electrode arrays, and in potential CNS auditory prostheses, retinal chip implants, semiconductor-based microphotodiode arrays placed in the subretinal space, visual cortex microelectrode arrays, and other neural implants intended for the mobilization of paraplegics, phrenic pacing, or cardiac assistance (Haggerty and Lusted, 1989; Niparko et al., 1989a,b; Lefurge et al., 1991; Burton et al., 1996; Heiduschka and Thanos, 1998; Guenther et al., 1999; Normann et al., 1999; Peachey and Chow, 1999; Kohler et al., 2001; Mayr et al., 2001; Pardue et al., 2001; Shoham et al., 2001; Freitas, 2003; Mannoor et al., 2013). Each of these electrodes interface with very diminutive and specific brain regions, and are always confined to the surface areas of highly localized domains.

Early "neural dust" proposals for providing BCI access to specific human–brain regions (e.g., neocortex) had several inherent limitations (Seo et al., 2013). Conversely, neuralnanorobotics technologies may possess the appropriate scale for optimally enabling BCI, exhibiting suitable mobility, being minimally invasive, imparting negligible localized tissue damage, and possessing robust monitoring capabilities over distinct information channels without requiring conventional surgical implantation.

Neuralnanorobotics may also be massively distributed, whereas surgically introduced neural implants must be positioned in one or several specific locations. These shortcomings suggest that neuralnanorobotics may be a preferred solution to the formidable challenges ahead in the development of B/CI technologies.

#### Brain-To-Brain Interface

fnins-13-00112 March 29, 2019 Time: 17:42 # 8

A BTBI involves inducing two distinct brains to directly communicate with each other (Pais-Vieira et al., 2015). BTBI systems were initially implemented in humans (**Figure 2**) using non-invasive recordings and brain stimulation. Information was transferred from the sensorimotor cortex of one participant (recorded via EEG) to the visual (Grau et al., 2014) or motor (Rao et al., 2014) cortex of the second participant (delivered via transcranial magnetic stimulation, or TMS).

A number of BTBI's involving different species have also been recently demonstrated, for example, by linking the brain of a human to the spinal cord of an anesthetized rat (Yoo et al., 2013). In another example of interspecies BTBI, a human brain guided the movements of a Madagascar hissing cockroach along an S-shape track, controlling the cockroach antennae via electrical stimulation (Li and Zhang, 2016). Human brains have also been connected to cell cultures, experimentally demonstrating that brain activity can control gene expression, using an EEG-based BMI to trigger optogenetic stimulation of designer cells, thereby mediating their genetic expression (Folcher et al., 2014).

## Brainet Systems

A particularly intriguing application of BTBI technologies, termed "Brainets," involve the interfacing and processing of neuronal signals recorded from multiple brains, to enable information exchange between interconnected brains (Pais-Vieira et al., 2015) in order to perform cooperative tasks (Ramakrishnan et al., 2015). While not yet particularly sophisticated, recently demonstrated Brainet systems have already provided several interesting insights, including verification of potential direct communications between the brains of two rats located on different continents, after the rats had been permanently implanted with microelectrodes in the sensorimotor cortex (Pais-Vieira et al., 2013).

Experiments have tested three different control systems using 2–3 implanted monkeys that shared BMI-mediated control of a virtual arm (Ramakrishnan et al., 2015). The first type of sharedcontrol, using two subjects, merged recorded neural signals to move a virtual arm on a computer screen. The extracted brain data were summed and observed to improve performance, using noise cancelation. Another system involved two monkeys with partitioned contributions. The first monkey controlled the X-coordinate of the virtual arm, whereas the second monkey controlled the Y-coordinate. The overall task performance was shown to be improved as each monkey made fewer errors. (Interestingly, each monkey brain adapted and responded less to the other coordinate). A third experiment involved three animals, which together operated and controlled the virtual arm in three dimensions. As the monkeys were unaware that their final task was three-dimensional (given that each monkey had a two-dimensional display) this Brainet might be considered as a rudimentary "super-brain," where the contributions of individual participants gave rise to higher-order operations that were not performable by each individual alone. Several cooperative BMI schemes have also been implemented in humans — for example, cooperative navigation of a spacecraft (Poli et al., 2013), cooperatively enabled decision making (Eckstein et al., 2012; Yuan et al., 2013; Poli et al., 2014), and movement planning (Wang and Jung, 2011).

A four-brain Brainet system was dubbed an "organic computer" for mimicking simple computer-like operations, such as information-input retention, in a memory-like buffer composed of four serially connected rat brains (Pais-Vieira et al., 2015). This experimental Brainet system always outperformed single-brain computation performance, particularly for discrimination tasks, in which the four brains "voted" to generate the response. This comprised an interesting advance toward the potential eventual emergence of very complex operations in systems with massive numbers of Brainet participants.

A three-human BTBI system, called "BrainNet," has been recently developed, which allowed three human subjects to collaboratively solve a task using non-invasive, direct brain-tobrain communication (Jiang et al., 2018). Similar to the twohuman BTBI system, the three-human BTBI system interface used EEG to record brain signals from the "Senders" and TMS to non-invasively deliver information to the brain of the "Receiver." The two Senders' brain signals were decoded using real-time EEG data analysis, extracting their decisions to rotate, or not rotate, a block in a Tetris-like game. These decisions were then uploaded to the cloud and subsequently downloaded and applied to the Receiver's brain via magnetic stimulation of the occipital cortex. Once this information was received, the Receiver, who could not see the game screen, integrated the information and decided to rotate, or not rotate, the block. The experiment was repeated with five groups with an average accuracy of 0.813. Such high reliability supports further research to improve multi-person BTBI systems that empower future cooperative multi-human problem solving.

Based on current elementary Brainet implementations, it is not yet clear if more complex Brainet systems might be employed for high-throughput information transfer between individual brains, although improved Brainet performance is expected with more advanced Brainet operations. With further progress in the field, the number of information transfer channels may increase, along with the number of subjects involved in each Brainet system. Clinically relevant Brainets that connect patients with therapists, or healthy to unhealthy individuals, would be a particularly interesting application.

## Limited Prospects for Current Techniques

Current technological trajectories appear to be converging toward the creation of systems that will have the capacity to empower a human B/CI. However, since the human brain possesses cellular (neuron) and sub-cellular (synapse) processing elements, any technology that is capable of establishing a

long-term and non-destructive, real-time human interface with the cloud must embody the following capabilities: (1) ultrahigh-resolution mobility, (2) autonomous or semiautonomous activity, (3) non-intrusive (ideally, physiologically imperceptible) ingress/egress into/from the human body, and (4) supplying sufficient and robust information transfer bandwidth for interfacing with external supercomputing systems. Current techniques, whether in present-day or extrapolated future forms, appear to be unscalable and incapable of fulfilling all of the temporal or spatial resolution requirements necessary for a properly comprehensive fully functional human B/CI.

## NEURALNANOROBOTIC BRAIN/CLOUD INTERFACE

Neuralnanorobotics is expected to provide a non-destructive, real-time, secure, long-term, and virtually autonomous in vivo system that can realize the first functional human B/CI (Martins et al., 2012, 2015, 2016). Neuralnanorobots could monitor relevant functional and structural connectome data, functionalaction-potential-based electrical information processing that occurs within synapses and neurons, and synaptic and neuronal structural changes associated with processing such electrolyticbased functional data (Seung, 2011). Monitoring the intracellular structural and functional connectome may be enabled by three classes of neuralnanorobots, introduced here as endoneurobots, synaptobots, and gliabots (Martins et al., 2016). They also constitute a non-intrusive, self-installed in vivo accessory highspeed nanofiber-optic network, which has been described elsewhere (Freitas, 1999b).

More specifically, endoneurobots are autonomous neuronresident neuralnanorobots that interface with all ∼86 × 10<sup>9</sup> human–brain neurons at the AIS to directly monitor and interact with action-potential-based electrically processed information. Synaptobots are autonomous neuron-resident neuralnanorobots that might employ multiple flexible stalk-mounted nanosensors to interface with each of the ∼2 × 10<sup>14</sup> synapses of the human brain to directly monitor and interact with synaptically processed and stored information. Gliabots are glia-resident autonomous neuralnanorobots that are endowed with the capacity to monitor human–brain glial cells and may further serve as supportive infrastructure elements of the system. Subsequent iterations of an initial high-speed nanofiber-optic network may also incorporate wireless transmitters (self-embedded at the periphery of the human brain or within the skull) configured as an evenly distributed network that can wirelessly enable an interface with neurons, axons, and synapses to receive/transmit data from/to the cloud.

To achieve a safe, reliable, high-performance B/CI system, a critical mission requirement is the initial establishment of intimate and stable connections to monitor the electrical firing patterns and waveforms of the ∼86 × 10<sup>9</sup> neurons and the ∼2 × 10<sup>14</sup> synapses of the human brain at a suitable repetition rate (400–800 Hz is the reported average maximum range) (Wilson, 1999; Contreras, 2004). Neuralnanorobots themselves, and/or other dedicated nanomedical mapping devices, such as an envisaged Vascular Cartographic Scanning Nanodevice (VCSN) (Domschke and Boehm, 2017) might initially generate an ultrahigh-resolution connectome map of the human brain. This would permit the acquisition and storage of detailed structural and functional connectomic data for each unique individual brain

and allow for reporting specific spatial coordinates of different classes of neurons, as well as their typical electrophysiological spiking pattern behaviors (i.e., regular-spiking, bursting, or fastspiking) (Seung, 2011).

For the purposes of a B/CI, interfacing with neuronal and synaptically processed action-potential-based electrical brain activity alone (without monitoring chemically based information) may be sufficient to facilitate robust human B/CI systems. For example, one recent study has found that quantum dots can function as voltage-sensitive probes for real-time visualization of cellular membrane potential in neurons (Nag et al., 2017). Optical interrogation of individual cells and organelles with a spatial resolution of ∼100 nm might be enabled through the use of carbonnanotube-based endoscopes that project from B/CI nanorobots (Singhal et al., 2011).

Here, synaptically processed action-potential-based information is regarded as fundamental information (Fuhrmann et al., 2002; Shepherd, 2003; Abbott and Regehr, 2004). Synaptobots would detect virtually all of the synaptically processed action potentials and their waveforms and report synaptically processed spikes into the data handling system. Consequently, neuralnanorobots would assist with the prediction of neurotransmitter bursts that traverse each synaptic gap. All these data would be continually processed at sub-millisecond resolution, enabling a virtually real-time data stream between the human brain and the cloud.

#### Endoneurobots and Gliabots

Neuralnanorobots might be transdermally injected, after which they would navigate the vasculature and anchor to the endothelial cells of the BBB. A 10 µm<sup>3</sup> volume of endoneurobots (**Figure 3**) would subsequently egress the bloodstream, traverse the BBB by methods that have been extensively reviewed elsewhere (Freitas, 2016), enter the brain parenchyma, and begin to navigate within the neuropil. Subsequently, they would enter the neuron cell soma and position themselves intracellularly within the

represent the actual neuralnanorobot design of the endoneurobots).

intra-glial regions to perform supportive B/CI operations. [Image credits: (A) Frank Boehm - Nanoapps Medical, Inc. (B) Julia Walker, Department of Chemical Engineering, Monash University]. (These conceptual illustrations do not represent the actual neuralnanorobot design of the gliabots).

AIS (Martins et al., 2016). Similarly, a 10 µm<sup>3</sup> volume of gliabots (**Figure 4**) would egress the bloodstream, enter their respective glial cells, and position themselves intracellularly at the most appropriate intra-glial region, which can vary. The synaptobots would also enter the human body via the bloodstream, cross the BBB (possibly assisted by auxiliary transport nanorobots), enter the brain parenchyma, commence navigation within the neuropil, enter the neuron cell soma, and then proceed intracellularly into the pre-synaptic or post-synaptic structure of a synapse.

The synaptobots would reside in the proper monitoring position within the neurons, in close proximity to presynaptic or postsynaptic structures. Once in place, these neuralnanorobots would monitor the action potentials and the structural changes initiated by the action-potential-based functional data. These data would be transferred from the synaptobots to corresponding endoneurobots (in some cases, with communications and other support from nearby gliabots). Once the data is received by the endoneurobots, it would proceed to the previously installed in vivo high-speed nanofiber-optic network, for subsequent transfer to the central units that are responsible for transmitting data to an external supercomputer. The auxiliary nanofiber-optic network system would provide essential support for the data that is transmitted by the endoneurobots and synaptobots, thereby minimizing their

onboard data storage capacity requirements. The external supercomputer would communicate with the cloud and handle data post-processing.

An optimal ingress strategy for all species of neuralnanorobots may employ the most rapid route to the human brain through the vasculature. Injection of the neuralnanorobots into the vasculature would be performed in the clinical environment under the supervision of medical personnel<sup>2</sup> . Once injected, the neuralnanorobots would have access to the dense microvasculature of human brain, which is composed of an estimated ∼100 billion capillaries, with a combined surface area of ∼20 m<sup>2</sup> and a total length of ∼400 miles. Intercapillary distances in the brain are typically ∼40 µm. Hence, each individual neuron within the human brain is at most 2–3 neurons away from a microcapillary (Pardridge, 2011).

The cerebral microvasculature is protected by the BBB, which comprises endothelial cells that are closely abutted as tight junctions. Cumulatively, they form a protective barrier for the human brain that is only naturally crossable by small molecules and lipophilic drugs. Neuralnanorobots can traverse the BBB by methods that have been extensively reviewed elsewhere (Freitas, 2016). For example, the potential uptake of nanoparticles (∼100 nm) through the BBB from the vasculature has been investigated, encompassing numerous strategies including passive diffusion, temporary disruption of tight junctions, receptor mediated endocytosis, transcytosis, and inhibition of p-glycoprotein efflux pumps (Kreuter, 2004; Lockman et al., 2004; Agarwal et al., 2009; Hu and Gao, 2010). Since the BBB consists of the endothelium of cerebral capillaries, the choroid plexus epithelium, and the arachnoid membranes (Talegaonkar and Mishra, 2004), it comprises one of the most impermeable ingress pathways for nanomedical devices (100 nm–1 µm) due to the presence of tight junctions.

Once the neuralnanorobots are distributed throughout the brain microvasculature, they could initially seek out any naturally present, randomly placed BBB junctional gaps or imperfections of various dimensions (Freitas, 2003). The BBB is not a perfect barrier, and perijunctional gaps of 0.5 µm have been reported (Stewart et al., 1987; Fraser and Dallas, 1993). Although various strategies exist for the traversal of nanoparticles through the BBB (Freitas, 2003, 2016; Grabrucker et al., 2016), further in-depth study would be required to precisely quantify the population, dimensions, and distribution of naturally occurring perijunctional gaps throughout the BBB network. This would be required if we are to consider passage through the BBB as the most appropriate method of ingress for some B/CI neuralnanorobots.

A process akin to "diapedesis" (the movement of leukocytes out of the circulatory system and toward the site of tissue damage or infection) might be employed by B/CI neuralnanorobots to traverse the BBB. As described by Muller, diapedesis is a multistep procedure by which leukocyte cells cross endothelial cell boundaries from within the bloodstream in ameboid fashion to access sites of inflammation within tissues. In humans, leukocyte transmission through interfacial junctions between tight, laterally apposed (≤0.5 µm thick) endothelial cells involves a number of sequential steps, including the organized activity of molecules upon and within the endothelial cells themselves. Additionally, the dual roles that endothelial cells must play, include facilitating the traversal of (∼7–10 µm in diameter) leukocytes, while sustaining tight apposing seals at the leading and trailing edges of these "passengers" as they are transferred through the junction to negate the leakage of plasma into the interstitial domain (Boehm, 2013; Muller, 2013). Further, it is conceivable that a certain class of facilitative B/CI neuralnanorobots with extendable/telescopic tendrils might project their nanoscopic appendages through smaller nanoscale perijunctional gaps to communicate with those neuralnanorobots that reside on the opposite side of the BBB, within the neocortex itself, or other relevant brain structures (Stewart et al., 1987; Fraser and Dallas, 1993; Freitas, 2003, 2016; Schrlau et al., 2008; Orynbayeva et al., 2012; Boehm, 2013).

Should large BBB junctional gaps be detected by the neuralnanorobots, they may be exploited to penetrate within the neuropil. However, in cases where there is a complete absence of large BBB junctional gaps, mission-designed strategies, including a combination of cytopenetration, cytolocomotion, and histonatation, would likely permit access to the neuropil (Freitas, 1999b, 2003, 2016). The BBB may also be opened using intravenous mannitol (an old method) and ultrasound, externally delivered (Samiotaki et al., 2017; Wang et al., 2017). In addition, "substances may cross the BBB by passive diffusion, carrier mediated transport, receptor mediated transport, and adsorptive transcytosis" (Grabrucker et al., 2016).

Once arrived at their designated neurons, the endoneurobots would autolocate and settle into their monitoring positions, intimately yet unobtrusively. Since action potentials might be initiated in different subcellular compartments, the endoneurobots would be anchored at the AIS (the most likely location for the initiation of action potentials), where they would monitor most action potentials. With some types of neurons, action potentials may be initiated at the first nodes of Ranvier or the axon hillock. Two synaptobots placed at these sites would ensure proper waveform detection of all action potentials. For example, the site of action potential initiation in cortical layer 5 pyramidal neurons is ∼35 µm from the axon hillock (in the AIS). For other classes of neurons, the action potential may be initiated at the first nodes of Ranvier, which for layer 5 pyramidal neurons is ∼90 µm from the axon hillock. The first myelin process is ∼40 µm from soma, whereas the length of the first myelin process is ∼50 µm (Palmer and Stuart, 2006).

All three types of neuralnanorobots (endoneurobots, gliabots, and synaptobots) would monitor action potential-based electrical information using the same types of FET-based nanosensors embedded in their surfaces (Martins et al., 2015). For the monitoring of neuronal structural changes (some of these triggered by the processing of action potentials), once they are securely anchored to the internal neuron membrane surface (with "typical" neurons having a "volume of 14,000 µm<sup>3</sup> or

<sup>2</sup>Alternatives to the regular injection of neuralnanorobots into the human vasculature include: intravenously, intranasally in aerosolized form, orally as a pill, via a dermal patch, or topical gel.

(∼24 µm)<sup>3</sup> ), endoneurobots and synaptobots might employ a tactile scanning probe to image the surrounding membrane surface area of (1.4 µm)<sup>2</sup> in ∼2 sec at ∼1 nm<sup>2</sup> resolution (∼1 mm/s tip velocity), or ∼50 s to ∼0.2 nm (i.e., atomic) resolution (∼0.2 mm/s tip velocity), assuming a scan rate of ∼10<sup>6</sup> pixels/s" (Freitas, 1999b). For their part, gliabots would utilize the same probing strategy.

#### Synaptobots

Synaptobots (**Figure 5**), the most diminutive (0.5 µm<sup>3</sup> ) of the three types of neuralnanorobots, are responsible for monitoring synapses, which are relevant sub-cellular structures of the human brain. Synapses (either of the 5–25% electrical or 75– 95% chemical variety (DeFelipe and Fariñas, 1992) are key components of the neural network that processes information. They play a crucial role in brain information processing (IBM, 2008) and are involved in learning and memory (Black et al., 1990; Bliss and Collingridge, 1993; Holtmaat and Svoboda, 2009; Liu et al., 2012), long-term and short-term memory storage and deletion (Kandel, 2001; Lee et al., 2008), and temporal information processing (Fuhrmann et al., 2002). They are also the key elements for signal transduction and plasticity in the human brain (Rollenhagen and Lübke, 2006; Rollenhagen et al., 2007). Synapses are so important that proper synapse formation during childhood provides the substrate for cognition, whereas improper formation or malfunction may lead to neurodevelopmental disorders, including various cognitive deficits and autism (Mcallister, 2007). The loss of synapses, as occurs in Alzheimer's patients, is intimately related to cognitive decline (Dekosky and Scheff, 1990; Terry et al., 1991; Scheff and Price, 2006). The monitoring of synapses is expected to be essential for a stable and robust fully functional real-time B/CI.

Synaptobots would be delivered via the brain microvasculature to avoid long-distance navigation within the brain parenchyma. Auxiliary transport nanorobots having a volume of ∼20 µm<sup>3</sup> (∼3.2 µm × 2.5 µm × 2.5 µm) might each convey cargos of 24 synaptobots (total of ∼12 µm<sup>3</sup> ) through the circulatory system and into the neuron soma. "The full complement of synaptobots would be transported by a fleet of ∼1 trillion auxiliary transport nanorobots, which perform ∼10 round trips to complete the insertion of all synaptobots" toward the implementation of the neuralnanorobotic system prior to the activation of the B/CI system. Individual neurons, on average, would obtain ∼117 such shipments, for an average overall distribution of 2800 synaptobots (≈2.42 × 10<sup>14</sup> synapses/86 × 10<sup>9</sup> neurons), which would assign one nanorobot per synapse (Martins et al., 2012).

The protocol for regularly updating the number of synaptobots in the brain (due to nanorobot damage, synapse elimination, neuron death, new synaptic formation, etc.) would be initiated by endoneurobots, which would communicate synaptic requirements to an external supercomputer. About 1 trillion auxiliary transport nanorobots may suffice to accommodate the workload of dynamically adjusting the physical deployment of synaptobots. Auxiliary transport nanorobots (∼2.5 µm) would adhere to a similar transit protocol for crossing the BBB and traversing the neuropil as the endoneurobots and gliabots, which are of comparable size (∼2.2 µm).

Once arrived at the neurons, the auxiliary transport nanorobots would release their cargo of 24 synaptobots into the cytoplasms of each neuron. Following deployment, each synaptobot would either remain within the neuron soma, or navigate (utilizing its onboard locomotion system) from the neuron soma along the axon or dendrite into pre-synaptic or post-synaptic structures — the sites at which synaptic monitoring would occur. To identify and differentiate presynaptic and postsynaptic structures of synapses, synaptobots would initially map (from within the cell) the surfaces of the axon (for axoaxonic, axo-somatic, and axo-dendritic synapses), the neuron soma (for somato-axonic, somato-somatic, or somato-dendritic synapses), and dendrites (for dendro-somatic, dendro-axonic, and dendro-dendritic synapses) (Harris, 1999).

Synaptobots would possess an independent propulsion system for traversing along the axons and dendrites in both directions and may also exploit existing biological neuronal axonic or dendritic transport systems. The process of locomotion may be biomimetically inspired by mitochondrial locomotion strategies within human neurons, to minimize any physiological damage to neuronal processes. Alternatively, oscillating piezo "fins" may operate in conjunction with a ovoid orifice to enable flow-through propulsion for synaptobots (**Figure 5**). The anticipated synaptobot deployment linear density would be ∼0.5 synaptobots/µm-length of axonic or dendritic processes, and the deployment volumetric

depiction (right) and calibrating at an axon (below). Oscillating piezo "fins" in conjunction with a central ovoid orifice might enable flow-through propulsion. In one configuration, ultrasensitive extendible/retractable "cuff" nanosensors might externally encircle synaptic gaps to monitor neurotransmitter traffic. [Image credits: (left) Frank Boehm, Nanoapps Medical, Inc. and (right and below) Yuriy Svidinenko, Nanobotmodels Company. (These conceptual illustrations do not represent the actual neuralnanorobot design of the synaptobots)].

number density would be ∼0.5 synaptobots/µm<sup>3</sup> of axonic or dendritic processes. Maximum synaptobot velocities of ∼1 µm/s may be required to respect biocompatibility requirements, given that the bidirectional movements of mitochondria within axons and dendrites are reported to have velocities of 0.32–0.91 µm/s (Morris and Hollenbeck, 1995; Macaskill et al., 2009), with mitochondrial motility in nontransgenic (NTG) neurons reported as 0.93 ± 0.55 µm/s for anterograde motion and 0.97 ± 0.63 µm/s for retrograde motion (Trushina et al., 2012).

Once securely emplaced at the monitoring positions in close proximity to presynaptic or postsynaptic structures, the primary synaptobot mission would be to monitor the exact timing and intensity of the electrical action potential information arriving at the synapses, and regularly monitor associated changes that occur in key structural elements of the synapse. With one synaptobot positioned near each synapse in the human brain, the action potential data might be acquired using ∼3375 nm<sup>3</sup> FET-based neuroelectric nanosensors (Martins et al., 2015), enabling monitoring of the synaptically processed 4.31 × 10<sup>15</sup> spikes/sec. Data collection would have a temporal resolution of at least 0.1 ms, which is sufficient for waveform characterization, even at the maximum human neuronal firing rate of 800 Hz. Facilitated and mediated by endoneurobots and gliabots, the synaptobots would subsequently transmit 5.52 × 10<sup>16</sup> bits/sec of continuous action potential data (Martins et al., 2012) via an in vivo nanofiber-optic network system, as described above (Freitas, 1999b).

Protocols for the application of the B/CI should include regular structural scanning of the human–brain connectome. The synaptobots, along with the endoneurobots and gliabots, could map and monitor relevant neuronal and synaptic structural changes using tactile scanning probe nanosensors (Freitas, 1999b) with special scanning tips that permit the synaptic bouton volume and shape to be measured, along with other relevant synaptic structural characteristics. This structural scanning process may include mapping the main ultrastructural components of a chemical synapse (whether located within the presynaptic axon terminal, the synaptic cleft, or post-synaptic terminal), the postsynaptic density (PSD), the active zone (AZ), synaptic vesicles (e.g., coated vesicles, dense core vesicles, and double-walled vesicles), endoplasmic reticulum, mitochondria, and punctum adhaerens (PA).

While scrutinizing synaptic structural changes, neuralnanorobots would also detect induced changes via monitoring synaptic plasticity and crosstalk, including longterm synaptic based potentiation (LTP), long-term depression (LTD), short-term plasticity, metaplasticity, and homeostatic plasticity. For instance, the activity-dependent modification of PSD proteins occurring over timescales of seconds to hours is believed to underlie plasticity processes such as LTP and LTD (Sheng and Hoogenraad, 2007). Longer-term changes in the PSD structure and composition (from hours to days) involve altered protein synthesis, either within the neuronal cell body, or dendrites (Sheng and Hoogenraad, 2007). The degradation of PSD proteins via the ubiquitin-proteasome system (Bingol and Schuman, 2006) also sculpts the PSD structure and plays a primary role in synaptic plasticity. Remarkably, recent evidence points toward the rapid exchange of PSD proteins, such as AMPARS and PSD-95, even between neighboring synapses under steady-state conditions (Sheng and Hoogenraad, 2007).

Neuralnanorobotic monitoring of the PSD appears to be an essential requirement. The PSD is a complex molecular machine that dynamically alters its structure and composition in response to synaptic activity. The PSD dynamically regulates its components through protein phosphorylation, palmitoylation, local protein translation, the ubiquitin-proteasome system for protein degradation, and redistribution of specific proteins (e.g., CaMKIIα, AMPARs) both entering and leaving the PSD (Kim and Ko, 2006; Sheng and Hoogenraad, 2007). Signaling pathways are organized by PSD proteins to coordinate synaptic structural and functional changes. These proteins also regulate the trafficking and recycling of glutamate receptors (which determine synaptic strength and plasticity), promote the formation and maturation of excitatory synapses by coaggregating with post-synaptic cell adhesion molecules, organize neurotransmitter receptors within the synaptic cleft, serve as a signaling apparatus. These proteins are also an essential component of an extraordinary synaptic signaling and regulatory assemblage. The "typical" PSD consists of a disk-like structure with an average diameter of 300–400 nm (range 200–800 nm), a thickness of 30–60 nm (Baude et al., 1993; Rácz et al., 2004; Okabe, 2007; Sheng and Hoogenraad, 2007), volume of ∼7.5 × 10<sup>6</sup> nm<sup>3</sup> , and mass of ∼1.1 GDa (Chen et al., 2005).

Events involving LTP and LTD structural changes to dendritic spines can alter spine number, size, shape, and subcellular composition in both immature and mature spines (Bourne and Harris, 2008). The dendritic spine neck serves as a diffusion barrier (controlled by neuronal activity) to current flow and diffusion of molecules between the spine head and the dendrite. The geometry of the spine neck determines the rate of calcium efflux into the dendrite shaft and hence the degree of elevation of calcium concentrations within the spine head, following n-methyl-D-aspartate receptor (NMDAR) activation (Bloodgood and Sabatini, 2005; Alvarez and Sabatini, 2007; Sheng and Hoogenraad, 2007). In experimental work, dendritic spines that received LTP induction increased in volume, from 50 to 200% (Alvarez and Sabatini, 2007), with this increase persisting for more than 1 h following stimulation (Alvarez and Sabatini, 2007). Sustained head enlargement in dendritic spines is induced by LTP, due to F-actin polymerization. LTD causes α-amino-3 hydroxy-5-methyl-4-isoxazole propionic acid (AMPA) receptor internalization with spine elongation and/or shrinkage of spine heads, due to actin depolymerization (Bourne and Harris, 2008).

There exists a clear and strong association between synapse bouton size/shape and the organellar and macromolecular changes that occur within the bouton. This provides some level of information redundancy and suggests that monitoring all dendritic spine organelles and molecular components is likely unnecessary. Synaptobots may deduce a great deal of useful information subsequent to scanning the gross volume and shape of the spine. This information redundancy is expected to significantly reduce synaptobot monitoring tasks.

The auxiliary nanofiber-optic system (**Figure 6**), coupled with endoneurobot and gliabot data transmission support, would likely serve to minimize the onboard data storage requirements for synaptobots. An onboard synaptobot nanocomputer might be manifest as a ∼0.01 µm<sup>3</sup> CPU device with ∼100 megaflops processing speed. The total internal volume of onboard synaptobot computation might be 0.11 µm<sup>3</sup> to fulfill redundancy requirements. Such volume allocation is similar to other nanorobot designs with comparable degrees of mission design complexity (Freitas, 2005b).

## Data Transmission Between Neuralnanorobots and the Cloud

Data from the three types of neuralnanorobots would be selected in real time, based on relevance to a specific use (such as auditory or visual content). The data would also be linked to other selected and related network activities, potentially with neurons in the prefrontal cortex and with mixed-selectivity neurons, which have been found to encode distributed information related to task-relevant aspects (Rigotti et al., 2013). Key design goals include: reducing latency, heat buildup, device size, and power for electronics; and tradeoffs for processing and latency between embedded/wearable/portable devices, local processing, and the cloud.

One key future technological advance in reducing latency will be 5G mobile telecommunication, expected in the year 2020 (AT&T Business, 2018). 5G promises to ensure a new way for mobile users to experience VR and AR, for example, via the cloud without latency artifacts. "To give you a sense of scale, the typical refresh speeds for a computer screen are approximately 80 ms" (Weldon, 2016). "However, for AR/VR, the industry is driving the conversation toward the Vestibulo-Ocular Reflex (VOR) the neurological process by which the brain coordinates eye and head movements to stabilize images on the retina. This is critical to synchronizing virtual and real objects to create a coherent view. The entire VOR process takes the brain 7 ms, a more than 10× reduction over screen-to-brain propagation. . . . Today's VR systems recommend a latency of <20 ms for standard performance, and very low latency (<7 ms) is even better. For this reason, developers and inventors want even lower latency to realize what they envision for the next iterations of VR." Similar performance increases may be found useful in B/CI neuralnanorobotic systems.

## Biocompatibility of B/CI Neuralnanorobotic Systems

Experimental data has provided a wide range of measured human intracranial volumes (1152–1839 cm<sup>3</sup> ) and total average cerebral spinal fluid (CSF) volume (82–125.3 cm<sup>3</sup> ), with a total brain-cell parenchyma volume of 1319 cm<sup>3</sup> , including 489 cm<sup>3</sup> of white matter and 786 cm<sup>3</sup> of gray matter (Vaidyanathana et al., 1997; Nopoulos et al., 2000). During B/CI operations, one ∼10 µm<sup>3</sup> endoneurobot would reside within every brain resident neuron, giving a total endoneurobot volume of 0.86 cm<sup>3</sup> , or only ∼0.06% of total brain volume. A similar volume will be displaced by gliabots, given one 10 µm<sup>3</sup> gliabot within each of the 84.6 × 10<sup>9</sup> brain-resident glial cells (Azevedo et al., 2009), displacing another ∼0.06% of brain volume. Thus, total volume displaced by endoneurobots and gliabots would be ∼0.12% of a "typical" ∼14,000 µm<sup>3</sup> neuron volume, which is orders of magnitude below the total 1–10% "safe" tissue and organ intrusiveness limit for nanorobots that has been recommended elsewhere (Freitas, 2003).

The synaptobot population represents a more significant neuralnanorobot intrusion on the volume of the human brain. Each synaptobot might contain ten 3375 nm<sup>3</sup> neuroelectrical nanosensors (Martins et al., 2015) for monitoring the action potentials of up to ten distinct synapses at adequate temporal resolution. Tagging all 2.42 × 10<sup>14</sup> synapses in the human brain with one robot each would require 24–242 × 10<sup>12</sup> synaptobots at 0.5 µm<sup>3</sup> per robot, giving a total fleet volume of 1.2– 12 × 10<sup>13</sup> µm<sup>3</sup> or 12–120 cm<sup>3</sup> and representing ∼0.9–9% of total brain volume-just within the "safe" tissue and organ intrusiveness limits.

Neuralnanorobotics for B/CI missions should include the capacity to navigate intracellularly (Martins et al., 2016), and even extracellular navigation might sometimes be required when intracellular navigation is deemed physically difficult, or impossible. For example, certain axonal and dendritic domains are less than 0.50 µm in diameter (Shepherd and Harris, 1998), and myelinated axons at three different sites of the corpus callosum in the human brain are estimated to have axon diameters of ≥0.50 µm, in 70–90% of cases (range 0.16– 3.73 µm, mean 0.73 ± 0.55 µm) (Liewald et al., 2014). Thus, a small percentage of the 0.5 µm<sup>3</sup> synaptobots might encounter difficulty in accessing distal axonic and dendritic regions via strictly intracellular navigation.

The biocompatibility of currently available BCI technologies has been a major challenge. Systems have performed well during acute recordings, but failed to function reliably over clinically relevant timelines, the result of brain tissue reaction against implants, making biocompatibility of implanted BCI systems a primary concern in current device design (Polikov et al., 2005; Winslow and Tresco, 2010; Tresco and Winslow, 2011). Biotic– abiotic interface cell biology has to take into consideration factors pertaining to various scientific domains, including chemistry, cellular biology, physiology, bioelectricity electrochemistry,

anatomy, surgery, and microbiology, as well as mechanical factors (Prodanov and Delbeke, 2016). The main reasons for biocompatibility problems with currently available BCI systems derive from induced acute injury, including: the breaching of the BBB to insert devices, the introduction of mechanical tissue strain from volumetric tissue displacement, mechanical tear of cells and the extracellular matrix, the activation of glial cells, the loss of local perfusion, vasogenic edema, secondary metabolic injury, steric blockade of signaling molecules, microglial activation, and locally induced neuronal degeneration (Gunasekera et al., 2015; Jorfi et al., 2015; Kozai et al., 2015). Some strategies have been proposed to address these biocompatibility problems, for example, the manipulation of BCI device surfaces that interface between intracellular and extracellular environments has helped passively reduce local inflammation, and consequently prevent numerous biocompatibility problems (Skousen et al., 2015; Oakes et al., 2018).

With proper design—respecting the limits of volumetric tissue displacement, minimizing residual impact on local perfusion, and ensuring no vasogenic edema—neuralnanorobots are not expected to induce localized acute injury and disruption to the BBB. Neuralnanorobots are also not anticipated to activate microglial immune reactions.

### FDA Protocols for Neuralnanorobotics

The development and implementation of a neuralnanorobotically mediated human B/CI will require that all hardware and software technologies involved in the process are extensively tested, verified, and certified by the appropriate technical and administrative organizations, to ensure compliance with the required protocols for biocompatibility, safety, redundancy, security/privacy, stability, and durability. Selected ingress and egress strategies will also be required to undergo highly detailed and rigorous scrutiny, in alignment with current/downstream FDA approval protocols for proposed clinical nanomedical technologies, particularly those that are to operate within the human brain.

The implementation protocols for neuralnanorobotics may be similar to those currently employed for the approval of any medical technology. The approval mechanism for neuralnanorobotics is expected to include testing the entire system, using (1) computational modeling, (2) laboratory testing, (3) in vivo animal studies, (4) robotic avatar testbeds, and (5) human trials. This step-by-step approach will comprise the proper clinical protocols, to be supplemented with detailed risk analysis and mitigation strategies. Once engaged in clinical trials, protection measures for human subjects may be instituted along with proper monitoring, in compliance with the requirements of a data–monitoring committee.

Aside from the FDA approval process, and prior to implementation, all stages of the neuralnanorobotically mediated B/CI system will require that each of its components and systems intended for ingress and egress undergo the comprehensive review of an ethics board. From an environmental perspective, all of the neuralnanorobots are expected to be made of diamondoid materials (likely produced in nanofactories via molecular manufacturing) with all nanodevices being completely recyclable, so they would impart no damage to natural ecosystems or the environment at large. Any disposal quantities should be of negligible volume and chemically inert.

The UN has recently condemned Internet access disruption as a human rights violation (United Nations Human Rights Council, 2016). Similarly, a neuralnanorobotics-based brain cloud interface might also, in the future, be considered a human right, given its profound relationship with the promotion, protection, and enjoyment of human rights on the Internet. The exercise of the human right to freedom of expression on the Internet has been considered of crucial importance, especially during a rapid pace of technological development, supported by the empowerment of individuals from all over the world to use new information and new communication technologies (United Nations Human Rights Council, 2016). In particular, the neuralnanorobotics based B/CI is expected to provide vast opportunities for affordable and inclusive education globally, consequently becoming an important tool to facilitate promotion of the right to education. A comprehensive analysis of the core ethical questions associated with implementation of the neuralnanorobotics-enabled brain cloud interface is expected to precede its implementation and mass adoption.

## HUMAN BRAIN/CLOUD INTERFACE APPLICATIONS

#### Significant Improvement of Education

Cumulative human knowledge doubled approximately every century until 1900. By 1950, human knowledge was doubling every 25 years. As of 2006, on average, human knowledge was doubling every 13 months, and the "Internet of Things" is expected to further lower the doubling time of human knowledge to 12 h (Coles et al., 2006). Such massive amounts of information increase the urgency to radically improve human learning capacities, which are currently limited by biological evolution-driven characteristics. The impracticability of keeping up with the modern rate of creation of scientific knowledge is clearly evident, assuming present-day human biological cognitive abilities (Larsen and von Ins, 2010). Contemporary approaches to this problem include limited strategies such as data mining and research maps (Landreth and Silva, 2013). Neuralnanorobotics may enable us to far surpass our presently limited cognitive capacity to learn in a world driven by exponentially expanding knowledge.

The ultimate learning process may be manifested as direct transfer of knowledge to the human brain, where neuralnanorobots empower practically instantaneous and nearly perfect learning. However, the injection of facts and accumulated knowledge may not necessarily translate to cognition, understanding, meta-analysis or meta thought that can inspire imagination and creativity. Complex skills such as playing the piano or performing a complex brain operation might be "injected" into the brain, which may reduce the time that it traditionally takes to learn the piano, or to be a proficient

brain surgeon. This may be possible, as these are specific manual skills that are imprinted in the brain. Access to the hippocampus and cerebellum for memory injection would also be required, as well as the cerebellum and basal ganglia for complex motor tasks.

This would require highly accurate data transmission, which would in some ways be similar to today's extremely precise computer data transmission, accompanied by instantaneous thought-activated Internet access, or B/CI. The first proof-ofprinciple of "instant learning" was accomplished using decoded fMRI, where human visual cortex brain activity patterns were induced to match a previously known target state and improve the performance of visual tasks (Shibata et al., 2011). Transcranial magnetic stimulation, involving the application of a strong pulsed magnetic field from outside the skull using a magnetic coil precisely positioned over the head, was also employed to induce new skills. Stimulating a "virtual lesion" of small regions of the brain either diminished or enhanced skills in some transcranial magnetic stimulation experiments, with approximately 40% of participants displaying remarkable new skills, such as drawing abilities (Mottaghy et al., 1999).

#### Enhancement of Human Intelligence

The brains of humans with high IQ are extensively integrated with neural pathways that connect distant brain regions, while the brains of humans with low IQ have less-integrated connectivity with shorter neural routes (Colom et al., 2007; Haier and Jung, 2007). Neuralnanorobotically mediated B/CI systems may enable significantly increased human intelligence, eventually superseding the inherent architectures of the brain's neural domains. Such systems could expand memory capabilities considerably, improve pattern recognition and cognition through the creation of novel hybrid biological/non-biological networks, and interface with non-biological networks as well as new forms of AI.

Neural prostheses are currently employed in cochlear implants to treat hearing loss, as stimulating electrodes to treat Parkinson's disease and other neurological diseases, and in "artificial retinas" to restore vision, among other applications (Dobelle, 2000; Mayberg et al., 2005; Perlmutter and Mink, 2006; Gaylor et al., 2013; Lewis et al., 2015, 2016). Brain implants employed in locked-in patients permit extraction of brain data into an external computer, enabling patients to communicate with the outside world (Hochberg et al., 2012). Since the hippocampus plays a critical role in learning and memory, damage to this small organ can disrupt proper electrical signaling between nerve cells, impeding the formation and recall of memories. This is something that artificial-brain-inspired prosthetics are currently beginning to treat (Berger et al., 2005, 2011; Lebedev and Nicolelis, 2006).

Computerized implants receiving signals from thousands of brain nerve cells may wirelessly transmit the data to an interfacial device that decodes intentions, with preliminary versions of these implants being used to control artificial limbs (Ferris, 2005; Au et al., 2007; Gordon and Ferris, 2007; Hargrove et al., 2013; Tabot et al., 2013). Neuralnanorobots may offer significant advantages over current surgically installed neural prosthetics, since they might be introduced through the bloodstream without surgery, via a fully reversible procedure that could be reprogrammed in real-time to permit instantaneous software updates.

## Artificial Intelligence and Existential Risk Prevention

Empowered by the exponential increase in price/performance of computational data storage and processing power, artificial intelligence (AI) algorithms are improving across many domains and demonstrating superior capabilities when compared to those of humans. Examples of the superiority of AI include: gameplaying (Jeopardy, Go, chess), driving cars, providing diagnostics for some cancer patients, and other examples in various domains (Ferrucci et al., 2010; Levinson et al., 2011; Chouard, 2016). Over the next decade, narrow artificial intelligence algorithms are expected to outperform humans in many other areas. Advances in artificial intelligence across machine learning, machine vision, and natural language processing domains, combined with advances in big data and robotics, are anticipated to empower robots to outperform humans in many, if not most, physical and cognitive tasks. However, in the future, we can expect far more powerful "artificial general intelligence" (AGI), a subfield of AI oriented toward creating thinking machines with general cognitive capability at the human level and beyond. (Minsky, 1985; Nakashima, 1999; Horst, 2002; Hutter, 2005; Goertzel, 2006; Adams et al., 2011).

Interfacing the human brain with the cloud via neuralnanorobotic technologies may be beneficial for humanity by assisting in the mitigation of the serious existential risks posed by the emergence of artificial general intelligence (Bostrom, 2002, 2013; Whitby and Oliver, 2000; Joy, 2007; Bostrom and Cir, 2008; Yudkowsky, 2008; Schneider, 2009). One such mitigation might involve the merging human brains with computers to prevent the dangers of unbridled artificial general intelligence (Dewey, 2015). Neuralnanorobotics may indeed be a suitable technology to assist with reducing human existential risk potentially initiated by rapidly emerging artificial general intelligence by enabling the creation of an offsetting beneficial human augmentation technology.

## Virtual and Augmented Reality

Fully immersive virtual reality may become indistinguishable from reality with the emergence of neuralnanorobotics, rendering many forms of physical travel obsolete. Office buildings might be replaced by virtual-reality (VR) environments in which conferences could be attended virtually, replacing today's VoIP conference calls and Internet-based video conference calls with highly realistic, fully immersive VR conferences in virtual-reality spaces. Immersive VR may enable long-distance communications in engaging ways within environments that are indistinguishable from reality. The economic and environmental benefits of significantly reducing travel requirements may be significant. For example, Cisco has reported savings of millions of dollars through the use of highly realistic telepresence systems.

Current systems for fully immersive virtual reality include VR headsets and haptic controllers (typically to facilitate immersive gaming) (Alkhamisi and Monowar, 2013; Tweedie, 2015).

In principle, fully immersive VR may benefit from advanced neuralnanorobotics to provide, for example, appropriate "proximal cues."

Neuralnanorobotically induced artificial signals may be indistinguishable from actual sensory data that is being received from the physical body. All brain output signals might be suppressed by neuralnanorobots to avoid the movement of real limbs, mouth, or eyes during virtual experiences; in place of this, virtual limbs would react appropriately while adapting the surrounding virtual world in the field of vision (similar to current immersive gaming). B/CI users might initially encounter a virtual dashboard in the cloud where they can select from an extensive menu that is replete with experiential pathways. The gaming industry provides virtual environments for humans to explore, from recreations of actual locations to fanciful environments — even environments that violate the laws of physics. Virtual trips in simulations of "real" locations will permit the equivalent of nearly instantaneous time travel. Ultrahighresolution, fully immersive VR might also enhance business negotiations and web-dating, among other applications. The "real" and the "virtual" worlds could evolve to become practically impossible to distinguish.

Another application of neuralnanorobotics might be manifest as augmented reality—superimposing information about the real world onto the retina to provide real-time guidance, explanations, or data on social events while traveling. Neuralnanorobotics might provide real-time auditory translation of foreign languages, or access to many forms of online information, which would integrate these augmentations into our daily activities. Some types of information might be presented by virtual assistants or avatars that overlay the real world to assist their human partners with the retrieval of information. These virtual assistants, running on the cloud, similarly to IBM Watson, might not even wait for questions if they can predict human desires based on previously registered behavioral patterns and other data.

## Ultrahigh-Resolution Fully Immersive "Transparent Shadowing"

Neuralnanorobotically empowered B/CI technologies accompanied by supercomputing technologies might permit users to experience fully immersive, real-time episodes of the lives of any willing human participant on the planet, via nonintrusive "Transparent Shadowing (TS)." In TS, an individual might literally experience another person's life, through their own eyes, for a predetermined duration via an "extra life" session. Such a capacity may be anticipated to elevate human collaboration, understanding, respect, and empathy to previously unimaginable levels (Domschke and Boehm, 2014). "We will be able to change our appearance and effectively become other people" (Kurzweil, 2005).

With neuralnanorobotically enabled B/CI, individuals might engage in the TS of voluntary or remunerated "spatial hosts." Under strict protocols, accredited spatial hosts would agree to allow single or multiple attendees (conceivably numbering in the millions) to literally experience portions of their life experiences over a predetermined timeline/schedule. These TS sessions might be akin to today's seminars or lecture series, where the knowledge or specific skills of the host would be experientially imparted to the "attendees." However, these TS sessions would offer exponentially higher resolution in every respect. The full sensorial realm (e.g., physical presence, tactile sensations, olfactory, visual, tastes, and auditory) would be experienced by the attendees, as if they inhabited the body of the spatial host. Although they would perceive the vocal instructions of the host, to temporally experience exactly what the spatial host is experiencing, for the sake of personal privacy, attendees might, by default, be completely blocked from any access to the thoughts, emotions, or self-speak of their spatial hosts (Ford, 2010).

From another perspective, access to some level of self-speak may be beneficial for attendees toward conveying the thought processes/intentions of a spatial host that underly their activities. However, it is likely that any self-speak of a spatial host will include their most private and intimate thoughts. Hence, this warrants a careful exploration of how these self-speak items might be screened such that the attendees are not privy to them, how will this be decided, and by whom. What self-speak will be allowable and what will be considered as out-of-bounds? For this assessment to take place via AI, the self-speak in question would have to somehow be processed in real time/on the fly, as any records of such in any form, would most likely be considered as highly unethical. This may translate to the establishment of a very brief (∼millisecond) latency, from spatial host to attendee, to allow for this virtually instantaneous self-speak screening.

Although the attendees would retain their own identities and experience real-time live-feed full immersion into a portion of the host's life experiences, these guests would have no capacity to control any aspect of the host. This particular B/CI application would be akin to an exponentially enhanced version of attending and viewing a movie, albeit one that is totally controlled by the spatial host. The host will have no mental or physical perception that they are being "shadowed" by the attendees. Hence, in essence, anyone on the planet (who is B/CI enabled) might be engaged as either a spatial host or an attendee.

Once established and potentially utilized by a growing demographic, this capability might have strong potential for conveying profound beneficial implications for human advancement across multiple domains, possibly assisting with the further development of human collaboration and empathy, perhaps eventually leading to the minimization or elimination of most armed conflict.

Given the prospect of virtual, fully immersive TS, issues pertaining to the possibility of immediate or residual (post-TS session) physiological and/or psychological transference arising from the interactions between a spatial host and any given attendee will require careful consideration. In addition to standardized TS operational procedures and safeguards, a range of prudent failsafe protocols should be explored and established for the protection of both the spatial hosts and the attendees. Potential physiological transference issues may arise when two or more individuals are engaged in shared TS activities. For example, in cases of unpleasant pain, the attendee might activate an instantaneous default auto-disengagement

protocol once a particular physiological threshold is perceived by the B/CI system.

TS may hold potential to facilitate understanding of the experiences of other people and to significantly increase empathy. Experiencing episodes of the lives of those in other cultures and ethnic groups could promote cross-cultural understanding and tolerance, improving prospects for the reduction of hatred and racism. For example, perhaps those of majority ethnic groups might be more sensitized with the issue of racism against those of minority ethnic groups, once they "experience" it for themselves through TS sessions. Similarly, minorities who experience a majority host might come to realize that many actions perceived by them as purposeful racism were entirely unintentional. Crossgender experiences might impact real-life relationships between genders, due to increased empathy and understanding. It might be possible that an eventual shift in gender attitudes could lead to decreased gender-related and domestic violence.

Although beyond the scope and space constraints of this paper, we acknowledge that there will likely be several "display modes" available to B/CI users once the technology matures. These may include optional text, imagery, and streaming video displays that are superimposed at customizable locations within the user's field of vision. It may be likely that via TS, all knowledge-based queries and responses, as well as fully immersive experiences within avatars and other users, could include an optional toggle mode. When this mode is engaged, internalized visualizations might be projected as on to a "float screen" that can be superimposed within the user's field of vision, where the float screen can be made more prominent in the user's experience via dynamic fading.

## CONCLUSION

Human knowledge is being digitized at an accelerated exponential pace for storage and processing in the cloud. Given our biologically constrained cognitive abilities, the impossibility of the human mind to keep pace with the increasingly rapid generation of human knowledge is evident. Hence, it is essential, and may indeed become urgent, that we develop a safe, robust, stable, secure, and continuous real-time interface system between the human brain and the data storage and processing systems that reside in the cloud. Neuralnanorobotics may provide a technology at the appropriate scale, with a suitable level of complexity to robustly interface the human brain with the massive volume of data that is stored and processed in the cloud.

#### REFERENCES


Neuralnanorobotics strategies involve the direct, comprehensive monitoring of the ∼86 × 10<sup>9</sup> neurons of the human brain, as well as its ∼2 × 10<sup>14</sup> synapses and ∼84 × 10<sup>9</sup> glial cells. Three proposed classes of neuralnanorobots (endoneurobots, gliabots, and synaptobots) may employ ∼3375 nm<sup>3</sup> FET-based neuroelectric nanosensors to detect and monitor virtually all individual action potentials and their waveforms. Neuralnanorobotic entities would transmit the nominal ∼5 × 10<sup>16</sup> bits/sec of synaptically processed electronic information, encoded in ∼4 × 10<sup>15</sup> spikes/sec flowing within the entire living human brain, wirelessly via a nanorobotic auxiliary 30 cm<sup>3</sup> volume nanoscale fiber-optic system that is capable of handling ∼10<sup>18</sup> bits/sec. This may permit real-time brain-state monitoring and data extraction into an external supercomputer that communicates directly with the cloud.

A human B/CI system mediated by neuralnanorobotics could empower individuals with instantaneous access to all cumulative human knowledge available in the cloud and significantly improve human learning capacities and intelligence. Further, it might transition totally immersive virtual and augmented realities to unprecedented levels, allowing for more meaningful experiences and fuller/richer expression for, and between, users. These enhancements may assist humanity to adapt emergent artificial intelligence systems as humanaugmentation technologies, facilitating the mitigation of new challenges to the human species. Human B/CI systems mediated by neuralnanorobots might also upgrade mutual human understanding and collaboration by making it possible to engage humans in TS experiences, which could enable considerably improved understanding and tolerance among all members of our diverse and amazing human family.

## AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

### FUNDING

ML's work was supported by the Center for Bioelectric Interfaces of the Institute for Cognitive Neuroscience of the National Research University Higher School of Economics, RF government grant, ag. no. 14.641.31.0003.






long-term implantation with subretinal microphotodiode arrays. Exp. Eye Res. 73, 333–343. doi: 10.1006/exer.2001.1041



**Conflict of Interest Statement:** YS of NanobotMedical, Inc. declares no competing or conflicting interests and FB of NanoApps Medical, Inc. declares no competing or conflicting interests.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Martins, Angelica, Chakravarthy, Svidinenko, Boehm, Opris, Lebedev, Swan, Garan, Rosenfeld, Hogg and Freitas. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

digital media

of impactful research

article's readership