Edited by: Catherine Carr, University of Maryland, USA
Reviewed by: Enrico Mugnaini, Northwestern University, USA; John P. Miller, Montana State University, USA
*Correspondence:
This is an open-access article distributed under the terms of the
The importance of neuronal morphology has been recognized from the early days of neuroscience. Elucidating the functional roles of axonal and dendritic arbors in synaptic integration, signal transmission, network connectivity, and circuit dynamics requires quantitative analyses of digital three-dimensional reconstructions. We extensively searched the scientific literature for all original reports describing reconstructions of neuronal morphology since the advent of this technique three decades ago. From almost 50,000 titles, 30,000 abstracts, and more than 10,000 full-text articles, we identified 902 publications describing ∼44,000 digital reconstructions. Reviewing the growth of this field exposed general research trends on specific animal species, brain regions, neuron types, and experimental approaches. The entire bibliography, annotated with relevant metadata and (wherever available) direct links to the underlying digital data, is accessible at
Neuronal morphology is a key determinant of information processing in the nervous system. The morphological diversity of axons and dendrites provides an essential substrate for synaptic integration, signal transmission, network connectivity, and circuit dynamics. The beginning of modern neuroscience is often associated with the first drawings of neurite arbors, and those early illustrations continue to appear in research articles and textbooks. However, computational approaches are necessary to quantify the intricate relationship between neuronal morphology (structure) and physiology (activity; De Schutter and Bower,
Light microscopy allows an optimal balance between resolution and field of view to routinely visualize every individual branch through entire neuronal arbors (Lu,
Digital reconstructions can be acquired from any animal species, brain regions, neuron types, and a variety of experimental protocols. Morphological quantification is central in some studies, such as to investigate chemical, genetic, or behavioral effects on branch complexity. In other cases, morphology is traced as an aid, e.g., to document the identity of intracellularly recorded neurons. Once digitally reconstructed, however, morphological data can be re-used in secondary applications beyond the scope of the original project, including computational simulations, comparative analyses, and large-scale data mining across labs or techniques.
Although recent developments promise to facilitate the future automation of digital tracing (Bas and Erdogmus,
Inspection of almost 50,000 titles, 30,000 abstracts, and more than 10,000 full-text articles (nearly 2000 h of literature search and information extraction) identified 902 articles describing ∼44,000 neuronal reconstructions. This count (and subsequent analysis) only includes the original papers reporting the first or most complete characterization of the reconstruction rather than use or application of data described in other reports. Independent sampling indicates that this collection represents a robust (two-thirds) majority of all actually published relevant material. This corpus provides a representative coverage of cellular neuroanatomy and a comprehensive assessment of the growth of the field. We thus analyzed chronological patterns and number of reconstructions by field of study, animal species, brain region, cell type, journal, and experimental method, revealing interesting research trends. Moreover, we stored the extracted metadata into a newly created literature database directly accessible from
The information collated through extensive
Drawings of complex neuronal arbors started to appear in scientific reports since the 1873 publication of Golgi’s staining (Senft,
The temporal analysis of the 902 identified publications describing digital reconstructions demonstrates a continuous growth starting from the mid 1990s (Figure
To examine the chronological trend in the usage of animal models, the numbers of publications for every species, and time period were normalized to factor out both the general temporal growth and the overall species distribution (see
Breakdown by reconstruction software, microscopy techniques, and staining methods also uncovered clear trends (Figure
Next we examined the most frequently investigated brain regions and neuron types (Figure
While interneurons are believed to account for ∼15% of neurons in the mammalian neocortex and hippocampus (Gonchar and Burkhalter,
For each of the identified publications, we annotated the number of reconstructed neurons and the general neuroscience subfield of study (see
Since the number of neurons reported in a study may be affected by the ease of the experimental process, the reconstruction system might constitute an important ergonomic factor depending on the user friendliness of different software programs. Nevertheless, we found no statistically significant differences in the number of reconstructions per publication among reconstruction systems (data not shown). In contrast, publications using mice or primates as experimental subjects contain more reconstructions on average than those using other species (Figure
We also categorized each identified publication by its field of study, based on the scientific question (e.g.,
To determine reconstruction availability for public
Independent sampling (see
Data availability varied substantially by species. Reconstructions from primates and invertebrates were on average five times more available than those from cats and other mammals (typically older data), while the availability from rats, mice, and other vertebrates was close to the mean (Figure
We posted the annotated bibliography at
Two important considerations are relevant to the literature mining approach underlying this study. First, the presented methodology is fully applicable to other subfields and techniques. Second, this process enacts comprehensive coverage of the published data in a given domain. This is a necessary and often overlooked step for densely populating the relevant databases. Dense coverage is of immense value for scientific resources as it elevates them to a one-stop portal for the user community. To ensure continuous maintenance of comprehensive coverage, the described retrospective strategy must be followed by an analogous, regularly ongoing prospective search to cover new data as they are published.
Our literature mining protocol consisted of five sequential steps: (1) pilot testing and progressive refinement of query terms; (2) bulk bibliographic search; (3) information identification; (4) metadata extraction; and (5) data request and follow-up author communication.
The first step aimed to select the search engines and query terms for minimizing missed relevant publications (false negatives) without excessively inflating the return of irrelevant publications (false positives). This process was continuous and dynamic; the results presented here were acquired over a period of 3 years, during which new bibliographic resources were introduced and existing ones evolved in their functionality or expanded their coverage, requiring adjustment of optimal query terms and search protocol.
A broad spectrum of search terms was initially tested in PubMed
Although PubMed is the largest dedicated database of biomedical literature with more than 20 million records (in early 2011), it only searches abstracts and keywords. The above eight selected keywords were thus searched in three additional resources: PubMed Central (PMC), which contains more than two million full-text records
The bulk results of these searches were progressively narrowed down to identify the actual publications describing neuronal reconstructions (Table
Search engine | Total hits | Unique hits | Abstract evaluated | Full-text evaluated | Positive fraction (%) | Fraction of identified positive (%) |
---|---|---|---|---|---|---|
8186 | 6652 | 6335 | 3105 | 5.9 | 23.6 | |
19306 | 11853 | 11853 | 11853 | 2.1 | 32.2 | |
13267 | 11799 | 11799 | 11799 | 3.2 | 48.9 | |
5208 | 2181 | 756 |
745 | 11.9 | 11.5 | |
45967 | 32485 | 30743 | 27502 | 2.4 | – |
To test the efficiency of this protocol and to estimate the proportion of possibly missed relevant literature, we compared these results with two independent samples. The first consisted of publications positive for reconstructions collected in our lab over the years. The second was populated by manually identifying positive publications from a random 4% subset of the results of a full-text HighWire search
The database of positively identified publications is continuously updated with monthly literature searches. The methodology for this ongoing process is similar to that described above, but the exact protocol is constantly adapted to evolving scientific knowledge and new resources. In particular, we currently employ two search engines allowing full-text searches:
The following keywords were chosen for the 2011 literature: “Neurolucida,” “neuron reconstruction,” “neuronal reconstruction,” “3D neuron,” “digitally reconstructed neuron,” “neuron morphology,” “dendrite structure,” “axon structure,” “axon arborization,” “axonal arborization,” “Imaris,” and “Amira.” After checking the PMIDs from all keywords for redundancies, full-texts are evaluated for presence of digital reconstructions. Positive PMIDs are then annotated with information on species, strain, brain region(s), cell type(s), staining method, reconstruction software, and number of reconstructions. Data are requested from the authors and the availability is noted. The resulting “Literature Coverage” database is updated monthly.
When annotating positive publications with relevant species, brain region, and other metadata, if multiple categories apply (e.g., a report describing reconstructions from two species), they are weighted based on the number of reconstructions in each.
In temporal trend analyses (e.g., Figure
Reconstructing complex neuronal branching in digital 3D format may help map brain circuitry with its billions of connections. This technique enables
Highly anticipated advancements in high-throughput data acquisition such as robotic specimen preparation, fast microscopic imaging, and automated digital reconstruction, would enable large-scale morphological studies. Reverse engineering the brain requires representative sampling of cell type diversity. The DIADEM Challenge demonstrated that, despite recent progress, full automation of 3D tracing is still a major bottleneck (Gillette et al.,
Sharing reconstructions through designated databases, such as
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 authors are grateful to Sriraman Damodaran, Anirudh Raghavan, Maryum Ilyas, Kaitlin Grainger, and Juhi Saxena for assisting with the retrospective and ongoing literature search; and to Dr. Diek Wheeler for help with critical evaluation of available full-text search engines. Supported in part by NIH R01 NS39600 and ONR MURI N00014-10-1-0198.
1
2
3
4
5
6
7
Representation of neuron morphology generated using tracing software that captures the 3D neuronal structure as a list of connected vectors.
An ordered series of 2D images captured at regular
Free online repository of digital reconstructions of neuron morphologies with extensive literature database of publications describing digital reconstruction data.
Descriptive information about the data content of a resource. Relevant metadata for neuron morphology include the experimental subject and protocol, location and classification of the neuron, and reconstruction methods associated with the digital reconstruction.
Extensive search of available literature to identify and collect information related to a specific topic from large corpora.
Exchange of scientific data to reach a mutual goal of research integration enabling secondary discovery.
The study of neuronal structure commonly used for identification and classification. In particular, axonal and dendritic arbors are key functional components of neural processing and fundamental determinants of neural circuits.
Statistical studies of neuronal geometry that are most effectively performed on digitally reconstructed data.