The complexity of aging biology creates challenges for the study of age-related disease. Aging takes place over long timescales, in a complex systemic and local tissue environment. This means that the study of cells in isolation likely overlooks the in vivo context (Figure 1). For example, signaling between immune and parenchymal cells drive aspects of tissue aging (Buechler, Fu, and Turley 2021; Yousefzadeh et al., 2021), the mechanical stiffness of the local tissue niche impacts the aging of central nervous system progenitor cells (Segel et al., 2019) and development of fibrosis (Wells 2013). Such factors can lead to very different outcomes when studying disease in vivo versus in vitro (Miller et al., 2017). Meanwhile, complex secretory and immune phenotypes (Covarrubias et al., 2020), or even the activity of specific neurons in the brain (Zullo et al., 2019), can alter cellular metabolism at distal sites (Camacho-Pereira et al., 2016). In addition, different organs, tissues and cell types within those tissues are found to age differently (Rando and Wyss-Coray 2021), and the causal factors of aging are likely multifarious (López-Otín et al., 2013).

While we have seen an explosion of multi-omics analysis in recent years, allowing us to generate large amounts of high dimensional data (Tabula Muris Consortium 2020), distinguishing correlation from causation is a major challenge. Any given metric that changes with age or in an age-related disease could represent a causal driver, could be a compensatory response, or could be an ancillary effect. Ultimately, discriminating these scenarios requires the ability to apply causal perturbations, as a means of hypothesis testing, in the in vivo context.

Ideally, to cope with this complexity, many targets could be compared in separate cellular experiments occurring in an intact, living animal, i.e., in vivo. Operating directly in the in vivo setting means that complex aging biology can be probed directly, without the need to first understand and then translate it to an in vitro model. Of course, in vivo screening has been a core feature of aging research for decades. Model systems like C. elegans have been a mainstay of this approach since they enable many perturbations to be applied, and thus causal hypotheses to be tested, in parallel by a single lab at a reasonable cost (Lucanic et al., 2018). Extending these approaches to larger, mammalian models, however, entails a much greater operational footprint, unless a means can be found to test many hypotheses in the same animal.

In this review, we describe how the emerging paradigm of in vivo pooled screening can contribute to overcoming these challenges, with a focus on technical considerations that will enable researchers in the aging field to apply these methods in their work. In vivo pooled screening builds on two co-occurring revolutions: in multiplexed single-cell analysis, and in multiplexed in vivo gene-specific perturbation. In turn, these technologies have only become possible in the past decade due to advances in microfluidics, molecular and cellular barcoding, nucleic acid programmable genetic and epigenetic modulators made possible by CRISPR technologies, as well as multiplexed gene synthesis and the increased availability of gene delivery vectors. We will focus on the two primary challenges to pooled screening in vivo: delivering perturbations in a targeted and resolvable manner, and readouts that measure effects on multiple biological processes in individual cells, while giving examples of past and potential applications to understand the biology of aging and age-related disease.

The essence of an in vivo pooled screen is to administer a pooled library of different perturbations to a single, living model organism, with a screen readout that can both distinguish the effects of each intervention on the cells that the intervention ended up in, and identify of each intervention. Successful outcomes, meaning accurate and interpretable data, depends on methodological choices at each step of the process shown in Figure 2. This hypothetical example walks through how a published in vitro screen for ways to treat age-related calcific aortic valve disease, with in vivo validation (Theodoris et al., 2020), could have been designed as a direct in vivo screen. The rest of this section provides detailed discussion of the challenges and considerations at each step of this process.

Utilizing the scalability of pooled screening depends on having phenotypic readouts that associate perturbation barcodes with the outcome of perturbation. Most in vivo pooled screens thus far have looked at enrichment or depletion of perturbation barcodes after genome integration/editing, representing gain/loss of cellular proliferation (Zender et al., 2008) and/or survival (Wertz et al., 2020). Of these, enrichment is more popular because gain of signal is easier to measure than loss of signal. Enrichment can be measured at a single time point, or after multiple rounds of directed evolution (Ruozi et al., 2015). Enrichment readouts are cost-effective because signal is inherently boosted by the phenotype screened for, and results can thus be robust even when a subset of perturbed cells are sampled.

Nevertheless, there are many sources of noise and artifacts for such screens [covered in detail elsewhere, e.g. (A. Schuster et al., 2019)], and signal amplification is often useful or necessary. In vitro, it is common to kill non-transduced cells with e.g. puromycin, but this strategy clearly does not work in vivo. Instead, a fluorophore or other sorting tag can be used to select transduced cells. This has the additional advantage of selecting for successful gene expression, where DNA rather than RNA barcodes are used, but note that this selection process will introduce false positives and negatives, and sacrifices some transduced cells through inefficiencies. Typically, PCR amplification of perturbation barcodes prior to sequencing is used to improve depth of sampling. Here, it is essential to use methods that eliminate differences from differing PCR efficiency between barcodes. A standard option is to tag each molecule with a unique molecular identifier (UMI) in the first round of amplification, which then distinguishes the parent molecule after additional cycles (Kinde et al., 2011). It is likewise essential to normalize apparent enrichment to representation of each perturbation in the input vector. Since representation changes at each step of generating delivery vectors (e.g. bacterial expansion and vector packaging), it’s rarely appropriate to normalize to the initial plasmid mix. In principle normalization could be relative to earlier time points. But since assays are typically destructive this relies on comparisons between different animals with physiological variability that is likely to affect readouts and bias the normalization.

In the context of aging, one could imagine using enrichment screens to identify protective factors in degenerative disease, or introducing reporters for e.g. senescent cell states alongside perturbations that induce or prevent that state. But many aspects of aging biology cannot be studied in terms of cellular abundance, and the variety of biological processes involved calls for multidimensional readouts. Single-cell RNA sequencing (scRNAseq) can provide this additional layer of interpretability for the effects of perturbations, similar to multivariate imaging screens (Liberali et al., 2015) but in a more realistic in vivo setting. Indeed, bulk transcriptomics has already been used not only to identify transcriptomic signatures of aging, but also to identify small molecule interventions that revert transcriptomes to a “younger” state and extend organismal lifespan (Baumgart et al., 2016; Janssens et al., 2019). Single-cell sequencing not only enables in vivo pooled screening using such transcriptomic readouts of aging, but also is able to distinguish age-related changes that occur heterogeneously in different cell types (M. J. Zhang et al., 2021; Kimmel et al., 2019) and even in cells of a single type (Enge et al., 2017; Martinez-Jimenez et al., 2017; Kaufmann et al., 2021).

Perturbation screens with single-cell transcriptomic readouts were originally reported in lentivirus-based in vitro screens (Adamson et al., 2016; Dixit et al., 2016; Jaitin et al., 2016; Datlinger et al., 2017), but have recently been applied to in vivo systems (Jensen et al., 2020; Jin et al., 2020). In addition to revealing heterogeneous responses of different cell states to a given perturbation, in vivo pooled screens allow transcriptomic effects to be evaluated relative to controls within the same animal. Although it is still important to consider environmental factors that affect physiology, such as housing temperature (Seeley and MacDougald 2021) and group size, circadian rhythms, and diet, perturbation effect vectors within outlier animals can still be interpreted. This also greatly facilitates analysis for animal models that manifest a range of disease severities, including naturally occurring rather than experimentally induced disease models (Figure 3).

But, for in vivo screens in particular, it’s important to ask whether the sampled cells are representative of the biology one intended to study. Both cell type and state affects viability after dissociation, so the cells that end up sequenced cannot be assumed to be a random sampling of the in vivo perturbed population (Meyer and Schumacher, 2021). When integrating vectors are used, cell division during the course of the experiment produces progeny cells that share barcodes and perturbations but may be quite different in terms of aging processes. Similarly, barcode molecules may be detected in immune cells that have consumed the originally transduced cells. And depending on how many cells in a tissue are perturbed, the overall state of the tissue may be shifted, such that the default cell state is no longer representative of the biology one wishes to model. Thus, in vivo pooled screens should always be designed with attention to how the host tissue will respond to both delivery, expression, and isolation of perturbations.

Another challenge with droplet-based scRNAseq is that not every RNA molecule in each cell is captured (X. Zhang et al., 2019), including the barcode molecules required for analysis of perturbation effects. Methods that improve expression and/or capture of barcode molecules improve the number of cells where the perturbation can be identified (Replogle et al., 2020). But more sensitive barcode detection can lead to the inverse problem: multiple barcodes detected in droplets, due to RNA from lysed cells or other contaminants. Different methods have been used to distinguish such droplets from “true multi-perturbations”, including UMI cutoffs and Bayesian modeling, but as yet there is no consensus on the best approach.

A related issue is confirming the efficacy of perturbations, e.g. whether a knockout or knockdown has occurred in a cell containing barcodes for a specific perturbation. As a first step, it’s essential to avoid technical artifacts (e.g. PCR template switching) during perturbation library production that lead to inaccurate pairing of barcodes and perturbations (Adamson et al., 2018). But even correctly identified perturbations may be inactive in some cells. Methods used to quantify genome editing (Clement et al., 2020) are difficult to apply to single-cell data, and variation in gene expression between cells makes loss of expression difficult to measure with confidence. One approach is to have barcodes contained in the RNA transcript of the perturbation itself, although this places some constraints on both elements and does not strictly measure activity. Recently, tools have been developed to detect perturbation based on effects on the broader transcriptome (L. Yang et al., 2020a; Duan et al., 2019; Papalexi et al., 2021). This approach can be combined with redundant perturbations for each target, such that response to perturbation of a given target can be distinguished from (in) activity of a specific perturbation.

Using -omic readouts such as scRNAseq in combination with a large number of perturbation generates a wealth of data, which describes comprehensive cell states and regulatory networks responsive to perturbation. This type of data has been used to generate machine learning models predicting the effects of drugs (Lotfollahi, Wolf, and Theis 2019; Theodoris et al., 2020; Burkhardt et al., 2021) and attempt to identify effective therapeutics. In vivo pooled screening takes this approach to the next level, by directly measuring the efficacy of perturbations in the physiological environment where they would be needed. As with all computational approaches, machine learning models are powerful only when the data used to train them is appropriate for the attempted predictions. With the right data, machine learning approaches are a promising tool for understanding the interplay of processes that drive age-related decline.

In vivo pooled screening as a technology is still in its infancy, and all the more so when applied to the biology of aging. Transcriptomic changes have been used to identify aging therapies using in vivo screens in individual model organisms like C. elegans (Janssens et al., 2019) and N. furzeri (Baumgart et al., 2016; Janssens et al., 2019), and this methodology can be extended to large mammalian species using pooled screens. As described in prior sections, this is already possible and starting to happen using current technologies (Jensen et al., 2020), and ongoing technology trends promise to overcome many current limitations for screen readouts. Table 2 summarizes these emerging prospects. Ultimately, in vivo pooled screening approaches could be applied broadly across aging research to interrogate causality in a highly multiplexed fashion, in naturally aged mammalian models.

Although pooled screening approaches only multiplex cell-autonomous (also known as “cell intrinsic”) perturbations, and cannot measure organismal lifespan, they could also be used to understand the mechanisms of systemic perturbations that do affect aging and/or lifespan. For example, the use of system-wide perturbations like the application of bloodborne factors or caloric restriction could be combined with pooled CRISPR inactivation screening to determine genetic effectors of such systemic perturbations on particular cell types. This could be done, for example, by comparing the single-cell transcriptome of calorically restricted and ad libitum-fed animals, as has been done in bulk (Hong et al., 2010), and then using single-cell transcriptomic readouts to identify perturbations that eliminate the response to caloric restriction. Such approaches could reveal not only cell-type specific or context dependent regulation, but also common pathways mediating effects of such perturbations across cell types, furthering our understanding of their basic mechanisms of intercellular aging. Previous studies have combined single-cell transcriptomics with bespoke single-cell measurements like optical detection of plasma protein uptake by cells within the brain (A. C. Yang et al., 2020b). While this study suggested potential genetic regulators of plasma protein uptake, the approach was only correlative; in vivo pooled screening allows an extension to probe causality, i.e., which genes, in which cell types, modulate plasma protein uptake by CNS cells. Aging research has recently begun to identify many proteins in blood plasma that change with age (Tanaka et al., 2018, 2020; Lehallier et al., 2019; Orwoll et al., 2020), or that modulate aging (Castellano et al., 2017; Vinel et al., 2018), and the possible secretion sources of these proteins from various organs are beginning to be identified based on transcriptomic expression profiling. However, the causal, functional roles of these proteins in the aging process are poorly understood. A pooled screening approach could knock down specific cell surface receptors for these plasma factors and look for reversion of transcriptomic signatures induced by parabiosis (Ferenbach et al., 2016), to determine mediators of parabiosis factors in a massively multiplexed fashion.

While conventional assays physically dissociate a tissue before biochemical analysis, in situ assays operate by imaging directly in an intact, often chemically fixed specimen, maintaining access to the three-dimensional spatial context of each measurement. In situ readout of pooled screening inside intact tissues could have key advantages, beyond just restoring cells to their spatial contexts. For example, some in-situ readouts are able to detect subtler properties of cells, such as the subcellular localization of specific proteins (Reicher et al., 2020) or nucleic acids (Alon et al., 2021) within specific organelles or compartments, or morphological features relevant to aging, like cell membrane shape or nuclear membrane integrity (Faulkner et al., 2021). Such techniques could be used to identify, for example, cytoplasmic DNAs that have been linked to inflammatory processes in aging (Simon et al., 2019). While in-situ readout of pooled screening has so far been applied only in isolated cells (Feldman et al., 2019), in situ sequencing and multiplexed immunohistochemistry methods are already applied to intact tissues (X. Wang et al., 2018; Alon et al., 2021). In situ readout approaches could also be especially powerful for understanding the effects of cell intrinsic processes on the surrounding local environment of that cell (Yousefzadeh et al., 2021). For example, a pooled screening approach with an in-situ readout could interrogate the effects of genes on the extracellular matrix composition or accumulation of extracellular aggregates in the neighborhood of a given cell or cell type, or the role of locally secreted factors (C. Wang et al., 2021). When applied to senescent cells, such approaches could probe genes that modulate local effects of the senescence associated secretory phenotype (SASP). Other examples would include the effects of gene perturbations on the local infiltration of immune cells, or the accumulation of lipid droplets, tau tangles, amyloid fibrils or synapse loss characteristic of neurodegeneration. Ultimately, in situ readout should allow any assay accessible on fixed tissue via an optical microscope to be used as an output for pooled screening.

The application of in vivo pooled screening to aging biology would benefit from extension to proteomic readouts as well. Single-cell transcriptomics can now be combined with single-cell surface affinity proteomics, and extensions to affinity proteomics of intracellular proteins are on the horizon (Mimitou et al., 2020; Rivello et al., 2020; Swanson et al., 2021; Stoeckius et al., 2017; Chung et al., 2021). At the simplest level, this could be applied to verify the direct effects at the protein level of the perturbations used in pooled screening, i.e., whether a knockdown of a receptor actually eliminates that receptor from the cell surface, in the example above. But more broadly, much of aging biology likely occurs at the protein level. Long-lived proteins, like nuclear pore complexes, are important in aging (Toyama et al., 2013), and mapping their abundance, and ultimately their subcellular distribution with multiplexed in-situ approaches (Goltsev et al., 2018), could provide key phenotypic readouts not accessible to transcriptomics. A pooled screening approach could search for the factors that most impact the states of such long-lived proteins, distinguishing, for example, roles of genes involved in reactive oxygen species (ROS) production, chaperones, or autophagy.

Moreover, key aging regulators like FOXO or mTOR are heavily regulated by protein-protein signaling networks which depend on post-translational modifications (Smith and Shanley 2010; Y.-X. Lu et al., 2021), with key biology likely invisible at the transcriptomic level. In this context, protein based, rather than RNA based, barcoding of pooled screening perturbations should allow the use of immunohistochemistry, rather than RNA sequencing, as a readout of pooled screening. This would open up integration with multiplexed immunohistochemical stains (Wroblewska et al., 2018; Goltsev et al., 2018), or possibly with emerging genome-wide single-cell proteomic approaches (Specht et al., 2021).

Epigenetic clocks measuring DNA methylation provide a putative readout of the biological age of a tissue (Haghani et al., 2021), and such methylation measurements and clocks have recently been extended to the single-cell level (Trapp et al., 2021; Linker et al., 2019), which could provide a direct readout of biological age in pooled screens for aging modulators. Organ-specific and cell-type specific transcriptomic aging clocks are also a possibility, to invoke some integrated measure of cellular age as a readout (Meyer and Schumacher 2021). The sensitivity of single-cell transcriptomics to small variations in gene expression, or to low abundance transcripts, is considerably lower than that of bulk transcriptomics, but targeted sequencing approaches like HyPR-Seq can achieve much greater sensitivity for a defined, yet still diverse subset of genes, such as aging clock gene arrays (Marshall et al., 2020). More complex clock-like signatures computed from comparisons across species or from responses to known aging drugs (Tyshkovskiy et al., 2019) could also be used as readouts for pooled screening. Methylation measurements can be combined with single cell RNA sequencing in a manner that should allow pooled screening (Linker et al., 2019) to identify factors driving the response of epigenetic clocks to aging interventions and explore causality for this metric of biological age.

Relatedly, future application of in vivo pooled screening could have special relevance to the biology of in vivo cellular reprogramming (Y. Lu et al., 2020; Ocampo et al., 2016), which has recently come to prominence in aging research due to its apparent reversal of epigenetic age and several other aging hallmarks. The current set of transcription factors used for in vitro reprogramming may not be optimal in vivo, given the risk of teratoma formation following complete reprogramming (Abad et al., 2013). Indeed, there is some evidence that dedifferentiation and rejuvenation are biologically distinct processes (B. Zhang and Gladyshev 2020). In vivo pooled screening could be used to screen for combinations of perturbations that reverse age related changes while preserving cell type identity and without the formation of teratomas. Recently, such an approach has been taken in-vitro but could be extended to the more complex in-vivo setting (Roux et al., 2021). These approaches would synergize with new approaches on the perturbation side, such as improved epigenome editors (Nuñez et al., 2021), in support of combinatorial perturbations, e.g., generalizations of the famous combinations of Yamanaka factors used for induction of pluripotent stem cells. More advanced future approaches could potentially modulate the duration of expression of different such factors (Chassin et al., 2019), and barcode such timing patterns at the single cell level as well, or vary the relative intensity of expression of the factors using barcoded promoter libraries. On the readout side, single-cell lineage analysis may make it possible (Biddy et al., 2018) to relate reprogramming perturbations to the proliferative properties of cells, although this method has not yet been applied in vivo. Also, combinations of single-cell transcriptome sequencing with chromatin accessibility (scATAC-seq) and chromatin modification (Bartosovic et al., 2021), could be used to refine our understanding of how such perturbations impact epigenetic cell state specification.

More speculatively, it is interesting to ask whether more dynamic properties of cells can be encoded as a readout for pooled screening. At present, readouts for in vivo pooled screens are destructive, and longitudinal data is only available when biopsies (e.g. of liver or blood) can be acquired without sacrificing the animal. Recently, researchers have found ways to encode cell-cell interactions into a DNA readout (Clark et al., 2021), and others are studying prototype molecular recording systems (Tanna et al., 2020; Loveless et al., 2021) that could encode aspects of the history of dynamic gene expression into static readout (Rodriques et al., 2021), including with single-cell precision (Chan et al., 2019).

While previous studies have mostly focused on common laboratory mouse models, a key advantage of the pooled screening approach is its ability to obtain a large amount of information from a single animal, and thus it could be applicable to less common animal models as well. Genetically diverse mouse strains would be a simple example, but long-lived species like naked mole rats should be accessible as well, as both the viral vectors used for delivery of perturbations and the CRISPR proteins used as effectors tend to generalize well across mammalian species (Salganik et al., 2015). A pooled screen in naked mole rats could, for example, knock out putative tumor suppressors (or the whole genome) unique to this organism in the presence of tumorigenic agents, and identify actual tumor suppressors by sequencing barcodes present in tumors, and thereby help explain by what mechanisms naked mole rats are so resistant to tumors—even at high age.

Overall, by directly perturbing causal mechanisms in their intact-system context, while reading out high-dimensional signatures at a level of resolution that can reveal both shared features as well the heterogeneity of cell, cell-type and organ specific responses, in vivo pooled screening approaches are likely to become a general platform for accelerated study of the biology of aging, the diseases that result from it, and perturbations that seek to reverse it.