Perspectives on defining cell types in the brain

Abstract

The diversity of brain cell types was one of the earliest observations in modern neuroscience and continues to be one of the central concerns of current neuroscience research. Despite impressive recent progress, including single cell transcriptome and epigenome profiling as well as anatomical methods, we still lack a complete census or taxonomy of brain cell types. We argue this is due partly to the conceptual difficulty in defining a cell type. By considering the biological drivers of cell identity, such as networks of genes and gene regulatory elements, we propose a definition of cell type that emphasizes self-stabilizing regulation. We explore the predictions and hypotheses that arise from this definition. Integration of data from multiple modalities, including molecular profiling of genes and gene products, epigenetic landscape, cellular morphology, connectivity, and physiology, will be essential for a meaningful and broadly useful definition of brain cell types.

One of the basic roles of theory in biology is to identify meaningful groupings of individuals. Although descriptive, this work provides an organizing framework and conceptual foundation for hypotheses regarding causal mechanisms and organizing principles. The classification of organisms by natural historians such as Linnaeus, together with deep investigation of particular cases such as the Galapagos finches, set the stage for Darwin’s theoretical paradigm of evolution. Similarly, a taxonomy of cell types in the brain is ultimately required to understand how neural circuits evolved to underpin complex behaviors. Here we argue that recent developments in high-throughput single cell molecular analysis will enable a new classification of brain cell types that is unprecedented in its completeness (comprising all cells across all brain regions), quantitative precision, and integration of multiple modalities of molecular regulation (e.g. transcriptome, epigenome), anatomy, connectivity and function. There are both conceptual and practical difficulties standing in the way of a comprehensive and accurate neuronal cell type atlas, and a universally accepted and fixed taxonomy may remain an elusive goal. Nevertheless, the impact of a high-quality cell census will be broadly felt across developmental, molecular, and even computational neuroscience.

Neuroscientists have identified vastly different neural circuits across brain regions and species, yet it is difficult to think of a functional circuit in any complex organism that does not involve multiple clearly identifiable, distinct neuronal cell types[1–4]. The striking diversity of neuron types in the mammalian brain suggests that complex behavior relies not only on expanded cell number but also on an increasing functional specialization that allows particular neuron types to play specific information processing roles. If that is the case, then information processing in the brain can only be understood by identifying the specialized roles and interactions among brain cell types. Indeed, it is not unreasonable to suppose that the evolutionary processes by which sister cell types duplicate, differentiate and specialize are driven by key functional needs [5], and that understanding the relationships among brain cell types will help to organize our understanding of the information processing functions which these circuits evolved to support.

Cell types as self-stabilizing regulatory programs

Different research communities use the concept of cell type in divergent ways, ranging from highly specific notions of identifiable single neurons in Drosophila [6,7] to formulations based on the connectivity [8] or the functional or computational role of a neuron [9]. Although different cell categorizations may be appropriate for different purposes, this diversity of definitions raises the question of whether there is a general, if not universal, notion of cell type that can be useful for a broad range of neuroscientific questions. Indeed, gene expression, cell location, morphology, connectivity, and physiology entail independent measurements but these properties are rarely independent variables.

Here we propose to define a cell type as a self-stabilizing system composed of specific genetic and developmental processes. We are inspired by Waddington’s original concept of canalization, i.e. the process by which a phenotypic outcome is produced and stabilized in the face of a range of environmental or genetic perturbations [10]. Accordingly, we suggest that each cell type corresponds to a self-stabilizing regulatory program, which acts to maintain and restore the cell type-specific program of gene expression (see Sidebar). We focus on self-stabilizing gene regulation, i.e. interactions among cell-intrinsic factors (principally genes and epigenetic marks) that form recurrent functional networks with feedback loops that preserve their structure. It is likely that, in some cases, structural factors (e.g. laminar position or connectivity) or physiological interactions (e.g. thalamic innervation) could be important for maintaining cellular identity as well [11]. In contrast with the core cell type-defining features of a cell type, other downstream properties, such as the expression of effector genes [12], may vary over time or in response to extrinsic signals but will generally revert to a canonical pattern induced by the self-stabilizing program.

The concept of dynamical stability is familiar in computational neuroscience, where it forms the basis of attractor networks for memory storage such as the Hopfield model [13]. A network of neurons, interacting via excitatory and inhibitory synapses, can encode multiple attractor states which can be retrieved through appropriate dynamics [14,15]. Similarly, networks of DNA binding transcription factor proteins can regulate their own expression to form stable, recurrent attractors [16]. An attractor of a dynamical system need not be a static fixed point, but could correspond to a limit cycle with periodically repeating properties as in a mitotic cell type. Importantly, an attractor network provides homeostatic, self-stabilizing interactions that ensure that perturbations which move the state of the network away from an attractor will be compensated to restore the equilibrium. At the same time, such networks afford the opportunity for state changes when a signal or perturbation pushes the network out of the basin of attraction of one attractor and into a different equilibrium; experimentally, this is the basis for recent progress in transdifferentiation and artificial induction of neuronal cell types [17]. Despite the intuitive simplicity of the concept of a stable attractor, dynamical systems theory implies there may be important distinctions among cell types in terms of the type of attractor (e.g. fixed point, limit cycle, or chaotic strange attractor) [18] and the types of transitions (bifurcations) they may undergo [19].

Single cell assays advance the search for cell types

The recent development of high-throughput single cell transcriptomic (RNA-Seq) and epigenomic assays (mC-Seq for DNA methylation and ATAC-Seq for open chromatin) has raised expectations for a new and more detailed empirical analysis of brain cell types [20]. Indeed, pioneering studies, reviewed by [21], have been followed by a dramatic increase in the number of single cell transcriptome studies and in the number of profiled cells per study [22–26]. An example of the high resolution for fine cell type distinctions is a recent description of 133 cell types in two mouse neocortical regions [26], a level of complexity that was hardly recognized before the advent of single cell transcriptomics. We expect that these data resources will enable fine-grained analysis of self-stabilizing cell type regulation across the diversity of brain cells.

While our proposal seeks a unifying conceptual basis for cell type, in practice we must acknowledge the limits of objectivity and the inevitability of disagreement: there will always be lumpers and splitters, each with valid empirical data and arguments to support their views. The seemingly simple question, “How many brain cell types exist?” is more difficult to answer than it is to state. Instead of attempting to directly estimate the number of cell types, we propose to focus on objective empirical criteria that could contribute to meaningful discussion of cell type distinctions. One proposal would follow the tradition of systematics, which classifies species by cross-referencing as many traits or features as possible [27]. According to this view, any cell type classification should simultaneously account for the similarities and differences between cells along multiple dimensions, including gene expression (transcriptome), epigenomic state, anatomy (laminar location, dendritic and axonal morphology), connectivity, as well as electrophysiological properties. To this list, we would add that a comprehensive understanding of a cell type should include an account of its role in processing information in the context of local and distributed circuits. By emphasizing self-stabilizing features, we focus on those cellular features most likely to be central to the functional role of a cell type and therefore to co-vary with many other cellular properties.

Cell type definitions: predictions and testable hypotheses

Our proposal to base the definition of cell type in the concept of a dynamically stable state implies several key empirical predictions. We would predict that some features of a cell type, such as expression levels of core transcription regulating genes and corresponding configurations of epigenetic marks, are critical for stabilizing the cell’s identity. These cell type-defining characteristics appear to emerge after the final mitotic division, as the cell enters a developmental ground state [28]. Other features, such as the expression of activity dependent genes [29], may vary across cells or over time without altering the cell’s stable type.

Dynamic gene expression across cells of the same type might be observed by using single cell transcriptome data to estimate the rate of change of gene expression, or “RNA velocity” [30]. Computational and statistical modeling of transcriptomic measurements from a range of neuron types could indicate which transcription factors are the core regulators of cell type identity [7]. Such models might take inspiration from biophysical simulations of neuronal electrical dynamics, which showed that specific conserved electrophysiological behaviors such as central pattern generators can be produced by a variety of combinations of molecular components [31]. Similarly, the core transcriptional regulators of cortical neuron identity were inferred from single cell sequencing data [22–24,32,33]. Such information could provide testable causal hypotheses, for example predicting combinations of transcription factors which may be used for direct reprogramming [17,34].

Perturbative experiments - such as transgenic manipulations - provide a powerful approach for demonstrating the causal role of specific molecular regulators in generating self-stabilizing cell types with distinct functional properties. For example, a network of key transcription factors comprising Ctip2, Fezf2, Satb2 and Sox5 were shown to determine Layer 5/6 projection neuron fates in the developing mouse neocortex [35–37], while Sox6 is critical for specifying interneuron cell fates from spatially distinct progenitor cells [38]. Similarly, a single transcription factor, Brn3, was shown to both establish and maintain neuronal identity of a population of medial habenula neurons [39]. Recent studies suggest that post-transcriptional regulation, including alternative splicing or RNA modifications such as methyl-6-adenosine, play a role in shaping neurons’ overt phenotype [40,41]. Given the recent progress in single cell transcriptomic and epigenomic assays that can subdivide neurons into types with increasing granularity, we may soon be able to propose causal experiments that could validate fine cell type distinctions by perturbing their putative (combinations of) regulators. In the interim, it would be useful to distinguish between “provisional cell types” defined by single cell -omics alone, and “verified cell types” for which more detailed causal data and mechanisms of stabilization are available.

Epigenetic regulation and self-stabilizing networks

Self-stabilizing gene regulation can be achieved through direct interaction among transcription factors [42,43], and these connections may be discerned in transcriptomic data through co-expression analysis. However, epigenetic modifications of histone proteins and DNA play important roles in regulating gene expression as well. Therefore, understanding the epigenetic landscape at the single cell level, in parallel with single cell transcriptome analysis, provides an independent evaluation of the cell’s molecular identity. Epigenomic data are particularly useful for neurons, which have uniquely abundant non-CG DNA methylation as well as hydroxymethylation [44,45]. These epigenetic marks can have complex consequences, including roles in both repressing and enhancing transcription factor activity [46], which could impact the self-stabilizing interactions among core cell type defining factors. In some cases, the same gene is expressed in multiple neuron types as a consequence of different, cell type-specific epigenetic drivers. The use of unique combinations of enhancers in each cell type can lead to differences in the regulation of gene expression, isoform usage [47,48], or plasticity and activity dependence of expression [49]. Recent advances in single cell bisulfite sequencing [50,51] and chromatin accessibility profiling [52–55] make it possible to measure these epigenomic signatures in thousands of cells, directly complementing large-scale transcriptomic studies [20,25,51].

Currently, larger datasets and more studies are available for single cell transcriptomes than epigenomes (DNA methylation and open chromatin). Although single cell epigenomic assays do not necessarily resolve a greater number of cell types, they provide an additional layer of regulatory information that is not available from RNA-Seq alone. For example, single cell DNA methylomes and ATAC-Seq data can indicate cell type specific active enhancer regions [50,56]. Such epigenomic information is valuable as a stable marker of cell identity that may be less sensitive than gene expression to changes in cell state, including circadian rhythms and activity-dependent gene induction [29]. New methods for joint measurement of epigenomic and transcriptomic information in the same cell could allow linking these complementary measures [57,58]. Given the more recent development of single cell epigenomic methods, we expect continued rapid improvements in these assays to enable recognition of cell types with similar resolution compared with the more mature single cell RNA-Seq methods.

Cell type classification based on these molecular data remains challenging, in part due to the sensitivity of results to different experimental and analytic parameters [29]. Computational methods for combining information across experimental batches and even across data modalities hold promise for reaching broad consensus [59,60]. Methods for statistical cross-validation of cell types from independent datasets are also critical [61,62].

As rich, high-quality molecular data resources become increasingly available for neurons as well as other cell types (e.g. Human Cell Atlas [63]), we expect increasing opportunities to apply sophisticated machine learning and artificial neural network based analyses to the challenge of unsupervised and semi-supervised learning of structure in these datasets [64–66]. Directly complementing the transcriptomic and epigenomic data, assays of chromosome conformation (e.g., Hi-C) at the single cell level can identify intra- and interchromosomal interactions that may represent the 4-dimensional physical manifestation of gene expression regulatory networks in the nucleus [67,68].

Integrating data from multiple modalities: benefits and challenges

Despite the power of high-throughput single cell transcriptomic and epigenomic assays, these modalities likely do not fully reflect all the relevant differences between cell types. For example, transient expression of genes during differentiation can lead to situations in which mature neurons with different projection patterns show indistinguishable transcriptomic profiles [6]. Such distinctions may be evident at the epigenomic level, but in some cases information about anatomy, morphology, connectivity and/or physiology will be needed to fully distinguish neuronal cell types. Indeed, the concept of a self-stabilizing network need not be limited to genes and epigenetic regulators, but could include physiological feedback loops [69].

From these considerations, we argue that there will be substantial benefit from integrating empirical data across multiple modalities, both as a means of cross-validation and to provide greater precision and accuracy in cell type assignments. Methods that can provide spatial context for transcriptomic data via computational analysis [70] and in situ sequencing [71] can help connect molecular and anatomical information. Moreover, molecular data can provide a stable platform on which to build and assemble more detailed information from other modalities that are needed to complete a comprehensive characterization of cell types. For example, data that define a cell type at the molecular level can be used to create transgenic lines based on genes or enhancers that are specifically active in that cell [72,73]. Such tools provide the experimentalist with genetic access, allowing analysis of the cell’s functional properties. They would also provide a means of perturbing the cell genetically, optogenetically and chemogenetically to test predictions about the cell’s identity and function based on its transcription and signaling factor networks, connectivity, or physiology. Alternatively, unbiased experimental techniques for simultaneous measurement of physiology and gene expression (e.g. Patch-Seq) may help to link these modalities [74–76]. Such multi-modal data showed, for example, that an apparently continuous distribution of parvalbumin expressing neurons in the dorsal striatum has a corresponding continuous gradient of electrophysiological properties [74].

In contrast to molecular modalities, however, quantitative measurement of anatomy, physiology and connectivity is less straightforward. Analysis of these data requires first defining the key, cell type-defining features and further estimating these from the data, tasks for which no settled consensus is available. For example, optical microscopy can provide information about dendritic and axonal morphology and connectivity, but extracting these parameters is challenging due to the high dimensionality of image data. Here again we expect that sophisticated computational tools, in particular computer vision algorithms such as deep convolutional neural networks (CNNs), will play a key role in objectively quantifying traditional parameters of cellular morphology and connectivity with high-throughput and across the entire brain. Indeed, neural networks have already proved useful for automating and improving the annotation and low-level quantification of anatomical data from high-throughput brain-wide anatomical data [77,78]. These computational approaches may also help to discover new regularities and features of neuronal anatomy that may not be easily extracted and quantified in traditional, manual analyses.

Outlook

The challenges of measuring and integrating data about neuronal cell types across multiple modalities are large and exceed the capabilities of any one research group. Moreover, the value of a cell type atlas is directly tied to its broad acceptance and utilization by the neuroscience community. For these reasons, collaboration is critical, and collaborative consortia, including the NIH BRAIN Initiative Cell Census Network (BICCN) to which we contribute, aim to build consensus through joint development of cell type resources. Building on the experience of the previous BRAIN Initiative Cell Census Consortium [20], BICCN members contribute a broad range of expertise and techniques, including high throughput single cell transcriptomic and epigenomic assays which will form the basis for a comprehensive survey of molecular cell types in the mouse brain. In parallel, complementary data from human and non-human primates, though less comprehensive, will enable comparative and evolutionary perspectives on mammalian brain cell types [79,80]. Importantly, the BICCN is also investing in large-scale anatomy, morphology, connectivity, as well as physiological investigations. The long-term, aspirational goal is to integrate information from all of these modalities to provide a comprehensive taxonomy of brain cell types. Although this collaborative and coordinated strategy is critical, organizing a multi-site, multi-platform analysis does bring challenges of coordination, consistency in experimental protocols, analysis, metadata and data formats, and effective distribution of data to the larger scientific community. In addition to the BICCN, other consortia such as the HCA and the Human Biomolecular Atlas Program (HuBMAP) are also attempting to address these challenges and to build large-scale cell type resources [63,81]. These consortia represent experiments in organizing neuroscience research, and the sociological experience and lessons of these endeavor will be valuable dividends of the projects.

Here we have outlined a conceptual perspective emphasizing the self-stabilizing nature of cell types, which may help to guide otherwise subjective debates about the appropriate division between cell types. As Darwin recognized, it is healthy for lumpers and splitters to challenge each other over the empirical and theoretical validity of their frameworks. Their arguments will not go away, but with new molecular and functional data and more data-driven theoretical constructs we look forward to a productive discourse in the years ahead.

Highlights:

• Cell types correspond to self-stabilizing regulatory networks that robustly establish and maintain a core phenotypic program

• Single cell molecular profiling of transcriptomes and epigenomes provides comprehensive and unbiased cell type information

• Molecular datasets must be integrated with each other and with complementary data about cell morphology, location, and physiology

• Despite continuing experimental and conceptual challenges, collaboration among a wide range of specialists will lead to a more mature and data-driven understanding of the brain’s cellular parts list

SIDEBAR.

Self-stabilizing regulation of olfactory neuron identity

Our proposed definition of cell type emphasizes self-stabilizing regulation. Here we explore the relationship between feedback, stability and deterministic vs. stochastic processes. Feedback is a key feature of recurrent networks, i.e. systems in which each element causally affects, and is reciprocally affected by, activity in other elements of the network. erhaps the simplest paradigmatic example from computational neuroscience is the interaction between excitatory and inhibitory neural populations, as in the classic Wilson-Cowan model [82]. Here, increased spiking activity in a population of excitatory neurons stimulates increased activity in locally connected inhibitory neurons, which in turn leads to inhibition of the excitatory neurons. Depending on synaptic weights, such a system can have different attractors, i.e. static or periodically varying (oscillating) trajectories that are stable against perturbations [83]. Negative feedback between the two neural populations ensures that any external or internal stimulus, such as afferent input from another neural population, may temporarily move the network away from the attractor but will ultimately be reversed. This basic model explains the dynamic stability of neural circuits which maintain activity levels within a physiologically normal range, and it can explain the breakdown in equilibrium and the generation of epileptic activity following the loss of inhibition.

Complex networks with more than two elements can stabilize multiple stable attractors, i.e. alternative configurations which are each stable against limited perturbations. A simple type of attractor network appears to govern the choice of a single odorant receptor (OR) out of >1,000 OR genes during olfactory sensory neuron differentiation [84]. Here, the initial expression of an OR gene results from stochastic demethylation of overlying repressive histone methylation marks. Expression of a functional OR leads to expression of Adcy3, which in turn promotes the OR expression (positive feedback) and also prevents derepression of other OR genes by inhibiting histone demethylation (negative feedback) [85]. This “epigenetic trap” mechanism ensures that only a single OR is expressed in each mature olfactory sensory neuron, and it further enables a cell that fails to express a functional OR to select a new OR for activation. This simple network shows how thousands of distinct attractor states, each corresponding to expression of a single OR, can be encoded and stabilized through epigenetic feedback.

Our concept of dynamic stability should be distinguished from the question of deterministic vs. stochastic regulation of gene expression in single cells [86]. A self-stabilizing system can have stochastic state transitions, as in the stochastic selection of a single OR gene during olfactory sensory neuron differentiation. Moreover, random fluctuations in gene expression, e.g. due to transcriptional bursting, can cause ongoing stochastic differences between the molecular state of individualcells. However, the self-stabilizing dynamics of the core regulators of cell identity serve as a buffer that will prevent stochastic fluctuations from altering the critical, self-reinforcing pillars of the cell’s machinery.