Scientist Stories: Hongkui Zeng, Understanding Brain Cell Type Diversity

Today I'll tell you about our recent work in trying to understand the cell type diversity in the brain. To start with, it is the collective goal of neuroscience to understand how brain works. This is a challenging task. This is a daunting task given the complexity of a brain.— with millions to billions of neurons and a similar number of non-neuronal cells in a mammalian brain and extremely complex interactions through synaptic and non-synaptic connections among all these cells. In order to tackle this complexity, it is essential to take a multi-level approach to try to uncover principles of brain organization at these different levels. At the Allen Institute, we have established a series of approaches trying to address basic neuroscience questions at all these levels as well, just like what the community is doing. Our primary focus is on the cortex, especially on the behavior and functional side in visual cortex. Of mouse. However, the largest program at the Allen Institute is actually focusing on the first level, trying to understand the diversity of cell types and individual cells in mammalian brains. So why studying cell types and how to do that? I think we all agree that in order to understand a complex system such as a brain, an essential first step is to create a parts list, namely a cell type composition of that system, the brain. Once we have a parts list in hand, an important advantage is to be able to gain genetic access using molecular genetic approaches to label and target those cell types. And those genetic tools would allow reproducible investigation of the same populations of cells or cell types across different labs. By having this kind of effort will not only allow us to categorize known types but also present the opportunity to discover new cell types as well as the organizing principles regarding those types. Given that cell types arise from elaborate evolutionary and developmental processes, knowing about cell types themselves will also help us to understand evolution and development. And finally, of course, it is important to have the parts list and see how different cell types change under diseased conditions, which will allow us to understand the disease themselves better. Given all this importance, how do we generate a comprehensive, if not complete, parts list, cell type category for the brain? We believe that it is important to take a systematic and quantitative and objective approach to profile individual cells and use quantitative phenotypic measurement to perform data-driven classification approach to come up with the category. Importantly, we need to profile the individual cells at different phenotypic levels. For example, at gene expression molecular level, at physiological level and at morphological and connectional level and seeking congruency or consistency or correspondence among those different properties in order to arrive at a unified definition of cell types. So with this kind of understanding, at the Allen Institute, we have established really a multiple set a set of technology platforms for single-cell characterization to use the information to define cell types and then use this kind of a cell-type-based knowledge to investigate connectivity and circuit function. So our front-end approach is the single-cell omics, especially single-cell transcriptomics method, in which we can profile hundreds of thousands to millions of individual cells and generate a transcriptomically defined cell type taxonomy for a particular region or even at a brain-wide scale relatively quickly. At the other end of the spectrum of our approach, we investigate connectivity among cell types and individual cells at both anatomical and physiological levels, at both a brain-wide scale as well as at a local, high-resolution local circuit level. As I said earlier, we also study the in vivo activity of the different cells or cell types using both calcium imaging and electrical recordings. Approaches in awake behaving animals and try to correlate the properties of the individual, the activities of the individual cells with their behavior responses. Now, how to bridge those two very different types of information? We use three approaches to do that. The first is to continue to generate more refined genetic tools. We started with generating a large battery of transgenic driver lines using genes marking specific molecular types. In more recent years, we have switched into developing viral vectors that utilize enhancer elements discovered in our single-cell transcriptomic and epigenomic dataset to target cell types in a more efficient manner because viruses provide more flexibility and versatility in labeling different cells in both genetic species as well as non-genetic species. The second approach we have been taking is to use a multimodal single-cell characterization method called PatchSeq. In which we patch into individual cells, obtaining their biophysical properties and intrinsic firing properties, their morpho— their local morphologies using biocytin fill, and then extracted the content of the recorded cell and perform single-cell transcriptomics. Then we perform integrated classification by combining all three modalities of information together. To come up with what we call a unified definition of cell types. The importance of this kind of approach is that it then provides a Rosetta Stone that bridges— directly links molecular identities of the individual cells with their physiological properties and their local morphological properties. And then this kind of information can be used then in other data sets, for example, in the EM connectomic data set that allows us to decode the molecular identities of the cells in the EM data purely by their morphological features. And the third approach we take is a spatial transcriptomics approach, for example, using multiplex FISH. This approach allows us not only to obtain the spatial distribution and relative proportions of the different cell types, molecular cell types, it can also serve as a post hoc method to decode the molecular identities of the cells on the tissue examined that have been— that have gone through connectional or functional studies. So to again assign molecular identities that we have obtained connectional and functional properties. So I should say that now this set of comprehensive approaches has been integrated into the community-wide cell type atlasing efforts, specifically in NIH BRAIN Initiative Cell Census Network. We are a major player in the BICCN, in which we collaborated with a number of PIs across the country and the world to use the same set of approaches that I just told you about to profile cells in several different mammalian species, including mouse, non-human primate, and human itself, with the goal of generating molecular taxonomies across species, spatially resolved molecular taxonomies, and multimodal integrated definition of cell types, and finally, input-output connectivity wiring diagrams of those cell types. And we have now expanded this effort into profiling cells across— as a first step across the entire mouse brain. And we are making really great progress in that. We hope to generate a whole mouse brain-wide cell type atlas within the next 2 years. So with that, I want to tell you a few specific studies to tell you about what we have learned from this kind of a systematic effort. So our work started with profiling cell types in the cortex, in the mouse cortex. So here is just a very quick introduction of the major classes of neurons found in the cortex. The first one is, of course, the glutamatergic excitatory neurons. Those neurons are long-range projection neurons that can be divided into several major subclasses. The largest subclass is the intratalamic cephalic, or IT, neurons, which are distributed across all the layers of the cortex from layer 2-3 to 4 to 5-6. The second major subclass is the cortical-subcortical projecting pyramidal tract neurons or subcerebral projection neurons. Those are the large layer 5 neurons that send extensive axonal projections to multiple cortical and subcortical regions. And the third major subclass of long projection neurons are the corticocelamic neurons. That are located in Layer 6. In addition to those well-known ones, through our transcriptomic studies, we also uncovered two other major subclasses of excitatory neurons. They are the Layer 5-6 near-projecting neurons as well as the Layer 6b subplate neurons. So the second major class is the GABAergic local-projecting inhibitory neurons. The cells within this class are also tremendously diverse in their morphological properties, especially their local axonal projection patterns and their electrophysiological properties. And they are marked and classified traditionally through a set of neurochemical markers. So as I said earlier, our front-end approach is to use large-scale single-cell transcriptomics to establish an initial framework for cell type classification. At the institute, we have established two separate pipelines to do this. One is the full-length deep sequencing SMARSeq approach. And the other is the highly scalable 3'-based shallow sequencing 10x. Approach. In the first set of studies that we published, we used the SmartSeq approach to sequence tens of thousands of cells in two mouse cortical areas and one human cortex. We found that the transcriptomic types derived from those studies follow a hierarchical organization. And there's a high degree of similarity or conservation of major cell types at the higher branches of those hierarchies. We also observed extensive species differential variation of gene expression, especially at a lower branch level. We also performed cross-area comparison of cell types in the mouse. And this is the study we did using two cortical areas, at the extreme ends of the cortical sheet, anterolateral motor cortex and primary visual cortex. The major finding of this comparative study across two cortical areas is that we found that the GABAergic inhibitory neuron types are shared between the two cortical regions, whereas the excitatory glutamatergic neuron types, the projecting neuron types, are distinct between those two regions. This difference is not only reflected in how the cells, individual cells, contribute to each of the clusters, but also the number of differentially expressed genes found in the same or the homologous pair of cell types between those two different regions. So now, what happens if we extend this study across the entire cortex. So recently, we have completed a much larger study of single-cell transcriptomics that now covers the entire isocortex as well as the neighboring hippocampal formation region. So here shows all the regions that are covered. The isocortex, also called neocortex, is basically this entire neocortical sheet We profiled all the sensory areas and motor areas, as well as all the associational areas in the frontal, medial, and lateral part of the neocortex. For the hippocampal formation, which is also considered as an archicortical structure, we profiled not only the hippocampus itself, including all the subfields, of hippocampus, but also its neighboring subiculum and para-pre-postsubiculum complex, as well as the lateral and medial entorhinal cortex. So this is really an extremely comprehensive coverage of the entire cortical sheet. The study involves using both PanEx and SmartSeq methods with more than 1.3 million individual cells. So we integrated those two different data sets together, generating a consensus transcriptomic taxonomy tree. And as you can see here, the number of clusters has greatly expanded from the two cortical area studies that I just showed you earlier. We found that the rules that we identified in a previous study still apply in this new dataset. At the same time, the rules have also evolved. On a GABAergic branch, we found that most of the cell types, transcriptomic types, are still conserved across all the regions as shown by this rainbow color. Representation of where the cells come from. At the same time, as you can see, that we also identified some highly specific clusters that appear to be region-specific, and all of them are shown in this kind of greenish color, indicating that they come from the different regions of hippocampal formation. So in addition to the shared GABAergic types, we also found a set of GABAergic cell types that are unique to hippocampal formation. On the glutamatergic side, the number of clusters dramatically increased. We still see a high degree of regional specificity, especially all those green-colored clusters which come from hippocampal formation. At the same time, as you can also see here, now there are many cell types that are shared. Among regions, especially among neighboring regions. And I'll show you more about this in the next few slides. So our further investigation is on the excitatory glutamatergic types. First of all, we compared those excitatory neuron types between cortex and hippocampal formation. And this is shown by this UMAP. That's colored either by the major subclasses of the glutamatergic neurons or the regions where the cells come from. As you can see here, the major continents of this UMAP are all occupied by isocortical or neocortical neurons, and the cells are largely actually intermingled across all those different areas. The major subclasses I mentioned earlier are all represented here. The largest group is the IT neurons. The PT neurons are here. Layer 6 cortical thalamic neurons are here along with the Layer 6 B and MP neurons. There's a special subclass of IT neurons which contain cells present in the lateral cortex and also shared with the claustrum. A small cortical nuclear region underneath the agranular insular cortex. In addition to those major continents, there are many now highly distinct islands of cell types that we identified. And all those cell types, as you can see, come from hippocampal formation. Nearly all of them come from hippocampal formation, this greenish color. And those islands of cell types are strategically located nearby isocortical major subclasses, suggesting a degree of homology between hippocampal neuron types and the cortical neuron types. We further investigated this homology by doing a direct homology search for each hippocampal formation cluster to find their nearest neighbor within the isocortex. And we found that amazing— there's a highly significant correspondence of specific hippocampal formation clusters and subclasses with specific projection types or subclasses in neocortex. And we are able to find the corresponding cell types in hippocampal formation for all the 6 layers of major subclasses of cortex. And this kind of homologous relationship is not only demonstrated at this molecular gene expression level. We also identified the major marker genes for each of those major cell types and found that those marker genes have similar laminar distribution, corresponding laminar distribution as their homologs in the isocortex, suggesting that through both molecular and anatomical evidence suggesting that they are indeed bona fide homologous pairs. The second major finding from the study is that in addition to the many individual-specific cell types that we identified, We also now found a large-scale gradient-like continuous distribution of cell types in both cortex and hippocampal formation. So here, first illustrate a vertical gradient distribution of all the IT neuron types across all 6 layers. We found that the transcriptomic clusters or subclasses are actually continually distributed from superficial to deep layers, and in many cases without following a clear boundary between layers. In addition to this vertical gradient distribution of cell types, we also found large-scale horizontal distribution of cell types. As I showed briefly in the UMAP earlier, that there are many now cell types that are shared across different cortical regions. And here is a detailed examination of the cell type distribution colored by the anatomical regions, the cortical regions they come from. As you can see here, all major subclasses of cortical neurons have this kind of continuous distribution. —with only some regions that are highly distinct from the others. And those regions include the medial anterior cingulate, retrosplenial cortex, and the lateral temporal, perirhinal, and ectorhinal cortex. Other than that, there's a really great anterior-posterior continuous gradient distribution of cell types, even including the prefrontal cortical areas. The gradient distributions of the different cell types are not all exactly the same across the different subclasses. Each subclass has its own distributed patterns, and they may be more continuous or more discrete. In particular, we found that the layer 5 PT neuron types, the cortical-subcortical projecting types, appear to be more region-specific than all the other types. And the combination of this kind of gradient distribution of shared types as well as the region-specific expression of the common types, we are able to perform a meta-clustering analysis to derive this kind of hierarchical relationship actually now across different cortical regions.. So we can define the relatedness of the different regions, and we found that we believe this shows that the cortical regions can be organized into what we call modules. There's a sensory module, there's a motor or frontal module, there's a medial associational module, and then there is a frontal associational module, and finally there's a lateral associational module. And these modules are defined by the molecular— the transcriptomic types. And we believe these molecularly defined modules are very consistent with the cortical modules that have been identified in connectivity studies as well as in brain imaging studies. Finally, we also uncovered multidimensional gradient distribution of cell types in the hippocampus and its nearby subiculate region. So the gradient distribution or graded distribution of different cell types in hippocampus has been known for a number of years now in a number of individual studies. However, the nature of this kind of a distribution was not clear in small-scale individual studies, and there's actually much controversy and debate about the exact nature of the gradient distribution. Here, using the large-scale transcriptomic data that encompasses all those different regions, we are able to visualize all those different gradients together in one picture and identify individual gene sets mediating the different dimensions. Of gradients. So overall, we uncovered 4 major dimensions of distribution of cell types or gene expression patterns across hippocampus and subiculum. And those 4 major dimensions are proximal-distal, dorsal-ventral, superficial layer versus deep layer, and finally, high activity versus low activity. We think this kind of— this really demonstrates the power of this kind of comprehensive study. So in summary, here is a diagram showing all the major cell types, excitatory cell types that we identified across all different areas of isocortex and hippocampal formation. And we derived two major findings from this study. The first is the identification of a parallel cell type organization between hippocampal formation and isocortex, which encompasses all six layers and all major subclasses of neurons in hippocampal formation itself as well. And we believe this is the first time identification of a near-complete repertoire of cell types in hippocampal formation, revealing information that was previously unknown before. Secondly, through the combination of region-specific cell types and gradient distribution of shared types, we are able to define cortical modularity and regional specificity the isocortex or neocortex. And we believe that this kind of comprehensive single-cell transcriptomic study in the adult stage is able to shed light on the potential evolutionary and developmental relationship among all these different individual types. So now I'll move on to the second study that I would like to talk to you about. Which is a multimodal integrative analysis of cell types by patch-seq. In this approach, we integrate electrical recording, anatomical localization of individual cells, transcriptomic profiling, and morphological reconstruction together. And in our first large-scale study, we profiled more than 4,000 individual cells, GABAergic interneurons, in one cortical region from the mouse visual cortex. We are able to map all those individual cells with our established transcriptomic taxonomy tree. And we uncovered a high degree of, again, layer-enriched distribution of many different types, especially the smart-stating and parvalbumin interneuron types. As well as a gradual transition across the different clusters through those layers. So a correspondence of transcriptomic types and their layer distribution. We are also able to classify the cells based on their electrical properties as well as their morphological properties. So how do morphoelectrical properties correlate with the transcriptomic properties? Here, we performed the independent classification or clustering analysis using only morphoelectrical information to derive this kind of ME types. And we compared the ME types with the mapped T-type identities of the same cells. This correspondence diagram shows that the correspondence between T and ME types is excellent. The cells are all falling, the clusters are all following along the diagonal. At the same time, it is also important to point out that this correspondence is not perfect. There's often no single one-to-one correspondence at all. It's often one-to-many, one-to-several, or several-to-several correspondence, suggesting a degree of fuzziness. Among the similar or neighboring types, either in the ME space or in the transcriptomic space. So then in order to integrate all the different information together, we used the mapping statistics to look at the different ME or T types and found that they form this nicely tight groups of cells which we then define as MET types. So through this kind of analysis, we're able to identify— define 28 MET types for the mouse visual cortex. And this is a reduction from the 61 transcriptomic types that we identified. And we showed that the MET types, the cells within each MET type displays high degree of congruency of all three modalities of properties. And importantly, high degree of mutual predictability between transcriptomic identities and the morphoelectrical properties of the individual cells. And as I said earlier, this mutual predictability is important, is actually critical for us to use the MET types as Rosetta stones to bridge disparate sets of data and investigation. So this is GABAergic interneurons. How about the long-range projection neurons, which we cannot study their morphologies fully in brain slices? So for the long-range projection neurons, which we knew have extensive transcription atomic diversity already, we used a single-cell full morphology reconstruction approach to characterize them. So this is yet another platform or pipeline that we established in collaboration with multiple groups in China for both whole-brain high-resolution fluorescent imaging using the FMOST imaging method as well as registration of the whole brain data to a common coordinate framework that we developed and the reconstruction of individual cells through two other groups in China. So this entire pipeline allows us to now reconstruct and classify now more than 2,000 individual neurons, excitatory neurons, across a number of different brain regions. Here I just show you some examples of extensive morphological diversity that we uncovered in several major subclasses of cortical neurons. Here shows the layer 2-3, only the layer 2-3, intratalamic serpentine neurons, the cortical projecting intratalamic serpentine neurons. As you can see here, just by By this simple example of 10 neurons from the motor cortex, every cell has a different projection pattern. And it doesn't seem to follow any kind of rules. It doesn't follow— within layer 2-3, the pattern does not follow the depth distribution within layer 2-3. And see that each neuron appears to select a subset of the projection targets from the population projection, overall population projection patterns. And this selection of the subset appears to be random in the adult stage. Similarly, for the long-range projecting layer 5 cortical-subcortical projecting PT, neuron types, we again uncovered a tremendous degree of diversity among individual cells. At a high level, the layer 5 PT neurons from the motor cortex can be divided into a modular projecting and a non-modular projecting type. And this is consistent with our previous study showing that those two major projection types correlate with two major transcriptomic types. So there is an example of correspondence here. However, beyond that, there's additional individual variation within each of those major projection types that cannot be explained and that do not really correspond with transcriptomic types yet. Of course, further investigation will be needed, for example, by combining morphological, full morphological studies with spatial transcriptomics multiplex FISH or BARseq type of studies to further investigate this relationship. However, at least for now, we don't see an obvious correspondence between morphological diversity at a refined level and transcriptomic diversity. And here's one last example. Showing a group of highly distinct cortical neurons. This is the CAR3-type IT neurons that I showed you earlier that are shared between a number of lateral cortical areas and the claustrum. Those cells collectively form a highly homogeneous transcriptomic cluster or type. And yet within this transcriptomic cluster, each individual cell exhibits a completely different projection pattern. There's tremendous diversity in the individual cell projection patterns and morphologies. And we can use the morphology to classify them into multiple clusters that are largely organized by topography. Or their positions in the anterior-posterior dimension in both cortex and a claustrum. So this is another just amazing example showing the divergence between the homogeneity of transcriptomic properties and the high degree of heterogeneity of the morphologies and projection patterns at the individual neuron level. So finally, in the last two slides, I will tell you about one last piece of study or a direction that we are going. And this is really to establish a correspondence between single-cell transcriptomics and single-cell epigenomics. And from there, to try to understand Cell Type Specific Gene Regulatory Landscape. This is a collaboration project within the BICCN with several groups at the Broad Institute as well as UCSD and Salk Institute. So in this study, we generated 8 different single-cell datasets, 6 of which are single-cell transcriptomics using a variety of different methods. And the two datasets are the epigenomics profiling. One is ATAC-seq, and the other is methylation-seq. Again, using an integrated computational approach, able to integrate all data sets together into a unified molecular taxonomy in which every individual dataset map very well into this integrated taxonomy. This shows the cells from each of those 8 datasets. Importantly, by deriving this integrated taxonomy, we are then able to correlate chromosomal landmarks, methylation sites, and chromatin accessibility peaks with the gene-specific gene expression patterns in a cell type-specific manner. And through this integration, we are then able to identify chromatin accessibility peaks or putative enhancer motifs that we then use to generate viral vectors that target— that allow us to target different cell types. So we just have two papers published today showing the proof of concept of this approach. We generated a series of enhancer-based viral vectors that allow us to target major excitatory and inhibitory neuron types, actually across multiple mammalian species. Those enhancer elements work not only in mouse, but some of them also work well in non-human primate and even in human. Cortical slices as well. So with that, I want to conclude my talk by providing the summary here. I have shown that large-scale single-cell transcriptomics provides a framework for cell type classification. And we found that different cortical regions have shared interneuron types. However, hippocampal formation also has its own unique set of interneuron types. We found many highly distinct excitatory projection neuron types in both cortex and hippocampus. They also exhibit large-scale gradient distributions across the neocortical sheet and gradual but discrete transitions between neocortical and retrohippocampal regions. And we showed that a multimodal classification refines a set of cortical interneurons into MET types that have congruent transcriptomic, morphological, and physiological properties and a high degree of mutual predictability. On the other hand, the complete brain-wide single neuron morphology reconstruction reveals extensive morphological and projectional diversity even in relatively homogeneous transcriptomic types. Combination of transcriptomic and epigenomic profiling enables enhanced discovery and the generation of novel cell type targeting tools. And to summarize all this work, as well as many other works from the BICCN that I didn't talk about, we found that a comprehensive multimodal characterization of cell types from a single region, brain region, and demonstrates that transcriptomic, epigenomic, spatial, physiological, morphological, and connectional information can be all integrated together and revealing diversity at the same time. This kind of integration allows us to bridge analysis in different modalities as well as across species. And I hope I have convinced you that this kind of cell type knowledge and tools will be really foundational for uncovering the principles of brain organization and function in the future. So with that, I want to thank a large team of scientists and investigators at the Allen Institute who have contributed to this truly unprecedented study of cell type diversity. I have— put a lot of people's names in my previous slides, so I will just not name names in detail here. Just wanted to say that very quickly. Ed Lin and I led the Cell Types Program. Ed focuses on the human cell types part. I focus on the mouse cell types. The transcriptomics work in the mouse are led by 3 investigators here. And PatchSeq study by— those investigators and full morphology studies by these investigators. And I want to thank our entire team for contributing to this effort and our funders, Paul Allen, for his generous support, and also many of the grants that we have obtained from NIH. So with that, I will stop sharing now and be happy to take any questions you may have.