Scientist Stories: Evan Macosko, Single Cell Genomics in Neuroscience

Hi, everyone. My name is Parisa, and I want to welcome you to the Harvard Brain Science Initiative's 101 series. So today we're really excited to have with us Dr. Evan McCusko, who's going to tell us about single-cell RNA sequencing and its impact on neuroscience. Dr. McCusko is an institute member at the Broad Institute and an assistant professor of psychiatry here at Harvard Medical School. And Mass General Hospital. He has training both in genetics and neuroscience in basic research and in medicine. He has an MD and a PhD. So he did his PhD work at Rockefeller with Corey Bargmann and did postdoctoral work here at HMS and the Broad with Steve McCarroll. He's also a board-certified psychiatrist who continues to go into the clinic on a weekly basis. And during his postdoc, Dr. McCusko developed drops Seq, which is a technique for high-throughput single-cell droplet-based gene expression analysis. He started his own lab pretty recently in 2017. The lab designs and also applies different genomics technologies for studying cell types in the nervous system. Recently, his lab invented SlideSeq, which is a tool for localizing gene expression in intact brain slices. And they're using this method together with the high-throughput single-cell sequencing to study mechanisms underlying psychiatric diseases, which I think he'll tell you a little about today. Before we get started, just wanted to share, we have an hour and a half. We're going to have about a half hour at the end for questions. Thank you to everyone who submitted questions before. We have that and you can either raise your hand or type questions in the chat. Feel free to jot questions down in the chat during the talk. And with that, thank you, Dr. McCoskill, for joining us. I guess we can get started.

Great.

Sure.

Thank you for having me. And thanks to all of you for coming. So yeah, I saw some— one second. Let's put this up. I saw that there is sort of a wide range based on the survey that you guys sent me, I guess yesterday, that there's a wide range of interest. Some people are quite experienced with single-cell genomics. Others are sort of more just beginning to think about it and just wanted to learn more. I've tried to hit both the big picture stuff and some more practical introduction to using the technology, and then also some more specific ways in which I can think about using the technology specifically in neuroscience. Even though the initial topic is single-cell analysis. I think increasingly the technology field is moving towards a more sort of general framework outside of single cell. One term that I like is contextualized genomics. Another is tissue genomics, in which we're really trying to take decades of measurements that have been made in vitro and in purified solutions from tissues and actually do them in tissues themselves in situ and with all the spatial and also often temporal context that would allow us to really interpret what's happening molecularly in ways that we couldn't from the original ways in which those measurements were made. There'll be a combination here of talking about single-cell analysis, but also some about spatial genomics as well. I think neuroscience, of course, Many fields have these kinds of issues, but I think neuroscience especially has what we might call a keys under the lamppost problem, in which we've built a suite of tools and understanding of neuroscience that allows us to look very deeply and critically at specific circuits and specific cell types and specific contexts, but sometimes we kind of miss the big picture. And from the perspective of cell types, this is the most sort of obvious one. You know, there are sort of really thousands of cell types that can be appreciated by modern single-cell analysis, as I'll talk about. But the Cre lines and other tools that we have, promoters that we have to study these cell types, really only allow us to deeply study a few dozen. And so what you see a lot in papers that do, you know, critical or, you know, functional analyses of circuits is the same circuits are sort of queried over and over again in the same— in different mutants and different behaviors because Those are the circuits that we can access, but we need to understand the full range of cell types that are present in the brain so that we can actually start to access those cell types and get a bigger picture view of how the brain is operating. From the perspective of connectivity, you know, we know that the brain's connections are obviously— they're, they're both highly stereotyped, but they're also just extremely complex. And yet the tools that we have available to us to actually query connectivity are relatively limited. And so even today, You know, in the last couple years even, you know, there are very high-profile papers from, you know, mouse connectomics in which particular circuits, particular projections are newly discovered. You know, oh, there's this projection from the cerebellum to the ventral tegmental area. These are two areas of the brain that have been heavily studied for decades, and yet we didn't know about connections between them simply because the connectivity tools at our disposal are relatively limited. And we think that genomics has ways of potentially addressing those problems. And then finally, there's, you know, sort of challenges associated with brain diseases, and that, you know, we— the substrate of disease, the human brain, is very hard to extract intact, and postmortem tissue has a lot of, you know, confounders and challenges associated with it. But we need to learn from it because we need animal models that properly quantitatively reproduce aspects of the disease that we need to model. And Right now, a lot of the models that we have at our disposal are, you know, imperfect in different ways or just incompletely validated. And so I think that that's another place in which the genomics technologies that I'll talk about today can really have a lot of impact in neuroscience. So first, I'm just going to begin with sort of an introduction to single-cell analysis and where all of this stuff began, which is, you know, 5 or 6 years ago. Well, I guess it kind of began before that. But as I mentioned, you know, as with many other genomics measurements, our ability to do those measurements really required large amounts of tissue. And so in sort of traditional gene expression analysis, we would grind up a large amount of tissue into a smoothie and sample that smoothie for sort of the average concentration of genes. But we know that cells are the substrate of gene regulation. They, you know, obviously there's ways in which you can get patterning across space, and so you could see some similarities between cells that are neighboring each other, but there's also a lot of heterogeneity. And so the analogy of what single-cell analysis brings is that we're really sampling genes more like a fruit salad, in which we preserve the context of a cell and we look cell by cell to see how gene regulation is different across each cell. And, you know, 10 years ago when I started thinking about this problem as a postdoc, the language around which we thought about these kinds of analyses really required this sort of trade-off technologically. You either had to sample a lot of cells with few genes, technologies like in situ hybridization where you had maybe a few channels at your disposal, you could look across a whole tissue, or, you know, a cell suspension. Or you could sample all the genes but very few cells, technologies like DNA microarrays. And that early DNA microarray era was actually sort of what we sort of think of as like the primordial single-cell analysis. And actually, a psychiatrist, James Eberwein, was one of the first people to do single-cell transcriptomics. The way he did it is he did patch clamp on individual cultured hippocampal neurons And he aspirated the material and he did T7 amplification multiple times on that RNA to get enough material to actually put it on a microarray. And actually this era, I just highlight this one example of a particular insight. It's not from neuroscience, but it just sort of shows you that you can do interesting biology even at this relatively low throughput number of cells. This is an example from Midori Saitou. He mouth pipetted cells, primordial germ cells from the gonad, and then neighboring somatic cells as a contrast. And he just did differential expression between, you know, 10 somatic cells and 10 germ cells. And he found a set of factors that actually, when he expressed them in embryonic cells, he could actually specify them to become primordial germ cells. So you can learn something with a small number of cells, but you have to do sort of very careful experimental design. And also The process of doing this is extremely arduous, right? You have to do one microarray for every single cell and a lot of amplification to get enough material for that microarray. So the reason that single-cell analysis was really born was because of high-throughput DNA sequencing. Because now with RNA-seq, you could barcode individual RNA-seq libraries and sequence many of those libraries on a single sequencer. And so the sequencer itself becomes this sort of like highly multiplexed microarray that we didn't have before. And so the initial era of this single-cell analysis 10 years ago just leveraged plate-based sequencing to do this. You would put a single cell into a 96 or 384-well plate. You'd do some molecular biology to get the RNA out and amplified. And then you'd make libraries and you'd do sequencing of those libraries. But of course, it's just hard to do thousands of cells like that. Many of us were aspiring to get to thousands of cells across the entire transcriptome. The way that we accomplished that with a technology that is quite scalable and has become quite popular because of commercialization is using droplets. Microfluidic droplets are an amazing technology that were developed by multiple groups, but David Whites especially at Harvard pioneered ways of taking an aqueous solution and partitioning it into droplets that were equal in size. And based on the way that you use— you tune the microfluidic channel sizes, you can make those droplets anywhere from picoliters to small number of nanoliters in size. And so we wanted to know if we think about whether we could put cells inside of these droplets along with beads that would have a unique barcode on them. And that was a technology that we called DropSeq, in which you take beads that are coming in on the right here and cells that are coming in from the top and bottom, and they come into individual droplets. And because the droplets are equal in size, you can, you can very finely tune the average concentration of cells in these droplets. And if you drop it below 1, you get Poisson loading, and that allows you to effectively encapsulate a single cell in a droplet. And you can do the same thing with beads so that there's a multiplicative probability based on concentration that you'll have a bead and a cell in the same droplet. And it's in those droplets in which the single-cell analysis is possible. So as important as this— all this technology was, and we've had to develop these barcoded beads, which had never been built before, that enabled this technology. The other thing that was really important and something that I continue to emphasize because it's not something that you really see universally in technology development fields was the need for a really good QC assay. And this assay is— we actually originally called it the barnyard assay because we gathered a whole bunch of different cells from different species together, but then we realized that actually just two species would be perfectly fine for this analysis. And we just chose a set of a cell line from mice and a cell line from human. And the idea is that because there's so many SNPs between— well, they're not really SNPs, but there's so much sequence difference between humans and mice that transcripts mostly just align when you align them to a joint genome. They'll only align to one species or the other. And so if you're actually doing single-cell analysis, you'll know it because when you demultiplex your sequencing assay, you'll look at the species specificity of each barcode. And if there's highly species specificity, then indeed you know that your technology worked. And that's just illustrated here graphically, the smoothie where you get a mixture of species between in each barcode versus a smoothie in which you get them all uniform. So this was one of the first assays we ever ran, and it was what showed us that, that single-cell analysis with droplets actually works. You get, you know, thousands of transcripts in individual cells from one species, but the barcodes are highly specific, so you don't get very many from another species. But based on the statistical loading, you will have a very infrequent number of doublets. So in this experiment, for example, we had one very obvious doublet right here that was indicative that the two cells happened to come in together. So this was really proof of principle that it works, because we did this experiment in a couple hours. You know, took a little time to actually actually like make the library and sequence it. But the actual hands-on time was extremely simple, and we were generating thousands of libraries in that single experiment. So we knew that we had something here that was highly scalable and effective in delivering single-cell data. So that's the sort of principle. And now there's many different technologies out there that do similar things, some, some in the exact same way or similar ways, some in different ways. I'll tell a little, talk a little bit about that in a minute. But I got a couple questions from a group from audience members yesterday that I thought I would just mention a little bit. And kind of go through some of the sort of detailed— more of the technical details about single-cell analysis that you might be thinking about for your own experiments. So one question that often comes up is like, well, why should I bother with single-cell analysis? Why shouldn't I just do bulk sequencing? It's much cheaper and it's much easier. And of course, this is a hard question to answer in a universal way because there's obviously great reasons to just do bulk sequencing. I think considerations that would point you towards bulk sequencing would be, one, if your tissue or your process is highly uniform. So if you're just looking at a very, very uniform set of cells in culture and you're looking for changes in regulation, transcriptional regulation across a lot of different conditions, then probably bulk sequencing is going to get you the answer, and it may be easier to do that. Another consideration, you know, for bulk sequencing might be— and although this is actually changing rapidly, but it's something I used to say a lot— is if you have a high number of samples that you need to run. Until recently, it's been much harder to do single-cell analysis on, you know, 100 or 1,000 samples the way that you could do bulk sequencing. And by samples, I don't mean cells. It's easy to do many cells from one sample in single cell. But to multiplex many samples, that's something that only became possible recently with several different techniques that we call hashing. So, um, but otherwise, you know, there's a lot of reasons to do single-cell analysis, and I think you'll see that from the rest of the talk. Another question I get a lot is, um, is about sample preparation, um, uh, related to, you know, uh, sample quality, uh, and cells versus nuclei, and, and questions about sample enrichment. So I just thought I'd give a few little didactic slides on this topic. So sample quality is really, really important. First, in initial single-cell analyses, it was all single-cell, but now we're also doing single nuclei. And in both cases, if cells or nuclei are disrupted in your suspension, they will leak out RNA and you'll get very, you know, you'll get problems. So For example, this is a plot that you'll see a lot in, in droplet-based assays. So because we're critically loading cells at such a low number, most droplets are empty. And you can actually appreciate that in our mixed species experiments, because if you look at the— if you stack each barcode that you sequence by the number of unique genes or transcripts that's in that barcode, and you progressively go to lower and lower numbers, you'll see an inflection point. And that inflection point corresponds to where there are no longer any cells present, and you're just getting a very, very small number of unique transcripts per droplet. Because now each one of these is, you know, order of magnitude or more smaller than these, these, these on the left here. And you can see also that the species specificity drops And that's because this is ambient RNA. It's RNA that's just been released into the medium. And it's not encapsulated in a cell. So you're just mixing between, you know, human and mouse in this initial experiment. But what would this curve look like if cells were bursting or dying? What happens is much more of the RNA gets shifted into the empty droplets. And so instead of getting this crisp inflection point, you might just get a straight line. And that would be suggestive of, you know, really low-quality data. And you can appreciate that, you know, that's one of the first QC things we do is look for the kind of crispness of this inflection point. Cell concentration can also affect data quality in a couple different ways. The higher the concentration of cells, the more likely you are to have collisions between cells and droplet-based assays. But also the library impurity rises. And that's just because you have sort of stochastic bursting of cells or nuclei. And if you have more per unit volume, that's going to increase the concentration of that ambient RNA. With modern techniques, especially computational techniques, a lot of this is actually easy to correct out. There's a really terrific computational tool called Cellbender that was developed by a colleague of mine at the Broad, Mirtosh Babendi. And you basically just run your droplet-based assays through that computational tool, and it subtracts out a lot of this ambient RNA contamination that that can be a problem in these assays. This is a little bit of an antiquated slide. It's just— but it's useful for didactic purposes. You know, how do we know how much RNA we're capturing? Well, one of the original ways that many of us did this in these early assays is by using a reference set of RNAs at fixed concentrations. And we just ran that through the droplet-based assays to understand what the relative capture is of these RNAs compared to what's actually known to be there, because these are kind of a reference set of RNAs called the ERCC spike ends. And this gave us data that told us that DropSeq was on the order of about 10% capture of RNAs. I think in modern, let's say, 10x genomics droplet-based sequencing, this number is definitely higher. I think it's probably more on the order of 20 to 25%. Nuclei are increasingly attractive alternatives, especially in the brain, to single-cell analysis. And actually, my lab doesn't do any single-cell analysis anymore. There's a couple of projects in which it kind of has niche purposes. But for pretty much any routine assay, we are running single nuclei. And there's multiple reasons why nuclei are superior, in my view, to single-cell analysis. The first is that it's extremely reproducible. We have an SOP in the lab and we have no batch effects across tissues, conditions, people running the assay. And just because, you know, it's much more robust to sort of subtle differences that you can get from single-cell assays. There were some— there are some protocols that give you, like, add transcriptional inhibitors to try to reduce the effects of dissociation on the cells. That's not necessary with single-nuclei analysis, and I think that's an advantage. We see that there are much better representations of cell classes because the dissociation process seems to selectively kill some cells more than others. It's compatible with frozen tissue, so you can exchange tissues across labs. You could potentially sack a huge cohort of animals like we've done for different mutant assays, and then you just keep the animals in the freezer and, oh, I'm interested in the amygdala today, so you can go dissect the amygdala and do that. And FACS is very useful in this case for debris removal because you can stain with DAPI for nuclei and just get rid of the debris with a simple FACS step. You do get a somewhat smaller number of transcripts per profile, but with modern 10x genomics technology, this is less of an issue. And there are some artifacts. There are, of course, artifacts with single-cell analysis too, but single-nucleus analysis does have this artifact that's associated with different amounts of shearing of the cytoplasm around the cell. So it's something you just need to think about. So this is a little bit of a walk down memory lane, but a few years ago we published an atlas of cell types in several major brain regions in the mouse brain. And one thing we noticed— this was all single-cell data— one thing that we noticed is that there was a digestion time-related artifact. And also, just to kind of highlight these differences in cell proportions, that Some cells were really highly lost. So for example, this population of interneurons was much more abundant in their SM-FISH counting of cortex than the single-cell analysis by DropSeq. So it's just something to keep in mind and another reason that I really favor the nuclei at this point. So we've actually used single-nucleus analysis. We've done more than 10 million profiles in the lab. 6 million of which are for this very large mouse brain atlas that we've built for a consortium project through the NIH BRAIN Initiative. And, you know, again, we don't really see much of an evidence of batch effects. We can do very standard single-cell analysis computational tools. We don't need to do batch correction on the data to get out cell types and see that consistently across replicates and animals and technicians and the like. So now I'd like to just talk a little bit about some of the kind of considerations about experimental design, and one of them relates to this question of batch effect, which is within versus across conditions. So it is definitely the case that when cells are processed together, they have minimal technical differences, and when you start to compare across batches, you do have issues with potential batch effects. One of them I think I've already talked about, which is the cell dissociation effect. But other aspects have to do with sequencing depth and also just, you know, differences in washing and the like. So it's really important to be careful when you try to compare across batches, and ideally anywhere you can, to try to pool cells as early as you can up front with multiplexing techniques. I'm just going to pass over this because I don't think it's that important. But even so, with many different kinds of analyses, especially when you're doing analyses where the batch effects may be baked in— so for example, with postmortem tissue where people die in different ways, there's different postmortem intervals— these affect the overall data in such ways that you can never remove batch effects completely. You need some sort of computational tool that can find shared axes of variation and integrate cells by the appropriate cell type. We developed a tool for doing this called LIGER. There are many other tools now, and mostly, you know, especially for sort of subtle or simple batch effects, they work in relatively analogous ways. Many of these tools like LIGER also aren't just for batch effects per se, but they're also for doing comparative analyses. So we've had a lot of really— we've made some really nice insights into, let's say, species specificity of particular cell populations By taking homologous regions of the brain, doing integrative analysis between mouse and human and primate, and seeing what's similar and different in these cell types across species. And that's something that you can do with these kinds of computational strategies. So another question that's often asked, and from an experimental design perspective, is like, how many cells should you sample per condition? And usually I don't know the answer to that because the— it's so dependent on the tissue and also on the experimental question. But there are just some kind of considerations that you can think about that are specific to the technology. The advantage of more cells is obviously that you can appreciate rarer states, and it may be easier to do batch correction because if you— especially if you stack more replicates, it's easier for you to appreciate what might be just something that's unique to a particular batch or replicate versus something that's consistent across replicates. There may also be other reasons why you might prefer to do more reads per cell. Some cell states are clearly driven by a small number of genes, and the way that we do clustering analysis is driven on a gene-by-gene basis. You scale every gene to contribute the same amount to the variance, and so if a particular state is only driven by, you know, 5 genes, it's going to be much harder to appreciate that if you can only sample 1 or 2 genes, right? So that could be a reason to sample more deeply. Another reason, which I'll get back to in a minute, is there's always people who are very gene-focused for their particular— you know, certain experimental questions are very gene-focused, and you may actually just need to see genes frequently in cells. And so that would be a reason to sample more deeply. It is the case though that your return on investment beyond a certain amount of sequencing gets to be very low. This is one of our first, you know, single-cell experiments with DropSeq, but it's equally true for 10x Genomics data as well. Beyond, you know, let's say, you know, 90,000 or 100,000 reads per cell, your return on investment is extremely low. You can only really get maybe 10% more even if you go 10-fold more sequencing. So it just underscores that on an efficiency basis, you're really better off here, like just sampling more cells. But again, there may be niche reasons why you might want to just allocate your sequencing space to fewer cells at a higher depth. So this favorite TF question and lowly expressed transcripts question, which I get routinely, I like to just try to illustrate this by comparing single-cell analysis to bulk RNA-seq. Because a lot of times people will say, well, I see this transcript in my bulk data, but I don't see it in my single-cell data. So the first thing to appreciate is that the absolute capture efficiency of single-cell analysis and RNA-seq analysis are equivalent, or single-cell is actually better now than bulk RNA-seq. So if I just have a bunch of RNAs in a in suspension, I'm going to capture a smaller absolute number of them by sequencing with bulk RNA-seq than with single cell. It's just that I start with a lot more RNA material when I do my bulk sequencing. So let's take this example where I'm sampling 200,000 striatal cells. I just cut out a piece of striatum. It's constituting 200,000 cells. And I'm really interested in this very rarely expressed transcription factor, which maybe on average is only found in 5 copies per cell. And it's only in a very, very rare interneuron population. That's only 2,000 of this total set of cells. So if I try to sample this thing by bulk RNA-seq, I'll be making a single measurement from these 200,000 cells. Maybe I detect a total of 200 million transcripts from those 200,000 cells. That's just based on, let's say, like a back-of-the-envelope calculation of the number of transcripts per cell that are present and the sampling efficiency of bulk RNA-seq. So maybe in my bulk RNA-seq, I'm going to see 10 transcription factor X detections. But let's look at single-cell analysis. Let's say we just, for the purposes of equality, even though this is no longer true anymore, it's now 25% for single-cell analysis, I detect 1% of all transcripts. Now I'm making 200,000 measurements on these 200,000 cells with the same total number of transcripts. And the same total number of transcription factors. But what have I learned that's so much more valuable than what I was able to learn from that single measurement of bulk RNA-seq? Now I have 200,000 measurements, but I can cluster based on all of the gene expression relationships that I've also similarly measured from those single cells because I have 200,000 measurements. And now let's just say for argument purposes that I can now identify this rare interneuron population in my striatum And I have two other populations. And for those of you who are thinking, know something about the striatum, that's— I'm, you know, sort of basing it off the absolute actual histology. This would be, let's say, you know, indirect spiny projection neurons, and these are direct spiny projection neurons. And now I've clustered out my rare interneuron population. And now I can appreciate that this transcription factor is actually relatively highly expressed in those 2,000 cells. It's much less is much more highly expressed relative to the total number of cells than what I would have expected from my bulk RNA-seq. So the point is, is that the power of the technology is the ability to partition cells into smaller groups that give you much, much better ability to say that actually maybe your transcription factor is not as rarely expressed as you thought it was. So finally, A lot of people asked on the— or several people asked about platforms and differences amongst different kinds of single-cell analysis tools. And this is getting an increasingly complicated question to answer. And that's because single-cell analysis is— and spatial and single-cell genomics are just exploding both in the academic literature, but also in the industry space. So I'm just going to kind of highlight some particular examples here. The first sort of major class is the thing I've talked about already, which is high-throughput single-cell analysis. I think there's a lot of value to droplet-based tools because of their ease. They are somewhat expensive, unfortunately. And then there's this other set of strategies called— that are based on plate-based strategies where you do the indexing by split pool indexing of the RNAs themselves as they stay within cell context. And Parse Biosciences has commercialized a particular version of this, which was called SplitSeq. It was developed at the University of Washington. It's a little bit cheaper per cell, and it may be useful, especially if you need to do a lot of cells, because it scales very nicely to many, many hundreds of thousands of cells. The efficiency is a little bit lower in terms of the capture efficiency than 10x Genomics. But now there's all these other technologies, and I haven't even mentioned the single-cell epigenetic technologies as well. But of course now there are a lot of ATAC-seq strategies that have been developed for single cells that are also very valuable depending on the application of interest. There's also now opportunities to do spatial profiling. And this is work that we're very involved in. We've developed a technology called SlideSeq where you can look at gene expression on an array, a highly dense array of barcoded beads. And by doing that, you can appreciate gene expression in 10-micron pixels. But there are also microscopy-based tools for doing this. One that has been commercialized is called MERFISH. It was commercialized by a company called VisGen. And because it's microscopy-based, it's actually subcellular in its resolution, which is very exciting. And it's becoming increasingly possible to multiplex those assays to hundreds of genes. And there are many, many more of these tools on the way that are kind of combining these tricks for doing additional single-cell measurements that kind of are more deeply informed by, with sort of other aspects of biology. And so I think with that, I'll turn to just taking the remainder of the, you know, 15 or 20 minutes to talking about particular applications of these technologies to neuroscience. Sorry, I didn't remove the animation there. But, and I'll just give you a couple vignettes from my lab and also from another lab on these 3 different kind of keys under the lamppost problems that I described: the cell types question, the connectivity aspect, and also more disease-driven studies. So this is an example of where just like doing a lot of single-cell analysis turned out to give us something, teach us something kind of interesting about a tissue that's very heavily studied, which in this case is the mouse cerebellum or the cerebellum more generally. So as part of a large-scale NIH initiative, we sequenced a huge number of cells in the striatum and the cerebellum, 650,000 profiles. We tiled those profiles across every lobule of the cerebellum, both vermal— we've divided vermal to cerebellar cortex so we could appreciate the regional variation in the cerebellum. So it's a very descriptive study at its outset, and it was led by two very talented RAs in the lab, Carly Martin and Valina Kozareva. And once we did our clustering and our analysis, we made a few very interesting insights that were— had not been previously appreciated, despite the fact that, as many of you know, the cerebellum is an extremely heavily studied tissue. So the first thing is that there's actually a lot of regional variation in gene expression within Purkinje neurons. Purkinje neurons had historically just been divided into two major groups, those that express ALDOC and those that don't. ALDOC has another name in the protein literature called Zebrin. The reason that name is given is because in the cerebellum in which Zebrin was stained with anti-Zebrin antibodies, you see these patterns of stripes that go across the longitudinal axis of the cerebellum. And those stripes are actually ALDOC expression patterns that are, that are kind of alternating in this longitudinal axis within Purkinje cells. But besides that variational axis, it wasn't really appreciated that there was much variation. But you can see that we found a total of 9 different populations, and especially within the ALDOC positives, there was a lot of regional localization. There are Purkinje cells that are unique to the paraflocculus and flocculus. Also unique to the nodulus, the area 10 of the vermis. And what's interesting about this is that the areas that had the highest variation were actually also ones that are known to have significant differences in their connectivity. Most Purkinje cells project to the deep cerebellar nuclei, but a select number actually project to other brainstem structures, and those particular areas are the ones that have the highest molecular variation. So we're working with our collaborators Wade Regeer's lab to explore the relationship between these molecular types and the connectivity. But one thing that really popped out for us, which was quite surprising, was that in the molecular layer of the cerebellum, which one of the first, you know, set of cells that were actually described by Cajal— he called them basket and stellate cells— we found molecular variation that was actually not— had not been identified before. So there were— so in our clustering analysis, there were two major groups of molecular layer interneurons, MLI-2s and MLI-1s. And then within the MLI-1s, there was actually this more subtle variation that we saw between them. So a total of three sort of distinct populations there. And initially we thought, oh, the MLI-2s and the MLI-1s, that must just be basket and stellates, because that was what had been sort of canonically labeled as such. But actually it turns out that MLI-1s and MLI-2s have both basket and stellate phenotypes. So there are very clearly basket MLI-1s and sort of basket-looking MLI-2s. And then if you look more distally at the upper layers of cerebellum, both MLI-1s and MLI-2s look relatively identical to each other. They both look like basket cells— I'm sorry, but like stellate cells. But nonetheless, even though they're morphologically relatively similar and they have the same similar axis of variation, They're just electrophysiologically quite different. They have different spike frequencies, they have different firing rates and the like, and also they're differentially gap junction. It turns out that Wade's lab, our collaborators there, Tomas and Stephanie, found that the mLI-2s are completely not gap junction, whereas mLI-1s are gap junction. It was always thought that gap junctioning was a consistent feature of all cells in the molecular layer, but in fact, that's not the case. A full 30% of them are not gap junctions. So it remains to be seen exactly what their roles are functionally, but it was a kind of a surprise because it's just this area that's very heavily studied. So now turning to brain cell connectivity, I think this is a really exciting place for new technology development, but there are some existing technologies that for those who are really focused on cell circuits in the brain, I think it's just really important for you to be thinking about, because it could be really useful for the kinds of biology that you're doing in your own work. And I'm just highlighting a couple different technologies, but there actually are— I was looking the other day to prepare for this talk. There's actually like 20 or 30 papers with sort of subtly different ways of doing this kind of thing. So I just am highlighting two. But the basic idea here, the premise, is to combine single-cell analysis with viral tracing technology. So in this example, which I think is relatively simple but clearly works very nicely, a collaboration between Ed Calloway's lab and Joe Ecker's lab, they took a Cre virus that is AAV retro, so it's a virus that very nicely traffics back to the soma of cells whose processes are at the site of injection. They injected it into a nuclear GFP reporter mouse, a floxed nuclear GFP reporter mouse. When this AAV is injected and it's taken up by cells, the Cre is expressed and it turns on this nuclear GFP reporter. What they do is they inject it in a target, a candidate target of a particular projection neuron, And then they look back at the source to see, like, well, which particular areas were sending projections there, because those will have GFP-positive cells. And then they can actually take those out, the nuclei out, and they can do, in this case, single nucleus methylation analysis, which is another reporter of cell type diversity that can be used as well as similarly to RNA-seq to look at differences among cell populations. To see, in this case, what are the transcriptional or methylation signatures of different cortical populations to different both cortical and extracortical targets. The data look really nice. Here you can see they've profiled a bunch of different— again, they're just sampling cortex, but they're looking to see where these different cortical cells project, both within the cortex and outside of the cortex. You can see that, for example, the intrathealencephalic cells that are expected to just send projections to only thalamic structures, indeed, if you look at these colorings of the targets, are only sending cells to those particular locations. You can identify some really interesting distinctions that are highly specific to different cortical regions. And so that's just a really nice way of connecting the gene expression features and the methylation features to actual, you know, connectivity information. Another strategy has been developed by Tony Zador's lab at Cold Spring Harbor, and they seek to map projections by infecting cells at the source. So with a virus that is hugely overexpressed at the area in which it's injected, And because of this high overexpression in the, in the soma and the processes, it gets sent out to the processes. And if you actually sequence these terminal areas, you get— you can actually read out transcripts that were delivered through the virus. And a nice thing is that the synbys can be barcoded. So you can, with one source injection, match barcodes at various target regions back to a single cell source if the diversity of the barcode is high enough. And they've now started to combine this with both single-cell analysis but also spatial analyses like in situ sequencing to try to sort of match up gene expression data with these barcode projection maps. So I don't think that either one of these methods is the be-all and end-all. They have some limitations. So for example, this method is very cool in many ways and could potentially be very high throughput, but one major challenge is that the Sinbis virus is highly toxic, and so it highly alters the gene expression patterns of these cells. And also the animal dies after a certain number of days after infection. But it's sort of a path forward for thinking about how one can marry targeted barcoding strategies with these high-throughput genomics readouts to really learn something about connectivity in a very systematic and high-throughput way. So finally, I wanted to just tell you a little vignette about brain disease and using single-cell analysis in brain disease, and this comes from my lab. We're really excited about thinking about spatial and single-cell tools as a next generation of neuropathology. There was this amazing ebullient era in neuropathology about 100 years ago in which scientists just started basically throwing different dyes down on tissues and looking under the microscope and seeing things that they couldn't before. One example is Hideo Noguchi who stained brain tissue from people with psychosis for— and a particular form of psychosis called general paresis— for a stain for syphilis. And so he identified syphilitic organisms inside these lesions and was able to basically, you know, prove something that had been hypothesized before this, but he proved that syphilis was the causal organism of this particular disease. And then more famously, Louis Alzheimer put silver down on brains, and he found that a subset of dementia patients had these amyloid plaques, and he called this disease Alzheimer's disease. So, um, I guess we've cured one of these diseases, but we haven't cured the other. But nonetheless, it's obviously extremely— these are like foundational experiments, and they moved us into this, you know, this era where we could actually study these diseases because we knew something about them. And we're really hoping that genomic data acquisition kind of leads us towards that same kind of, uh, exciting, you know, era where we can actually learn much more in detail about how diseases of the brain arise and how we could potentially fix them. So what can we learn? So from the perspective of single-cell analysis, we can characterize dysfunctional cell states, and there are many now nice papers doing this in several different diseases. We can identify relative cell population abundance changes. This is especially valuable for neurodegenerative diseases where we know certain cell populations are dying. And critically, we can connect this data to human genetics. And this is exciting because human genetics is an influence on disease, a risk factor for disease that arises at birth. And so it could potentially teach us something about the causality of disease. And finally, especially with spatial duals, we can start to connect these disease states to histology. So for this particular vignette I was going to tell you about, it comes from Parkinson's disease, and it's work that was led by Tushar Kamath, a student in the lab. We started with Parkinson's because it's a disease that is highly cell type specific. We know that certain brainstem structures, especially the substantia nigra pars compacta, is highly vulnerable to Parkinson's disease, and the loss of substantia nigra cells, dopaminergic cells, leads to the pathognomonic symptoms, the motor symptoms that are associated with Parkinson's and diagnostic of Parkinson's. One thing that you see grossly is that there's this depigmentation of the tissue, and that's because the dopaminergic cells themselves are pigmented. But even though many of these cells die, not all the dopaminergic cells die. It's imperfectly known what the molecular phenotypes are of the cells that actually are dying. We know spatially that actually they are restricted preferentially to the ventral areas of the substantia nigra versus the more dorsal areas which seem to be spared in sporadic Parkinson's. We wondered if there could be gene expression differences amongst these DA subtypes, molecularly defined types, that could help to explain the relative vulnerabilities of these different populations. So again, Tushar led this analysis work, and Abdul Abdul was a talented RA, former RA in the lab, who did a lot of the data collection. And the data collection in this case was not trivial because dopaminergic neurons are quite unique and also rare. And so when you just take out a piece of the substantia nigra and you profile it, you get very, very few dopaminergic neurons because they're just not a very common cell population. So Abdul identified an antibody, NeuR1, which could be used to enrich for these cells by FACS, and then we could actually profile many, many more of them per donor. So in total, we profiled 202,000 nuclei, and we also actively selected and sampled DAPI-stained cells as well, just so that we could get all the cell types as a comparison to the dopaminergic neurons. Of which 22,000 were dopaminergic neurons. And we did both controls and also PD cells. Just to give you a sense of where the technology has come from some of those earlier slides I showed, with modern, you know, 10x genomics technology, we're able to sample within dopaminergic nuclei from postmortem tissue more than 10,000 median unique transcripts per nucleus. So get a very, very deep view of these cells. When Tushar did his analysis— so this is a representation of the data we call a tSNE plot or a UMAP plot. They're two different algorithms that are very similar where cells here are individual points and they're being grouped and positioned by gene expression commonalities. So when we cluster the data, we're looking for cell populations that are sharing particular gene expression relationships with each other. 2STAR was able to identify a total of 10 from the controlled donors. And they followed an axis of variation that was appreciated previously in the substantia nigra between calbindin expression on this set of cells over here and SOX6 expression. Those are the sort of two major classes that had been described in the literature before. And there were a set of unique markers that we could identify for each of these 10 populations. I won't talk about it today, but one of the One of the things we identified was that this particular population that expresses this gene GEM is actually primate-specific, and we can figure that out by doing alignments of this data with data from mice, tree shrews, and primates. Now we wanted to position these cells within the substantia nigra because I told you that in Parkinson's disease, substantia nigra cells that are preferentially lost are in the ventral area. We wanted to see whether particular populations were more localized to the dorsal versus ventral areas. So to do this, it's an illustration of— I actually put this in in response also to a question because somebody asked me how you map spatial data and single-cell data together, and I'll just tell you how we do it. So we take single-cell data, in this case this single-cell dataset that I just described, and we collected slide-seq data from this macaque, fresh macaque. These are— this is a coronal section, and we took a set of coronal sections and 3 individual 3-millimeter arrays to measure the gene expression across the full medial-lateral axis of the substantia nigra. And we used a technology— a technique called RCTD, which was developed by Dylan Cable, a student with our close collaborator Fei Chen. And Dylan's strategy, it's basically a tool that does factorization, a Poisson-based factorization And then some maximum likelihood analysis to give you assignments of particular single-cell profiles to particular beads. So this is what the RCTD algorithm does. So in practice, so we took these 27 arrays, and I'll just give you a couple examples from this particular study. We then took our single-cell data, but now zooming out to all the major cell classes. The dopaminergic neurons being one, which we could see here matched the Nissl staining quite well actually, because you can see that because these cells are so big, the Nissl staining really pulls out the dopaminergic neurons along this axis and you can see them following the same axis here, along with some of the other major cell classes, oligodendrocytes, astrocytes, and the like. But now we zoomed in on the dopaminergic neurons and did RCTD at a lower level of analysis, or I guess a higher level, more granular level, comparing SOX6 to calbindin, and we could see that calbindin was overall more abundant in the dorsal part that's more marked by the upper part of this, from this dotted line, versus the ventral parts that are below the dotted line, which are more SOX6 positive. But you can see there's still some mixing. So we wondered at the level of individual cell populations of those 10 populations that Joshua identified in the single-cell data, whether some of them were even more selectively found in the ventral tier, which is the one that really degenerates. And we did find one population which is marked by this gene AGTR1, which is highly enriched in the ventral tier versus the dorsal tier. And that's just summarized by these ridge plots where we're looking at the perpendicular distance of each one of these mapped beads to this medial lateral— sorry, this dorsal ventral axis. And you can see that the AGTRs are the most highly enriched in the ventral versus these others, actually the species-specific one, the primate-specific one is actually highly dorsally enriched. Just now going to the final couple of minutes, just going to the disease data. Now we actually put our 10 Parkinson's patient profiles back into our single-cell data to see if we could confirm or refute whether these AGTR cells were the ones that were most heavily lost. You indeed can see that when you do a proportionality analysis, you find that the cells that are most heavily lost in Parkinson's disease are these AGTR cells, and actually get preferential sparing of the ones that are highly dorsally enriched. This really told us that we had found the population that is the most vulnerable to Parkinson's disease through this molecular analysis. Now we wanted to know what's interesting about these cells. We took two different approaches to this, and I think we came up with some pretty interesting insights into Parkinson's disease. The first is that we looked at the intersection of genes that are known through familial studies or through GWAS to be implicated in Parkinson's disease with genes that are unique to this population versus the other dopamine populations and the other populations in the substantia nigra. So, on the familial side, we found that there was an overabundance of expression of the 20-some-odd familial variants in dopaminergic neurons generally. We did not see amongst the familial variants whether there was selective expression, more enrichment of expression in any one subpopulation or another. But on the common variant side, which has much more, many more variants and many more genes associated with Parkinson's than the familial variants, we were able to show enrichment in DA neurons of the PD GWAS signal compared to Alzheimer's where we saw enrichment in microglia. This is a very well-established enrichment. In Alzheimer's disease. But at the subtype level, when we tested individual subtypes for enrichment of PD GWAS genes, we found that the only population that showed high enrichment was the cells that are preferentially dying, the SOX6 AGTR1 cells. So if you wonder why these cells die, I think the answer is, well, they're just expressing a lot more of the risk genes associated with Parkinson's disease. And that means probably that these are cell-intrinsic processes that are driving this degenerative disease, much more so than in Alzheimer's disease, where the processes are much more cell non-autonomous. A neuroimmune process is clearly mediating the cell survival. So now we just also wanted to see whether we could identify particular regulatory pathways that are upregulated or downregulated in the cells that are dying. So what Tushar did for this analysis, which is my final slide, He looked for transcription factor target enrichment. So these are all transcription factors, but he looked for a significant enrichment of individual transcription factor targets in the AGTR1 cells compared to the other dopaminergic populations. So the p-values for enrichment on these non-AGTR1s are on the x-axis, and then on the y-axis are the specific AGTR1 population.. So, the ones in this upper left corner are selectively enriched in these cells versus the other dopaminergic populations. And we found some really interesting transcription factors here. I mean, p53 was really interesting, and p53 more generally, its enrichment tracked with the vulnerability to Parkinson's disease. So, the cells that are the least vulnerable had the least enrichment of TP53 targets. TP53 itself is transcriptionally upregulated in these cells. It really suggests that the process of cell death may be mediated by p53 itself. We also found some other interesting TFs. This gene, TFAP4's activity was decreased in the PD cells that are of this particular type. TFAP4, when its activity decreases in cancer and also in other disease contexts, it actually leads to apoptosis. So the fact that this one was selectively decreased, we also thought was very interesting. Several of these others also have some interesting biological implications. But one final analysis to leave you with is Tushar tried to unite the GWAS data and the GWAS finding with this transcription factor analysis, and he looked for the enrichment of p53 targets within the GWAS data itself. And you can see there's sort of an intriguing trend here in which there is some some significant enrichment of p53 targets in the GWAS set itself, not quite as much as prostate cancer, but much more so than schizophrenia and Alzheimer's disease. So it may really suggest again that, or I think it really strongly suggests that cell-intrinsic processes in Parkinson's disease are really what's most important for mediating neurodegeneration. And I hope it just as an overall sort of vignette gives you a relative flavor for the kinds of analyses that are possible now with modern single-cell analysis, that you can combine GWAS data in very clever ways. You can leverage existing transcription factor datasets, other genomics datasets, and to inform the gene expression data that you're getting or the epigenetic data that you're getting. And you can also leverage spatial data as a way of further contextualizing what you're finding with your single-cell data. So I'll leave it there. I know we have some time now for questions. And I'll just thank you all for your attention.

Thank you so much for that wonderful talk. So if people want, feel free to raise your hand or type in questions. We have a question in the individual chat that came in to me from Gregor that I can get started with. It says, I have a two-part question. You mentioned that the absolute efficiency of single-cell RNA-seq is around 25% of all available transcripts. How does this compare with capture-based profiling methods like SideSeq and Visium? He also says many capture-based approaches are moving towards smaller and smaller beads or pucks, seemingly with the aim to get subcellular resolution. Does this impact absolute capture efficiency and how?

Great questions. So the first answer is that SlideSeqs— we carefully measured SlideSeq efficiency in our second paper. That's a Nature Biotech paper from this year, Stickles et al., and we found it to be something on the order of 3 to 5% efficiency. The comparison with Visium, which is a little bit trickier because the feature size is a little different, but when we made as careful of a comparison as we could, we found that the efficiency of Visium was a little bit lower, perhaps 20-30% lower than the SlideSeq data that we generated. It is definitely true that smaller feature sizes are going to lead to sparser data. Some very clever people— I mentioned Murtaz Pabendi, who's a really terrific machine learner here at the Broad. He— I told you about that Cell Bender tool that he used to— he developed to subtract ambient RNA, but he's also very interested in this question of imputation of sparse data in spatial data, which actually I think is potentially much more tractable than in single cell because you actually have the spatial context to appreciate potential relationships. So I think that's definitely a possibility, and actually that could be increasingly important because when you get down to subcellular feature sizes, you'll actually want disciplined computational ways of segmenting those into individual cells, right, for downstream analyses. I actually also am very I very strongly believe, based on work that my lab is doing, also my collaborator Fei Chen's lab, that the capture efficiency of spatial technology is by no means complete. We're sort of at the era that we were at with single-cell analysis, sort of pre or at the time of DropSeq in 2015. And just like we moved from 5% efficiency for cells to 25% efficiency now for cells, I think that similarly we will get to levels that are that high or even higher with spatial data.

Thank you. So we have another question from Jenny. Thank you so much for the informative talk. Does the enrichment of transcription factors involved in cell survival in the degenerating population you were talking about suggest a compensatory mechanism which may explain why those cells have not yet degenerated?

Yes, I think in general, this question about what's a compensatory mechanism versus what isn't is one that is very, very hard to imagine answering directly from the data. I mean, first of all, I just was completely blown away that we found any of this in the data. I mean, you know, if you look back at sort of more traditional transcriptomic analyses of major diseases, it's very hard to find similar kind of crisp findings as you could with this. And I think just shows you that the single cell is giving you information that's very valuable. But causality is obviously a really hard one. And so I think the only strategy there is models. I mean, I think you have to go to animal models to figure that out. There may be clever things you can do that could help you to uncover it. But I think primarily that has to be a question that's addressed in models. Thank you.

And then I wanted to read some of the discussion questions. One that had been interesting in particular— there are a lot interesting— but was about development. How would you annotate cells that exist in a developmental continuum? That person gave examples of neural stem cells, oligodendrocyte precursor cells becoming mature oligodendrocytes.

Great question. And it's, you know, there are trajectory tools that have been developed computationally. That was one of the first areas of computational development in single cell, you know, 5 years ago. Tools like Wanderlust and Slicer. You know, and most especially Monocle, which is heavily used. Those are— those continue to be ways of defining trajectories and sort of traditional strategies for doing so. I think that there also is a lot of— continues to be a lot of opportunity to better summarize data, because I think the really challenging part is not so much the kind of purely developmental data that that you may be describing, like a differentiation to a particular cell type. But when you have these very complex mixtures of continua and discrete patterns, and we see that in the mammalian brain and the mouse brain dataset that we generated, where you have these super discrete differences. But then, for example, those Purkinje populations, there are some discrete differences, but there's also these interesting continua which are much harder to summarize with existing tools. And there's a computational biologist in my lab who's very interested in that problem, but I think it's something that we really need to do better at.

Thank you.

Jill? Oh, hi. Great talk. So I was curious about generalizations between the difference of the RNA profiles that are in the NUC-seq data versus the single-cell sequencing data and whether anything's coming out, for example, with the in situ sequencing that you can make some generalizations about the differences in those two profiles? Yeah, it's definitely the case that nuclear transcripts, as you might imagine, are ones that are sort of more being more dynamically regulated because the cell is like actively transcribing them. So one thing that when we started to look at the differences between RNA in the cell body and RNA in the nucleus is that the cell body RNA appears to be stuff that's kind of just more A lot of it tends to be stuff that just, you know, needs to be there for kind of active translation. So for example, ubiquitin is hugely— the transcript is hugely enriched in the soma because presumably, like, when the cell first differentiated, it just threw out a huge number of ubiquitin transcripts. They're sitting on the ribosomes because ubiquitin is used constantly, and they're just being actively transcribed— actively translated, but the transcription is not as dynamically regulated. But in contrast, cell states that may be transitioning actively at the time of sample collection are probably better captured in nuclei because that's where those actually are being made at the time of sample collection. So I think there is some temporal component to the differences. But I also know— but also there are distinctions that have to do with cell morphology as well. So in our SlideSeq paper, we did some analyses of SlideSeq of the hippocampus, and there the orientation of the cells is highly uniform, and all of the processes are extended into a particular layer which is highly sparse for cell bodies, right? And so what we could do is we could actually look for the enrichment of transcripts in that layer versus in the soma. We had to do a little bit of clever computation to filter out, let's say, glial transcripts, because of course glia are also sitting in that that molecular layer as well. But we actually were able to recapitulate a lot of the findings, let's say, of Aaron Schumann's lab for decades has been looking at, you know, RNA enrichments and their functional roles in plasticity in that particular sub-compartment of the cell. So I do think that there is value, especially as the resolutions of the spatial data gets better, for these kinds of sub-cellular localization experiments that, of course, nuclei are not going to be able to do. But I would argue that single cell is not going to be able to do it very well either because dissociating neurons, for example, you're not going to get processes out anyway.

Isabelle?

Hi.

First of all, thank you so much for a really fantastic talk. So my question is regarding changes in cell composition. So particularly in neurodegenerative diseases, you showed Parkinson's. How may neurodegeneration affect the interpretation of the data? And actually, for example, in Alzheimer's disease, you also have increase or because of changes in microglia and astrocytes because of neuroinflammation. So how could we account for this in particular at the single-cell level?

I'm sorry, I really apologize, but my—

I—

you broke up for some reason. And so I actually was not able to fully parse your question. Would you mind just repeating it? I apologize.

All right, I hope it was not my connection. No, I don't know. It's related with changes in cell composition. So particularly you showed Parkinson's and another one being Alzheimer's, and in Alzheimer's you also have the increase or activation of microglia, astrocytes. So how can you— does that— would you expect— I would expect that it would— that would affect the data interpretation? How so, and how could we control for it?

You mean that because certain populations are proportionally upregulated that that could lead to a downregulation of other populations?

So, my main concern is actually even with neurodegeneration, because in both Parkinson's and Alzheimer's you'll have cells dying. So, when you, for example, do single cell, how are you certain you're losing those? You may be losing those cells because they're just dying. So it's more—

yeah, that's— yeah, yeah. I think that that's where the diversity of the populations really helps you, because if you see a selective upregulation of just a couple glial populations but not like the vast majority of glia, then it really can't be a compensatory increase that's only a consequence of cell death from the neuron side, let's say. But if you do see that everything is proportionally going up except for a couple populations where they're even, that could obviously give you a sense of what might be happening there. But most of the tools that we use, I mean, I can say that we use a very simple tool. It's called MASC from Shoma Raychardouri's lab. And the model itself accounts for those kinds of different issues. Sort of intrinsically. But I will say that it isn't possible to tease apart— proportional differences only report on kind of the relative rate of degeneration, let's say. So one thing we're very mindful of is that even though we see that the AGTR population proportionally is the only one that goes down, it's probably the case that these other dopaminergic populations are declining to some extent. They're just dying much more slowly, right?

Thank you.

You're welcome.

Thank you. Do we have more questions? Are people able to type things in the chat to everyone? Because I'm getting questions that are only to me. Sorry, one sec, having a little slide share problem. Okay. Anyway, so we had a question in the discussion earlier, survey earlier, that was interesting, which was about integrating datasets. How does your lab think about integrating public information from, you know, public datasets?

Oh yeah, that's a great question. How? So I can tell you how our lab is thinking about it, but first I'll just tell you what's available now in case this is a question from somebody who like has a dataset that they really want to integrate. So there's two strategies. The first strategy is to use a tool like ours, Ligur, to bring the published dataset together with your own. That works very frequently well. It sometimes can be challenging if there's high imbalances in the relative sizes of the datasets. Another tool that can also be useful for that sort of thing is Rahul Satija's tool in Sirot. Which also allows you to integrate data from different sources. So I think from a practical perspective, if that's what— if you need to do that like right now, those are the tools that I would recommend. But I do also want to just emphasize that this is like an active area of computational development, something that my lab is very interested in. More generally, like, it's just not easy to work with single-cell datasets. Like, every time somebody comes up with an idea like, oh, we should like check for this or that, it takes such a long time to just like bring the data up from, you know, and load it in and, you know, sort it out and sort of munge it in ways that are necessary to actually get your answer. So we're very interested in ways of summarizing the data in a much more flexible way. I don't think that we need to always have, you know, 6 million cells on hand anytime we want to answer a question. A lot of that information is highly redundant. So we're really working on sort of highly faithful summaries of the data that still recapitulate all the things we just talked about— discrete differences, continuous differences, differences across samples. But I think that's something that just remains to be done.

Thank you. So if there aren't any final questions, I guess I wanted to maybe end on the note of just your, your big picture overview for the next, say, 5 years or so. Like if you had to say, you know, what is your lab most excited about? What developments in this field in general are you most looking forward to?

Yeah, I think there's a couple different, there's a couple different things. So the first thing is I think there's just immense opportunity on the disease side. I just have been really sold on the Parkinson's work. I think that it really reveals how much you can learn from just very well-designed case control studies. And I think that that will be enormously powerful with the technologies we have. But then I also think there's just a lot of opportunity for more technology. And so I mentioned the connectivity part. I think that's a place where we're going to just see a huge explosion of what's possible, you know, that you may be able to just routinely phenotype animals by looking at all their projection patterns. Like, that would be— that could be possible in the next 5 years. Similarly, there are lots of other avenues in which technology is really advancing, especially in the spatial domain, looking at interactions amongst different molecules, interactions amongst different cells. And I think that you'll start to see this emergence of what I just described as contextualized genomics, in which the same kinds of measurements that we're used to seeing, ChIP-seq, RNA-seq, ATAC-seq, DNA-seq will now be made in ways that allow us to actually relate it to what's happening functionally in cells, and I think that will be quite transformative.

Thank you so much. Thank you so much for all you've covered in an hour. It's just astounding from the history of how some techniques were developed to the specifics of different methods to the big picture overlook for disease research. We really appreciate your time.

My pleasure. Thanks for everybody for your interesting questions.

Yeah, thank you very much everyone for joining.