Skip to content

2026-04-20-single-cell-multi-omic-technologies

Date: 2026-04-20 Primary topic: single-cell-spatial-omics Cross-referenced from: single-cell-spatial-omics Word count: ~13,326

Auto-generated captions

This transcript is Vimeo auto-generated English. Expect misheard names and technical terms. Use the ripgrep/search approach for exact quotes.

It's my pleasure to be the chair for this session, which is on single-cell multi-omic technologies to understand genotype and phenotype relationships. And I'll kick things off with the first talk, and that'll be followed by Charles Gewirtz and Andrew Attie. I don't have any disclosures that are relevant to the stuff that I'm talking about today, but I just wanted to mention some of the SABs that I serve on, just to be transparent. So first I'll talk about some of our efforts to develop single-cell DNA and RNA sequencing methods to understand how mutations and copy number alterations lead to phenotypic changes in cancer cells. Chuck will be giving the second talk on high-content single-cell multiomics to resolve molecular determinants of cancer cell phenotypes. Chuck has been working in the single-cell DNA field for quite a long time on hematopoietic cancers and has founded a company, BioScribe, which many people use and collaborate with these days. And then Andrew's going to give the third talk on advances in single-cell DNA methylation and multi-omic technologies analysis. And Andrew's really a pioneer in technology development, especially around combinatorial indexing methods and studying epigenomics both in cancer and in normal development and neurological systems. So we have a really great lineup of talks today. I'm excited to hear from all of them. So just to introduce the topic a little bit, this is a PubMed search I did for single-cell sequencing papers. It's a field that's just been growing tremendously over the years. It's a very common practice in research today, and there's a severely outdated analysis I did of papers in the field from 2015, but I think they're still about correct today. The cancer field is still a very large one where these technologies are being applied a lot, as well as fields like neuroscience, developmental biology, and other areas like immunology. So one of the limitations for single-cell technologies is that although many of them have been commercialized now, we have great methods to do single-cell DNA sequencing, especially copy number analysis, but also to look at mutations. We have great methods to do single-cell RNA sequencing that are scalable, high throughput, 10x genomics, also combinatorial indexing methods. And many of these have been commercialized now as well, Pars Bio, Scale Bio as well, and even great methods to do single-cell epigenomics, DNA methylation, ATAC sequencing, CUT&Tag methods. Most of these are unimodal. There's a few now that have been commercialized, like there's a DNA and ATAC-- Sorry, RNA and ATAC method from 10x Genomics. But the majority of methods are unimodal, and they do great. But that's a limitation because a lot of times we do want to understand how these modalities relate to each other and lead to biological phenomenon. So I just wanted to brainstorm a little bit about some of the questions we can ask if we have really great multimodal technologies, and the biology that we can study with them, especially in cancer. So one of the first ones is really just understanding the central dogma of biology. How do things like mutations and copy number changes result in changes in the transcriptome or the epigenome, affect protein translation and protein abundance, and then ultimately affect the phenotype of the cell? Another one is really to understand the relationship between the genotype and the phenotype. And so how genetic alterations like mutations and copy number can lead to changes in the cell's proliferation, properties, the cancer hallmarks like EMT or invasion, hypoxia migration. And so really if we can get multiple measurements from the same cell at scale, we can understand these relationships. Another important question has to do with cancer evolution, and we know that we can reconstruct genetic lineages of cancer cells and understand how different subclones evolve in the tumor over time. But a real question is whether these phenotypes that we just mentioned in cancer, are they encoded by the genetics or which ones are plastic? We know things like EMT, for example, doesn't have a lot of known genetic mutations or copy number alterations. And so with this type of data, you can reconstruct lineages, map phenotypes onto those lineages, and understand about plasticity in cancer. Another important question is to understand kind of cis and trans impacts of genetic alterations on things like gene expression. In other words, when you have a copy number gain or loss or even a mutation, how does it impact the genes in those regions, and how does it also impact genes across the genome? And how can that affect regulation that affects the phenotype of the cell? And so by looking at things like subclonal alterations in cancer, you can start to get and study at some of these relationships. Another one is understanding cells of origin in cancer. So in many cancers, we don't actually know what the cell of origin is. It's usually an epithelial cell, or in sarcomas, it might be another type of cell. But with these methods, we can link the programs of the cells directly to their genetic profiles, and that can teach us a lot about cells of origin. In breast tissues, we have three different types of epithelial cells, luminal cells that form the lumen inside, the myepithelial cells around the outside. And there's big questions about whether these give rise to ER positive or negative breast cancers, which are the two main subtypes. And a final question that I think we can address here is are there DNA alterations in non-cancer cells? In other words, can you acquire mutations, copy number events in cells like stromal cells or immune cells, and can they also help promote a cancer progression? This is also very understudied in the cancer field. There was a nice paper by Fuqiao Tang's group not too long ago showing that fibroblasts in colorectal cancer often have this chromosome 7 amplification across patients. And they suggested that that might contribute to more of a CAF-like phenotype that promotes cancer progression. But for most cancers, we really don't know whether these alterations exist or not and whether they promote cancer progression.So, some of the first techniques that were described to look at DNA and RNA in a cell, which I'll talk about a lot in my presentation today, included things like GNt-seq. This was by Thierry Voet's group, and DR-seq from Alex van Oudenaarden. These important studies showed that it's technically feasible to measure both modalities from a cell, but they were quite difficult to perform. GNt-seq, you have to physically separate DNA and RNA from a cell before amplification. DR-seq, the bar coding is quite difficult. So they are low throughput. They were expensive and very high costs, but they were pioneering for the time. This is really around the time that the first single-cell DNA and RNA methods were being described. There have been another series of papers, work from Fuchou Tang. Also more work from Thierry Voet's group. Also DEFENSEC from Peter Sims' group, and some methods like GOT from Dan Lundahl's group. And these are all methods that allow you to look at DNA and RNA simultaneously in cells. They still have some challenges. For example, DEFENSEC doesn't completely lyse chromatin, so the complexity of the DNA is low. GOT is really looking at mutations in transcriptomes, so that's affected a lot by expression, RNA editing. And these methods are still relatively low throughput. So there's still a lot of work to be done here, I think, to be able to scale this up. And so we really want to be able to scale these co-measurements up, have better high-quality DNA library complexity that you can look at, better genomic resolution, and then streamlined protocols that you can perform in the same reactions or tubes so that they can be adapted for micro droplets or nanowells. In my own lab, over the years, we've developed a lot of unimodal technologies. So, starting with some of the early single-cell DNA sequencing methods that use things like DOP-PCR, methods that use things like MDA to look at mutations across the genome at base pair resolution. We've also developed some methods to do single-cell RNA sequencing. And then more recently, we've been using tagmentation chemistries to look at single-cell DNA alterations. But they've all been unimodal so far. More recently, we've tried to really develop multimodal methods, and so we have a few methods developed now that are in papers that are published or coming out. One of them is WELL-DAC that looks at DNA in ATAC. It does essentially two rounds of tagmentation, one with the chromatin ATAC, one with it removed. This is now on bioRxiv. I won't talk about this today, but you're welcome to look at this paper. Another paper called WELL-DR or WELL DNA and RNA sequencing. This was published in "Cell," and this is one of the first methods we published to do the chemistries. It's a DNA first chemistry, so we do tagmentation of DNA initially, followed by the RNA chemistry. And today, I'll talk a lot about a new method called ARCH-DR or archival DNA and RNA sequencing. And this is a new method where we can look at copy number alterations, mutations, and RNA simultaneously in thousands of cells at the same time. So I won't talk about the WELL-DR study too much because it's already published. It does allow you to scale up things quite a bit. In nanowells, you can look at thousands of cells, usually about 1,800 to 2,000 cells per experiment. It did have very good genomic DNA resolution. The RNA quality, I would say, is moderate, and so that's something that we've been improving on a lot. But it requires fresh cells, so that's quite a limitation for clinical samples. It can't really look at mutations. And yeah, as I mentioned, the RNA quality was kind of moderate. So, today, I'll be talking about the new method, ARCH-DR, and how we apply this to study esophageal cancer. And this is a collaboration and work that was done through Rui Yi, a postdoc in my laboratory, Kyler Wang, former postdoc, now a PI in Shanghai, and Steven Lin, who is an oncologist in the Radiation Oncology Department. So esophageal cancer or EAC is one of the most common forms of esophageal cancer, esophageal adenocarcinoma. The major risk factors are acid reflux that you get, as well as smoking, alcohol use, and there's about 21,000 new cases per year in the US. The five-year survival rate is only 20%, so we really need better treatment modalities here. And, interestingly, it has a much higher incidence in China, Iran, and Central Asia. These cancers are interesting from a genomic perspective because they have lots of extrachromosomal DNA. We also see more intratumor heterogeneity and aneuploidy in these cancers and many other types. And there have been some driver genes identified, including things like CDKN2A, KRAS, ERBB2, TP53, EGFR, and PIK3CA. So the questions we're interested in studying here is really what's the cell of origin for EAC, and is there switching of cell lineages during cancer and clonal evolution? Which are the ecDNAs that arise, and how are they inherited through the genetic lineages? And which driver and copy number and mutational events lead to gene expression differences? And finally, are different phenotypes inherited in the genetic lineages, and which are plastic? So, some of the technical challenges then is that this extensive intratumor heterogeneity exists, and so, we need to use single-cell DNA methods. We want high genomic resolution, so we can detect really focal events that are ecDNAs that might only be a few hundred kilobases. We need a co-assay method to look at DNA and RNA, and we want to be able to link the cellular epithelial lineages and identities to the genotypes. And we also want to be able to look at frozen and FFPE materials because fresh materials are not always available and don't have long-term clinical outcome data available. And we want to look at mutations and copy number events. So in this method that we've developed, archival DNA and RNA sequencing, we can look at RNA species, copy number alterations, and mutations all at the same time in the same cells. We've done a lot of work to look at QC metrics, and we've shown that things like overdispersion is very low, similar to our previous methods and things like DLP from Saurabh Jha's group. The breadth of coverage is very high, so similar to what we see in our unimodal methods for single-cell DNA sequencing.And for our RNA-seq, we see very high gene counts, which is great, higher than what we see for 10x version three, and much better than our older Well DR chemistry or what we see with Defense EQ. This is just an example of a human tumor sample where we took the same tissue, we made a cell suspension, we snap-froze the tissue, and we embedded it in FFPE, and then we ran the same assay on all three samples. And at least at the cell type level, you can see that the cell types overlap very well by the fixation method or preservation method, and it suggests that we're still getting a really great RNA quality from these tissues, even despite fixing them or freezing them. So in this study, we looked at 12 patients. We dissociated the fresh tissues, we ran ArcDR, and then we were able to look at RNA copy number and whole exome mutations at the same time. We looked at about 34,000 cells in total with these co-measurements and investigated some of the questions that I mentioned earlier. Just looking at the RNA data, you can see that we typically see that the patients have patient-specific clusters. That's largely due to copy number alterations, which create gene dosage effects since every patient has kind of a different overall global transcriptome profile. In the other cell types, we see that they overlap quite a lot between patients, so that's good. So first, we wanted to look at the different cell lineages. And so in esophageal cancer, the three major lineages are esophageal, there's actually two separate ones, gastric, and then intestinal, we have three different types. We leveraged some data from the Human Atlas projects for the GI as a reference, so we can really distinguish all these different lineages very well. And then in our own data, we saw that the patients clustered a lot by the lineages, which were different across different patients, as shown over here. And this plot over here shows you that within one patient, there typically isn't a single lineage. You see one patient will often have multiple lineages that are present. And I'll show that in a few slides after this as well. So interestingly, when we look at copy number data from a single patient, this is one of the patients here, we can look at all genomic alterations and identify subclones in this heat map. In this particular patient, we had a separate subclone that was very different, didn't share any of the copy number alterations and really looked like an independent lineage. Now, when we looked at the mutations, actually found that there were a few shared variants, although that lineage down here has a lot of mutations that are different as well. And in a subclonal tree, you can see that that divergent lineage or independent lineage was kind of out here, an outbranch, and it had gastric epithelial cell type, really looking like a different cell type. On the other hand, the other subclones were over here, and they have kind of intermixed lineages, as indicated in the colors here, some of the intestinal populations over here, and some of the esophageal populations over here. But interestingly, when we see this in patients, there are about three or so. When there's a divergent genetic lineage, it typically has a different cell of origin as well. But as you can see, most of the subclones have multiple lineages associated, suggesting that there's pretty high cellular plasticity here for the epithelial markers. Next, we looked at extrachromosomal DNA events, and in EAC, we see quite a lot of these. The easiest thing is if you look at this line plot down here, you see these very thin lines. These are high copy number events. Interestingly, in this patient, there were two separate subclones. One subclone had high EGFR amplification, the other one had high ERBB2 amplification, and that was indicated in the lineages here. These subclones over here have ERBB2, these have EGFR. If we look at gene expression, it correlates very well. The subclones on the top have high ERBB2, and they have low EGFR. And on the other lineage, they have low ERBB2, high EGFR. This is actually important because in the ecDNA field, it's still unclear how these things are inherited in genetic lineages and whether it kind of follows what you would expect for a kind of mitotic inheritance and segregation, or whether it's just random dispersion. And so we really see that it does follow the segregation pattern. We use methods like Amplicon Architect to reconstruct and assemble the reads across the subclones and show that these are really circular species that exist at a high copy number level in the cells. Interestingly, we looked at the gene expression for different ecDNAs and overall across all the patients. And we found that for ecDNAs, the gene expression really saturates around 20 copies or so. You kind of see this linear increase with copy number as you go from zero to 20. But once you approach 20 copies, it's fairly flat. And so that allows us to understand and interpret that perhaps cells just can't tolerate really, really high levels of ecDNA or expression. We also looked more globally at the correlation of gene expression and copy number events. And we found that overall, there's about 51% of the genes that when you have a copy number gain or a copy number loss, follow that pattern. And there's about 50% genes that don't agree with increased expression or lower expression in copy number regions. And based on that, we can classify genes into dosage sensitive or dosage insensitive. And it's interesting to look then at a lot of the EAC driver genes. Some genes like CDKN2A really has this linear increase and gene dosage relationship, as well as KDM6A, another event that's common in EAC. But other genes, even ones that you in other cancers you know correlate a lot with expression, in EAC like MYC doesn't seem to correlate with copy number and increased expression levels. And that's also true for ARID1A. And so we can classify all the genes based on whether the data is sensitive or insensitive. Next, we can also look at mutations. And so mutations, it's also very interesting to understand the relationship with gene expression. Overall, these were the variants that we found and looked at. We found that about 20% or so correlate with differences that happen in gene expression. And so we're showing you some examples here. So interestingly, if you look at P53, we compare wild type subclones versus mutant subclones. The ones with the mutations have much higher gene expression levels. But if you look at copy number in EAC, actually, there's no difference.CDKN2A as well, when you find a mutation, there's a very high difference in gene expression. And then SMAD4, this is an inactivating loss of function mutation. And so when you have this one, you see that actually the expression goes way down, whereas there's no difference in copy number. So overall, we found that about 20% of somatic mutations actually had an impact on gene expression, which is different, of course, than what we see for copy number. Finally, we looked at these questions around genetic hardwiring and plasticity in the lineages. And so what we do is we take all the single cell profiles. We can take a consensus and look at different subclones as shown over here, and then we can reconstruct a copy number lineage. This is with a method called MEDIC2, which is based on copy number events. And it has a bifurcation here into these two major lineages, and then within those, there's differences in the subclones. And so we found that some genes like MDM2 are very heritable, so you see them only express high in one lineage. This is MYC over here. It's also expressed mostly in one lineage, but other genes like cyclin E1, you can see that it's expressed across the lineages. And so we would call this a more plastic type of gene. So we can do this across all the patients. So we can look at pathways now and signatures that follow the genetic lineage or are independent of it. And so we find some pathways and signatures that seem to be very genetically heritable. In this case, I'm showing you this example of wound healing, where you can see it's very constrained to the genetic lineages across multiple patients, and following here, also this lineage. And other phenotypes, things like mitoses. We also saw DNA repair. These look to be much more plastic, and so you can see that they exist across multiple lineages. It's high over here as well and here as well across the lineage. And so, those seem to be much more plastic or maybe more transient programs that occur within lineages and are not always tied directly to genetic events. And then finally, we looked at the non-cancer cells. And so, in this study, we were looking across all the cell types that are essentially not the tumor cells or aneuploid cells. And we found that there are chromosome X losses, and they're very common, mostly in the T cells. We don't see those in other cell types. And of course, we looked at this mostly in the female patients, where there's two intact chromosomes. Interestingly, when you lose the X chromosome, there's really no difference in expression. That's probably because you're losing that inactivated allele that's been inactivated by Xist. In contrast, for the Y chromosome loss, when we see that male patients lose the Y chromosome, there's a big difference in the expression of genes. And we can also map both of those events very granularly to different cell states. We did this in T cells, and so we see some different distributions across the T cell populations, but actually Y and X are occurring pretty much across all of the T cells. And I think the loss of Y chromosome is very interesting. There's been more attention to this paid recently in the cancer field, showing that sometimes loss of Y chromosome can affect cancer progression and also T cells in patients. So, to summarize this data, we developed a high-throughput method that's compatible with fresh, frozen, and FFPE materials. It's pretty scalable. We can look at thousands of cells. We find that even though EACs have this clonal origin, so at the genetic level, they really start from just a single cell. However, they're able to switch, and they're very plastic during progression in terms of the lineages they represent from the expression level. ecDNAs are common. They're segregated, usually in genetic lineages. And we see that they saturate in expression around 20 copies. And about 51% of the genes are dosage sensitive, so they correlate with gene copy number, but many genes are also not. And about 20% of mutations result in expression changes that we can see in the single cells. And so the EAC patients have both plastic and genetically inherited genes. And as we look at more patients, I think this will give us a better understanding of events that are caused by genetic differences versus non-genetic differences. And just on my last slide here and to set the stage for the next talks. Really, I think the ultimate goal here is to be able to measure many modalities from single cells. Hopefully, in the next five to 10 years, we'll get there. This is a fast-moving area. So in addition to DNA and RNA, which I mostly talked about today, we'd like to be able to look at epigenetics, look at methylation, ATAC, Hi-C conformations, certainly protein abundance modifications as well. And I think what's missing a lot too right now is imaging data. So if we have ways to do high cellular imaging, we can also get morphometrics, look at subcellular differences as well, and to be able to do this all modalities at high scale. And then we can really learn a lot about cancer phenotypes, cancer hallmarks, how these things ultimately interact together to affect cancer and cancer progression, and then how we can target that to improve cancer treatment and diagnosis. So I think I'll end it there. And a big thanks to everyone in my lab that did this work. I especially want to recognize the work from Rui Yi, who did a lot of the technology development here, Kyler Wang, and our collaborators in the clinic, especially Stephen Lin for this project. And I think I have a little bit of time, so I'll take some questions. Thank you. Thank you for the presentation. John Schatz, University of Miami. I'm wondering, when you looked at the non-tumor cells, if you did any analyses of the state of the microenvironment relative to the tumor genotype/phenotype relationships that you established? For example, loss of antigen presentation machinery leading to depletion of CD8 T cells, that sort of thing. Yeah. So with this RNA data, we can look at all these cell states in quite granular detail. And so I was just showing some examples for the X and Y chromosome losses that occurred. And-We see a lot of changes at the RNA level, obviously. So, for fibroblasts, we see a lot of CAFs in tumor, in endothelial cells, we saw a lot of tumor endothelial cells. But we don't always see genetic events that are associated with those phenotypes. So far, I think, at least the data we have in EAC doesn't show a strong correlation with those tumor-associated cell states and genetic events that we have. But it may be cancer type specific. Certainly, the work Fuchou Tang did in colon cancer showed that a very common gain on chromosome 7, I believe it was, leads to CAF states in those cancers. So I think we have to look at a lot of cancers and try to understand, and see if there are genetic events that also drive things over in the stroma or the immune cells. Thank you. Andrew? Hey, Nick. The gene dosage work is super cool. Do you see instances where you have, in some patients, gene dosage sensitivity and then others where it's not? So, like a patient-specific, whether that gene is affected by dosage. Yeah. So those are really interesting questions, and I think what you're getting at is differences potentially at the epigenetic layers that might be regulating those things, knowing your background. But, yes, no, we try to really classify things across patients. So there was also a lot of patient variability, too. I think we're missing the extra layer. It'd be great to look at methylation or ATAC data that could help us understand those 50% of transcripts that are not correlated. When I try to think about things more longitudinally, I imagine that there's a lot of aneuploidy occurring early on in progression, and so at that time, probably gene dosage is very highly correlated, I think, with the copy number events. But cells have to respond. You can't tolerate thousands of genes being upregulated or downregulated. And so I think that's where the epigenome comes in and regulates things. And so by the time we see the tumor, which of course is after millions and millions of cells, I think there's a lot of biology that's already happened to try to do gene regulation and get these cancer cells into their high fitness programs that lead to the expansion and everything. But yeah, thanks for the question. Thank you. Over there. Sandy Markowitz at Case Western. I'm wondering whether you know in the instances where you're seeing sex chromosome loss in T cells, is that specific to the T cells in the tumor, or is it systemic? That's a great question. We haven't looked a lot at many normal tissues. We have seen these losses in several cancer types, so I think we would have to look a bit more at normal tissues and see if these events are common as well. Yeah. So whether it's a cancer-specific phenomenon or whether it contributes to cancer in some way, I think that would be an interesting question. Even though you lose the inactivated allele, I think there have been studies showing that even when it just inactivates one of the X chromosomes, there can still be gene expression that might be important, and so it's not like it's completely inactivated. So yes, I think it'd be interesting to look at normal tissues in more detail with methods like these. And can you also comment on the state of technologies for doing single-cell methylation studies? Yes. Actually, I think we'll be hearing a lot about that in the next talk, so I'll save that for Andrew. But certainly, there's a lot of work been done in that area. There's also some commercial platforms now. ScaleBio, which was recently bought out by TenX Genomics, has a comet tail indexing-based method. And I think you'll see a lot of new technologies being developed, especially over the next year or so, that are coming out soon. Thank you. Hi. Yashvin Manickam from UCSD Medical Center. Great talk, Dr. Navin. I was really interested in the slide you showed showing the differences in pathway variation across cell lineages, and I was wondering if you could speculate as to the effects of differential cell proliferation and DNA repair on the distinct lineages that you see across cell types in the study. Yeah. To me, those programs make sense to some degree because, of course, cells, they're going into cell cycle, but then they're also going out and going into G1 or G0. And so it's a transient state. So I'm not surprised that we see certain subclones that may be proliferating more and others that are proliferating less, and that we can capture that. Of course, there are genetic events that are associated with proliferation, like in breast cancer, cyclin D amplification, RB1 mutations, those things. Those occur earlier on, typically, in cancer progression. So that's not really what we're looking at here in terms of classifying things as plasticity, because the lineages I show you are branching after the MRCA. So you can think of it a little bit like the later stages of cancer progression. But yes, and DNA repair as well. Again, it's usually something that might be going on and off and may have higher levels in certain subclones than others. So, yes. But those are the programs that we see that, at least in the subclonal data, seems to be a little bit more independent at that time in the tumor from genetic events. Yeah, it's very exciting work. And I think that the slides you were showing with the correlation of specific mutations correlating with gene expression is very exciting as far as expanding that into the multi-omics scene, which I'm sure we'll get into in later talks. So yeah, thank you very much. Great. Thanks. Maybe last question. Thanks. Hi, Kara Maxwell from University of Pennsylvania. So I was wondering if in your data, if you've looked to whether or not the ecDNA load correlates with TP53 mutation status or expression levels yet? Yeah. Great question. We didn't look in detail at the correlation, I think. I think that's interesting. A lot of these patients do have P53 mutations. We do have that data. It might be something that we can go back and look at some more. So thanks for the suggestion. All right. Great. Thank you. So, let me introduce our next speaker then. It's Charles Gewirtz. He's coming from Stanford University. And his talk is High Content, Single-Cell Multi-omics Can Resolve Molecular Determinants of Cancer Cell Phenotypes. Thanks, Nick. So, great first talk to set the stage here, and I'll continue with a lot of the same kind of themes. But I do have one disclosure, which is relevant to what I'll be presenting. So the way I'd like to frame this is the field's split in a couple different areas right now. So, as you'll hear from Andrew, he's developed technologies where you can do lots and lots of cells. But typically, because of sequencing cost, you have to choose, am I going to do a lot of cells, or am I going to do a lot of the genome? And so we've focused on doing small numbers of cells, but trying to get very, very high content information. I think the field, in general, is moving towards lots of cells at very high content. And that's really, I think, going to be driven over the next couple of years by the significant drop in sequencing costs. So my strategy, similar to Nick, I think, as well, has been to really focus on the DNA first. And so we're trying to make really the highest quality DNA measurements from a single cell. And then on top of that, continuously add layers, the transcriptome, epigenome, other things that Nick mentioned. And then use that same cell data to try and train models that can connect genotype to phenotype. So what do you get from going to single cell? I get asked this question sometimes. So there are very sensitive methods for looking in bulk, and you can detect really rare variants in that. But you don't know if they occur in the same cell. You don't know the phylogenetic relationships. You don't know the associated phenotypes with those, so connecting genotype with phenotype is not possible. And then when you get down to the single cell, you can actually do things like compute the number of mutations per cell, look at signatures per cell, look at early versus late signatures, a number of different things. So how have we measured-- What's the most accurate ways of measuring mutations in single cell? So this is work that's been going on for some time at the Sanger, looking at normal cells. And they've developed protocols where they isolate individual cells, they'll expand them over a month or two, and then when they have enough cells, they'll sequence them, and with that, be able to get very accurate numbers of mutation calls. And you can see naive B cells, memory B cells, and they can look at a person's age, how many mutations they acquire. So this, I think, was a great review on somatic genomics. And the thing that really struck me was when you look when they actually quantitate the number of mutations in normal tissues, in the liver, we have 410 trillion mutations by the time we're 50. Even in our brain, we've acquired across all our cells about 74 trillion mutations. And so there's a whole level of genetic diversity that's hidden there that we don't know about it. So the issue, though, with being able to do this type of study is that you're actually selecting for specific cells that you can grow. The phenotype can drift over time. Some cells you can't grow at all. And so it's really quite limited in connecting genotype to phenotype. So the more direct way to do it would be to amplify the genome. But the bottleneck for a long time has been the quality of the data you get back when you go from one copy of the genome to thousands of copies, you need to make a sequencing library. And as I'll show you, you typically get these biases where you get regions of lots of coverage and other regions where you don't cover very well at all. And so again, the issue's been poor uniformity or coverage. That leads to changes in your single nucleotide variant sensitivity. You get false positives introduced depending on which enzymes you use. And there's really been lack of reproducibility, which I think has limited the ability to scale to lots of cells. You really need a very robust, reproducible chemistry to scale to lots of cells as well. So what we came up with in my lab, so the traditional workhorse has been this method called MDA. Generally, how it works is you have a random primer, you add Phi 29 polymerase, it runs on for tens of kb, it has strong strand displacement activity, and then it starts to amplify with branching structures in an exponential manner. And what that does is it leads to this high coverage of some regions wherever the primer's bound first and low coverage of neighboring regions. With PTA, we wanted to take advantage of the low error rate of the polymerase and the processivity, the easiness of the protocol, but we added irreversible terminators to the reaction. And what that does is limits the processivity, slows the reaction down, and then we end up displacing those, but we get small amplicons that in theory would cover the genome in a much more uniform manner. And so if you look at number of reads or amount of sequencing and fraction of genome covered, in the initial paper, we compared to all the other methods that had been out there at the time. So you can see we're approaching what you would cover in a bulk sample within a single cell. And then when you look with MDA, as you walk across chromosome one, this is the centromere, which isn't covered very well in both methods. But you see, again, these peaks and valleys, which you'll get a lot of information from here and very little from the things that are down there. So with PTA, again, with the same amount of sequencing, same windows, you can see we're covering the genome in a much more even manner, not as even as bulk, but much more even, which improves our ability to detect variants and do things like call copy number alterations. And so using the bulk samples as the gold standards, we asked what percent of mutations, germline variants, actually in this case, in the single cells can we detect that were known to be in the bulk sample as kind of the gold standard. And so with PTA, we're detecting about 90% of variants, which is a significant improvement over other methods. And then the precision, which this was a few years ago, things have actually improved significantly, but was kind of on par with MDA. And so doing things like lowering the temperature and adding things like deep variant models for variant calling has helped us improve the precision quite a bit. One thing to notice is methods that use a lot of PCR, so use a lot ofthermostable polymerase, which has one to two orders of magnitude higher error rate, you get lots and lots of false positives as a result of that. And so another important variable is looking at the allele frequency of heterozygous variants. So you want to maintain the two alleles as even as possible. And in a single cell, they should have an allele frequency around 50%. And this is one of the key features that's really enabled PTA to be able to call heterozygous variants quite accurately or variants quite accurately. And so you can see much better than the other methods, we retain that allele frequency distribution of variants that were known to be heterozygous in the bulk sample. Again, compared to the other methods, you can see we don't see that. Again, not to belabor the point but the false positives we see with PCR-based methods are much higher just due to the copying and the heating. One other interesting feature of PTA compared to MDA, so Qiagen is MDA, and in this experiment actually was Qiagen with terminators. And just with the terminators, it changed the reaction and the kinetics in such a way that the small mitochondrial genome is able to amplify much more efficiently in the context of PTA compared to MDA. And so that's relevant to mitochondrial genomes, but it's also relevant to ecDNA, and Dan Landau presented some of his work on ecDNA at the plenary session a couple of nights ago. Also relevant to things like bacterial genomes. So we've worked with a group at the Joint Genome Institute and shown we can amplify and assemble much better single bacterial genomes as well. So this method's been picked up by many groups now including Peter Park and Chris Walsh, who've been doing single-cell sequencing of neurons for some time. And so again, you can see bulk, you can see MDA going across the genome, the dispersion. And then you can see PTA, again, not bulk, but much better than MDA. They've quantitated this with the MAPD, so the lower the better. Again, you can see PTA has lower MAPD with much less variance. And then these are the heterozygous variants, the allele frequency distribution. Again, you can see PTA much more centered around 50%. And so then we wanted to ask, what can we use this for? And as a pediatric oncologist, typically go back to looking at childhood leukemias. So part of this story is that when people have done bulk sequencing of childhood leukemias, they've kind of concluded the genome's not very interesting. It's pretty genetically silent. We only find hundreds of mutations per genome. And so there's really not much interesting there. But what we did is we actually took single cells from leukemia samples. We sequenced the bulk sample and then took single cells from each of those bulk samples and then used SCAN2, which is the variant calling tool that Peter had developed here to get very accurate mutation numbers per cell. And we found each cell actually has 2,000 to 4,000 mutations, where there's only less than 1,000 mutations on average detected in the whole bulk sample. And then you imagine a child when they're diagnosed, they have about 100 billion leukemia cells. So you're missing about 2,000 mutations per cell, but that's going across billions of cells. And so the genetic diversity there is just much, much greater than we can detect with bulk sequencing. We can then do things like look at signatures, whether they're single cell-based or, and then, as I'll show in phylogenies, you can also separate those signatures out over time. In this case, we created three de novo signatures, one which looks like SBS2 and SBS13 from these samples, which are APOBEC signatures, which are known to occur in this type of leukemia. So the other thing we can start to do is look at different regions of the genome and what's enriched for. So in this case, what we found was an enrichment in intronic, exonic, and upstream mutations from the single cells. And so this suggests the enrichment isn't as high for the upstream, but it suggests there are a lot of mutations in regulatory regions upstream that we're really not looking at right now that are important. And as I'll come back to, those could be important when you connect DNA and RNA to be able to connect a mutation to a change in expression in the same cell. And then I've been working with Dan Landau and Tamara as well on this. They've built some phylogenies. They used a totally orthogonal method to estimate the number of mutations per cell. These are normal cells. It looks like there's some pre-leukemia cells here. But again, you can see in the 1,000 to 3,000 range, so quite similar. These were actually different samples, but in the same range of mutation numbers per cell. And then we wanted to ask, getting to phenotype, can we connect genotype to phenotype? And in this case, phenotype is drug resistance. So if we take cells, we expose them to chemo in vitro or in patients. And in this case, what we did was took samples before and after a month of treatment, take the surviving clones, and then say which clones became enriched, and can we learn what was different about them that enabled them to survive the chemotherapy? And so, again, for heme malignancies, patients typically go into remission, but a number of a significant percentage are MRD positive. And so what we did was took a sample here, took a sample after a month of treatment, and then did whole genome sequencing on those, built phylogenies, and asked how does the phenotype map onto the phylogeny? To do this, we did index sorting, so we also recorded six surface markersSo in each well, we know the expression of six surface markers. We also have whole genome from them and can put all of those together. And so this is just showing you what you're missing in bulk. So at diagnosis in bulk, there were two different activating JAK2 mutations in this sample. At day 29, they were a small percentage of the cells, so we didn't detect any of those. At diagnosis, there was, looks like a clonal copy number alteration on chromosome five or six, and a subclonal chromosome alteration. And then again, the copy number profile's flat at day 29 because the tumor cells are quite rare. So when we sorted out the individual tumor cells, in each of them we found one of the two, almost all of them we found one of the two JAK2 variants. But then we started to see different variants we didn't detect in the initial bulk sample. And then we saw a whole range of copy number alterations. So in each one, we saw this founding event, this loss on chromosome four. We saw this other subclonal event that was restricted to this clone, and then we saw a whole range. This one was restricted to this clone, was the most common that we didn't detect in bulk before. But then we saw many different copy number alterations, mostly deletions, that we didn't see at all in the bulk sample. So we then had the ability to think about how do we integrate CNV surface markers, SNVs, and then also, again, map pre- and post-treatment samples and ask, going down the phylogenetic tree, which clones were enriched for? So where do you see... In these two clones in particular, we didn't see them prior in any cells of those in the pre-treatment sample, but we see those clones in the post-treatment make up significant percentages. We didn't see a change in the major copy number changes. But again, we saw these two clones that really stood out, and we have coding SNVs here that were associated with those clones that weren't detected prior that you would think are playing a role, or we would hypothesize are playing a role in the resistance. So we did some work with Dan on this as well, and Dan really developed this tool called PATH, where, again, this concept that Nick introduced, whether or not something's heritable or plastic. So is it randomly distributed on the tree, or does it associate with a specific clone, suggesting that genetic determinants within that part of the tree are driving the phenotype? And Dan and Josh developed a quantitative framework to really measure that. And so as a first case study, we have JAK2 mutation, which we know is heritable, and then we have a copy number alteration, which we know is heritable. We have something like sorting time, which we know should not be heritable. And we use those as controls, and then we looked at the surface markers we had and said, "Which ones appear to be most heritable?" And again, this was a first case. We only had six markers, so kind of limited in what we could look at. But really the powerful part is the quantitative framework for being able to do this on larger data sets. And so the larger data set would be being able to do this with genome and transcriptome. Again, being able to connect a really high-quality genome with a high-quality transcriptome and ask the questions that Nick introduced. How are the two interacting? How does genotype influence phenotype? How does the context of the genotype change the consequences of it? And so we developed Resolvome, which is a single-tube reaction where you lyse the cell. We do a reverse transcription first and biotinylate it, and then we lyse the nucleus and do PTA, and then separate the DNA and RNA after the amplification, and then do paired sequencing. And so we're able to maintain the DNA quality with that and then get good quality RNA libraries as well. So this was the first paper that came out, but a number have come out, several have come out since then. So Joe Gleason was using this to do lineage tracing within the brain, and then look at the phenotypes of the cells in which he was doing lineage tracing. So Dan, you can look here in his preprint, but has taken this to another level. So this is on normal esophagus, but they've done 3,500 whole genomes and transcriptomes, and so built these phylogenetic trees of the normal esophagus. And on top of that, we're able to find all sorts of oncogenic events that occurred, could characterize which cell types they were in, a whole range of things. So you compare that to what you would get with just a single-cell RNA-seq readout, where you would just get cell identity and maybe some differences between them. I think this is a rich data set for mining deep understanding of the correlation or the cooperation of the genetics in the cell state. So then lastly, methylation was brought up, so we've been adding methylation as the next layer. I don't have a lot of detailed information to present today, but I'll present to you some data that looks very promising. So the idea is quite straightforward. We have PTA, which can amplify the genome. So what if we could specifically deaminate 5mc? And so we found some enzymes that really like 5mc and don't deaminate unmethylated cytosine, so they're marking the methylated. And so we just treat with this deaminase, and then we do PTA. And with that, we get the conversion, and then we also get mutations as well. And so this is percent of CpG sites covered. These are different cells, single cells. So you can see we're covering 90% of CpG sites. With that, we've been developing deep variant models that can differentiate methylation from variants using the context.But we found 80% of CPG islands we can call, and then we still have 90% of the genome left. So we can then add variants on 90% of the genome in addition to that, and then this is a genome in the bottle cell line, so we have a really good gold standard, and when we look at the correlation between the single cells and the genome in the bottle, you can see very good correlation. So still early with this, but very promising and exciting. In the last minute, as I mentioned, really excited about scaling to thousands of cells as the sequencing prices drop. So for example, Ultima has announced an $80 whole genome. They do 30X whole genome. Most people using our technology do a 10X whole genome, so that gets to $26 or $27 per genome. And so at that rate, you can think about doing thousands of cells, and really getting deep information if we had genome transcriptome, and then we think we can do genome transcriptome and methylation in the same cell as well. But number of other sequencing offerings, so it will be interesting to see how we can really scale as that limiting factor gets dealt with. And so again, getting back to AI, I think just a couple take-home points real quickly. So I think there have been a lot of things on virtual cells, just looking at low content data, perturb-seq studies in cell lines, and I think to get really deep models of tissue, we need really deep information and we need primary samples. And so that's what we're working towards. And again, the rare cells is what we care about, which are lost in bulk. And so if you have a treatment-resistant cell, you can build a phylogenetic tree, you can understand the context of it, you can layer on phenotypic information. But ultimately, I think that deep information is what's going to lead to biomarkers that are directed towards the cells you really want to kill or to-- because they're the cells that are actually going to determine the outcome of the patient. And so with that, I thank the organizing committee again for the invitation and happy to answer questions. Hey, Chuck. A really nice talk. I was wondering if you could tell us a little bit more just about the scaling things up. What are the challenges? Are there any strategies you're thinking about? Yeah. Because you have really wonderful data sets, very high quality, very deep data. But it'd be nice to scale things up if there's- Yeah, no. It's something I've thought quite a bit about. Typically, droplet-based approaches are problematic for us because we have multi-step chemistry. And so most droplet approaches are one or two steps, whether it's 10X or Mission Bio. And so those aren't really compatible. The other thing is we like to have bigger volumes. So having teeny, teeny volumes is, I think, problematic when you're trying to amplify a whole genome. And so the first step we've done is really to make the chemistry so it's completely additive. So you can't have cleanups as you go to high throughput. So being able to have the chemistry tolerate a barcode within the chemistry, which we've been successfully able to do. There are other methods out there with combinatorial indexing is a powerful tool for trying to do that. We've been thinking about ways to do that but haven't come up with a real solution. But I think the first thing is getting the chemistry so it's deployable within a completely additive environment. And then scaling. The plan we have right now is working with HP. They have a-- or Tikan, they have a liquid dispenser that's based on inkjet printing, and it can do teeny, teeny volumes, and it can do 1536 plate. There are other challenges with that, but I think we may have solved that. But I think we'll be able to get to 1536 pretty quick. But being able to scale to tens of thousands of cells is going to take a little time. But again, the sequencing cost. To me, if you look at doing an exome capture, you want to do a methylome and exome capture, it almost becomes cheaper at this point to do whole-genome sequencing. And so I think pushing-- And as the sequencing keeps coming down, we can continue to ramp up the number of cells we look at. Thanks. Thanks. And our next speaker is Andrew Addy from OHSU, and he'll be talking about single-cell DNA methylation multi-omic technologies. All right. Thanks. Yeah. So, first, thanks for the invitation. I'm excited to be here, especially in a technology session. I get to talk about one of my core passions, which is technology development. So it's going to be technology-heavy. And talking about what motivated us to develop these technologies, how we've gotten to where we're at now, and really what our current capabilities are to look at DNA methylation and multi-omics. Some disclosures which are relevant to my talk today. So first, DNA methylation, this is the property that I've really been pursuing the most throughout my career. And what's the motivation for that? AndMany of you are aware, this is review, but in mammalian genomes, DNA methylation is typically this methyl group on the five carbon of cytosine bases in this CG or CpG context. We often represent it as kind of like these lollipops, and the CG dinucleotide context is the rarest in the genome because it has this regulatory role. And typically 60% to 80% of these CG sites are methylated genome-wide in a typical mammalian cell type. Whereas the non-CG sites are typically completely non-methylated. There are some exceptions that I'll touch on a little bit later. So generally, DNA methylation is associated with repression. I like to think of it as this initial gatekeeper of genomic permissiveness, where the DNA methylation patterns really dictate which regions of the genome have the potential for activation. So often, DNA methylation of a promoter isn't going to correlate well with expression. It's a common kind of misconception and an over-assumption. It really correlates with the capability of a gene to become activated, that the next layers of epigenetic machinery, such as open chromatin, actuating that open chromatin with a transcription factor to then lead to active transcription. And because it forms this real masking role of the genome, it really makes it the most basal layer of epigenomic regulatory control. So it's kind of that first layer of gene regulation when you're looking at the whole spectrum of gene regulatory dynamics. We know a lot about some of the nitty-gritty mechanisms of how it works. DNMT1, for example, is the methylation transferase that does maintenance. So when you have a new cell division, it copies over the methyl groups at these CG sites. And then we know Tet proteins can target loci for removal of DNA methylation. And then de novo deposition of methylation is with DNMT3A or 3B. And many of you will recognize these genes if you study any blood cancers, whereas Tet2 and DNMT3A are two of the major top driver genes in AML. Which then brings us to the role of DNA methylation in development and disease, where cancer genomes-- And I have the title of this slide is "DNA methylation is almost always altered in cancer." I have yet to find a paper that's really studied DNA methylation in cancer and has not found significant alterations. And pretty frequently across a broad spectrum of cancers, you get this hypomethylation phenotype, where large swaths of the genome end up being hypomethylated and a reduction in that methylated level. This can lead to things like retrotransposable element activation, which can then lead to genome instability, or hypomethylation of promoters that can lead to activation of oncogenes. And more broadly, this really makes a lot of sense with cancer because you get global hypomethylation, which then leads to this reduction of that filter of regulatory control and repression on the genome that gives a lot more pathways that are available for activation. Cancer cells want to adapt and evolve, and removal of that methylation kind of frees things up to enable them to explore those other pathways and increase the cell state plasticity and ability to adapt to external extrinsic pressures. It's also a massive signal for cell-free DNA. If you've looked at a lot of the work from GRAIL, for example, you see a lot of cell-free DNA methylation that's being used for diagnostic purposes. So like other properties, this is cell type specific, so it really requires single-cell resolution to enable looking at it and understanding this property. But it's been traditionally much more difficult to look at methylation than some of these other properties. A lot of this comes down to the conversion approach. So like what Chuck was saying, there's multiple steps that are required to do this conversion, whether you're doing the traditional method where you convert all cytosines to uracil and then methylated cytosines retain their cytosine identity, or flipping it with some of these newer chemistries with methylcytosine-specific deaminases. It still requires multiple steps, harsh chemicals, and it makes it trickier to enable high-throughput single-cell technologies. As a result, most techniques previously have been developed where it's single cell, single well. And some of these can produce great quality data, but it is inherently lower throughput. And there's a place for low throughput, high coverage, as Chuck was just showing. We're shifting things in a different direction. We tend to do a little lower coverage, but much higher throughput, and it's just not amenable with these approaches. Though, there is an exception. This method, SNm-C-seq from Joe Ecker, who's at San Diego at the Salk Institute, who's really a pioneer in the methylation space. They just brute forced it with automation. And here's an example from one of their studies from the BRAIN Initiative. Over half a million cells that are really high coverage, high-quality methylomes. But when you look at what it took to generate this data set, it's almost 5,400 384 well plates with a large portion of those actually 384 well column cleanups. So just in plasticware alone, that's pretty inhibiting for most groups. And I'm not saying this to take away from the paper. I want to emphasize just the amount of work that went into producing this beautiful atlas. But I can't do that. Most labs don't have the storage space for that much plasticware. So that's motivated us to develop our technology, SKY-Met. And this is a combinatorial indexing approach. So we do two tiers of cell barcoding, the first being tagmentation using indexed adapters. We then pool these intact nuclei together, split them out, and then append a second barcode. So this is the multi-tier barcoding. So you can see in the kind of bottom well of this examples here. There's multiple cells in the well, but they each have unique barcode combinations, and it's that combination that gives us our unique index. So for methylation, though, if we wanted to do this where we do in situ tagmentation, the first challenge is that this is the first step for single-cell ATAC-seqSo you're only going to be getting reads specifically at open regions of chromatin, which we don't really care about with methylation. We want to get those regions, but we want to get the regions where it's actually repressed. So you're almost exclusively going to get hypomethylated regions in these open chromatin loci. So we developed a strategy where we take our intact nuclei and then actually remove and deplete these nucleosomes across the genome to provide genome-wide access for this transposase enzyme. So we can get genome-wide coverage. So it's essentially ablating that ATAC-seq signal and getting uniform coverage across the genome. We initially applied this for single-cell whole-genome sequencing, and this is an example where we identified some subclones in pancreatic ductal adenocarcinoma, and showed that we could take that strategy. But one of the challenges here, and one of the things that we've really wanted to move forward with in my lab more and more is making these technologies more accessible. No one was doing this technology when we published it, and the reason is because you can't get these reagents. You had to home brew, custom make a lot of the reagents, which is really a barrier for a lot of research programs. So not too long ago, we wanted to ameliorate this and find a strategy that makes it much more accessible. So to that end, we adapted our nucleosome disruption approach and made it so it's compatible with the 10x Genomics single-cell ATAC-seq workflow. So you can use that off-the-shelf kit with our pretty simple, straightforward approach for nucleosome disruption to enable you to get single-cell whole-genome sequencing in really high throughput. Now, I also want to note this is kind of apples and oranges compared to what Nick and Chuck were talking about. This is lower coverage. It's not deep coverage. You can't really do SNP profiling with this, but you can get copy number profiling. And we applied it to this control cell line and then pancreatic cancer as well. So can we apply this nucleosome disruption approach to get genome-wide methylation using our techniques? And indeed, we can. The first version of this was developed by Ryan Mulqueen, who's now a post-doc in Nick's group. This was definitely a proof of principle. It was a brutal, I think, 16-hour protocol with no stopping points and not super robust, but it was that initial step that showed that it was possible. And then we since did a lot of optimizing with Ruth in my program to develop a really robust, really high coverage, high throughput single-cell DNA methylation technology that we're really proud of its really robustness and ability to be able to be applied to even tricky tumor samples. And then since then, we've scaled this up with our latest version. This one has three levels of cell barcoding. And as an example Vignet data set, we produced 140,000 single-cell methylome data set in the human cerebral cortex. The total library was over 600,000 cells, but we just selectively sequenced a subset of it, which is kind of a nice aspect of it. It works with different conversion methods, and as well it's compatible with different sequencing platforms. We did a lot of this on Ultima because it was a more cost-effective option for us. So most of the data was Ultima. And then we have a multi-omic version that I'll touch on a little bit at the end. But I do want to make a quick plug for our single-cell methylation analysis tool called AMETHYST. One of the things-- and this is developed by Lauren Rouleau, who's staying in my lab. There's a GitHub with lots of tutorials, and the goal of this was really to make it a kind of one-stop shop for single-cell DNA methylation analysis. Our goal here was to get the analysis tools in the hands of the biologist that's doing the data. So to that end, we really put a lot of work into the visualization tools so that biologists can get their raw data, take it through this workflow, and start to look at the genes of interest and understand the biology of the methylomes. I like these plots in the middle here. There's PBMCs on the left where the rows are the different cell types, and then brain on the right. And the blue are these hypomethylated regions. And you can see there's high cell type specificity, and these black boxes are the differentially methylated regions that were called. So it's a really nice way to visualize and interact with the data. And I think it's really important to be able to develop those tools to really allow the biologists to really be digging into the data. Using AMETHYST and SKI-MAPv3, I'm going to touch on a vignette here. And I mentioned non-CG methylation is not present in most cell types. That's true, except in neurons. So I know this is a cancer meeting, but this is a cool vignette where seminal work by Joe Ecker showed that when you're born, your neurons, you actually don't have any of this non-CG methylation, but through the first several years of life up to adolescence, it actually accumulates in your neurons. So it's this postnatal maturation of your neurons during this peak window of synaptogenesis. And we really wanted to study this. So in mice, this corresponds to P7 to P28. So we took these regions in the mouse brain, five microdissected regions, sorted for neurons and profiled them using our SKI-MAPv3 technology. And we were able to produce this really comprehensive atlas of postnatal neuronal maturation. So this is about 240,000 high-coverage single-cell methylomes, where we're able to show all of these branching lineages of early neurons and then differentiating out into these very specific neuronal subtypes. And just a vignette that I think highlights how you can use this single-cell methylation data that's applicable across all sorts of tissue types and certainly cancer, where there's a lot of evolution in cell state dynamics. As an example of looking at the pseudo temporal ordering of the progression of methylomes in this branch from MGE-derived GABAergic neurons. So these are inhibitory neurons branching either into parvalbumin or somatostatin. And you can see here's a couple genes here where in the parvalbumin branch, we're getting dramatic increase in methylation as we go through pseudoage.With a lot less in somatostatin. And conversely, here at this gene, we get a clear hypomethylated window that is not the case in progressive increasing in methylation in the other branch. So you get clear branch-specific methylation patterns. And these concepts can be applied across any dynamic system. And we have lots of other kind of really novel, exciting discoveries in this that really get into the neuroscience, but I'll leave that for those of you that are interested in neuroscience and can look up the preprint, which will be coming out hopefully in the next few weeks or so. So finally, I wanted to touch on how we've taken this to the multi-omics space. So we have our process for doing our nucleosome disruption and taking it through combinatorial indexing with tagmentation for single-cell whole genome sequencing, right? Well, work also by Ryan Mulqueen showed that we can actually tack on this step. It's almost like a freebie processing step for generating Hi-C, which could probably be compatible with Chuck and Nick's technologies too, I would imagine, where you add on this restriction digest and religation, and it allows you to capture 3D chromatin structure. So we showed this is, again, in pancreatic cancer. We can get our single-cell genome plus 3D chromatin, and this is showing the 3D plots here. And we were actually able to identify a translocation. So this is a balanced chromosomal rearrangement. So you can't pick it up from copy number alterations and copy number profiling, but it's a subclone-specific balanced translocation. And indeed, we can apply this to methylation. So this is actually data from a very small snippet of a very large data set on mouse embryo development, where we can take a single cell and look at our 3D chromatin structure. But we also get our whole methylome profiles, and that's what this UMAP is looking at the methylation. And then our chimeric reads give us this 3D structure. And where we're really driving this, the ability to look at the 3D structure in Hi-C from single cells is low because it's not super high coverage. But what you can do is aggregate over these methylation clusters and really capture that. And then lastly is coupling ATAC with methylation. So first ATAC with whole genome. We did this stepwise tagmentation. So we found if you do initial tagmentation with one index complex and then do the nucleosome disruption, and then do another tagmentation with a different index, you actually get whole genome sequencing where a subset of those reads are actually tagged with a different index that allows you to get chromatin accessibility as well as whole genome profiling. And also if you use a different set of indexes, you can actually multiplex different samples together, which is kind of nice because you often don't want thousands of cells for a single sample when you have to sequence really deep for whole genome sequencing. So it allows sample multiplexing. And in general, though, the copy number profiles are a little more noisy than what you would get. I don't know if you're seeing that too, Nick, with your dual tagmentation. Maybe not, because you're doing it where you're isolating the DNA fully after. But we do see more noise. And then same with the ATAC, we get a little more noise than the assay on its own. But we can also apply this to DNA methylation, where we're profiling DNA methylation and ATAC from the same cells. And this was in our publication on the SkyMapv3 approach. We were able to look at nice high-quality ATAC profiles for our neuronal subtypes, as well as methylation from the same cells. Slightly less quality for each of these compared to the assay on their own, but still pretty good and able to do de novo analysis on each of these modalities. But typically, we're using this in a bridging sense. So in new data that we're working on in collaboration with Ted Brown and Geronimo Dimja, we're working towards a large AML project, where we can look at DNA methylation and ATAC. This is the co-assay. If you analyze it for methylation, we get nice cell type separation. Same with ATAC. Thank you. Yeah. Sorry. But we can get nice profiles for each of these modalities on their own. And this is actually a small subset of data from human hematopoiesis, where we've got these different lineages that are branching out from bone marrow samples. There's 140,000 cells, where about 8% of these are actually done with the co-assay. So we have a subset of these cells that are actually co-assayed and have that ATAC signal as well. And we can use this to integrate with other unimodal ATAC data. So it allows us to do that bridging. And we're starting to expand this into bridging with RNA as well. And with that, I want to thank those involved and take any questions. A simple technical question. Will the methods work on frozen tissue, or does it need to be fresh tissue? What's really nice with the methylation is that the first step is fixing cells. So it works with fresh, it works with frozen, it works with banked fixed cells. We haven't done FFPE yet. But I imagine there are ways you could certainly get it to work. It's really robust to input, which is one of the things that's really nice about it. Yeah. I'm kind of fascinated by the non-CpG methylation. Yeah. Have you looked more broadly in your data sets? Could things that were neural crest-derived or somehow related to the nervous system or maybe just other cells within tissues have non-canonical? So there's a great study on that from Joe Ecker's group. It's, I think, in bioRxiv. I don't know if it's published in full yet, but they look at that. It's really, really low levels in other tissues. I think it's hard to delineate between just off-targetMethylation. So, the DNMTs, it's DNMT3A that does it in the brain, and that's also active in a lot of tissues. And if it can do non-CG methylation, chances are if it's really active doing CG methylation, it can do some off-target CH methylation maybe. But the levels are very, very low in other cell types, except embryonic stem cells, where it's a very different patterning than in the brain, and that's a cell type that's turning over, and we have no idea what is going on with that. So that's another area that we're starting to explore. Okay, cool. Andrew, really great talk. I was wondering if you could tell us a little bit more about, I guess, using methylation data to separate out cell types and cell states and when you show the data on the brain, it looks beautiful, like you have very differentiated cell types and they separate very well. But then I think the last slide or two you showed, you can see that- Yeah ... cell types are really close together. You can still distinguish what they are, but there's a lot of overlap, and it's something I think we've been struggling with a little bit too with breast cancer type of data where it's actually quite hard to distinguish cell types. And I was wondering any ideas why that is and- Mm-hmm ... whereas at the RNA level, it's usually very straightforward. And strategies, I guess, to choose subset specific markers, do more supervised analysis or... Yeah, absolutely. So a couple things. One, in the brain, we can leverage that CH methylation, which is a really, really powerful high information content kind of modality. It's almost like a separate modality. With CG specific cell types, yeah, it's more challenging. In this one, this is a progression. This is actually not the most up-to-date UMAP, and we actually have more sequencing on these cells now. And it separates a bit better. But we do expect a continuum in this context. And we are capturing what we expect. But yeah, it's definitely challenging. I think a couple things. One, we need to get better at some of the computational tools to address this. A tool like MethScan I think we've seen works really well. We have our AMETHYST tool, but the dimensionality reduction is just one small step, and we actually encourage people to use a variety of tools to see what works best with their data. But I think there's more that can be pushed forward in that space. And it is relatively sparse data, so accommodating that I think is important. And then, yeah, doing the cell typing, I know that's a challenge. One of the challenges there is that a lot of the genes that are great marker genes for RNA might not be great marker genes for methylation. And the best strategy I have is look at the genes that define the lineage and if they're lineage-defining as opposed to where that gene gets maybe activated earlier on in the lineage specification. Those tend to be more indicative. But it's a variety of factors. But yeah, it's more challenging. It's a newer modality. I remember going through all these headaches with single-cell ATAC-seq when that was new. But now that's much more straightforward, and I think we're moving towards that now. Thanks. You mentioned that the front end of this is the 10X ATAC-seq technology, which is incorporated into their multi-ome technology where you get ATAC-seq plus expression. I'm wondering if you've been able to do expression plus methylation at the single-cell level. Yeah. So we have some ways where we're working towards that. To get it to work with the multi-ome, there's potentially ways to do that. One of the challenges, doing the nucleosome disruption well can really ruin your RNA. So you have to find a balance there. There's the Peterson's, the DEFENSEQ, they do that with RNA and whole genome where they capture that co-assay where they do a nucleosome disruption. It's one of the other ones that we had worked on a long time ago, and it's trickier, but it's possible. The one we like to use for methylation, it tends to destroy the RNA, so it's not really compatible with something like the multi-ome kit. But it is compatible with the combinatorial indexing. If we have some upfront processing of the RNA in certain ways, we can retain it. So we're working towards that, and we do have some datasets where we have it, but it's not as robust as we would like right now. Thank you. Thanks. All right. Thanks, everyone. Let's give the speakers another round of applause. And that ends our session. If there are any extra questions, just come up here and talk to us. Thanks.