2026-04-19-spatial-and-single-cell-omics-at-scale

Date: 2026-04-19 Primary topic: single-cell-spatial-omics Cross-referenced from: bioinfo-tools, single-cell-spatial-omics Word count: ~13,751

Auto-generated captions

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

Okay. Thank you everyone, and welcome to this both technology and biology-focused session on spatial and single-cell genomics at scale. And it's really a session focused on, I think you'll see, decoding the ecosystem of tumors. And I think many of you guys will appreciate that there's been a wealth of new technologies that have really recently come to bear on the ability to measure transcripts at scale within tissues. I'm a big science fiction fan, and Arthur C. Clarke has these three laws, and one of them is, "To discover the limits of what's possible, you have to venture into the impossible." And you might be able to tell that many of these technologies that we have, being able to measure the whole transcriptome within cells and tissues would have been maybe impossible to consider just maybe a decade ago. And so it's my pleasure to chair this session and introduce myself. I'm from the Broad, Dana from Sloan Kettering, and then Itay from Weizmann, who will talk about three different aspects of this. And actually, unfortunately, due to some of the geopolitical circumstances, Itay couldn't make it out here, but he pre-recorded his talk. All right. So we'll just dive right in. If I can move... Okay, great. Yeah. So, my lab has really been focused on developing tools and technologies and thinking about what questions we can answer with these technologies, especially with respect to measuring the transcriptome within tumor ecosystems. And one, we've both been interested in the spatial context and the temporal context. And this is going to be clear to many of us, but cells really only make sense in context. For example, this is a zoomed-in H&E image of one cell, and you might be able to see some features. It has this funny-looking nuclei. But really, it isn't until you see it in the context of the tissue, for example, this is the Hodgkin lymphoma's tissue H&E specimen, you might appreciate that it's this large, multi-nucleated cell surrounded by many other cells of the microenvironment. And what we've been doing in genomics for a long time is a little bit like zooming in to that cell and looking at each cell in its individual context. And it's been focused on dissociated cells. And it's only been recently that we've been able to bring the full power of molecular measurements directly within the context of tissues. The other thing I would just say is that these measurements are snapshots. You have to kill the sample, or maybe the sample's already dead, when you make the measurement, and they're destructive snapshots. But in fact, to really understand maybe mechanisms, you might need to understand the progression and the cellular history to understand how we arrived at the present state. For example, this Hodgkin's lymphoma sample might have been a healthy lymph node in the past, but there are some evolution of lineage and exposure that arose this. And we really have also been quite interested in developing tools to measure this cellular history directly in the context of tissue. So I'll touch on both of these in my talk. And we've been building a variety of technologies for bringing genomics to tissues, including ways to measure transcriptomes, genomes, and analysis tools. I won't talk about this. I just want to say one thing. A class of these technologies have been built on these bar-coded arrays. You can see that here. And this is pretty old now. This is something we developed six or seven years ago, actually eight years ago, where we put tissues onto these arrays that have bar-coded beads, and the beads have XY locations that are encoded by DNA barcodes, and RNA or DNA can be captured onto these beads. And then you know where those RNA and DNA came from by the DNA barcodes. And this, obviously, many of you guys have used technologies. Takara has commercialized this under this seeker technology we call the Slide-Seq, and then Visium HD is very similar. But most recently, a couple of years ago, I feel like we've developed a version of this technology that we use regularly, and I want to just briefly touch upon it today, called Slide Tags. And we think it has many advantages. In fact, we think of it as a kind of a universal platform to move any single-cell measurement, whether that is DNA, RNA, epigenomes, directly into the spatial context. And so how does Slide Tags work? It works by using these same bar-coded arrays, but instead of trying to capture the molecules onto the arrays, we invert the process. Now, we put the tissue onto these arrays, and these beads have photo-cleavable DNA barcodes. And then when you cleave these DNA barcodes, they diffuse into the tissue, allowing these barcodes to be associated with the nuclei. And this is a little bit like GPS for cells, because each cell picks up more than one barcode. It picks up a distribution of barcodes nearby. And with that, we can localize the cell with very high precision, actually higher than the 10-micron resolution of the beads. And it's true single-cell resolution. We can encode basically the location of each cell that we sequence with an XY, or the majority of cells we sequence with an XY location. And it's compatible with any single-cell approach. And so a lot of the data that I'll show today is with 10X, either Phi Prime or Flex. And so here you can see this fancy UMAP, where each individual cell in this UMAP, this is from mouse hippocampus, is placed into the context of space. And what does this really give you? Well, one thing is it gives you really high-quality, true single-cell data. You don't have to deal with segmentation or other sorts of things thatLike result in mixing of cells. And actually, this is comparisons between single-cell RNA-seq or single-nucleus RNA-seq and Slide-Tags, and you can see the cell type proportions, the number of molecules, which is in the thousands, and the counts are all the same. And most importantly, this technology has been commercialized by Tecan Tecara. I don't have any financial interest into Tecara, but they licensed the technology, and they've made it compatible with FFP through 10x Flex. And we've just been consistently very impressed by the data from this Flex in this spatial tagging of FFP samples. Here is a melanoma specimen. And on the exact same paired specimen, we ran it on Visium HD, and you can here see the median UMIs per cell, and genes per cell, for the single-cell technology versus, well, Flex versus Visium HD. And I think you can really appreciate that the true single-cell resolution in combination with molecular resolution gives you much, much better separation between cell types and states. That's all I'll say about the technology. We've been using this as a workhorse and it's readily available for you all. And so, I think a question that many of us, and you'll see in this series of talks we'll try to answer is, how are tumor cells and immune cells organized and shaped by their microenvironment? We're quite interested in how cell-intrinsic states are shaped by their neighborhood and the niche of the cells that they live in. And for example, tumor cells obviously have some dynamic evolution through the course of the tumor, whether that's genomic evolution through genetic and epigenetic means, as well as transcriptomic adaptations to their local environment. And that evolution is actually, through the course of the tumor, is in some equilibrium with the environment, right? It's driven in many part by the immune cells and stromal cells. And many of you guys here know this and are appreciative of this because it's very therapeutically relevant. Because many of our ways of modulating the immune system actually needs to take advantage of what our understanding of this equilibrium and evolution. Also targeted therapies as well. And so for example, response to checkpoint blockade in melanoma is highly bimodal, and it's really unclear what drives this bimodality. We speculate it is something to do with the immune microenvironment. There are many papers around it. But how do we actually understand if cell states within the cancer are associated with the microenvironment? And so we've been looking at spatial niches within the melanoma microenvironment, in particular in metastatic melanoma, using Slide-Tags. We actually did this in our original Slide-Tags paper. I'll show you one tissue as an example, and then I'll generalize these observations over many samples. So here is one of our first Slide Tag specimens on a melanoma. You can see there are these two histological regions of the tumor. And when you perform single-cell RNA-seq, here is the UMAP. You can see very fine-grained resolution of cell states, and then you can place the majority of these cells back into space here. And actually, if you just look at, this is tumor cells, you can see that they adopt a couple of transcriptionally distinct states that are also spatially distinct. And actually, these distinct transcriptional states, they turned out to actually also have kind of different CNVs that represent different evolutionary histories of these two clones, and they have very distinct cellular neighborhoods. Here's just a very, very simple analysis where blue represents the kind of depletion of cell types in the local microenvironment, while red represents the enrichment. And you can see generally, tumor cells in state one are relatively depleted in immune cells in their neighborhood, and then there's a particular enrichment in T cells in tumor cells with state two. And actually, you can go and use, here we used 10x Phi Prime to map the TCR sequences in single cell, but now we have the spatial localizations of those TCR sequences. And you can see here this particular T cell clone that's actually driving this enrichment to state two and depletion in state one. Here, the gray represents all T cells in this spatial context, and purple represents this particular clone. And you can, without any statistics, you can just see by eye that there's this strong tumor state and T cell clone association. So what's driving both this clone dependence and this T cell depletion? And we had a couple theories, one of which is that melanoma dedifferentiates during the course of progression, and as along to this more mesenchymal-like or neural crest-like phenotype. And along the way, then it can lose both its endogenous antigens or maybe it can drive a more immunosuppressive niche. And you can kind of already see that in the H&E. And this is something that clinicians see all the time that melanomas are often pigmented before treatment, but during treatment they lose their pigment. But in this particular sample, you can see that the two lobes, histological lobes of this metastasis actually have differential pigmentation. And indeed, when you look at just the transcriptome scoring for these mesenchymal-like states, you see this difference where tumor one has adopted this more mesenchymal-like state and lost endogenous antigens. And so, this is just one sample. We've then have been systematically profiling many metastatic biopsies, 27 or approximately 30 so far actually now. And I'm skipping a lot of the analysis and data, and I just want to show you two highlights for what we see. One is that this dedifferentiation trajectory actually promote different niches correlated with the cancer cell state. Here you can see-On the rows of this heat map, different states of tumor cells, whether they're melanocytic or this neural crest transitory phenotype, and I haven't shown the fully undifferentiated here. But in particular, we observe that the neural crest transitory state is strongly associated with a particular immune suppressive neighborhood consisted of this SPP1 positive macrophage and this fibroblast that's post-N positive. So this is an immunosuppressive macrophage population, and then a tumor-associated fibroblast population. And association with this is exactly anti-correlated with many, basically effector CD8 T cells and also associated macrophage populations that we term an immune effector population. And when you're next to an immune effector population, we see that the cancer cells generally actually are interferon stimulated and dying. So, basically de-differentiation to this neural crest state causes this association with this immunosuppressive niche that also is anti-correlated with the effector status. And actually, not only is it anti-correlated with these cell types, the T cell clones are suppressed in the niche associated with SPP1 max and post-N fibroblasts. So, in those niches, if you look at T cell clonality, most T cells are singletons, or many more are singletons versus inside of this effector niche, many more are expanded, meaning they're dividing. Okay. And so what's going on with this? Well, we know SPP1 is one of these wound-associated macrophage factors, and it's probably a wound-healing program that also suppresses the local immune response. And we did some functional interrogation of... These are secreted factors, right? So post-N is a secreted factor from this fibroblast. And you actually see that in in vitro, melanocytic cell line, when you apply post-N to these cell lines, it de-differentiates them to a more neural crest/undifferentiated state. And in addition, actually when you add SPP1 to cells, it seems to also protect, only in the neural crest cell line state, it protects cells from interferon and TNF cytokine killing. So it could be that there's this de-differentiation driven by this fibroblast secreted factor, and then SPP1 both protects the cells and then also, I haven't taught you about maybe some immune suppression, but it also prevents the cancer cells from being killed directly by immune cytokines. And so there's this model of these two niches where actually there's this positive feedback where post-N fibroblasts might actually drive cells into this neural crest state, which then positively interacts with SPP1 max and prevents those cells from being killed by cytokines. And then I didn't talk about the immune effector niches, but these are places where you get the right coordination of immune cells to kill the melanocytic or maybe other interferon-stimulated cells. All right. So that's kind of measuring what's happening at a snapshot. We'd like to also understand, this is a clear example of one of these de-differentiation trajectories where you see one lobe of the tumor is melanocytic and one lobe has lost the melanocytic identity. But how did it arrive there? And how do you record the history of basically the evolution of this tumor and its interactions with cells? And so we've been thinking about this a lot. And you can imagine that you can do this through genomics. It turns out that genomes, transcriptomes, epigenomes, and genomes provide three complementary windows in the past. DNA might really be a good way to read out lineage and clonal structure. The epigenome might read out exposure and gene regulation. And the transcriptome can tell you something about the path transcriptional state in a local timescale. And we've actually been building methods to read out all three, but today I'll just tell you about lineage and clonal structure. So, for a long time, we've been wondering, can we use DNA to infer the lineage context directly within tissues? And, obviously, as cells divide, they can accumulate single nucleotide variants or chromosomal aberrations. In normal cells, actually, you can use DNA divisions. Divisions between cells actually still accumulate approximately tens to hundreds of single nucleotide variants per division, although that might be challenging to read out. We actually thought that CNVs and chromosomal number alterations would be very easy to read out with these spatial genomic technologies. And so many years back, actually, we adapted these beads to measure DNA directly through tagmentation, and then capturing the tagmented fragments using these arrays. And using that, you can actually directly visualize CNVs from chromosomes within tissues. Here you can see a capture in H&E of a tumor, and then we've actually captured this onto the array, but instead of capturing RNA, we captured DNA. And these heat maps represent relative amplifications and gains, for example, a chromosome 8q amplification here, and then loss of 15. And you can see that these chromosome 20 might highlight a subclone of these tumor cells. And in fact, when we do simple PCA, we can assign within this very small region of the tissue, these cells to different subclonal assignments.And we proceeded to examine in this paper how the environment is associated with clonal identity, which I won't have time to talk about. But I just want to say that these CNVs show that tumor clones are coupled to extracellular contexts, but the clonal resolution is low. And so in parallel, we've been using these engineered barcodes in synthetic systems to enable higher resolution lineage tracing, and we've been collaborating with Jonathan Weissman's lab and Fernando Carmago's lab, where they have these approaches to introduce Cas9 to edit parts of the genome to increase the mutation rate within those regions, and then you can read those regions out via RNA or DNA sequencing to assess the lineage of cells. And so, in collaboration with Jonathan's lab, we've been doing this with this KP tracer model where you can induce tumor formation in the lung, using pre-recombination of KP. And then you can put these onto slide tags or slide seek arrays and capture both the cell state as well as the lineage information with higher resolution. And this allows us to ask basically how does lineage map to the cell state context beyond just the clones that you observed in the snapshot. And then we were also curious about, are the clones which seed metastasis in this model, are they related in cellular neighborhood to the primary? And so, this sort of thing generates very cool data. Here is spatial data of where we've -- each tumor shot is a cell in space, and we've marked, basically by colors, the primary clones. These are the initial seeded clones. But within this red clone of a tumor, you can derive a lineage tree from those mutations and actually find sub-degradations of the lineage tree actually reveal, for example, the subclonal spatial heterogeneity of the evolution of this particular primary. And you can see, for example, that there's spatial coherence in these subclones, although there's also some slight intermixing. Beyond just looking at where the subclones are, you can use the tree to ask how fast the tumor cells are expanding. Right? Because if they're expanding more, you have a different shape of the tree than if it's static. And you can see here we've marked regions of this particular-- These are three primary tumor regions that are seeded, and only over here are there really, basically cells that are really actively expanding. And if you look in that neighborhood, these are macrophage populations. There's a particular tumor-associated macrophage, Arg1-positive macrophages. These are actually very similar to the SPP1 population that's associated with these expanding regions. And actually, when you perform differential expression, you find many markers of immunosuppressive macrophages in these actively expanding regions. And in particular, you also find that associated with hypoxia. So there are these hypoxic cores that are associated with these suppressive macrophages in regions with high rates of expansion. The other cool experiment that we did, actually, with this is that we sampled from one lung, many sections so that we can capture basically all of the primary tumors. And then we also sequenced the metastases. And so, then we can map on the primaries what their distance to the metastases are. So areas with low allelic distance to the metastases are where the metastases came from. And so you can see here that, basically, the metastases that seeded in this animal all came from this particular tumor in this region. Not even this clone, but this subclone within tumor one. And if you look at gene expression signatures of the neighborhood next to the seeding metastasis versus the mets themselves, you can actually find these other regions are the other regions of the same primary. They have a different gene expression profile within the tumor cells and nearby compared to the seeding metastasis. But the seeding metastasis shares some key microenvironmental factors, such as this hypoxic signature and this carryover of this immunosuppressive macrophage into the environment of the met. And then you can see that the met subsequently has much more fibrosis and collagen-rich signature. So it's actually quite interesting. You can see that these metastasis-initiating lineages actually share tissue contacts with the seeding clone, carries this over. Okay, so, just to summarize this history section, I'm happy to talk about epigenetic state and stored transcriptomes as well. So these are things that we're quite excited about. And one thing that I think history might be important is that differences in past state might create natural perturbations that allow you to separate cause from consequence. That's often how we infer causality in the real world. We observe something happened in the past, then see the consequences in the future. So that's why we're particularly interested in being able to measure past history. Okay, so for the last one minute, I just want to say something we're quite excited about how this sort of data can be used. So I think what we've got now, we have these measurements that allow us to measure both the space and temporal context inside of tissues. But what we really need are ways-- These generate hypotheses, but we need ways to systematically test hypotheses in vivo. And we're doing that through CRISPR screens. I think that's a very powerful way, but we're also thinking about ways where we can pair down-These hypotheses through in silico simulation. And in that vein, we've been trying to build tissue simulators, things that can simulate a tissue through time. And we drew some inspiration from self-supervised generative AI, the things that, for example, behind DALL-E or ChatGPT. And here they're trained by basically self-supervision, by filling in masked inputs. And that's a very scalable way to train the data. And so we've been doing that with tissues, basically making generative models of tissues, using some basically state-of-the-art transformer models from computer vision, where now we can mask the locations of cells or transcriptome states, and we ask the model to fill it back in. So like train it in the same way that maybe DALL-E or ChatGPT are trained. And this is interesting and useful in many ways. One, you can get tissue representations and learn tissue niches, but we're particularly interested in this generative aspect, where it can simulate new effects of perturbations. And so, this is pretty early days, but here we've actually... This is a piece of colitis. This is a healthy colon epithelium, but we've tried to get it to simulate the cell type changes that will get it to... Okay, well, actually, the video won't play. This is interesting. So that will get it to advance to colitis. If the video was playing, you would see that the model can simulate changes in adaptations of cell types that actually capture the disease progression from healthy to inflamed. And actually, from that model, you can measure the effect of perturbing any given cell type on the effect of cell type compositions and signaling of all these other cell types. And so the simulation isn't perfect, but it allows us to rapidly test and generate new hypotheses in silico for these sorts of high throughput CRISPR-based screens. And so, I think the view is that, in the past, we had classical genomics where we can measure molecules. Now we are already at the state where we can measure the organization, interaction, and history of cells directly within tissues. And we're quite excited about these tissue simulators and perturbations that will allow us to modulate these tissue circuits and test these hypotheses. And with that, I'll thank the lab and funding and you guys for being patient and listening. Okay. And so for questions, since I'm moderating the questions as well, you need to use these mics, so that people online can hear you. Guess I'll go first. Hi, Fei. Wonderful talk. I really enjoyed it. I wonder if you could tell us a little bit more about how you could estimate the actively expanding areas, and did that correlate with proliferation markers like Ki-67 and other things that you would typically stain for? Yes. The second part is, yes, it does correlate with Ki-67. What we did was, actually, Dan Landau has this graph algorithm to infer actively expanding areas from the shape of the lineage graph. And I think there's some mathematics behind it, but maybe I'll talk about that with you offline. But we used his algorithm. Hi. Super cool talk. I was wondering about those T cell clones that are associated with specific tumor sub clones. Have you been looking at the T cell receptors and what they recognize? And are you able to track the T cell clones together with the tumor sub clones? Yeah, so that's a great question. So can we figure out what these T cell receptors recognize? I think we would love to be able to do that. It is probably still challenging in the sense that we've been trying. We've been collaborating a lot with Cathy Wu on that. There's computational predictions of the antigens, and then in models where we know what the antigens are, you can use the computational predictions to nominate them, and then you have tetramers, and you can validate. But in human samples, it's much more challenging, because the computational prediction models aren't that great, and we don't quite have a good sense of what all the antigens are. And it's HLA dependent on all these things. But it would be very interesting, for example, if you could use the clone information and the local mutational profile to try to nominate what the pairs are. We'll see if that's science fiction or reality in the future. Thank you. That's great work. I also have a question. Hi. I also have a question on the tissue simulation. So, did you put a hypothetical factor there to try to simulate the gene expression network and then the tissue dynamics? Or basically, you have a history of the colitis process, and you simulate? So how we're doing the simulation is that we train the model on many different observations of what the tissues look like, both inflamed and not inflamed. And so it has learned some cell-cell interactions. And then, basically, we self-supervised train it by masking different parts of the tissue. And then when we simulate it, the model makes certain perturbations by deleting or adding certain cell types, and then it models what the effect of that perturbation is by adding or deleting other new cell types. So currently, it is a cell type simulation. But then, coupled with the cell types is their underlying gene expression, and that's how we get these receptor ligand profiles that are associated with the simulation.Is this two minutes and 40... Okay, great. Thank you so much for that riveting talk. My question is more of a theoretical one. So you have a tissue section that's about six millimeter, right? Yeah. That you use for your spatial transcriptomics. How do you generalize that data that's so limited? How do you square that with the intra-tumoral heterogeneity- Yeah ... of most tumors? Yeah. So I think this is a question that comes up a lot. It's a question about sampling. And, in single cell, when we dissociate the whole tumor, like with a lot of tissue, then you're sampling the heterogeneity in some way, even though your sampling might be shallow. When you take sections, you only have a thin cross-section, but you might actually be able to have a large area. And, I think the answer has got to be like, you empirically learn from the set of many samples and also single cell what the distribution of states is, and then you can answer if you've actually sampled that distribution of states in these sections. Thank you. Okay. I guess one last question. So, I was wondering... That was wonderful. I was wondering whether in your simulations for the clones, have you considered any biomechanical or biochemical- Yeah ... local cues into your niches, and are you interested in doing that? We haven't done that on the simulation side. We are very interested in measuring those things and seeing how that changes. For example, like post N is a ECM factor, and it probably is... There was this talk this morning. It probably is manipulating Yap/TAZ signaling, for example, for dedifferentiating those cells. And so we're quite interested in being able to make those measurements in combination. I think it's a great area for technology development. So we're- So could you infer on those based on the transcriptional signatures of the local stroma? I think you could. Yeah, I think you could, but you need the ground truth. You need to measure a bunch of mechanics in combination, and I don't think we have the ground truth to train it. So that's maybe the call to action if folks want to do that. Okay, I think we'll have to... Okay, one last question. Thank you so much. So one follow-up question regarding sampling. During those simulation works, do you have a magic number? What's the range of a sample size that... And have you experienced when you have a good amount of sample or compared to little amount, do you see the changing of the landscape and the- Yeah ... changing of the interpretation? Yeah. I think it's going to be really dependent on the sample you're simu- Mm ... like the amount of heterogeneity- Mm-hmm ... in the process you're simulating. There, we do certain number of measurements, and then we have a lot of held-out, unseen samples. And we can down-sample the measurements to understand how many measurements we actually need to see before we get accurate, and it's not a lot. But to generalize the totally unseen tissues, for example, it might be many. And so, it's a nuanced answer, I think. Thank you. All right. And next, yeah. That's it for me. And so now we have Dana Pe'er- ... from Sloan Kettering Institute. And she's going to talk about de-coding tissue architecture in the context of tumors. Good afternoon. Thanks for being here. For a while, we have been really talking about technologies and technologies, and now I really want to see, can we get these technologies to do something and understand biology? Oh, God, this screen is huge, so I can't see all these lovely people on that side really well. And so really, cancer was always a disease of mutations, but cancer is really also a disease of the microenvironment. You see benign neoplasia versus full-blown PDAC. It looks completely different. The tissue completely rewires. And this new rewiring of the tissue, that is function. So how can we decode that? And here, this is a really striking image where you see how these pre-malignant cells, particularly these progenitor-like pre-malignant cells that are grabbing stuff and stealing regenerative programs, how they completely change their morphology. These red cells look disorganized and are completely different. So we need to understand this changed morphology, this changed microenvironment, and how does the progenitor cell environment look like? So we dive deep into 10x. This was from a while ago. And one of the most important things when you're using 10x, if you really want to get real biology, is to design your own panel and pack it with the genes that matter. But then even a simple task of assigning transcripts to a cell, what's called segmentation, we realized that's not really working well. And in cancer, this really matters. You don't want to give all sorts of epithelial transcripts to your immune cells. And we don't even have a ground truth because cancers are promiscuous in their expression. So we really want to be able to do this accurately. So I'm going to dive in and sort of interweave a little bit of an explanation of some of the methodologies we build to actually get biology from these rich data sets. And this is SEGER, our solution to assigning transcripts to cells from any spatial transcriptomics data set.And here's the fundamental issue. It's not a geometric problem. Here we have this sort of boundary between stroma to epithelial, and right here, we have contamination. And all these purple genes, they are stromal genes deep into the epithelium, and you could just see how far they go. And likewise, the other way around, we have clearly epithelial transcripts into the stroma. We're just not going to solve this by using a boundary, by using segmentation. This is not a draw boundary. No membrane staining is going to save you here. So our approach is to basically say, "Okay, what do we have?" We have nucleuses. We can look at the transcription of the nucleuses, we can look at their X, Y positions and build this local graph. But what's really important is we're not just going to look at the geometry, we're going to look at the biology, we're going to look at the features of these cell states. We're going to look at the features of these genes and try and take the biology of it and use that to drive our assignment. Specifically, we're going to take this graph that we build and try and use this sort of AI business. But really, we're going to tell it to do something very specific. We're going to build an embedding where a cell, it's close to the transcripts that belong to it. So we're going to teach the AI to know which transcript belongs to which cell. We're going to use this sort of noisy, the nucleus also has misassigned transcripts. It also has this diffusion and 3D stuff going on, but it's close enough. So that's going to sort of teach it the biology of what transcripts belong to what cell states. And then we're going to use that to link everything else. And really what we're going to infuse the full biology. Our AI is going to understand the phenotypic manifold, cell states, how they relate to each other, how they exist, not cell types, but the full rich manifold and how these genes belong to this. And this really gives you a much cleaner thing and better components. You have a third decontamination of 10X nucleus and just as much capture of transcripts as 10X cell. So taking this, I'm going to show a similar animation. We have this spatial data now mapped to cell states, to the UMAP that we're used to, and it's not just cell types, it's the full rich continuum. We really captured pretty much anything you'd capture in single-cell RNA-seq with this. Here's a QR code for the software. And so again, looking at these epithelial cancer premalignant cells, they're not quite cancer yet. We see this continuous axis of a gastric to progenitor state and this continuous axis, it's not cell types, it's continuous. We have continuous change of gene expression along this axis, which is why discretizing cell states isn't enough. But we have this in the spatial context. So we map rich mapping of cell state with exactly the cell and what cells surround it, and we basically take its environment, its nearest neighbors, its environment of K near cells or radius R near cells and see what do we have in that environment. And if we look at what we have in that environment, the yellow cells are these anchor cells. We see that actually cell types don't change that much along this continuum. Again, really pushing against the cell type approach to analyzing data. But we see that actually the stroma, the immune cells, everything else changed dramatically. So what we have along this continual axis of gastric to progenitor-like cells is a pseudo-time trajectory of niches, a cancer cell state, and its surrounding microenvironment. And again, we see that the microenvironment, in this case, the fibroblasts, are changing continuously between these different cancer cell or premalignant epithelial states. And this matches the axis of variation in fibroblasts. Likewise, the progenitor niche is rich in these myeloid cells, and these are immunosuppressive myeloid cells, very, very different than the myeloid cells around the gastric cells. And really, if you want to see it, it really resembles the immunosuppressive fibrotic nasty thing of full-blown PDAC. This is a progenitor niche in a premalignant settling versus PDAC. And we really see this fibrotic TNC and these immunosuppressive macrophages, and this maps all the way to human, where we see a premalignant niche looking exactly with the same features in human. So what I'm going to show you really quickly is that p53 loss enables this nasty fibrotic environment to stay after injury. Usually, these progenitor-like cells clear the tissue, but so that's cleared versus not cleared. You see these yellow premalignant cells just remaining three weeks after, it hasn't been cleared. You see how many more of these premalignant cells there are, and not only do they remain in the tissue with p53 loss, they actually become more progressed, more EMT-like with expressing more vim. So really p53 is keeping this progenitor state, this plastic progenitor state in bay. And by knocking it down with KRAS inhibition, the state that's highly dependent on KRAS, we can really see that we return and lose this premalignant progenitor-like state and the environment which it builds. And so really you have this KRAS mutant progenitor-like states which orchestrates this environment, this cancer-like immunosuppressive fibrotic environment. Activation of either p53 or KRAS inhibition removes these progenitor cells and thus dismantles the niche. Therefore, p53 is really the guardian of plasticity, not just the guardian of the genome.So really, I want to dive into how these cells are organized into space. I showed you one example of how a really, really brilliant postdoc really pieced it apart with these methods in a joint collaboration with Scott. But not everyone is as brilliant as Jose, and we want to generalize these tools. So we need to be able to take many of the units and many of the reasoning, which I just showed you, and try and automate it. And that's what we've been thinking about it a lot. And the critical thing to do any reasoning about these tissues, about these neighborhoods, about how they change under condition, for example, KRAS inhibition or P53 knockdown or before and after therapy, is we have to find a good way to represent spatial units. We're really good about representing cells, the sort of UMAP, the phenotypic manifold, all these things. That's why we've got so much out of single-cell. But, for neighborhoods, we really haven't figured it out yet. And I'm not claiming that this is the be all, end all, but this is something that we really need to solve. And so this is work by Dorum, really trying to think what is a good representation for these niches? And again, most of the current approaches out there, because it's so complicated, because you have these little neighborhoods, you have an entire neighborhood in the tissue of rich transcriptomes. I showed you a data set with 10X with 480 transcripts. Current technologies are measuring 5,000 and 6,000 transcripts. So to pay all that money to measure 6,000 transcripts and then collapse it all to cell type, which is what most approaches are doing because it's easy, really loses most of the information. I hope I showed you just how rich the information is, how continuous it is, and how we cannot collapse the cell type. So we need to do something else. So we want something where we actually capture the full richness, the full transcriptome. But the reason that's hard is that these matrices are complicated entities. And for example, these two matrices are identical. Really, they are just a permutation of one of another. I swapped around the row. And so you need to be able to reason for that. And this type of computation is actually really heavy. So to have a representation of this matrix that understands and is invariant to permutation becomes a computationally heavy problem. And one way to try and overcome that partially, is to do something called optimal transport. And the simple way to think about optimal transport is how much would it take to move one matrix to the other if I pay some type of cost to move each and every transcript from one to the next, but I'm allowed to permute the rows in any which way I want. So take the best permutation that you have and then move around transcripts to get a perfect match, and an optimal transport allows you to do that in a pretty powerful way. But even that takes two weeks for 50,000 cells. The data set I showed you from Jose is 16 million cells, so this doesn't scale. And that's one of the reasons people resort to cell types, scalability. So here we use the awesome power of AI in order to actually create scalability for the function that we want. We take these complex niches, these complex matrices, and we send them through something called a transformer, which is actually a very, very, very powerful way to learn very, very, very complicated functions. And the specific function that we want to learn in this case is we want to learn a representation where the distance in Euclidean space, in this really easy to work with geometry, is similar or as close as possible to this rich, complicated, hard to compute, heavy to compute optimal transport distance that actually captures this biologically meaningful distance between a cell. And because we can train a transformer to build this new latent, and we can do it in a very efficient way, that's the power of AI, what we have at the end is a latent similar to your UMAP, but now this is a latent over not cells, but niches. And we could take this latent and now do all the things we love to do on cells, trajectories, comparing conditions, looking for archetypical averages and typical behaviors. Everything we once did with single-cell RNA-seq, now we have a UMAP-like representation for niches where all this geometry now holds. And so here we took Jose's data, applied now automatically this Wormhole. This is not engineering. It didn't know what progenitor is. It didn't know what gastric is. We did this completely automatic and said, what is the strongest axis of variation through this data? And once we had the strongest axis of variation through the data into this Wormhole transformer, without even knowing any biology, it could pick up that the gastric to progenitor axis and the resulting niches is the strongest signal of the data. And I put it here in a tissue that you could sort of see these organized, nice, round, well-behaved gastric niches are colored as gastric. And these disorganized, like from the first slide, progenitor niches are these disorganized mess. But this was found now more automatically. And really what matters is you want the embedding to encode meaningful biology. And the thing that is similar between the Seger algorithm and the Wormhole algorithm is that the latent matters. We want latents that are a microscope, that encode biology, that teach us biology, not just a black box. So in order to get valueFrom these things, we don't want black box AI. We want to build latents that encode biology, in this case, encodes meaningful spatial distances and meaningful spatial representation of all these small spatial units of neighborhoods in the tissues. So now we're going to try and use this approach to really try and interpret awesome data. This is collaboration with Thomas, who's hiding in the back, and a really stellar, hardworking group. It's actually mostly Thomas's group with Anu, Kate, and Young, and Sam from my lab trying to compute on all the fire hose of data that these guys generate-- gals generate for him. And so what really powers this is this, what I call the supermouse. And the supermouse is really taking genetic mouse engineering to the next level and putting this cassette in a very specific way. So now we have moved to understand full-blown, nasty PDAC. We're no longer in this progenitor situation. We want to know what drives full-blown, nasty, fully formed, fully advanced pancreatic cancer, and the basal cells, they are the bad guys. And so we want to understand what their role is in PDAC. And so we, the royal we, put the cassettes specifically in these basal cells, allowing us to visualize these basal cells, allowing us to lineage trace and see what happens to them and how much plastic they are and how they respond to therapy. And I won't go in it, but indeed these basal cells are insanely plastic and do all sorts of crazy things with and without therapy. But the thing I'm going to focus on is we can kill them with DTR, so we can send them a suicide signal and just at once or close to at once remove all the basal cells from the tumor and see what happens to a tumor that no longer has basal cells. And here we see right at the beginning, four hours after this DT kill signal, there's lots of apoptosis going on. The basal cells are indeed dying. This system works. This magic powerful mouse actually does what Thomas said it would do. But what is really amazing is two days later, we've killed all the basal cells. There's no more basal cells left in there. Nevertheless, cells continue to die. Now the classical cells die. Other cells die as well. And in fact, the entire tumor collapses. So we sent a kill signal to the basal cells, and boom, there is a collapse of the entire tumor nearly five days, seven days later. And this is remarkable because the cancer cells are about 20% of the tissue, but the basal cells, they're only 4% of the tissue mass. So we send a kill signal to only 4% of the tumor cells, and the whole entire rest of the tumor cells are dying. And so how can this happen? And so we asked, is it unique to the basal cells? So Thomas and his team built the magic mouse now for the classical cells. So we have the entire system of visualization, ablation, and tracing in classical cells, and so we can ablate the classical cells. And when you ablate the classical cells, well, some stuff dies, but really nothing really happens. You still have plenty of tumor left. And only in the basal state do we have this entire collapse. So basal cells really build and sustain this tumor-suppressive, protective niche. The tumor really depends on them. So how and why? So that's, again, understanding the tissue, understanding the how and why is where spatial transcriptomics comes in. And here we see time point zero, the nasty full-blown PDAC. And we see indeed this highly tumor-suppressive Arg1 macrophage environment really there. This TNC fibroblast, this same state that we saw built early on in the talk, it's still there in advanced PDAC. But five days later, we see that the immunosuppressive macrophages are gone. This TNC fibroblast and the ECM that it deposited nearly melted away. And what's really remarkable is suddenly we see T cells and NK cells and other lymphocytes just infiltrating the tumor. So there was absolutely no yellow cells in the top, in the control, and suddenly you see these lymphocytes infiltrating the tumor. And so what recruits the immunosuppressive macrophages is the first question. So here we want to look at this sort of baseline, time point zero, and really build and look at these sort of niches and these spatial neighborhoods. And we're going to really look at these heterogeneous niches, and we really want to map their difference. And we're going to take wormhole here, and this is, as I said, it's a UMAP. But what's critical is each dot in this UMAP is not a cell, it's an entire spatial neighborhood. It's an entire niche of about 100 cells. And you can see differences. The biggest variation, the biggest diffusion component that changes along data that's colored is the classical to basal component. Up at the top right here, this is actually hypoxic cells. So we can really map different regions where hypoxia acts, where the basal cells are, where the classical cells are, but we didn't need all this fancy stuff to do that. But now that we have an organization of our niche space that we can really compute in, we can build trajectories in it and ask what are the strongest signaling featuresThat are associated with the basal niche and change along this structure. And so we can look at all the cytokines and receptors that are changing on this niche axis, and we see that CSF2 is the strongest one, and it's very, very, very specific. You see how red it is in the basal end of the thing, and then it just very quickly disappears. And you could look at these trends to see how both the myeloid compartments, and specifically, both the CSFR receptor, the CSF ligand, and the sort of arginase nasty tumor-suppressive macrophage are all very unique to this basal niche. So, the CSF2 might be doing something here. But how do classical cells die? Because we sent this kill signal only to the basal cells. And I showed you some hints because you see there's no more protective myeloid cells. There's the T cells. But really, we want to understand what happens. So now this is, again, a Wormhole latent, the same type of latent, but now we're doing it for the entire seven-day time series. And now you can actually see this animation of how the time series changes in microenvironment space. So this is how the niches are remodeling over time. This is niche remodeling. This is an embedding of 10 million spatial niches and how they remodel over time. And now we can again look at a temporal response trajectory of this remodeling. We're going to really focus on the classical cells because we want to know what happens with classical cells. And you can really see that this embedding really captures a temporal time point signal, as you can see from this diffusion component. As you could see, we really have, again, there's asynchrony, so this is not a perfect thing, but you could really see that within the asynchrony, we're capturing a temporal cascade. And we could ask what, again, are the big things that are changing along this temporal cascade? We see the cancer cells are expressing all sorts of chemokines, these stress chemokines, these interferon-induced chemokines, and importantly, this is exactly what the lymphocytes are trained to sense. The lymphocytes, the NK cells, and the T cells have the receptors for exactly these chemokines, and they emit the interferon signature. And interferon gamma really hypercharges this entire thing, really bringing in the T cells and the NK cells. So now, Thomas didn't want to stop there. It's like, let's push our functional mice even crazier. So I don't know how to explain this crazy cassette, but Thomas is back there, so don't ask me questions about this. But I know what it can do for me, and what it can do is my dream. It can take the basal cell and do a CRISPR for a specific gene just in the basal cell. So we're taking in vivo, and we're saying let's limit ourselves to that 4% of basal cells and remove CSF2 just from that 4% of cells in vivo inside a tissue. And lo and behold, this phenocopies the entire basal ablation phenotype. So just CRISPRing out CSF2 just in the basal cells phenocopies collapsing and killing the basal cells. And lo and behold, we have suddenly CD3s coming in if we remove this CSF2 signal and just remove that specific signal from the basal cells, collapsing the tumor by either removing the basal cells or disrupting their CSF signaling, that really builds this immunosuppressive niche. So our model is this basal cell, the master architect of the tumor niche, that PDAC is so well known for, and that pushes all therapy away. It calls the proliferative MYCAFs that really protect it, the fibrotic one, the immunosuppressive myeloid cells, and T cells cannot enter. The basal cell goes away, the immunosuppressive niche goes away, and in come these hero lymphocytes to clear the tumor. And so across both early and late disease, malignancy is really sustained by these plastic states. The oncogenes tumor charge the plasticity, but it's really the plastic states that organize and maintain the tumor ecosystem that makes cancer the nasty thing it is. So finally, acknowledgments. The first story was Jose and Scott. Andrew built SEGER, and I recommend you use it. Doron built Wormhole, and we're using it, and we love it. Sam did the computation in my lab with the amazing team from Thomas's lab, and Sail in our lab really helps collect all this amazing data. And thank you. Hi there. Impressive methods and results. So I have one question about the cellular neighborhood map you're showing, the latent plots where you mention one dot is one cellular neighborhood. Mm-hmm. So I have a two-part question. First is that how is each dot being... So like for example, for a cell neighborhood, are you separating, for example, 100 cells, grouping 100 cells, and then call that as a cell neighborhood? Or are you looking at one cell and the 100 cells around it to be considered as a neighborhood? Yeah. One cell and 100 cells around it. And I'm not collapsing at the cell type, so I'm basically imagine having a matrix. Mm. A little sub-matrix of 100 cells-By all their transcripts, so 100 cells by 480 transcripts. There's a little bit of collapsing to gene programs, but 100 cells by their rich gene expression, a little mini matrix. And as I showed in the wormhole algorithm, we are reasoning about that entire unit at once without losing the information. Yes. So in that case, if two cells are similar or close to each other, will those two, the neighborhood overlapping then? So the neighborhood, yes, there is some overlapping features in these neighborhoods, but that gives robust statistical signal. As people who've used spatial transcriptomics knows, these things are very sparse. So actually having many similar units gives a little bit more robustness. But yes, this is not completely disassociated neighborhoods. But basically, we have an anchor cell, and much of our analysis, we actually reason about, here is the anchor cell, here's its neighborhood. And of course, many of these neighborhoods overlap because, just the math of it. Got you. Thank you. Hey, Donna. A very interesting talk. I had a question about the wormhole data you showed on the gastric tissues, where you estimated the progenitor cells and the more differentiated cells. So differentiated cells were more in the ducts, the circular structures. But the progenitor cells were in the ducts, and they were also out in the stromal areas. And so I was wondering what you thought about those distributions. So again, I discretized it and said progenitor and gastric because unlike computers that can learn complex functions, we're humans, and I can't say in one word, "Here is all the continuum of cell states." But within the progenitor cells and within the gastric cells, there's a very rich variation, and that's why I'm really pushing for this continuous representation. And we learn a lot from the continuous representation. And we actually played a bit, and we said, "What happens if we discretize the data to progenitor cells and gastric cells?" And that doesn't cut it. You really need this continuum, and I just can't talk continuum, so I give it names. Okay. Hi. Really great talk. And just to, a little bit of clarification, when you present a matrix for the niches, you said only the rows are permutated, right? Yeah. So would it be helpful to consider the SVD-based methods to do the reduction instead of doing optimal transport? Well, again, any way you try and collapse it-- And you could do SVD with inside each row, but if you try and do SVD for the entire matrix, you lose the cell identities. Anything that sort of averages across cell identities and collapses and says, "Okay, let's take an average of a myeloid cell, a fibroblast, and a T cell," there's some people in this audience which would not be happy if I tried to do that. Thank you so much. Hi. I was just curious if you had benchmarked SEGER against other transcript-aware methods like PROSEG, for example. We've benchmarked it like crazy. I want to give you the intuition. And so one, in performance, it performs better by the metrics you care. But also, PROSEG, for example, it's the second best. It still tries to solve a geometric problem. It still tries to find a boundary. So you look at the PROSEG cell membranes, and you have all these deep notches that goes like this to try and avoid including transcripts in these-- It looks like a macrophage that's just trying to pull this thing to pull that transcript. It's really psychopathically shaped. And PROSEG is pretty good because it actually uses a lot of very similar reasoning to us about learning cell states and what genes go together, but it still insists on drawing a boundary. And we just say forget the boundary, assign transcripts to cells as link assignment, and that's our big difference. But actually, PROSEG works really well. We work better, but PROSEG gets a lot of good marks, except for the fact that it has psychopathically shaped cells. Hi there, and thank you very much for your talk. Always inspirational. So my question is a bit naive maybe, so apologies for this. I really like this concept of working in this embedding space with niches, no longer cells. So I was wondering, if we want to go from cells to patients, does it make sense to think in the embedding space for patients where we start even to use the geometry and do some clustering according to these common niches, to ways that these patients possibly get together, not in the canonical imagination that we have at the single biomarker level? So I'll give a quick answer to that, even though I have a longer answer because Faye is trying to pull me off the stage. Multiscale is the answer. You have small niches, you have entire lesions, you have entire tissues, and you have entire humans. And I think the way we need to build embeddings is these multiscale thing where you also have an embedding at the level of a patient-- But this patient is a multiscale embedding that you could go deep into finer and finer spatial resolution. We're working on that, but not quite ready with that. Thank you. Thank you. And last but not least, we have Itay Tirosh joining us from virtually. And he's going to talk about spatial transcriptomes and single-cell transcriptomes in head and neck cancer. And he's very sad that he can't make it today, but due to the geopolitical climate. Hi, my name is Itay Tirosh, and first I want to apologize that I was unable to come to San Diego and present this in person. But I want to thank the organizers for both inviting me and for the flexibility to allow me to record this talk. So I will tell you about our work on using spatial approaches to understand the organization of tumors, both in glioma and in head and neck cancer. And we are generally quite interested in the diversity of cells within tumors, trying to understand the entire ecosystem, but in particular, we are focused on the states of the cancer cells. And after doing a lot of single-cell RNAseq analysis to define those states, we've been working in the last few years to better understand how these cells are organized in the tumor. This allows us to better understand what generates the different states and perhaps also to start to understand their functions. And we are using two main technologies. The first one is Visium, spatial transcriptomics. Here we have a glass slide in which each spot has its own unique barcode. And this is producing very good data, and we learned a lot from it. But of course, one important challenge is that the size of the spots is still 55 microns in the first version, which means that in a given spot, we have somewhere between 5 to 20 cells. So understanding the pattern within a spot is still quite challenging. The second technology that we've been using is spatial proteomics, the CODEX platform, also known as PhenoCycler, and this is highly complementary to the Visium, both because it gives us the protein information, but also because this produces a real single-cell resolution. Now, this is antibody-based, and so we've constructed a specific antibody panel for glioma, as you see here, as well as another panel for carcinoma. And we've been using both of these technologies to understand the organization of both gliomas and head and neck cancers. And I'll start with glioblastoma. So in glioblastoma, our single-cell analysis in the past have shown that in a given tumor, there's typically four different states of cancer cells. And many of the cells are somewhere in the middle, but the diversity is really captured by these four different cell states. And these cell states recapitulate many patterns from development of the brain. So some of the cells, one state resembles neural progenitors, another resembles oligodendrocytes progenitors, a third state resembles astrocytes as well as their progenitors, and finally, we have a state that resembles mesenchymal-like cells, and this is the state that is the most responsive to the microenvironment. So we've shown in the past that this mesenchymal-like state can be induced by macrophages as well as by hypoxia. And when we see cells in a mesenchymal state that also express many hypoxia response markers, we call this mesenchymal hypoxia, and we know that this indicates that the cells are in hypoxic regions. So we recently applied Visium as well as CODEX to 32 glioblastoma samples, and here you see two of them. In this data, we can identify all of the major expression programs that we saw in the single-cell data. So you can see on the left all of the programs that we defined. The bold ones correspond to the cancer cell states, and the others correspond to the microenvironment. And we can assign each spot in the Visium to the main signal that we see in that spot, and this is how two of those tumors look like. But of course, there was still this challenge with the resolution of the spots. So each spot contains approximately 10 cells, and we wanted to better understand what is happening within those spots. So to address this, we went into the CODEX data to see what happens within a given spot. And what we saw is that even though a typical spot is indeed a combination of multiple cell states or cell types, typically, and more specifically in 64% of the spots, we see that one cell state or one cell type controls the vast majority of the cells in a given spot. So the Visium data still gives us a good amount of information about the main signal in a given spot, which is often sufficient to really understand the organization of the tumor. This also tells us a lot about the biology of glioma, that the cells are really clustering together. And we wanted to better understand to what degree the cells are clustering, not just within one specific spot, but in larger areas than that. So we expanded from analyzing individual spots to looking at adjacent spots, and we defined this measure of spatial coherence, which is basically asking, after we assign each spot to the dominant state, how many of the neighboring spot for one specific spot are still dominated by the same state? So we see variation in spatial coherence between different cell states, for example, higher coherence for this mesenchymal hypoxia state in yellow. We see variation between different regions of the same sample, and we see a lot of variability between different samples. And surprisingly, this was actually the strongest effect that we saw. So there were some samples in which overall, we saw high spatial coherence. So almost wherever you look in this particular sample, you see that spots are surrounded by other spots, often with the same color. This is what we call a structured sample. And in other samples, we see very limited spatial coherence. This is what we call a disorganized sample.And overall, we could divide the vast majority of the samples to either mostly structural or mostly disorganized, and there were very few cases that showed both of these behaviors within one sample. So we next wanted to better understand what is creating this dichotomy of structured versus disorganized. And the main difference that we saw between those samples was, again, the presence of this mesenchymal hypoxia state. And this suggests that whenever we have a high degree of hypoxia, not only the immediate cells in that region are affected and express this mesenchymal hypoxia state, but also the entire tissue becomes more organized, suggesting that hypoxia might have a long-range effect on the organization of the tissue. So to better understand the organization that emerges in these cases, we next asked what was the colocalization between pairs of cell states or cell types. So in this graph, each node is a cell state or a cell type, and the edges corresponds to a significant degree of colocalization across all of the data sets. And as you can see, this analysis divides all the cell states and cell types into five different groups or five different layers, such that all the edges are either within a given layer or between adjacent layers. And this, again, is consistent with the role of hypoxia in organizing the tissue. So the first layer is where we see the actual mesenchymal hypoxia spots and also the necrotic spots, the necrotic regions. The second layer corresponds to other mesenchymal states and inflammatory macrophages that are adjacent to the hypoxic regions. Then in the third layer, we see a lot of vascular cells, macrophages that secrete VEGF, and so this appears to be the layer that is trying to resolve the hypoxia. And then in the fourth layer, we see many more of the cancer cells in a neurodevelopmental state that is presumably more dependent on oxygen. Finally, in the fifth layer, the furthest away from the hypoxia, we begin to see the infiltration of the tumor to the normal brain, where there are many normal brain parenchymal cells, like oligodendrocytes, astrocytes, and neurons. So this suggests a model for the organization of GBM that consists of five different layers. Of course, we are not suggesting that they are very clean layer. It's more of a probabilistic shift in the abundance of different cell types and cell states. And here you see an example of this from the Codex data, where we start in the left from a region that is highly hypoxic with this mesenchymal hypoxia state. And then gradually, we see more of the other mesenchymal state, macrophages, vascular cells, and towards the end, we begin to see more of the neurodevelopmental cell states and the normal cells. So overall, what we see is that in GBM, there are many regions that are quite disorganized, but as we get close to the hypoxic regions, we begin to see more structure emerging, both in terms of cells being surrounded by cells that are similar to them, as well as this approximate structure of five different layers. So this work in GBM was already published, and after finishing this, we began to look at the second major type of glioma, which are IDH mutant glioma. And so we profiled both by Codex and by Visium, 34 IDH mutant samples covering both astrocytomas and oligodendrogliomas of different grades. And when we analyze the degree of structure in those samples, we see that there's overall less structure in the IDH mutant cases than in the GBMs, and there is more structure in the high-grade cases, which are more similar to GBM compared to the mid-grade cases. So this is still very consistent with the role of hypoxia, because hypoxia occurs mostly in the high-grade cases. And so far, this is quite consistent with what we saw in the GBM work. But what then surprised us is that when we looked at low-grade cases, we saw much more structure than the mid-grade and even the high-grade cases. And in the low-grade cases, there was almost no hypoxia and almost no necrosis, suggesting that something else is driving the organization of these samples. So in high-grade IDH mutant glioma, as well as in GBM, we see these approximate layers in which, as you can see in the bottom, the tumor really consists of these small patches that are dominated by specific cell states or cell types. In mid-grade cases, we begin to lose this organization. There is still some degree of organization, but much lower. And then in the low-grade cases, again, we see more structure, but it's not the same structure. We don't see again those patches. Rather, we see pairs of cell states that are correlated in their localization. So there's a different type of structure, and we are trying to understand what is driving this structure, which it probably not hypoxia. And what we are starting to see is that in those low-grade cases, the normal structures of the brain are still maintained to a large degree. These normal structures have been eroded in the high-grade cases, but in low-grade, they are still there, and they appear to influence the organization of the tumor. And I'm going to show you one example of this, which is the division of the tumor to white matter, gray matter, and the junction between them. So the normal brain is divided into white and gray matter, as well as other regions. And in the tumor, we can still see that division quite nicely in the case of low-grade tumors.Not only do we see white and gray matter, we also see a very sharp junction that separates them. And the fact that this is such a clean and sharp structure suggests that this has been maintained despite the growth of the tumor. So next, we wanted to see how this influences the cells around that junction. So we counted the number of cells moving from the white matter through the junction and to the gray matter, and we saw that this actually serves as a barrier for the growth of the tumor. So the density of cells drops dramatically as we switch from the white matter to the gray matter. And this was surprising for us to see because in the context of glioma, we normally think about tumors that diffuse in the brain and that expand almost with no interruption. And here we see that this junction between white and gray matter is in fact a barrier for the growth of the tumor. So not only the overall density of cells dropped at this junction, but also the fraction of cells that are malignant further dropped, suggesting that this is really stopping the growth of the tumor. At the same time, there are changes in the microenvironment. We see, of course, many more neurons in the gray matter, and we see almost no macrophages as we pass from the white matter to the gray matter. We can go deeper into this and ask not just about the overall amount of malignant cells, but about the specific states of the malignant cells. And we see that as we go through the junction, we see an enrichment of OPC-like cells, a further depletion of astrocyte-like cells. And just around the junction, we see some enrichment of undifferentiated cells. And even for the state that is enriched after the junction, the OPC-like cells, we can go deeper into this and look at specific markers, and we see that almost all of them are changing significantly. So even for a given cell state, we see more subtle changes in the degree of differentiation as cells move from the white matter to the gray matter. And so to summarize the analysis in glioma, we see two different behaviors. We see in high-grade cases, as well as in GBM, that hypoxia and necrosis appear to organize the tumor. But in low-grade cases where there is almost no hypoxia and necrosis, it seems like residual brain parenchymal structures are organizing the tumor. And specifically, this junction between the white and gray matter serves as a barrier for the growth of the tumor. And around this barrier we see significant changes in the composition of cell types and cell states, and specifically the OPC-like cells are the ones that appear to be most capable of migrating along this border, which is very consistent with the ability of OPC-like cells to migrate during development. So now I want to switch from glioma to head and neck cancer. So also in head and neck cancer, we see specific cell states that are commonly observed in many tumors. Here I'm showing just one tumor as an example. These main cell states include cell cycle related states, stress states, states reflecting epithelial differentiation and senescence. And one state that I'm going to come back to because it's particularly important clinically is a state that resembles epithelial to mesenchymal transition with dementin, podoplanin, TGF-beta responsive genes, and others. Now, this is a very partial form of EMT. The cells do not lose expression of epithelial markers. They also do not gain expression of canonical EMT factors like SNAI, ZEB, and Twist. But this partial EMT is still associated with metastasis as well as with poor overall survival. So we applied Visium as well as Codex to 28 head and neck cancer samples covering different sites and different genetic classes. And again, with the Visium data, we are able to identify all of the main patterns that we expect to see in those tumors directly from the Visium data. We divide the spots into the malignant spots, those where we see a strong signal for chromosomal aberrations, and to the non-malignant spots where we don't see this signal. And again, overall, we see high consistency between the single-cell data and the spatial data. But there was one interesting exception, which is that in the malignant spots, we see a program that we call the inflammatory response that consists primarily of genes that are often cytokines and chemokines that we know from the single-cell data are expressed not by the cancer cells, but actually by myeloid cells, either macrophages or neutrophils. So this indicates, and we can validate this with the Codex data, that there is a strong colocalization of cancer cells with myeloid cells that are active and produce many cytokines and chemokines. And I will get back to this program later on. So next, we wanted to better understand the organization of head and neck tumors. So again, we asked about the colocalization of these different cell states and cell types. And what we saw, unlike glioma, is a global division into immune-rich compartments, stromal-rich compartment, and cancer-rich compartment. Again, with this exception of inflammatory response of myeloid cells embedded in the region that is mostly cancer-rich. And next, we wanted to better understand the organization within the cancer compartment. So we defined another analysis that we called a zonation analysis.In this case, every spot that is dominated by immuno stromal cells is considered in the non-epithelial zone. After that, all the spots that are dominated by cancer cells but are bordering the non-epithelial zone are considered as the tumor edge. The spots that are dominated by cancer cells and that then border the tumor edge are the mid-tumor. And finally, the spots dominated by cancer cells that are only bordering the mid-tumor are called the core of the tumor. And so basically, we are dividing all the spots based on how far away they are from the stromal regions. And here you see an example of how it applies to one specific tumor. So we apply this to all of the tumors, and then we can ask for every cell state, how is it distributed across these different zones. And I'm going to focus on the analysis of this partial EMT that I highlighted a few slides ago. So here you see in black the overall pattern of partial EMT across all samples, and the thinner lines in the background correspond to individual tumors. So you can see that the average pattern doesn't really reflect what's happening in individual tumors. And the individual tumors can actually be separated into two main patterns. In some samples, we see EMT enriched at the edge of the tumor, and this is consistent with a lot of literature about EMT. But then in other samples, we see EMT mostly enriched at the core of the tumor. And we could validate this by staining for LAMC2, which is a top marker for this partial EMT program. And importantly, when we analyze multiple samples of the same tumor, we see that this result is highly consistent. If a tumor has EMT at the edge in one sample, it's usually going to have the same pattern in other samples of the same tumor. So this suggests that we are seeing two different subtypes of tumors. And so we asked what else is different between them. And some of the main differences that we saw include more lymphocytes in tumors with EMT at the edge and more of this inflammatory response, more of these myeloid cells in tumors where EMT is at the core of the tumor. And this suggests that perhaps this inflammatory response might be inducing partial EMT at the core of the tumor. So to ask this, we went back to this colocalization analysis, but now we cleaned each of the edges based on whether this colocalization is seen in all samples or whether it's seen only in cases where EMT is at the core or at the edge. And what we see is that partial EMT actually has two opposing interactions. When the EMT is located at the edge of the tumor, it's colocalized with the fibroblast, as you see with the yellow edge. And when partial EMT is located at the core of the tumor, it's instead colocalized with this inflammatory response. And this suggests that there might be two different mechanisms for microenvironment-based induction of partial EMT. And we search for potential ligand receptor interaction that might explain this. The top prediction from the fibroblast was TGF-beta, which is, of course, a well-known inducer of partial EMT. And for the inflammatory response, the top prediction was oncostatin M. So next, we treated two different head and neck cancer cell lines with both oncostatin M and TGF-beta. And we saw that both of these cytokines indeed are able to induce many of the partial EMT genes. Now, TGF-beta was more effectively inducing partial EMT in this in vitro analysis. But in vivo, we see the same partial EMT program when it's at the core and at the edge, suggesting that perhaps there are other mechanism and that there is convergence to the same partial EMT program. So to summarize, we see in head and neck cancer a global division to immune-rich, stroma-rich, and cancer-rich compartment with this exception of myeloid cells in the cancer-rich compartment. We see two main patterns of partial EMT. At the edge, it seems to be induced by fibroblasts, and at the core of the tumor, we believe that this might be induced by the myeloid cells and specifically by oncostatin M. And in those cases, we see an even lower amount of T cells and B cells, suggesting that perhaps EMT at the core of the tumor might have a stronger effect of immunosuppression. So right now we are trying to better understand what determines the separation into patients with EMT at the edge and those with EMT at the core, and whether that would have any clinical implications. For example, it's easy to imagine that EMT at the core might have a stronger immunosuppressive effect, going back to what I mentioned about lower amount of lymphocytes. And perhaps EMT at the edge might be more clearly linked to invasion and metastasis. So this is some of the questions that we are trying to answer now. So finally, I want to thank everyone that contributed to this work. In my lab, the glioma work was led by Alisa, Reuven, and Noam, and this was a great collaboration with Mario Suva's lab and specifically with Chris Mount. And the work on head and neck cancer was led by Dor Simkin in my lab, together with Sid Pouram's lab and specifically with Tom Barrett. And thank you again for listening. All right. Thank you. Well, I guess you can email your questions to Yitai, and thank you for joining this session.