Scientist Stories: Elham Azizi, Probabilistic modeling of dynamics in the tumor microenvironment

making it this early on a rainy Friday morning. I really appreciate it. And thanks to the organizers for this fantastic opportunity. It's really great to be back at MLCB in person. And MLCB has become one of the symbols of the comp bio field and how it's matured over the years. And looking at the poster session last night and your talks, I really see how far we've come in really blending and merging machine learning concepts with biology. And it makes me really excited about the future of the field. So first, to give some background, so a lot of our work is focused on dissecting the tumor microenvironment in, and specifically in specimens from patients, the patient samples. And because of that, because of the complexity of patient samples, we have to take, you know, additional care and rigor in designing methods that are adapted to answering questions in this context. And the tumor microenvironment, just briefly, consists of this system of diverse tumor cells that are interacting with support cells, endothelial cells, fibroblasts, as well as a wide variety of— I don't know if you see my mouse, maybe not— as well as a variety of immune cells, but I don't see it here on my screen, but that's okay. The immune cells that are infiltrating the tumor and try to eliminate the cancer cells. And so, why is the tumor microenvironment important? We know now that there's a lot of great strong evidence that the genetic alterations are not the only driver of response and outcome of patients. And the microenvironment also plays a significant role. And this heterogeneity of microenvironments across patients and within the tumors is important to dissect. At multiple molecular layers, phenotypic, epigenetic, and so on. And so, the problems that we focus on, and Bianca in her beautiful keynote yesterday highlighted and introduced these concepts of dynamics, but some of the major problems we focus on is using single-cell resolution multi-omics to understand the dynamics and altered cell state dynamics that happens in early stages of cancer initiation., as well as combining time series single cell datasets to look at temporal dynamics and how cell states change and evolve over time, in particular in response to treatments. In parallel, we also look at spatial dynamics and modeling how the spatial organization of cell types, diverse cell types in the patient tumor microenvironment, also affects the response and outcome. And through this modeling of both temporal and spatial dynamics, we also want to derive principles about the interactions between diverse cells. And the major methods and models that we build are of the probabilistic flavor, and this audience does not need any convincing about the advantages, the theoretical guarantees, the interpretability, and the adaptability of these frameworks, especially for integrating multiple data modalities and ability to to hypothesis testing. So, in the first part, I want to talk about some of our work on modeling spatial dynamics. And this project was started with my— some of my previous postdoc work where we were interested to know what's the impact of the tumor microenvironment in the diversity of immune cell states. So, what— how do they change once they encounter the tumor cells? And in particular, we wanted to compare these diverse immune cells in the tumor microenvironment. How are they different from those found in the healthy tissue, especially from the same patient? And to do that, we looked at breast cancer patients and matched samples from the tumor, from healthy tissue, from bilateral mastectomies, as well as lymph node and peripheral blood and sorted CD45-positive immune cells, and did single-cell profiling and generated the first single-cell breast immune atlas that revealed diverse set of distinct cell states. And in particular, some of the important highlights of the study was that there's indeed a significant increase in the diversity of the immune cells when going from the healthy tissue. Once they encounter the tumors, we see a significant increase in all major cell types of T cells, myeloids, and NK cells. And for those interested, how did we quantify this diversity? We devised this metric called phenotypic volume, which looks looking at the pseudodeterminant of the gene-gene covariance matrix. And so this intuitively, the geometrical interpretation is the volume that spans the eigenvectors characterizing the gene-gene co-expression matrix. And so if you have, for example, in the tumor some higher numbers of co-expressed genes, modules of genes that are independent from those found in the healthy tissue, you expect additional eigenvectors and more independence of eigenvectors to be adding to this volume and showing this increase. So, in addition to this increase in the diversity, one of the surprising results from the study was about the trajectory of T cell differentiation and activation. So, if you look at the T cells going from the bottom of the arrow, from more naive central memory T cells, it's going to effector memory cytotoxic T cells. We expected that, you know, just like textbook models of immunology, there should be distinct and defined states of differentiation along this axis, as well as activated T cells and exhausted T cells. But we were surprised that this distribution actually looks very continuous. If you look at, for example, the diffusion component, an axis explaining the variance across all T cells, you see that it's the least activated in blood and normal tissue, as expected. Highly activated in the lymph node, but in particular, in the intratumoral T cells shown in the light blue on the right, you see a very continuous distribution. And so that was surprising for us. And we thought, could it be explained by the fact that there are so many antigen-specific T cells that were aggregating? Maybe the TCR, the T cell receptor profiling of that can help deconvolve them. So we did pair TCR and single cell, and the answer was no, T-cell clonotype identity actually doesn't fully explain this continuity. When we looked at defining, you know, rough clusters, even though they're not distinct, and looking at the expression of signatures shown here in rows of these heat maps related to responses to microenvironmental stimuli, we see that the clusters show different combinations of levels of expression representing responses to these stimuli. The hypothesis that we took away from this was that maybe we have T cells exposed to different niches in the tumor microenvironment that are also differing in the levels of oxygen, nutrient supply, its activating ligands and cytokines, and so on. And because of that, we're looking at the aggregated from dissociated tissues in single-cell profiling, and therefore seeing a continuous distribution. So the answer was, okay, let's do spatial. And now we have spatially resolved technologies that can, you know, resolve the phenotypes in the spatial context of the tissue. And again, thanks to Bianca for introducing spatial methods. We continued our collaboration with Sasha Rudensky and George Kliotis at Sloan Kettering to do Visium spatial transcriptomic profiling of breast tumor tissues and to try to understand if the spatial organization defines immune differentiation and stromal response. We were also particularly interested to compare spatial organizations across tissues. For example, in metaplastic breast cancers, these are a type of triple negative breast cancers that are more aggressive. They actually grow on chemotherapy, unlike other triple negatives, and they have different morphologies, very heterogeneous morphology. You can see on the histology slides, some of them look like soft tissue and some connective tissue, and So there's a lot of demand tailored for this very aggressive subtype that, you know, is adapted to the unique biology of this breast cancer. But to be able to compare spatial organization of cells across samples, we need new methods as well. Also, in addition to that, as many of you know, the spatial transcriptomic technologies, which look at sampling the entire transcriptome, are limited in their resolution. They're not exactly single cells. So you have a mix of multiple cells within every spot shown here in the zoom out plot here on the top left corner. And so this is a problem when you're looking at complex patient samples because you have mixing of immune cells that you're trying to compare across tissues with patient-specific tumor cells. And because they're patient-specific, there's not going to be any overlap between spots. And so in particular, if we want to really deconvolve the refined cell states like activated T cell states from exhausted T cell states, cell states, we need methods that are able to deconvolve spots at that resolution. And when looking at— there are many great methods out there now for deconvolving spatial transcriptomic data. The majority of those that are able to get these refined cell states rely on a single cell reference that is from the same matched sample. And this has its limitations, but also it's, you know, something that might be limiting when you're looking at patient samples and you might not have enough material. So a number of brilliant students in my lab, Siyu, Yino, and Ashil, and also my postdoc, Lingting, worked a lot on the interpretation, thought together about how can we address these issues to deconvolve cell states without the need for a single-cell reference and also leverage all this information that we have on the histology imaging. This is what the pathology uses for, you know, the annotation and the diagnosis. So it has a lot of information about the density of the cells, the tissue architecture, So why not integrate these two data modalities to improve our understanding of the spatial dynamics? And most importantly, we wanted to build a framework that is able to do comparative analysis of multiple tissues. So this led to STARFISH. I'm not going to read the full name. I leave the naming of tools to my students, so they come up with these very long acronyms. But the idea is to combine spatial transcriptomics and histology. So to give an intuition on how we avoid the need for single-cell references, I'm going to show a cartoon that is, well, simulated data that is based on, generated based on mixtures from a single-cell RNA sequencing data with known mixture levels from a breast cancer tumor dataset with 5 major cell types, CAFs or fibroblasts, tumor cells, myeloids, normal, and T cells. And you can see the UMAP, just like the UMAP of any spatial transcriptomic data, doesn't show defined clusters of cell types because you have these mixtures. But at the corners, you see those are spots that are enriched for markers of known markers of cell types. Now we thought, okay, if a user has knowledge of the markers of these cell types, can we use those as guidance for the deconvolution? But in many cases, especially in patient-specific tumors, you don't have accurate markers for of those cell types. And so the idea we used is the concept of archetypal analysis, basically fitting a polytope to the low-dimensional representation of the data. And the extremities, the vertices of this polytope, are the extreme states that the spots can reach. And our major assumption is that those extreme points represent the purest spots in our data, meaning they're either consistent of one cell type or dominant in mainly one cell type.. And we recently tested this assumption, thanks to our reviewer as well, on xenium single-cell resolution data. And indeed, it is a correct assumption. And then finally, we align this. If there's information about the known markers, we align the archetypes to kind of adapt our marker set, any known marker sets, to the context of this tissue, and then use these anchors to kind of pull apart the rest of the spots, we get these starfish-like shapes, hence the name. And then we show with this data that it fits the data. For the more technical details, what exactly do we do? We have— so X here is our transcriptomic data expression of cell I— sorry, spot I in gene G. We have a variational autoencoder mapping it to a latent space Z, and Z is centered around weighted sum of our anchors, which we also learn, but we had this preprocessing step. They are guided by the definition of the archetypes. And then C here gives you the proportions of the cell types, which you also learn. Theta is just the gene dispersion. L is library size, similar to other generative models. And finally, we have this additional data modality, the histology image, as a second autoencoder, but a mapping from the same joint latent space. And again, beyond Mentioned the advantages and benefits of learning joint representations from multiple data modalities. We indeed see that it helps with improving the deconvolution of tumor cells and fibroblasts using this integration. So we show on simulated data, we do as good as methods that use a single-cell reference even without using it, but significantly improve over methods that don't. This is now has been expanded to higher numbers of cell types. We use datasets recently matched Xenium, so single-cell resolution, but lower dimensional, and Visium data from adjacent slices of the tissue to actually evaluate, are we getting the correct deconvolution? So here you can see two types of invasive tumor cells at the bottom and noninvasive ductal carcinoma in situ cells on the top. And you can see it does align with the inferred proportions from the Visium data. But what was really interesting was that we found the archetypes actually were able to deconvolve kind of in a hierarchical manner, further distinguish two different subsets of these noninvasive DCIS cells shown at the far right there that are exactly aligned with expert annotated regions shown in the red and yellow in the histology. And so without any knowledge about what are the markers that disentangle these to more refined subsets, we found archetypes that exactly match them. So what do we learn about the spatial dynamics? So here, this is an example of a triple negative breast cancer tumor. First, we define spatial hubs or neighborhoods by grouping spots that have similar inferred composition of cell types. And so now we can look across these spots that are high in tumor cells, define trajectories of how the phenotypes change, are changing and evolving across these spots. And we can see, for example, going from the cyan to the higher top left part of the tissue, we have a reduction in cell adhesion molecules, an increase in mesenchymal markers and EMT. So, really showing that there are cells, the tumor cells are changing from more basal states to more mesenchymal states within the same tumor tissue from a patient. So, now, how do, how are the immune cells changing if the tumor cells are adapting and changing in phenotype? We see that interestingly, as there's this shift in the tumor states, we also have a shift from more central memory to effector memory cells, and also, and then activated CD8 T cells at the margin of the tumor, and also shifts from monocytes to macrophages. So, this really kind of goes back to the hypothesis that we started with, that there are indeed different immune cells associated with different niches of the tumor microenvironment and the tumor cell they're interacting with actually does impact or determine the immune cell phenotype as well in the same spatial neighborhood. So, in addition to this integration of histology and spatial transcriptomics without requiring a reference, STARFISH is also able to integrate multiple tissues and create these atlases of spatial hubs or neighborhoods across multiple tissues and then do downstream analyses like just what I showed, this pseudospace, or how phenotypes change across the space of the tissue and then do cell-cell interaction prediction. So this is an example of an integrated space of 14 samples, 14 breast tumor samples, triple negatives and metaplastics. You can see the UMAP is just showing clusters that are patient-specific, no overlap despite the fact that there is common immune cells among them, whereas StarFish is showing a merged and blended and mixed latent space with more mixing in the immune cells like T-regs and less mixing in tumor-enriched spots, also quantified with entropy. And then you can also see the space colors by the spatial hubs. So now with the spatial neighborhoods defined, we can see what type of neighborhoods are, for example, enriched in the metaplastic breast tumors, shown as MBC here. For example, hub 0 is one case. Or what are the types of hubs that are common across all breast tumors? Tumors. And just as sanity check, we also compared to, you know, annotated, expert annotated histologies. We do see that, for example, that blue hub there in the middle is aligned with a healthy annotated tissue from, from a pathologist. Also between two adjacent slices at the bottom in the middle, we have reproducible hubs identified showing it works on replicate. And give same results. But most importantly, we can compare across samples. So how do the tumors— how are the spots in metastatic breast tumors in the intratumoral regions compared to the triple negative? So looking again at this joint latent space, we find that there's an increase in inflammatory response, but more interestingly, in immunosuppressive cell types such as regulatory T cells and myeloid-derived suppressor cells, or MDSCs, in the metastatic breast tumors. Tumors. In addition to that, we also see an increase in perivascular-like cells, which might indicate angiogenesis. And we also see interesting stromal hubs in the metaplastic tumors with a very concentrated organization shown in these red pockets across the metaplastics that are also exactly overlapping with the regions that are high in hypoxia. So low oxygen levels regions are forming these regions that are high in immunosuppressive cell types, like regulatory T cells. And we also see an increase in perivascular-like cells. So, we think that the regulatory T cells might be contributing to building these immunosuppressive niches and potentially adapting by angiogenesis. And we— if you zoom in, in the histology, we also see some red blood cells forming around the region. So, there's an interesting connection with the immune cells. We validated these immunosuppressive niches recently with CODEX data profiling based on markers that we found in our transcriptomic level deconvolution. And lastly, we looked into predictions. Can we predict from the deconvolved gene expression profiles what might be the receptor ligands that mediate interaction between the regulatory T cells and the tumors? And FGF2 pathway was one that is coming up a lot. Does show this decline by distance only in the metaplastics and not in the triple negatives and ERs. So, I want to wrap up this first part. I showed StarFish, a deep generative model that uses anchors guided by archetypes that replaces the single-cell reference, does image integration and comparative analysis of tissues. And I showed some results about insights we get in comparing metaplastics and triple negatives. The enriched immunosuppressive niches. I want to move on now to temporal dynamics as another perspective to look into cell-cell interactions and in particular longitudinal profiling, profiling of tissues at multiple time points, for example, before and after immunotherapy, which can tell us how are the immune cells changing once the patient is given immunotherapy in the tumor microenvironment, which ones are infiltrating, which ones are dysfunctional in the patients that are not responding to the treatment and how can we reverse the immune dysfunction. So, in our paper published earlier this year with, in collaboration with Ben Iser, we looked at anti-PD-1 therapy-treated melanoma patients. And looking at single-cell RNA sequencing at multiple time points before and after, we see, for example, cases like this red clone defined from Inverse CMV that is shrinking and responding to the treatment. So, it disappears in the on, on therapy samples. But we also see cases like this purple clone that you see it expands in proportion. It's seen at very small proportions pre-therapy and significantly expands. So, and this patient was a non-responder. So there are, you know, some clones in the tumor microenvironment that are resistant. And we wanted to know, is there any difference in terms of the spatial organization of the immune cells around these resistant clones compared to the responsive clones? And indeed, applying spatial Starfish2 slide-seq data now instead of Visium, we were able to disentangle activated from dysfunctional CD8 T cells and saw that the activated CD8 T cells are only co-localized with the responsive tumor clones, whereas we see dysfunctional T cell clones— T cells co-localized with the resistant clones. And so, this level of, you know, disentangling of T cells was not possible again with other methods, for example, RCTD methods that don't rely on single resolution. In other cancer contexts, for example, in leukemia, we try to draw principles about response and resistance to immunotherapies from established types of therapies. For example, donor lymphocyte infusion is an adoptive cell therapy that's given to leukemia patients that relapse after stem cell transplant. And we don't know exactly how they work. They used to work very effectively for chronic myelogenous leukemia patients before matinib and targeted treatments. Today, they are the standard of care for AML patients, acute myeloid leukemia patients that relapse, but they don't work as well. They work up to 20, 30% of patients. And so, we wanted to see, can we draw, you know, principles about DLI response from CML cases, CML patients, and use that to guide how we can improve treatments for AML? And in my previous postdoc work, we focused first on CML and using Gaussian process regression models that I won't talk about today. We uncovered two major subsets of T cells that are in particular this dark blue cell types that— sorry, light blue cell types that are exhausted but also have stem-like properties. And they expand after treatment, and we think they are shaping the response. And we wanted to see if they're coming from the infusion, the therapy itself, or is the therapy reinvigorating interactions among T— among immune cells in the tumor microenvironment using T T-cell receptor data, and our data was suggesting that it's rather the latter, and it's invigorating interactions across cell types. So, that really motivated looking and trying to model interactions between cell types in the tumor microenvironment from this longitudinal data. So, we started this collaboration with Cathy Wu at Dana-Farber, Katie, a postdoc in her lab, Cameron, PhD student in my lab, leading this project with Katie, and also Shovik Mani, who was a master's student in CS in my lab, now a PhD student at Stanford. We worked on this project, ambitious project of trying to compare two different cohorts, AML and CML, to try to understand how their DLI response is different. And so just creating this atlas of more than 400,000 cells, we find there's a lot of indeed overlapping immune populations while there's distinct diverging leukemia trajectories. But to get cell-cell interactions, you've probably heard of many interaction prediction methods, NicheNet, CellChat, CellPhoneDB. They essentially look at differential expression of receptor ligands in clusters of single cells. And the— you might kind of even look to talking to our biologists. It's really hard to prioritize which of these many hits predicted receptor-ligand interactions might be relevant to the context of this disease. And so, So, our idea was, can we actually learn from the temporal dynamics of cell types, which cell types might be interacting, and at least combine that with prior knowledge? For example, if we see a decline in tumor cells and increase in a T cell, maybe the T cell is inhibiting the tumor cell. So, that negative temporal correlation is informative about potential interactions. So, the idea of DISCO, Dynamic Intercellular Interactions in Single Cell Transcriptomics, is to exactly combine these temporal correlations with prior knowledge of receptor ligands to try to constrain the space of possible cell-cell interactions, but also get time-resolved interaction predictions. Because obviously, after therapy, we expect a different set of cell-cell interactions in the tumor microenvironment. The idea of the method is actually building on beautiful previous work from David and others using Gaussian process regression network models, where we have We model every cell type as a Gaussian process shown in Y, and we have latent features F1 to FK for K cell types. But we also have this W parameter here, which is also a Gaussian process showing the interactions between cell types. And so this model by itself is not identifiable. So our idea was to constrain Ws using this prior knowledge of receptor ligands to have, you know, improved identifiability and also get time-resolved interactions. So this is now public as well. Feel free to use it if you have time series single cell data. Let us know how it works. To show you, in our preprint, we show the performance on both simulated and also more controlled experiments, in vitro experiments with CAR-T and cell line interactions. But here I want to show some exciting results on the full AML cohort. We see that there's different types of T cells from what we found in CML actually potentially driving the interactions, more CD8 Temra cells that in addition to NK cells and other immune cells such as B cells that are also expanding in proportion in AML responders and inhibiting the proportions of leukemia. So, these lines here, the line is showing the mean of the Gaussian process that we fit to the data points and the The confidence intervals are shown in shaded areas, but we also learn about interactions and we can have time resolved interactions which cell types are interacting. For example, after therapy we see this, these immune cell types, the CD8 Temra NK cells increasing in proportion and strongly interacting to inhibit the leukemia cells. And this cool video by Shovik, if I can get it to run, also shows how the proportions of cell types and the interactions change. So you can have predictions from the Gaussian process at any, at any any time point. So— next one. Now it doesn't stop playing the video. Okay, there we go. So we're now looking into validating these interactions using spatial CODEX data that I won't talk about today. But just to recap this second part, DISCO is able to infer time-resolved cell-cell interactions, and in our AML patient cohort. It is identifying coordinated immune networks that are not found in more homogeneous leukemias such as CML, and this was possible with longitudinal single-cell mapping. All right, so if we have a couple of minutes, I want to talk about another exciting project now on the dynamics of cancer initiation, keeping the best— saving the best for the last. So this project is trying to get that, well, started with answering this question about tumor initiation. So, we know that, for example, chemotherapy, a lot of standard therapies eliminate the more, most mature tumor cells. And these early precancerous cells that have stem-like properties are able to grow back, for example, leading to relapse and metastasis. And we don't have very good markers for them in the majority of cancer types because they are rare and hard to profile and characterize. So, our question was, can we use, you know, trajectory inference methods and single-cell resolution profiling to try to reconstruct these tumor initiation trajectories? And we focused on AML. So, AML, a bit more background. Normal hematopoiesis is shown on the left. You have these hematopoietic stem cells at the root of the tree differentiating into progenitor cell types, and then differentiating into all your white, red blood cells continuously. And in the context of AML, we have a blocked differentiation on the myeloid arm. And we think what's happening is that the cells that are either stem or progenitor lead to precancerous cells, and then through acquisition of accumulation of mutations and alterations, they lead to leukemia cells.. So, towards the end of my postdoc, we collected some data with Vincent Lavallee, who now has his own lab at the University of Montreal, and Donna, looking at multiple patients, 12 AML patients that we tried to focus on particular types of mutations, TET2 epigenetic mutations and MPL1 driver mutations to at least have less heterogeneity. This is a very heterogeneous cancer type. And again, we see that all of the main clusters, major clusters are patient-specific. There is some overlap between patients in this lower left populations here in the T cells and B cells as expected because this is a myeloid malignancy. The lymphocytes should not be affected. But what was surprising was the circled population here, this immature population. Also, the Shannon diversity quantifies the high mixing of multiple patients. So it's a population that's highly correlated with healthy hematopoiesis. Found in 11 out of 12 patients. So that was interesting. We thought maybe these— this cluster includes some of the precancerous cells, and so we found markers for them, zoomed into this population by sorting for them, doing additional round of single-cell profiling, and we saw that it consists of two major populations, one more stem-like and one more progenitor-like. And what was really cool was that we when scanning for the mutation of the driver, NPM1 mutation, which happens on the 3' end, that's why we chose this mutation and leads to high expression. So, we could detect it from single-cell RNA sequencing 3' data. We saw that the mutation exists only on the progenitor-like cells and not on the stem-like cells. So, this really kind of suggested that there's a direction of mutation events accumulating and we really wanted to study what are the transcriptional consequences of these mutations along the full trajectory of tumor formation. So for that we thought, okay, how about comparing these cells to healthy hematopoietic cells? And that turned out to be challenging when you compare these now sorted cells in purple on the left with healthy CD34-positive hematopoietic stem cells in green. You see that integration method methods basically don't show any overlap or any, you know, idea of the global geometry. We also have this annotation of the state maturation stages of the leukemia cells from blast 1, blast 0, blast 1 to 3. We see that the ordering of these cell types doesn't make sense in any of these representations. It also doesn't make sense when we try to do batch correction because batch correction is designed for correcting technical artifacts. Between samples. So if you're comparing disease samples where the majority of the cell types are not found in the healthy tissue, it's just trying to force them to overlap. So these methods shown on at the bottom row are really showing trying to force the cells to overlap on top of each other, but they might be misleading in the interpretation. They can tell you, for example, in SURAT that maybe it's the healthy cells branching and bifurcating to the BLAST2 and BLAST3, which are the more mature leukemia cells. So obviously incorrect representation of the trajectory. Trajectory. So, two brilliant, talented students, Ashiel and Joy, in the lab thought of designing a new method that tries to address these issues. So essentially what we thought about is if we're trying to get an interpretable representation from two datasets, a, for example, normal and a perturbed or disease dataset, there must be shared mechanisms between these two trajectories. For example, cell differentiation is a a preserved mechanism between healthy hematopoiesis and leukemia, but they might also be diverging, so leading to very distinct cell states. So we thought, can we relax the assumption of independent and orthogonal latent components in our representation? And so this, you know, becomes challenging, again, dealing with identifiability, But the idea we came up, and I'll show the details in a second, indeed shows that if you allow the factors, the latent factors in this case, deep generative model to have correlations between them, they can exactly resolve and disentangle these cell types shown on the right that are distinct, whereas other VAE methods like SCVI completely overlap these cell types and do not distinguish. Them. So what does— how does the method actually work? What we have here is, you know, really building on methods like SCPI. So we have this VAE mapping X to Z, but the major novelty here is adding an additional layer where the Z, the latent factor Z, are actually a function of another layer V, which is even lower dimensional, and in our case, a 2-dimensional And so there are multiple advantages of this defining as a 2D space. We can directly use this model to visualize our data in a meaningful manner and not have to do UMAP or another visualization method that distorts the geometry of the data. And also we have this Z factors that are defined by this higher level characterization captured in the 2D space. 2D space, whereas the Z factors kind of give you more refined cell state characterizations. In particular, in the same sample that I showed, we see that it, you know, nicely aligns this green trajectory, healthy trajectory, cell maturation with healthy cells maturing to on this green trajectory, and also aligning that with the purple trajectory of AML maturation with the correct order of the cell types. And then the second axis is telling us what's most different and distinct between the two samples and exactly aligns with the accumulation of the MPM1 mutation on the decipher2 axis. In particular, we find this PROM1 marker marking the progenitor-like population consistently in all of our AML patients. So potentially a marker that could be used for guiding early detection of the disease. But now we can use also these axes, decipher components to understand understand what are the mechanisms that are disrupted along the leukemic initiation with MPM1 mutation. I'm going to skip this. We also compared to a wide variety of methods to try to assess how well Decipher gives a faithful joint representation by defining two metrics. We wanted to assess if these methods give us the correct ordering of cell types along the trajectories while also also, you know, preserving the divergence and bifurcation of the trajectories and not forcing them to overlap. And so, for example, we see some methods like UMAP, they're great at preserving the order but not as good in preserving the divergence, whereas other methods like SCVI are the opposite. And Decipher kind of has a good balance between the trade-off between these two. We also have a basis decomposition that then looks at— so the nice thing about this method is you also can get explicit reconstruction of expression directly from the model. So we can— any point in this 2D space, you can predict what is the gene expression of a cell that might be in that space. And you can basically look at now how genes change along any trajectories across the clusters. And then we have a basis decomposition technique that also quantifies difference between expression along a perturbed or disease trajectory and a healthy trajectory. Basically identifying the most dominant patterns as basis functions that we learn, and then using the weights to quantify the distance between them. I think I'll skip this, but very briefly to show that we found a lot of interesting insight into disrupted regulation during AML initiation. We see, while we in the healthy sample, we see a lot of coordinated TF expression, this coordination is completely lost in the leukemia samples, and we rather see especially multiple homeobox HOX transcription factors around the time that NPM1 mutation happens. We validated these with single-cell ATAC-seq as well. I want to end with saying that we started building Decipher with this question of trying to understand cancer initiation, but we found that actually the representation we get by allowing the latent factors to be correlated gives us a lot of meaningful insight into other contexts. For example, this is T cells infiltrating melanoma tumors from 33 patients undergoing PD-1 therapy. And you can see in the UMAP, they're just, you know, they're— when we annotate them, for sometimes you— I'm sure you've seen the CD8 T cells and CD4 T cells, they become mixed. They're not like, you know, partitioned properly. And you don't see any association with the timing of the therapy or response, whereas with Decipher, you get this nice alignment of CD8 T cells going from naive CD8 CD8 to effector memory and exhausted CD8 T cells. And in parallel, you have the naive CD4s going to more differentiated CD4 T cells and regulatory T cells as their own distinct cluster. Now, these axes are meaningful because we now see that they are actually associated with time of therapy, with response to anti-BD1, and the length, the duration of the therapy given to the patient, which really tells us, you know,, you know, can guide the administration of these therapies to be more effective. And we also get additional insight from the latent factors. This is very new. I just received this figure last night. This is new work that we just started with Bianca and led by Joy, extending the idea of Decipher to also integrate single-cell ATAC-seq data. And we just see very interesting results in the context of pancreatic cancer where you don't get the correct trajectories of ACNR cells transforming into pancreatic cancer cells, from the integration you are able to resolve this correct transformation through ADM and PANIN cells. So just summarizing, a deep generative model for the joint hierarchical representation of cells, for example, from two disparate contexts of healthy and disease, and this two-layer dimension reduction really helps with interpolating between different degrees of nonlinearity and gives you much more flexibility and therefore more interpretability. Ability. The biggest bonus is that you get a direct 2D visualization. Feel free to try it. We'd love to hear about your experience with this method. It is now online. And I want to end with thanking the many people who are actually doing all this work. Many people to thank in my lab, collaborators, especially at Sloan Kettering, Sasha and Dana Farber, Kathy's lab, funding sources, patients for their samples. And most importantly, we have open positions especially the postdoc position joint with David's lab, where we're looking for causal inference, cancer genetics, cancer genomics, any interest in at the intersection of those, feel free to talk to us and we'd love to chat more about it. And thank you for listening early morning. All right, I think we have only, yeah, a few minutes for questions.

Uh, great, great, uh, talk. So with— is there any time you wouldn't want to use Decipher versus using scvi or, you know, doing SerratUmap?

Yes, when you have technical bash batch correction. So in cases we have one dataset where we have 3' and 5' chemistries, we want to align them because we know they're similar. So technical artifact, but Decipher, at least this version that doesn't do a batch correction, does not fully align them. So in that case, something like that, something like a CDI makes more sense. I'd like to get a sense of the resolution of the cell types, how many cell types and what kind, and how did you do the mapping between cell and then cell type? Yeah, great question. So if you have prior markers you want to use, if they're refined, it will use those, try to disentangle them. It quantifies if they're overlapping a lot in terms of the archetypes, and then it will collapse them. How the number of archetypes are chosen is based on an intrinsic dimensionality metric that we built in. So it's really a data-driven calculation of number of cell types. But we have seen in all of our tumor samples, melanoma, breast, lung, that it is able to disentangle very refined immune cell states.

Yeah, thank you. Very nice talk. One more question on DECYFER. So I think you briefly mentioned the identifiability issue. So I was wondering, how did you make sure that the model would not, like, simply ignore one of the latent representations and degenerate to to a vanilla, like, variational autoencoder?

Yeah, great question. So we don't have theoretical guarantee that it is identifiable, but because the space of the latent space is really much smaller, so the z-space is really 10-dimensional by default, you can— we've tried up to 30 as well, similar to SCVI. It's just because it's compressing 20,000 dimensions to 10 dimensions, it will try to make them them as, you know, independent as possible, but also correlated if needed through this additional layer of V. So we just see through practice so far that we get very robust representations of Z factors and also the ordering of cells along the trajectories that it infers through this minimum spanning tree. The length of trajectories in the V space has also been meaningful, again through practice. We are thinking of ideas on how to actually try to understand what is— what are the theoretical backings for this. But that's a good question that we're following up on.

Any questions for that one? So, I just heard with Starfish, you know, you're deliberately not using a reference. If you have a reference, are you able to leverage that and then learn cell type archetypes on top.

Exactly. So it's— there's 3 options. One is if you only want to use your reference, you have high confidence in your markers, you can just use that. The second option is you have no markers, you just want to use archetypes. The third is combining both, so adapting your marker set. And it doesn't use the level of expression of the gene markers either, so that therefore you don't need really single cell data. Even just gene names, lists of genes, even a few marker genes are sufficient.