Scientist Stories: Bianca Dumitrascu, Morphogenomics and statistical representations of development

recently joined Columbia as an assistant professor in statistics at Irving Institute of Cancer Dynamics. She's affiliated with the Data Science Institute and the computer science department as well at Columbia. Prior to Columbia, she held positions at Cambridge, University of Cambridge, at the School of Mathematics and Institute for Advanced Study, and also at, in the Department of Statistics at Duke University. She completed her PhD in Princeton, actually with Barbara Engelhardt, who was one of the organizers of MLCB maybe 10 years ago now. And Bianca's group focuses on developing interpretable statistical machine learning tools for spatial transcriptomics and perturb-seq experimental design with application to development and regeneration. So really excited to have you here and take it away.

Thank you so much for having me here. It's a bit of an emotional time for me because last time I was in Seattle, I was visiting for graduate programs. So I was looking at the genome science program, and I met really wonderful people, and it was really inspiring. I didn't end up coming to UW, but I've always thought about, you know, like all the research that's happening here, and it's just really fantastic to be here on a beautiful campus. It's also one of the first times I am coming to a conference after the pandemic, so that's— it's really nice to reconnect with you in person. And if you find any of this interesting, or if you don't, like, please still, let's get coffee and chat. And today I'm very, very excited to tell you about something that I find very inspiring, and you will get to see if if I figure out how to use this slider. But so in the past 4 years, I got very excited about development. And one of the most beautiful things about development is that it's one of the first kind of like types of self-organization, the most important self-organization kind of systems that we encounter in nature. You know, this is a beautiful live sheet microscopy recording of a zebrafish embryo from Fengzhu Zhong's lab at Cambridge, at the Gordon Institute. And there's so much wealth of information that you can see here. You have like cells dividing, specializing. And one of the things that I'm very interested in my group is kind of studying the relationship between various modalities that could maybe lead to shape and form. So I'm very interested in how the cell identity emerged as a function of not only what cells contain, but also who they talk to and how do they look like physically. So I'm very interested in like 3 main areas of machine learning. So explainable machine learning, iterative design, and also spatial temporal prediction. And I think these are pillars a little bit different from what we've been talking about today, but I hope that you'll find them equally exciting. So I'm interested in control and self-organization. And today I'll probably try to talk to you about, okay, like how do we bridge across these modalities in ways that use either machine learning as an engineering tool and statistical data analysis. And, you know, I'm here to tell you that, you know, we all know that biomedical data analysis is very exciting. It's a prime time to be a computational biologist. We have access to information across multiple scales. Scales from the smallest to the largest one. We can look at population levels to predict risk scores. We can look at cell identity in terms of like what genomes are doing, what their gene expression, how is the gene expression organized. And this has really been possible by technological advancements. And the two technological advancements I'm very invested in are related to single-cell transcriptomic developments in the context of spatial data, in particular the light sheet microscope, where we can actually probe not only how these cells— what the cells contain, but also how do they change, exchange information with their environments, and where do they end up. And one of the things that makes me very excited to be part of the computational biology community is that we've been learning how to be collaborative and integrative in order to communicate with one another. We need to develop language that allows us to understand one another. And there's been a lot of initiatives that have enabled people like myself, a computational person, to engage with biological data. So Human Cell Atlas, UK Biobank, and GTEx Consortium. Actually, the GTEx Consortium was one of the first consortiums I engaged with as a graduate student when single-cell data wasn't so popular. I primarily started working with bulk gene And I think we have a prime opportunity to engage with each other across these communities in order to solve challenges that maybe we couldn't just solve independently if these datasets weren't available. And just to kind of summarize a lot of the themes that are appearing in this conference, I know we talked a lot about large language models, but there's a lot of other challenges that we face as a field. And a lot of these challenges, I would say they kind of like are represented by the 6 pillars. So first of all, data is extremely heterogeneous. In some cases, we might have a lot of it when we can train a lot of large language models, but in some situations, we might also have very few data points, in which cases we might need to be a bit more creative or even Bayesian. So we should try to think about how to model that data. So representation is another pillar that's very important. We've talked a lot about cell types and cell type prediction, pseudotime ordering. So thinking about how we go for very large scales to a way that allows us to visualize data and the diversity in the data is very important. And of course, we also want to understand how do we connect back to mechanism, scalability, and experimental design because we might want to be able to use this information to create new, to come up with new hypotheses, but also generate new experiments. And in this talk, I'll be primarily focusing on multiple data types, which some of, of them you might not be so familiar with. And I'll be focusing a little bit on how can we build in mechanism or understanding from biophysics in order to think a little bit about developmental problems. So talking about heterogeneity, of course, there's been a lot of interest in thinking about multimodality and deep learning in biology. So when I made this plot, you know, There were, like, thousands and thousands of results. So if you look at how deep learning in terms of, like, search results or representation in papers, this is quite a big peak. And multimodality also has been quite at the forefront of thinking about gene expression data. So in the context of recent research, you know, there's a lot to think about. And of course, like, we're maybe very familiar with the gene expression. We're thinking about cell identity as being the way the way each particular cell is employing different genes. But there's a lot of different wealth of information out there. And some of the talks today highlighted a lot of the unique challenges at different scales that these particular representations have. So we saw a lot of beautiful talks on protein sequences and TCR motifs and how difficult it is to generate data that comes from these discrete spaces where we need to optimize and maybe come up with new designs. And a lot of these challenges, you know, come from the fact that there's a diversity on how these mathematical objects can be represented. So oftentimes when you work with gene expression data, you work with tabular data. It can be either continuous or discrete, and actually deciding how to normalize it properly or what's the right way to normalize it is still a really big challenge in a lot of new emerging datasets as we collect more information across different technologies. Of course, there's also a lot of skills that have to do with the type of data we're collecting. So how do I encode time and how do I encode space and how do I think about cell-to-cell interactions? That also induces a lot of challenges when thinking about multimodal data. And I really like this part because it kind of shows a little bit the hierarchy from Sarah Tagman's paper, which shows the abundance of different modalities, but also the type of tools that we're trying to use to bridge across the scale. So some of the most popular tools that exist in our community are related to like matrix factorization and transfer learning, but also kind of like thinking about how do we go from gene regulatory networks in a context that might be either interventional or related to also chromatin accessibility and protein abundance. So, despite, so not despite, but motivated by a lot of these different modalities as statisticians or machine learning, as computer scientists, we have a lot of challenges that we have to address. So some of these can be quite low level. We think about sparsity or we can think about scaling and low-level processing, but there's a lot of challenges that we have to overcome when thinking about objects that come from very different sources. Recently, what I've been really thinking about has been how do you organize and represent spatial data in a way that retains the variability and it allows you to kind of relate one modality to another while thinking about the differences that could arise from the type of data one collects. What do I mean by what type of modalities? Well, so maybe we're like very used to thinking about what cells contain, but we're not so used to thinking about how cells look like and how that intersects with what we want to say about our system. So recently there's been a lot of new methods that have been allowing us to look at cells not only in terms of what they contain, but where they So there's been a lot of like challenges that come with the ability for us to segment these new datasets better and also kind of thinking about single or multiple object modeling prediction. So what is the right way to represent these datasets and how much can they teach us about biology is something that we haven't really formalized properly in the field. So going back to kind of like what does it mean to think about multimodal data is like it's important to recognize two things. Data comes across multiple scales and data also kind of tries to say something different about cellular identity. We heard something really nice earlier. I really like what Ashul was quoted saying that things in the genome happen asynchronously., and trying to assign synchronous identities to cells could be very different. So like, how do we learn what are the actual maps between gene expression and morphology is something that can become quite difficult. And one of the things that I found very interesting is that if you read across the multimodal literature that we are presented with, there's actually kind of a nice structural method that is present in a lot of situations. So if you look at a lot of the literature in machine learning where you are presented with objects that come from multiple modalities, one particular context where you might find yourself in is how do I find what is a common representation across multiple modalities? And Being a statistician, I take a lot of inspiration from the work of Hotelling. So Hotelling published a really foundational paper on canonical correlation analysis in the 1900s. And one of these tools really has to do is like if I'm presented with multiple modalities, one way to relate them to one another is to assume that there is some latent space from which I can sample and I can generate additional modalities in a way that allows me to maybe reconstruct one modality from another. So in our collective vocabulary, multimodal learning has been primarily seen through the lens of a variational autoencoder. In the context of a variational autoencoder, we can relate this to a simple factor analysis model where I'm assuming that my data is sitting on a low-dimensional space. There's different features that are associated with this latent dimensional space. And this can be combined according to my knowledge of biology to generate either gene expression, for example, or histology samples. So if you have a convolutional neural network that allows you to go from a lower dimensional space to an image, then that's really good. If you have chromatin accessibility, you can also use your favorite way of generating that. And similarly, you can use tools like SCVI to think about what's the right way to go from a latent space to a gene expression space according to this model. And this is neat, and this is really kind of a clean way to formalize a big class of models. What can they really tell you? Well, you can use the latent representation models to cluster different cell types, or maybe you can classify histology samples and say these are are coming from one particular condition versus another. You can combine this with maybe sequence variation models to find what are single nucleotide polymorphisms that would be affecting or mediating image phenotypes that would be related to gene expression. And at the end of the day, you can do a lot of this analysis, kind of like putting under the carpet the fact that there's a lot of complexity in the image in the gene expression space that you might not have much control over. So the question that I wanna pose is more like, what do we wanna accomplish with these representations? There are increasingly new architectures that allow us to go from modality to modality. There's a lot of really innovative work that's happening in the space, and we can accomplish quite a lot with them. So we can represent cells in compositional ways. There's no reason why I have to have 2 modalities, I can have 3, 4, or 5. If one modality is more expensive than another, then I can go and I can— I don't have to do an experiment. I can do feature selection and say which parts of my input are most crucial to a particular prediction. So I can do a lot of these tasks that are very useful in an engineering setting. But what about situations where we don't really have to pretend that we don't know anything about the biology?. So this really got me thinking about situations where if my underlying image has some physics that's preserved or that's very interesting, how do I think about incorporating those models in the context of multimodality? And this brings me to a quantity that's very dear to my heart because it really kind of bridges between local and global behavior, and this is tension. So in developmental biology, there's been quite a lot of work recently in trying to understand what is the effect of tension on the quality of development early on, and this also in the context of the quality of organoids and all sorts of kind of associations that we're trying to understand. So how do cells read in mechanical tension, and how is this mechanical tension further interpreted to determine cell identity? And there's a lot of work on kind of finding out what are the genes, or like protein-coding genes, that are most relevant to thinking about tension. But we wanted to ask, how do we how do we go about finding these particular elements if we just had no idea which elements would be at play? So going back to this problem of multimodality, we decided that instead of just thinking about how do we take multiple modalities and how do we represent them in a latent space, we asked, what if I actually know what the modalities that are coming from images are? So in other words, you know, if I have— so we worked with the seqFISH dataset from a developmental mouse embryo context, And we asked, okay, like, if I have multiple temporal slices at stages E8, E8.5, and 9, so this is kind of like a little bit before gastrulation is kicking in, can I think about these images in a new way? And can I find spatial correlations that might be determining the variability between gene expression and tensional profiles? So in order to accomplish this task, we had to think about 2 different fields that kind of were not really connected before. One of them was the biophysics of mechanical forces, and the other one was kind of what my expertise is mostly, which is gene expression statistical analysis. And I paired up with a very talented postdoc in Ben Simon's group, Adrian Hallou, who is a biophysicist and has a long-term interest in physics, and with one of my master's students, Ruyang He, then at Cambridge and now a PhD student in Sanja Vicovic's group and my group at Columbia. And we asked together, okay, can we put together our expertise and interest in biophysics and gene expression analysis to say something about early boundary formation in embryos? So what we looked at, we realized that there's innovation required at multiple levels. So people have been primarily reporting a lot of information about the analysis of seqFISH data, but a lot of the analysis has been mostly focused on cell type clustering and spatial clustering. So we asked, can we go deeper with this type of data and can we actually extract something more biologically relevant from such interesting datasets. So we realized that we had to actually do a different sort of image segmentation that would allow us to assign the nuclei that we would get from the seqFISH analysis appropriately. We had to do and had to develop new statistical machine learning tools for hypothesis testing for kind of trying to understand what are the genes that are most important to prescribing tensional variability. And in order to accomplish this, we put together ingredients from 3 different communities. So one of the communities is the very large community for image segmentation, and we built on an existing U-Net architecture for— in general, that was developed not particularly for sick fish data, but that's been recently repurposed for sick fish. And this architecture has been quite useful. There's a lot of interesting questions here. So one of the things that I should point out is like the number of annotated samples that one needs for this type of tools varies from dataset or dataset collection modality to dataset collection modality. So it's actually quite important to think about how can we develop as a field tools that would not require so much manual annotation. The second tool that we, the second community that we pulled ideas from is the community of biophysics physics and biomechanical modeling. So in this case, it was, it was really neat in the sense that we had access to this kind of really high school level physics, which is if I have a bunch of like bubbles in a 2D setting and I'm really thinking about force balance analysis, can I assume that when cells are in contact, there's kind of a little bit of a spring that's keeping them in place. So we really pulled in from a beautiful literature from Boris Strymon and Sebastian Streichen's work on early development of fly embryos, in which they came up with a lot of conclusions about how the tension that can be measured using a force balance analysis can actually go back and be tested in in vivo system and be correlated with, with actual forces that the cells are feeling. So what we did is we scaled this up and applied it to the zebrafish setting. And we also realized that there's another kind of piece missing. And this is, okay, like now I have all this information that I'm pulling from cells that are situated in space and time. And I also have information about what is the gene expression profile at this particular locations. And how do I go about testing the relationship between these very two different measurements, one of them being tension, which is in arbitrary units, and the other one being single-cell gene expression. So, in order— for those of you who are not familiar with Gaussian processes, Gaussian processes are a very neat way of incorporating spatial correlations. Maybe you've seen them in work on non-negative matrix factorization where people try to find ways of clustering data by assuming that there's some sort of spatial prior that enforces continuity between factors. You're definitely familiar with the normal Gaussian distribution. So Gaussian processes are just a way to extend that particular notion to stochastic processes. And they're quite neat. They've been used a lot in the medical literature for incorporating electronic healthcare records and kind of thinking about variability of continuous signals. And I like them for the fact that they're very interpretable in the sense that you can go ahead and design kernels that can think about how do you model periodicity or frequency, and you have really immediate control of your parameters when designing these processes. So you can have a lot of flexibility in terms of modeling quite complicated signals. So what we do with these 3 components, we're asking a question, it's like, can I ask whether tension has a true impact on causal impact on changes in gene expression in the context of development? And you can't test directly for causal assumptions without an intervention, but you can think about situations where you might have some sort of spillover effect from a confounder. So in this particular case, space could be impacting gene expression, but space could also be impacting tension because I told you the way we measure tension is actually using spatial correlations. I do this force balance analysis that's really affecting what will be the tensional distribution in my slice.. So I cannot purely fit a linear regression model. I have to control for the existence of these confounders. And luckily enough, there's quite a lot of solutions in the geo-additive spatial statistics literature. You know, people were thinking a lot of kriging in the early '40s and '50s. So they're kind of like, right, this comes from metallurgy. People were saying, okay, if I'm going to this location on Earth and I want to see like what's the distribution of metals how predictive can that be of a distribution of metals in another region of the, uh, of the Earth. So there's a lot of, uh, there's a lot of, uh, tools that are quite neat in that space. Um, they're called geoadditive structural equation models if you want to learn more about them. And, uh, one of the most recent papers is the one by Tadden and Knight. You can find it in the Annals of Applied Statistics type literature, which might be a bit different of a community for some of you here, but there's a lot of beautiful ideas that could be quite useful for us as, um, single-cell genomics. So what's happening here? Well, instead of fitting a regression model, you're taking into account the fact that space affects both modalities. So what you're doing, you're trying to pull out this information from confounders. So instead of testing directly for correlation between tension and gene expression, you're testing for correlation in a residual model. And this residual model is further enhanced by the fact that you're assuming that your errors will be spatially distributed. And this is where the Gaussian process comes in. This model, so we put this pipeline together and we're trying to make it accessible to different organisms. So not all this is for a mouse embryo, but we're trying to apply it to zebrafish and chick embryos to make it a little bit more useful. But in a nutshell, it's kind of like combining different ways of doing segmentation, mechanical stress inference, and for morphometrics, analysis and is performing all this kind of like statistical testing in order to ask what are genes that are most responsive to tension. And what are these genes that are most responsive to tension? So doing this analysis, we were quite excited to recover a lot of genes that were previously known to be associated with mechanical sensing properties. And not only that, but we are also able to kind of identify a lot of processes that were very important in early development. So for example, branching morphogenesis, and we were able to identify a lot of regions that were very important in muscle tissue development, which were kind of like the boundaries that we were looking for. And not only that, but we also found the fact that if we were looking at tension-enriched boundaries, you could find that you could actually read the boundary between different cell type compartments just by looking at tension and not by looking at the gene expression profile. Which was quite— maybe it's not surprising if I tell it to you now, but, you know, we definitely didn't necessarily expect this when we started the project. We also built a biophysical model to ask the following question. So if I give a boundary, how important is tension whether the cells of the different cell type are going to mix or they're going to stay together? Because this is very important in organ formation. So like one of the things is like once I established a boundary that has very strong tension across the boundary, will these cells ever mix up again? And we actually found out that not all boundaries are created equal. And we could identify between boundaries where there's a lot more cellular movement and boundaries that actually stay together as kind of a constraint later on in development. And that's something that I'm actually quite interested in exploring further, especially in the fly embryos where we have much more control on follow-up than in this setting. So what can we learn in this setting? Just to recap a little bit, we can recover non-mechanosensors, we can identify new mechanosensing gene signatures, and we can characterize boundaries between cell types in a way that's not only mediated by cell type clustering. And some of the things that are quite exciting in this space is thinking about how do we go from a 2D context to a 3D context so we can understand whether this type of boundaries are actually enforced in a 3D setting. So going back to this, what we really want to move forward to is kind of thinking about, okay, if I have this developing video, can I do force inference in this particular setting? Can I learn what are the cells that are under most pressure? And can I correlate that information to whether those cells are differentiating at a rate different than the ones around them? There's a lot of interesting recent work on tackling the first problem, which is how do you move from a 2D setting to a 3D, 4D setting? And this is some beautiful work called Embryo Mechanics Cartography. It's on bioRxiv now, not from my group, but from Eric Tullier's group and It's really beautiful work that requires a lot of computer vision and a lot of engineering solutions. One of the machine learning challenges in that case would be, could we use machine learning to actually solve this very large system of equations in a more efficient way without having to redo segmentations? But that's a bit of a— happy to talk more to you about this outside the talk. So, Going back to this, you know, like, sure, we can find genes that are related to tensional memory, but one of the challenges would be like, okay, if I want to truly study the variability and I want to think about tension as a quantity that I want to see how variable it is, how much, how precise it is specified genomically, what I might want to do is I would want to understand like first, how do I integrate live imaging data when I don't have access to all the gene expression profiles? How can I integrate this with spatial transcriptomics? So in seqFISH data, sure, like we have access to about 300 genes, But if I want to go to a live sheet microscopy setting, then I can only tag 1 or 2 or 3, maximum 3 genes. So until we get to that ideal dataset where we can have like full spatial temporal recording of developing embryos, maybe there's still something else that we can do. And in order to do that, we've been working together with John Marioni's group and a very talented postdoc, Jiawei Wang. Who's a Marie Curie Fellow, hosted in my group and in John Mariani's group. And one of the questions that we asked was like, okay, like if I have multiple light sheet microscopy recordings, can I find the right representation learning framework for integrating data if I have multiple embryos? So it turns out, you know, this is a very difficult question. If I have multiple embryos in sections, how can I train a similarity model that would allow me to recover the temporal ordering in such images. So to show you a little bit how the images actually look like, we use the light sheet microscopy data from Kate McDowell's group, who actually recovers this particular sick fish periods of E8.5 pregastrulation setting. And we ask, okay, like, what is the right machine learning tool that we should be using in order to be able to kind of stitch these slices back together. So if I give them to you, I slice my embryo and I give you all these images, could you tell me how I should bring it up, kind of like a giant 3D puzzle? And Jia Wei has been developing a really interesting contrastive learning-inspired architecture. We talked a little bit about transformers in the context of language models, but transformers actually are quite useful in the context of thinking about learning representations from images as well. And one of the key ideas are that in order to find good representations, so the key idea is if you use a convolutional neural network, you're not going to get very far. So for reasons that we still need to understand, what really helps is if instead of just training convolutional neural networks, you're training them in a contrastive way in which you give it positive samples from nearby sections and negative samples from very far away sections.. And this is kind of like maybe a recipe that keeps appearing also in interpretable machine learning techniques where actually a lot of these ideas from contrastive PCA that actually James Howe kind of introduced, and there's some beautiful work from Sui-In Lee's group also kind of thinking about how do you understand what are features that would give you some sort of contrast active behavior. So they're very, very powerful in helping you finding representations that allow you to perform well in downstream tasks. And what are these downstream tasks? Let's say I give you data from multiple embryos and I want to ask you, okay, if I provide you with a particular image, can you tell me where this image would be in a different embryo? This would be kind of like, think about it as a kind of single cell synchronization, but done at the level of light sheet microscopy. Microscopy. So kind of thinking about instead of doing batch correction from multiple single-cell atlases, you want to think about this— how do you do this in the image space? All right, so we're making quite a lot of progress in this space. We're not quite there yet. There's a lot of challenges related to how do you actually qualify the quantity of these methods? What is the right way to compute distances between shapes? There's a lot of— we're doing a lot of benchmarking efforts, and I feel like I want to take this as an opportunity to highlight that benchmarking kind of efforts are very important in our community. And I don't believe that the single method would do best at a particular challenge. It's just very important to kind of think about how do we put a lot of methods together, understand where they fail in order to move forward. Beyond representations, one of the challenges that I'm very excited as a field is like, you know, kind of if you see this beyond representing these images, you know, if I have data from developing organisms, how can I think about self-organizing systems that are actually giving rise to this property? So like, This is a neat paper from a few years ago where a group trained a set of like deep learning algorithms for seeing whether they can recover the morphologies that are present in a video game. So kind of like this is trained with video game data over millions of users and it's kind of saying, can I learn generative structures that would be able to recover this particular behaviors in the video game. And I, you know, of course biology is much more complex than that, but I think that coming up with models for 3D morphogenesis using diffusion models is actually kind of exciting and kind of understanding how do we include constraints like the ones in physics here in this 3D morphogenesis example, you kind of have cubes and they're independent from one another, but kind of thinking about how do you go if instead of cubes you have a bubble system, kind of like soap bubbles with a lot of physics can we think about diffusion models that are allowing us to recapitulate the trajectories of cells in a particular way? I think this is a very exciting area. So in my group, we're starting to think about computational morphogenesis. So kind of how do we use diffusion models in order to do inverse design? So one of the particular questions is, now that I know how to measure tension and see how tension evolves over time, can I create systems that would be having a particular tension profile? What's essential in specifying So think about this kind of like an inverse origami set of problem, you know, where you start with the image of a bunny and you're saying, what is the right set of folding tools or algorithms that I should be able to be creating in this space? So I'm really fascinated in this idea of thinking about embryos as materials. So kind of like not thinking about single cells in separate contexts, but thinking about them as, you know, kind of gooey materials that are inhabiting a space and they're reading in and reading out and contributing to the shaping of that environment. So I'm very excited about this area of work. In the context of making this work translational, I've been working very closely with my collaborator Yvonne Wapi, who's a specialist in organoids and skin and wound healing. So one of the things that we're trying to understand is that can we incorporate information about tension in order to enable faster healing? So we're trying to think about how do we incorporate spatial transcriptomic and to understand whether this has any sort of impact on how do we design Perturb-seq screens that could allow us to collect better data that would come up with maybe therapies for non-fibrotic wound healing. So, you know, if any of this is interesting, I would love to see applications from graduate students and postdocs, we're quite actively recruiting for this particular problem. And before I leave, I just kind of wanted to leave you with a perspective of what I think that we're— where we are and where we're going with this type of tools as a community. So biomedical data analysis, in particular computational biology, I talked about the challenges at the beginning. They're kind of like 4 pillars for me. There's the spatial-temporal and multimodal aspect. There's the interpretability aspect, directionality, and the translational aspect. So spatial-temporal multimodal, we've been thinking a lot about how do we learn pseudo-time analysis? How do we learn how to do clustering in a spatial context? We've been thinking a lot about spatial and time separately. So I'm very excited about tools that will be combining these axes jointly in a multimodal setting. So if you have chromatin accessibility and gene expression, For example, what is the arrow of time? How do we think about causality in the context of spatial data? I think there's a lot of innovation to say, how do we think about building interpretable models that allow us to pinpoint what are the key differences between different pathologies? So the nice thing in computer vision is that sometimes if you do the right thing, it's very clear that you did the right thing. So if you do a segmentation task, You can see a cow, you can say like, ah, I segmented the cow correctly, you know. What is the equivalent in gene expression data? I think rethinking interpretability in a gene expression-centric fashion is something that's very exciting to me and that will allow us to build better therapies. Actionability is another pillar that I think that it's very exciting and I'm looking forward to some of the talks later this afternoon that we'll be talking about how do we prioritize gene expression panels with problems like control in mind. And of course, there's numerous translational applications. For example, especially in the context of organoids, like are organoids good model systems for the biology we're trying to learn? So can I determine whether a cell in an organoid context is mature enough? How do I align trajectories from these datasets on the same trajectory? And is tension important in this context? Do we incorporate morphology that is consistent between, let's say, an organoid context and a fetal context? How do we assign quality to these models that we are getting more and more data from? And with that, I just wanted to kind of emphasize the fact that none of this work would be possible without us working closely with experimentalists, and we're just part of this loop. And I hope that, you know, as we're evolving as a community, like, we already saw so many beautiful talks today about how validation is important and how having wet lab kind of analysis is so important. So I hope that my own work is going to be going more towards scientific hypothesis and design and With that, I would like to thank you for your attention and my growing group. If you want to read more about this work, there's two papers that are very relevant to what I've just told you, and I'd like to thank my collaborators because without them, this work will not be possible. Yeah, thanks. Thank you so much. We have time for a couple of questions.

Very nice talk. I have to admit that I didn't think about morphology with that depth. So if I understand correctly, you are defining tension per edge. For each pair of touching or neighboring cells, you are defining the tension between them, right? So my question is kind of an extension to that representation. What would you think would be the best representation for, let's say, cell-cell interaction? So, so far we've learned a lot of representations for gene expression, for chromatin accessibility, but all of those representations are at single-cell level, right? But if we want to incorporate cell-cell interaction, what would be the right way to do so?

Yeah, I think that's a really fantastic question. I think that there's a lot of work work in, in the statistics literature when thinking how do you parameterize adjacency graphs with covariate information. So for example, I think in the context of 2D data, it's a little more immediate because you can think about each cell is a point in a graph. You have the tension, it's quite, quite discrete, you know, so you have an edge, the edge has a particular feature, And then you can think about kind of building these bivariate maps, you know, where you have like dot products with some sort of weighting function in the middle. So you can get creative and use a lot of the toolkit in that literature of latent models for graphs to get a little bit of that. And we started to think a little of how to do that. I think the main issue here is how do you validate any of these results in vivo? And this is why I've been starting to look at the fly embryos a little bit, because there's more biology known and you can do optogenetic screens. You can actually do a lot more in that space than in the mouse setting that I just described. So, so I think the testing part, it takes a little bit of you know, labor. But I think the hard part really conceptually is validating. Yeah. So thanks for your question. All right.

We can move on to the next speaker.

Thanks again. Oh, I have to go— how do I stop sharing the screen? No, we do want to keep sharing, right? You're using this laptop, are you? Yeah, great. Yeah, so I think it's right in here. I'm sorry, what's your name? All right, great. And it's a Google— it's a PowerPoint? Okay. Did you make it in Google Slides or did you make it in— okay, great. Sometimes it messes them up a little bit. I'm always A little bit more.

Okay.

And then I think, yes, we can try and figure that out. Let's, uh, I'll figure this once it dries. Okay.