Scientist Stories: Omer Bayraktar, High-Throughput Spatial Genomics

Hello, everyone. My name is Omar Bayraktar. I'm a group leader at the Wellcome Sanger Institute. And it's a pleasure today to tell you about my team's research on high-throughput spatial genomics. So the underlying biology that my group is interested in is the astonishing diversity of cell types in the human body. The system that we study with the most passion is the nervous system. Specifically the brain. And for many years now, we know that, you know, the function of the brain in health and disease is enabled by a significant diversity of different cell types. Neurons, the principal cell types of the brain, come in this really staggering complexity of different flavors. You know, different neuronal subtypes have different, you know, morphologies. They have different electrical properties, and these also travel with different transcriptional and molecular properties as well. But in addition to neurons, you know, there are many other cell types in the brain. So these glial cells, as they're called, also are made up of several different flavors. So astrocytes, oligodendrocytes, microglia, and, you know, many others and subtypes of each of these glial cell types. They all come together to basically support neuronal function in health and disease. So now one of the sort of things that my group would like to accomplish is to really take high-throughput approaches to dissect the biology of all of these cell types, both in health and disease. And the approaches that we use to be able to basically study the biology of as many brain cell types at the same time as possible is transcriptomic profiling. So I assume through previous talks that you may have heard from Sanger faculty or through your own research interests, a lot of you are familiar with recent developments in the field of single-cell transcriptomics. So this is now a fantastic field of techniques that enable us to take a complex piece of tissue like the human brain, dissociate it into a soup of single cells, and read the transcriptomic context of each cell. Now, this is really a fantastic method to take a complex tissue and really classify the different molecular flavors of cell types, which should give us a really great starting point to understand their function. But one thing that single-cell transcriptomics doesn't capture is actually the spatial arrangement of all these different cell types in the tissues. So as all of you would appreciate, both in the brain and other solid organs, all of these different molecular flavors of cell types are not really floating in space. They are found in very specific tissue microenvironments. Where they are found in particular regions of these organs, and at each location they are surrounded with a very particular set of neighbors. And, you know, the sort of conversations they have with their neighboring cells, the cellular interactions, really determine how the tissue functions both in health and disease. So to be able to really identify the spatial architecture of cell types in the tissues in high throughput, over the past 5 to 10 years, a new field has emerged. And we could describe this field broadly as spatial genomics. And the idea here is really use a different collection of techniques that sort of span both tissue sequencing but also multiplexed RNA and protein imaging to be able to study cell types in situ in intact tissue samples. And something that my team is really interested in is really, you know, taking these approaches, applying to, you know, both the healthy and the diseased nervous system. But, you know, ultimately we're most interested in really applying as many of these approaches in an integrated manner as possible to study the spatial architecture of tissues in a truly multimodal manner. And, you know, in the lab, we are really excited about how we can use this sort of multipronged strategy to understand healthy tissue structure and function, but also dig deeper into disease pathology as well as disease models, and also, you know, use this as an extensive sort of, you know, Swiss Army knife type of tool to perform in situ validation of spatial single-cell transcriptomic findings. So today I'll tell you mostly about our work in the space of integrating single-cell RNA-seq with spatially resolved sequencing. And the work I'll tell you today is one of the sort of major directions that we follow in this new initiative that I lead between the Sanger Institute and the European Bioinformatics Institute, EBI, that we call the High-Throughput Spatial Genomics Initiative. So spatially resolved sequencing perhaps sounds like a mouthful, but it is more or less performing next-generation sequencing of RNA molecules that are positionally captured from tissues. There are a number of different techniques in this field, but a lot of them sort of follow this, you know, layout where you start with a thin tissue section—this is 1 or 2 cell layers thick. These tissue sections are positioned on various types of tissue micro—various types of microarrays that contain consists of oligonucleotides that are built to capture RNA transcripts. These oligonucleotides are, again, arranged into a spatial grid in 2 dimensions. At each spot, each nucleotide is tagged with a position-specific barcode. And throughout these spatially resolved sequencing experiments, these techniques often permeabilize the tissue sections to extract RNA molecules These RNA molecules are then captured onto the array and then positionally barcoded and read out with next-generation sequencing. So these spatially resolved sequencing methods are quite attractive to us because right off the bat, they provide you an unbiased way to study the transcriptome. So, you know, they are not using probes to target specific genes. They are capturing more or less everything from the tissue material. A lot of these methods now rely on just only standard histology and molecular biology protocols. Methods like the 10x Visium assay are now commercially available as kits. So they're really easy to apply in a lot of different labs. And most importantly for us, you know, in order to sort of reveal the fine architecture of tissues, the spatial resolution of these methods have been rapidly improving. So at the dawn of their inception, these spatial transcriptomic methods provided a rather coarse spatial resolution where sort of each microarray location was, you know, about 100 microns in diameter, which corresponds to many cells. But over the past couple years, the resolution of these methods have significantly increased. So the currently commercially available methods, Visium spatial transcriptomics from 10x Genomics provides you this 50-micron resolution, which corresponds to several cells, but, you know, a much smaller number than before. And there are also homebrew methods such as slide sequencing that give you even higher resolution. So these spatial resolved sequencing approaches are quite exciting to us. And, you know, we think that, you know, they provide— as I'll show you through my presentation, they provide, I think, fantastic partner to single-cell transcriptomics. But the data that comes out of these methods is not quite like what you get from single-cell transcriptomics. And the sort of simple reason for that is that in each of these methods, what we have are, you know, basically these profiling grids that are superimposed onto the tissues. And, you know, our tissues across different organs, you know, do not have cells that are arranged into regular type of grids. So both in high sort of coarse resolution methods like Visium and high resolution methods like slide sequencing, what you often have is the lack of a one-to-one correspondence between individual cell types and these microarray locations. So in the case of Visium, each location will profile multiple cell types, and in the case of slide see each spatial tissue location on the array might require fractions of multiple cells. So to be able to get cell type-specific information out of these technologies, we actually need computational approaches that would allow us to deconvolve this cell type information from spatial transcriptomic data. That is precisely what my lab, in collaboration with Oli Stegel's lab in the FKZ and EMBL, aimed to build. So the computational tool that we built, we call it Cell2Location, and what Cell2Location does is that it combines combines single-cell transcriptomics with spatial transcriptomics in a probabilistic manner to deconvolve cell types in complex spatial transcriptomics datasets. When you have something like this, which is diagrammed here, you have the ability to immediately predict the locations of all of the cell types in your single-cell RNA-seq reference in the tissues, and that enables you to ask more sophisticated questions such as identifying tissue regions based on the composition of these different cell types, but also looking beyond that and try to sort estimate which cell types are actually interacting with each other, that are physically co-located with each other. So all of the work that I'll tell you about is mostly done by a fantastic graduate student at the Sanger Institute, Vitaliy Pleschevnikov, and it was recently accepted for publication. So I'll just give you a very quick look under the hood of what cell colocation does. What we are doing in our workflow is that we are using single-cell RNA-seq as basically a cell type dictionary. We are starting with the average gene expression profiles of all of the different cell types in our single-cell RNA-seq reference. We can estimate this using standard single-cell RNA-seq analysis workflows, but we also provide as part of the self-location package more complex methods to account for batch effects and technical effects and artifacts of single-cell RNA-seq data to extract these cell-type-specific gene expression signatures. After that, Cell2Location uses a Bayesian framework to take these gene expression signatures of all the different cell types in your single-cell reference, look at the mRNA counts in the spatial transcriptomics data— and this could be really any type of count-based spatial technology from Miseum to SlideSeq to NanoString— and then estimate the abundance of each of these cell types at each spatial transcriptomics location. And what this reveals at high throughput is the mapped locations of every single cell type in your single-cell RNA-seq reference. So those of you who follow this field of spatial transcriptomic deconvolution are probably aware that a fantastic number of tools were made available in this field over the past year. With Cell2Location, I think the advantage that we offer is that we can really map cell types and complex tissues with high sensitivity. So the kinds of things that we really, you know, like to do with self-allocation is not to just map broad, very well-defined distinct cell types into tissues, but actually identify fine-grained differences between different cell types that are often really important to tissue function. So as I'll show you throughout the rest of the presentation, we can really map these types of fine-grained flavors of cell types better than other tools, and we can also handle complex references that consist of hundreds of cell types. The tool itself is versatile across spatial technologies, so we're really presenting it as kind of a one-stop shop for spatial transcriptomics. And finally, really thinking about, you know, high-throughput application of these methodologies, we find that Cell2Location is very computationally efficient, so we can scale it to really large datasets. So to dive right into the application, when we developed Cell2Location, we decided to generate a pilot dataset to really test drive the algorithm. And for this, we picked the mouse brain. The reason for that is that the mouse brain is really a fantastic tissue that has many different neuronal and glial cell types that have very well-known locations across different brain regions. And to really create, I think, the kind of— I think the best starting dataset for cell colocation, what we did was that to generate a paired single-cell and spatial transcriptomic dataset. So we took these mouse brain samples, we cut consecutive thick and thin tissue sections from them. And the thin sections we use to extract nuclei and, you know, perform unbiased single-nuclei RNA sequencing to sort of get those cell-type gene expression patterns. And the thin sections, we place them onto these Visium spatial transcriptomics arrays, which allow us to profile the tissue at this 55-micron resolution. And finally, we put these two data types together with cell location. So on the right here, what you can see is actually a histology image of one of the very tissue sections mouse brain tissue sections that was assayed with our Visium approach. So, what we have done here is that we have sort of manually marked the different brain regions. So, you can see we have lots of regional complexity here. And if I were to sort of show you a high magnification of this image, what you can see here is now that by using the histology image, we have actually segmented every single nuclei on this tissue. So, they are, you know, colored nuclei that you can see in this image. And these big transparent dots are actually the locations of these 55 micron Visium spots where we have actually counted the number of nuclei at each location. And this really shows you that while with the Visium spatial transcriptomic assay we can profile large tissue areas and many sort of different cell types, at each location we are picking up more than 5, on average something like 8 different cells. So we really need cell type deconvolution. So to dive right into the results, on the left what you can see is our single nuclei reference of the mouse brain. In this plot, every single point— sort of tiny points— correspond to individual nuclei in our datasets, and they're clustered and analyzed according to standard single-cell workflows to identify about 60 different neuronal and glial cell types, where different colors indicate different cell types. So it's quite a rich dataset to start this type of analysis. On the right, I'll play you a little animation that actually shows you cell-to-location on the job. So, at the beginning of the animation, we'll start with our single nuclei reference, where we have marked 6 different neuronal and glial cell types. These cell types have very well-known spatial positions in the brain, so they provide us really good controls to understand if the self-location approach works. And throughout the presentation, you'll see that we will compute the average gene expression patterns of these cell types, and then we will switch to the Visium spatial atomic datasets where at each spatial position in the mouse brain, indicated as these different circles over the histology image, we will estimate the positions of all of these 6 target cell types. And with these patterns, what you can see is that we have taken these cell types and mapped them to these spatially distinct brain regions ranging from neuronal subtypes mapped to cortical layers or, you know, different parts of the thalamus and the hypothalamus, but also glial cell types such as oligodendrocytes map to regions like the white matter. So all patterns that we expect. So it's great to be able to show you this, that, you know, show you that these 6 cell types are being mapped to the locations that they're supposed to live in. But I think the true advantage of methods like cell-to-location is the ability to take all of the cell types in your single-cell reference and map them all at once. So with that at hand, I think it really opens the door to doing high-throughput cell atlasing. And the example in the mouse mouse brain dataset. You can see it here. So on the very left, you can see the sort of anatomy, the histology of this mouse brain. In the middle, you can see how we can actually identify distinct tissue regions in an automated manner. So here we take every single tissue position and we cluster them to different sort of, you know, regions based on their estimated composition of all of the 59 cell types in our reference. And on the very right, you can see our high-throughput cell type map where, you know, we can identify 29 tissue regions in an unsupervised manner using this automated mapping approach. And each region, we can read out the abundance of every single one of these 59 cell types. So I think, you know, for our research and others, it's really valuable to be able to build these type of cell types. In this mouse brain experiment, we could really sort of see, you know, lots of different interesting patterns. We could identify neuronal cell types that are enriched in very sort of specific brain regions, but we could also see sort of more homogeneously distributed cell types like, you know, vasculature and microglial cells that are located across multiple brain regions. But to really showcase you the sort of power of cell-to-location, I'll just focus on one set of results. And that is really identifying fine-grained differences between these brain cell types. And for that, the focus— that's the sort of like, you know, the population that we focused were astroglial cells, So astrocytes are, you know, one of the sort of principal brain cell types that my team is interested in. They are one of the major glia in the nervous system. And, you know, over the past 20 years, you know, they have really transformed from sort of passive brain glue to really active participants in the development and refinement of neural circuits. So— but one sort of aspect of astrocyte biology that is still not really well known is their regional heterogeneity. Are astrocytes in different parts of the brain doing different things? Are they regionally heterogeneous? So to really get at this, what we did was that we pulled all of the astrocytes from our single nuclei dataset, and we performed unbiased clustering to identify 10 different molecular flavors of these cell types. With these cell types, at the beginning, we didn't have much of an idea if they belonged to different brain regions. And when we actually mapped them to our Visium data using self-location, we were quite happy to see that many of these cell types are actually specifically coming from different brain regions, such as the cortex, you know, the hippocampus and the thalamus and the hypothalamus, as you can see here. So, you know, a lot of these differences were quite fine-grained. Only a small number of markers, genes, distinguished these astrocyte populations, but we could nicely map them across different brain regions, showing that we can really get in and identify where these fine-grained cell types lived. So, the one sort of population population that I'll focus on for this talk is really, you know, kind of, I think, the most challenging population out of these new astrocytes. So when we looked into these subtypes, we find that a very rare population of them, which correspond to less than 0.1%, map to a very specific brain region called the medial habenula. So, again, these habenular astrocytes were quite rare in our dataset, you know, measured in just, I think, under 100 cells. And with cell-to-location, we could actually see that these cell types were most abundantly present in relatively few tissue locations in this brain region, the medial habenula. And this was a pattern we could actually validate using orthogonal imaging technologies, such as single-molecule FISH, which is shown here. So I hope that this really demonstrates to you with self-location, we can do these types of, you know, complex tissue mappings, you know, map, you know, dozens of different cell types into the tissue space. But beyond just mapping the most distinct, the most obvious cell types, we could really dig deep into the biology and heterogeneity of cell types and map these fine-grade cell differences. So as part of our, you know, I think many fascinating collaborations with the Sanger Institute, we've been really excited to take cell phylolocation and really extend it to different human tissues. So as part of a couple of our collaborations that I've indicated here, we were quite happy to see that with our, you know, integrated single-cell and spatial transcriptomic approach, we could really go into multiple human tissues ranging from the human brain to the human lymph nodes to human gut, as well as, you know, other organs such as the human, you know, reproductive system, the endometrium, and we could really perform this integrated multimodal mapping to arrive at new biological insights on how different cell types are spatially organized in different tissues and how their interactions are are supporting the biology of these human organs. So, you know, in our future work, we are quite excited to take these types of approaches and extend them to other healthy and diseased tissues. So at the beginning, I told you that self-location was really good at mapping fine-grained cell differences. And while, you know, we had one set of results in the mouse brain with these habenular rare astrocytes, we were quite also interested in, you know, comparing the performance of self-location to other in a quantitative manner. So, to be able to do this, we collaborated with Sarah Teichmann's group, which recently published a fantastic single-cell atlas of the human gut throughout development. And what Sarah's lab did was that they provided a very comprehensive annotation of immune and non-immune cells throughout the human gut. And what we could do was that we could take this wonderful dataset from Sarah's lab and spatially map it to a recently available another published dataset, a Visium dataset of the human colon from Allison Simmons' lab from Oxford. And, you know, when we did this, our main question was whether we could take all these fine-grained cell types in this gut single-cell reference and actually locate them properly into the secondary lymphoid organs found in the gut. And sort of some of the initial results you can see in here. So, in the middle, you're looking at a histology image of the human gut that was placed on a Visium array. Just by the histology image alone, actually, you can predict where the immune cell types are supposed to be located. So, this sort of dense region here is a lymphoid follicle, and, you know, with cells location mapping, what you can nicely see is that some of these cardinal cell types of secondary lymphoid organs, such as cycling B cells, are highly enriched there. But beyond that, you can also see the spatial architecture of the colon revealed here, ranging from monocytes to epithelial progenitors to myofibroblasts. But what's really allowed us to do is that this allowed us— this provided us really a, you know, quantitative benchmark where we can sort of identify where we expect to see these different immune cell types and compare cell-to-location to other methods. So when we did this, we were quite happy to see that indeed cell-to-location, beyond just mapping, you know, these transcriptionally distinct cell types, we could take transcriptionally fine cell types that do not show lots of differences between them, such as these different flavors of B cells, T follicular helper cells, which are very important shepherds of B cell maturation in the lymphoid nodes, and other flavors of T cells and really map them, you know, accurately into the lymphoid follicles. And often when we looked at other methods, computational methods that perform this deconvolution, we found that while they were good at mapping these distinct cell types, immune cell types, into the lymphoid structure, they often struggled to accurately map the more transcriptionally fine-grained types. And the quantification of the data is shown here, where, you know, we can show you that when it comes to transcriptionally distinct cell types, most of the methods perform similarly. But then when it comes to fine-grained cell types, which are often quite important to understand, you know, the finer aspects of tissue biology, cell location provided a better tool. So with this, what I'd like to do is that I hope I demonstrated to you that cell location is a versatile tool that allows you to, you know, map different types of tissue architectures, but also do a really deep sort of analysis of cell types in situ and reveal, you know, the locations of fine-grained cell types and sort of start thinking, you know, provide, you know, I think, good starting points for biological hypotheses. So to give you sort of flavors of the types of, you know, projects we're interested in the lab now, pushing forward is that we would love to basically take self-allocation and apply it to dissect the pathology of different human diseases. And as aligned with this goal, a new project that we have started is working with a number of partners in the UK, including Michele Noceda in the Imperial College and the UK Coronavirus Immunology Consortium, to really start studying the pathology of COVID-19 in the lung. So, I think a lot of you are now familiar with the sort of, you know, severe tissue damage COVID-19 infection inflicts on the lung. You know, if you were to just use standard histological techniques and look at the lung tissue samples of patients who have passed away from severe COVID, you can easily see sort of severe alveolar damage. But, you know, what we would really like to do is we would like to apply these single single-cell spatial transcriptomic methods to start answering— to start asking more detailed questions about what's going on at these sites of pathology. Can we identify immune cell populations that are enriched in these specific pathological lesions? Can we decipher the pathological cell-cell signaling between these populations at once? And could these be useful to identify useful biomarkers and therapeutic targets? So to do this, we decided to generate both extensive single-cell RNA-seq Seq and spatial transcriptomic dataset of the human COVID lungs. With this, we started with these post-mortem autopsy lung tissue samples from donors who passed away from severe COVID versus donors who passed away from unrelated reasons to provide us a gene expression signature of cell types. As I'll show you in the next slides, we integrated a large number of published datasets to provide a very comprehensive single-nuclei reference, single-cell reference of COVID-19 the healthy lung and to profile, you know, specific tissue locations that show pathology. We used a new spatial transcriptomic profiling method called NanoString whole transcriptome assay, which I don't have, unfortunately, time to go into right now. But then we basically took the cell-to-location workflow to integrate these two datasets and sort of start looking at cell-cell interactions in the pathology. So the reference that we put together was quite integrated and comprehensive. Massive in that it spanned over 100 donors, both in the healthy and the COVID You know, it consists of many different cells that, you know, accurately represents multiple cellular compartments from the immune to epithelia to the endothelial to the fibroblast. And it actually allowed us right off the bat quantitative comparison of cell types that are selectively enriched in COVID-19 versus those who are selectively lost in COVID-19. But what we are quite interested in is really taking this single-cell RNA-seq reference and mapping it into lung tissue that shows COVID-19 pathology. So to do that, again, we took Cell2Location, we adapted it to work more accurately on NanoString data, creating an algorithm called SpaceJam. And with this, what we could do was that at one step, we could estimate the cellular composition of all of the different tissue locations that we profiled. On COVID-19. So I will not go through this slide where, you know, I— where, you know, we're demonstrating to you the sort of estimated abundance of different cell populations across different positions in the healthy and the COVID lung. But I hope that at the sort of a superficial look, you can appreciate that the cellular phenotypes that we're seeing in COVID-19 in terms of loss or gain of cell types are actually quite variable and complex. So to be able to really study them, I think we need these quantitative tools that can map the cell type precisely and look at their cell interactions. And what we are doing right now is that we are basically using this now. We are not just stopping there. We're using this, you know, high-throughput spatial mapping of cell types to lung tissue samples as a starting point to identify which immune interactions are spatially co-located. So we can actually take this, you know, sort of high-throughput estimation of cell types types in the lungs, and we could perform a second layer of deconvolution as part of cell colocation that reveals which actually cellular compartments are present throughout the tissues. And at the moment, we are looking through each of these compartments using both manual as well as automated cell interaction analysis to be able to identify which cell types are present there and which of these interactions could be interesting to COVID-19 pathology. So with that, I hope I demonstrated to you that we are putting together single-cell spatial RNA-seq in a manner that really allows us to dig deep into complex tissue architecture. I hope that the work that I presented both with Visium and NanoString demonstrates to you that we can actually start with low or coarse-resolution spatial transcriptomic data and use computational deconvolution to really extend our analysis to a higher resolution. And in terms of, I think, you know, next steps for the development of SALTOLOCATION, we are quite interested to sort of follow both the single-cell playbook, you know, we're interested in batch integration, scaling up SALTOLOCATION to deal with much larger data set. But we're also interested in digging deeper into spatial transcriptomic data sets, looking into things like tissue microenvironments and spatial gene expression variability in a more quantitative manner. So if you're interested in testing out SALTOLOCATION, both our preprint as well as our code is available. So please give it a Joshua Elkington. So moving forward, our lab, as I said, is interested in a number of tissue mapping projects. The final one that I'll tell you today is one that I'm quite excited about. It's a new project that we started to really apply the single-cell spatial transcriptomic tools and the cell-to-location approach we have developed and really use it to understand the biology of human disease. So this project is very generously supported by a new initiative from the Wellcome called the Valcum Leap as part of this Delta Tissue program. What, you know, us and our team of wonderful partners shown here— what we are quite interested in doing is to really integrate these unbiased single-cell and spatial genomics analyses to discover how the tumor microenvironment regulates malignant cell states in glioblastoma. And I think, you know, at this point, I think you are now— we are familiar with the type of approach that we are hoping to use here as part of this project. We are taking a large number of tumor resection samples from glioblastoma patients, and we are really applying all these different multimodal profiling methods ranging from joint RNA and chromatin accessibility profiling of cells to higher-resolution Visium profiling as well as spatial whole genome sequencing and really put these different multimodal types of information together using improved versions of cell location to understand the biology of GBM at a very high throughput level. And what we're really hoping is that as we put all these different multimodal genomic methods together and spatially chart the biology of glioblastoma, we're really hoping that that would give us very tangible targets in terms of cells and molecules that really, you know, carry out the communication between the tumor microenvironment and the tumor itself. And really help us understand the biology of glioblastoma and the control mechanisms of these malignant cell states and allow us to really devise better therapeutic targets. So with that, I'd like to acknowledge our fantastic team at the Sanger Institute. I think, you know, it's been really— our lab has been here for about 3 years, but it's been a fantastic journey to recruit people and really flourish in a large number of collaborators. Collaborations we have here. Beyond that, I'd also like to establish our fantastic set of collaborators, you know, like Stegel in the South location, as well as Michael, Michael and Aseda in Imperial, and, you know, our collaborators and, you know, the Leap Grant David Rowitch and many spatial mapping projects such as Sarah Tiedemann and Rosa Rendertormal. With that, I'll be happy to take your questions in the following Q&A session. Thank you very much.