Scientist Stories: John Marioni, Using Single Cell Genomics to Understand Cell Fate Decisions

Good morning, everybody. I hope that those of you sitting in the back aren't just overcoming your jet lag, though I wouldn't blame you entirely. So over the next 25 minutes or so, I'm going to tell you about work that we've done in the lab where we've used a combination of single-cell omics tools to understand cell fate decisions, not in human but in the mouse. And so you're probably all familiar, or at least you've all heard of, how some of these single-cell omics technologies— we'll hear more about these later in this session— are being used to generate atlases of different cell types, whether it be in model systems such as zebrafish or the mouse, or indeed much larger-scale projects such as the Human Cell Atlas. So this is this large international consortium that aims to map cell types across the human body to create, if you will, a periodic table of cell types that can then be used as a reference for different diseases. So, for example, If you're looking at the liver and you're wondering what cell type the particular disease that you have is affecting or the mutation you have is affecting, the Human Cell Atlas should be able to define those cell types and then map, and then that'll allow you to hone the downstream decisions about targeted therapies perhaps, or just in terms of an experimental protocol to decide what tissue you should be looking at for downstream work. However, there's an argument that such atlases, whilst nice, are only really useful when we begin to perturb them to understand what happens when things go wrong. So what you want to do is you want to map things to your cell types to understand what goes wrong in the context of development, what goes wrong in the context of disease. And what I'm going to show you about today is early work that we've done that tries to take some of these reference maps and perturbs them in different ways using both genetic perturbations but also looking at the epigenome to try and understand something about cell fate choice. So the system that I'm going to use to motivate this work is a very classical system. It's a system that's been studied for literally hundreds of years, and that's early development, and in particular, early development in the mouse. And we're going to focus around a time point called gastrulation. This is an absolutely critical time point of development. It's when the basic body plan is laid down, when you begin to see the first signs of organs. And it's a time point that's particularly exquisitely suitable for single-cell omics techniques because there are very small numbers of cells in the embryo. There are on the order of hundreds or low thousands of cells. They're highly heterogeneous, and so one really needs to be able to study these cells at single-cell resolution to understand both what they're doing and what they're going to become. So this is what we tried to do. We used the mouse as our model system, and together with wonderful colleagues in Cambridge— and the work I'm going to talk about today is collaborative with Bertie Goodkins and Wolf Reich—, as well as other colleagues in Cambridge who have also been heavily involved, we determined about 5, 6 years ago now that we could use some of these up-and-coming single-cell omics technologies to study early development. So we're going from E4.5— this is embryonic day 4.5, 4.5 days post-fertilization in the mouse— and we're going through to about embryonic day 8.25, so around 4 days later. And what you can see here is, if you just look at this little cartoon, is quite a big change in the morphology. But really the key point is here around E6.5. And what happens here is that cells move from the anterior portion of the epiblast, this green tissue. They pass through a structure called the primitive streak. And as they receive signals, particularly from the extraembryonic cells, that then commits these cells to different downstream lineages. And as they commit, you then see a couple of days later the first signs of an organism that in some way recapitulates the adult animal with the signs of cell types that are going to be present. So the first study we used using single-cell approaches to try and get to this was a rather small-scale study where we sorted 1,200 cells, and this is work with Berti's lab, focused on move from gastrulation through to the development of early blood cells. So the earliest time point at embryonic day 6.5, we just looked at cells from the epiblast. We didn't sort them for a particular lineage. But as we moved later, because we're profiling only on the order of hundreds of cells, we sorted cells that were committed to a mesodermal fleet, so those that are enriched for Flic1 expression or surface protein expression. And then later on we also sorted for a combination of Flic1 and CD41, with CD41 being an early blood marker. And so you can see in each of these cartoons, the little schematic illustrates the regions of the embryo from which we think we're sampling cells. After having sampled these cells, we sort them into 96-well plates in this experiment, and then each cell is profiled individually using RNA sequencing, here using the SmartSeq2 protocol, in order to assay its expression. So this experiment gives us the expression profile of 1,200 cells at these 4 time points. Having done this— and this is a very typical single-cell RNA-seq analysis now— you take these expression profiles, you find features that are of interest. They could be highly variable genes. They could be principal components that correspond to so-called eigen genes, so features that are a combination of multiple genes. But essentially what you want to do is you want to identify features that explain variation in the high-dimensional expression space, and then you want to cluster these. So you want to cluster the cells into groups. That correspond to different cell types. And that's what's illustrated here in this heat map where we've got cells on the columns of this heat map and then marker genes for different clusters are shown in the rows. So for example, if we look at this cluster 3 here, this cyan cluster, you can see that it's enriched for expression of Nanog and Oct-4, and also given its developmental time point, we know that these cells come from the epiblast. And similarly, we can assign identity to other clusters. So for example, this cluster 8 here, has high expression of the blood transcription factor Tal1, as well as the fetal version of hemoglobin. And so these are early blood cells. And so this is what you do. You define your clusters, you go through them, you typically look at marker genes, and then you assign that cluster a particular identity. There's been a lot of work in the field about how you visualize such data. So heat maps like this are useful, but they only capture a small amount of the data. And maybe in particular, they don't tell you about relationships between different types of cells. So there have been a variety of approaches visualized and developed to visualize these data in two dimensions. This is quite a popular one here on the right-hand side, so this is a t-SNE plot. You could think of this as effectively equivalent or trying to do something similar to what you would get with the first two principal components. In essence, what you're trying to do is you're trying to take that high-dimensional expression space and you're trying to project it into 2D. So here in this plot, each point corresponds to a cell., and we've just laid them out in two dimensions. The difference from PCA is that here you're trying to group cells by similarity rather than trying to find the axis that explains the most variation. That's essentially the difference between a t-SNE and a PCA. But what you can see here are data from the earliest time point, then we go through the mesoderm, and then up here we move towards the more terminally differentiated lineages with blood progenitor cells and then the early erythrocytes here in this brown population. In this region of the plot. So this is an atlas. How can we begin to interrogate it? So we decided to interrogate its function in the context of just a single gene here, Tal1. I mentioned before that this is a marker of early blood cells. Indeed, it's an essential gene in blood development. Embryos lacking Tal1 die around E9, and they don't develop blood, so it's critically important for development. And the hypothesis when we performed this work around 2015 was that Tal1 acts as a gatekeeping transcription factor. So you have this precursor prehemogenic endothelial population of cells, and then in the presence of Tal1 in the wild type, these cells will commit towards definitive blood cells, so the earliest blood cells in the embryo, whilst in its absence, these cells will instead commit in a stepwise way to cardiomyocytes, so they'll move towards a heart phenotype. So we reasoned that we can maybe interrogate the function of TAL1 in a little bit more of a detailed way because these— this function had been determined using lineage tracing assays where you mutate the gene, you see what happens to the cells in which it's mutated, and inevitably you're looking at its function a couple of days or at least 24 hours after the deletion, after the time point when the decision is being made is actually taken. Single-cell genomics offers a way of resolving this problem, so what you can do is you can capture cells from wild-type and null embryos at the relevant developmental time point, profile their cells individually, and then map them onto the reference. And this is exactly what we did. So we generated new wild-type and knockout data that were enriched for cells going towards a mesodermal or blood lineage. And so we did a wild-type and then a Tal1 knockout, so this is a homozygous null. And then we can computationally map these cells onto our reference set, so we can build a random forest classifier, and then map them to the different cell types that we have in our wild-type atlas. So we did this, and I would focus you particularly on this prehemogenic endothelial, this precursor population up at the top in red here, and the mapping is shown at the bottom. So this is our reference atlas with different cell types with different colors as before in the clustering plot, and at the bottom here, this is now the mapping of the additional wild-type and mutant data onto this reference., where the purple triangles here correspond to wild-type cells and the turquoise triangles correspond to mutant cells. And so you can see in this sort of proof of principle that things are working. In this region here, this yellow population is blood progenitors. We're only getting wild-type cells mapping to this region of the plot. We don't get any mutant cells. That's what you'd expect. These TAL1 mutant embryos die because they can't develop blood, so this makes sense. More intriguingly though, we were able to map both wild-type and mutant cells to this progenitor, this earlier progenitor population in red here. And then what we can do is we can compare the expression profiles of these cells between the wild-type and the knockout. And what you might expect happened, that in the knockout cells there was a depletion of expression of blood-related genes. But more intriguingly, and particularly in the light of the cardiomyocyte work that had been described previously, there was no upregulation of cardiomyocyte genes in this mutant population. And so that allowed us to refine the model of how TAL1 functioned, that instead of in the TAL1 knockout these cells committing immediately to a cardiomyocyte fate, we reasoned that they effectively get stuck. They don't immediately commit to a second fate. Instead, they're just a little bit confused, and they're unable to decide at that particular point of development what they're going to do. And we reasoned that perhaps secondary induction events, perhaps signaling later on in development might push these cells towards a cardiomyocyte fate. So this was the model we had around 2016, but this experiment actually had a number of problems. Perhaps most obviously, we selected a small number of cells at each developmental time point. There was a relative lack of granularity, so the space between the sequential time points here was moderately large. And perhaps most critically in terms of understanding the function of the particular gene Tal1, these embryos are very sick. So if you're going to knock out a gene that's developmentally essential, then of course the embryos are going to die, but lots of other things are going wrong. So is the function you're seeing of Tal1 actually attributable to its cell-autonomous intrinsic function, or is the lack of ability to develop blood, the lack of upregulation of cardiomyocyte genes, just a function of the fact that the whole embryo is completely screwed up and that the the signaling environment is messed up, there's lots of organs not forming properly, so it's really hard to disentangle the cell autonomous function of a gene from just the environmental defects. So we wanted to address both of these problems. The first is relatively easy. Technology— this is a technology session— has moved on tremendously in the past 5 years, and we're now able to profile tens of thousands to hundreds of thousands of cells very routinely at not terribly high cost., so we could scale up this experiment fairly straightforwardly. So instead now of looking at 1,200 cells at 4 time points, we can now look at well over 100,000 cells, indicated in this little table here, at 9 sequential time points. So we're sampling embryos every 6 hours through development. So this is now quite a fine-grained time course of development. We can use the system here. We used a droplet-based system developed commercially by 10x in conjunction with sequencing with Illumina. In order to profile all of these cells throughout this developmental time course. Of course, this does yield important computational challenges. My lab's a computational lab, and one particularly important thing that we were nervous about here was batch effects. So, just due to the way that the data are sampled, we cannot randomize cells from different time points together and thus control for batch effects with a nice experimental design and regress it out downstream. You've got to sample these embryos at the specific time points, profile them immediately to ensure the cells are of high quality and you get good quality data. But that does mean that you have massive batch effects. So how are we going to combine these different time points? And indeed, this is a more generic problem. Consider something like the Human Cell Atlas or any single-cell study. Suppose that we are sampling livers from everybody in this room and that we want to try and combine the data together, right? Well, just depending upon where you sample, the bit of my liver or Martin's liver, you're going to sample slightly different regions. You might have many of the cell types being the same, but depending upon the region you sample from, there's going to be differences in the composition. You might have 75% hepatocytes in the sample from me, 70% in the sample from Martin, and you really want to correct for that. The power of these atlases comes in combining data. So you might think this isn't a problem. Technology developed for bulk RNA-seq can be applied here. Actually, it doesn't work so well, and it doesn't work so well for a very simple reason. So this is a very simple simulation. We've got two experiments. So we've sampled, say, my liver, Martin's liver, batch 1, batch 2, and there's three cell types that we've identified here— cell types 1, 2, and 3, red, green, and blue. And the only difference between these two samples is the relative frequency of the cell types. So you might have 5% of cell type 1 in batch 2, 10% in batch 1, in batch 1. That might be the only difference. But then you think you could just apply things like SVA, surrogate variable analysis, or other batch correction approaches that essentially are trying to find principal components in the high-dimensional space that are orthogonal to the biology of interest and regress them out. This doesn't work. So, this is the batch correction that's determined by these bulk RNA-seq-based strategies, and you can see that ideally you'd have the all of the cell types coming together across batches, and this simply isn't the case here. And the reason why this isn't the case is purely composition effects. We can show that when we have a batch effect, yet the composition of cell types is the same across the different batches, that these bulk-based approaches perform very well. But when you have differences in composition, they really don't seem to perform very well at all. And so this motivated my postdoc Lally to come up with an alternative approach, came out towards the middle of last year, and it's very simple but I think very effective in practice. So we've got two batches as before, and we've got three populations in the first batch— red, blue, and purple— and in the second batch we've got two populations that are the same, red and blue, and this additional gold population here, this population Z. And what Lalli reasoned was that she could take each cell in the first batch and in the high-dimensional space look at what cell cells it was closest to in the second batch. So this cell here might be closest in that high-dimensional space to these cells around here. And then the thing that Chi had the insight was that we could do this in the opposite direction now. So we can take each cell in batch 2 and ask what cell is it closest to in the first batch. And if cells are mutually closest to one another, we define them as pairs of mutually nearest neighbors. And to give you an example of cells that are not mutually nearest neighbors. Consider this cell here, this purple cell. This is present in a population that's only present in batch 1. In the high-dimensional space, it's closest, for example, to this sort of cell here in batch 2, but the same is not mutually true. The same is not true going in this direction, so these are not pairs of mutually nearest neighbors. But after having identified these pairs of M and Ns, what you can do is you can take them you can compute a batch correction vector indicated by these solid vertical arrows here, and this is currently smooth so that it applies across the whole of the space. And effectively then what you want to do is you want to map these two different batches onto the same manifold, and that's what's shown here. So effectively you take the batch correction vector, you map the cells on top of one another from the different batches, and then you can iteratively add more and more batches as you go downstream. This works beautifully in simulated data. That's nice. More interestingly, perhaps it also works in real data. So this is 4 experiments, 4 studies of the pancreas by 4 independent labs with different single-cell RNA-seq technologies. They identified the cell types themselves. So the cell types are indicated by colors, the experiments by the symbols. And what you can see on the left-hand side here is the uncorrected data. So you have very substantial batch effects. Where depending upon the experiment you're getting cell types clustered by experiment. The bulk-based RNA-seq approaches don't do— seem to do a very good job here either because you're getting this sort of weird structure down at the bottom, whereas the strategy based upon peers and mutual nearest neighbors is doing a much cleaner job of mapping these cell types onto one another. And this work has since been taken and developed and integrated into other methods, so there's a paper coming out from Raul Satieja's lab where they combine this strategy with an approach based upon canonical correlation analysis that they developed that I think does a very nice job of beginning to remove some of these effects. But I would say that this is one of the major open challenges in single-cell RNA-seq analysis, how to combine across these batches. So I foresee quite a lot of additional work over the next few years. Anyway, having developed this approach, we could then develop— apply it to this single-cell RNA-seq data through mouse development. We can now get a much bigger tSNE plot, so This is now looking at over 100,000 cells, so instead of the 1,200 we had before. So again, each point corresponds to a cell here. We start from the epiblast, we go through to early blood, the endoderm, more neural tissues in the middle. The cell types are indicated on the right-hand side here. There's a huge amount of work goes into this annotation process, as this is very non-trivial in and of itself, but illustrates to you the sort of the complexity and the beauty in some ways of the data. A nicer way that I like to look at this data is via this ribbon plot. So here we have the 9 sequential time points from E6.5 through to E8.5 on the x-axis, and then on the y-axis is the relative frequency of different cell types at each time point. And you can see you move from a fairly homogeneous set of cell types at the earliest time point where most of the cells come from the epiblast to this beautiful increase in diversity by the time you get to 8.5, where really many of the cell types that are going to populate the adult animal are already present. And what you can do is you can take any one of these lineages, you can track how expression changes as you move along it, work out the order in which transcription factors are expressed, and you can really see how you can begin to use this data to get deeper insights that have been possible using more targeted approaches at early development. You can then take any one of these populations, you can divide them additionally into further subclusters. So here's the endothelial population divided into additional 20 or so clusters. So you could almost always find additional structure here. By so doing— and this is a different representation, this is now a UMAP from progenitor through to early blood cells here— you could identify new progenitor populations. We identified at these stages some new and quite rare immune types that were present. But what we were hoping to do was to expand our understanding of this transcription factor Tal1, and I mentioned to you before that the knockouts have some challenges. And so we decided that we use an alternative approach that didn't lead to such destructive embryos as the homozygous nulls to understand its function, and the approach we took was chimeras. So what are chimeras? Well, you take embryonic stem cells from the same genetic background as the wild-type mouse, you endogenously label them with a fluorescent marker, here tdTomato, You mutate using CRISPR the gene that you're interested in, here, here, TAL1. And what you do is you then inject these cells at the blastocyst stage into wild-type embryos so that you get these chimeras that are composed of both a mix of the injected and of the wild-type cells. These embryos then develop properly, so they develop blood, but critically, they're a combination of wild-type and mutant cells. And given that they're fluorescently labeled at the developmental time point you care about, you can take these embryos, you can dissociate them, and then profile the two populations separately. So this is what we did here. So this is at embryonic day 8.5. On the right-hand side are the wild-type cells from the chimera colored by the cell type labels in our reference atlas, and on the left are the mutant cells from the chimera. You can immediately see it doesn't take great statistical rigor to work out what's different between these plots. You've got this big, big red block lob here in the wild-type cells. These are blood cells. You're only getting blood cells from the wild-type M cells within the chimera. You don't get anything within the mutant. And this is as one might expect. But critically, all of the other cell types seem to be present in roughly the same proportions between the wild-type and the mutant cells. You can then take these cells and work out what's going wrong in development of blood specifically. So this is now the early erythroid development. So we've got the blood cells here on the top right. We've mapped our wild-type and mutant cells onto this reference, and you can see here we've got our wild-type cells going up towards blood here on the top right. And this is now the mutant cells, that Lactal1 being mapped to the same reference embryo. And you can see that they're stopping around here. So this allows you to pinpoint the precise time point in development where these cells stop progressing, where you need Tal1 endogenously within a cell that acts in a cell-autonomous way to push on towards blood. And so you can then say, well, Tal1 is necessary for the generation of these early immune marker— immune cell populations, sorry, that are present here. And then you can begin to work out what might be going wrong. So you could then take these cells here and ask, what are they actually doing? What's characterizing this population, this Tal1 mutant population that's showing markers of a progenitor population but not actually progressing towards blood more maturely? Well, we can ask whether this mutant population is enriched for a particular cell type relative to the reference cells. And what we actually observed was that this is indeed the case. So we had this set of subclusters that I mentioned, and we observed that the Tal1 mutants mutant cells were enriched in this endothelial cell C3 cluster here. So there's more of these mutant cells that are present at this cluster than wild-type cells, given what we might expect. And importantly, this EC3 cluster, when we compare it to our wild-type atlas, is present at a slightly earlier stage of development. So it's consistent with the fact that these mutant cells are a little bit slowed down in development. So kind of what we observed observed before, but quite satisfyingly we see at this stage at 8.5, these cells are subtly beginning to upregulate particular heart genes, but other mesodermal genes suggesting that at this point, a day or so after we'd done the study before, they are beginning to commit to different downstream lineages. And this allowed us to then further refine our model of how Tal1 acts in a cell-autonomous way during early development. Where in its absence cells don't immediately commit towards another state. So that was true, what we observed previously. They do seem to get blocked in an earlier endothelial precursor-like population, and only secondarily via these signals that we still don't quite know where they're coming from, but we speculate they could come from extraembryonic signaling, that these cells then will begin to slowly acquire an alternative cardiomyocyte-like fate or other mesodermal cell fate beginning at around embryonic day 8.5. And what we're doing now is we're taking— using other chimeras with transcription factors that are relevant all the way through early blood development, so starting with BRCA1, going all the way through to transcription factors that are necessary fairly late on in the process, to try and understand when their role is essential in the developmental process of developing chimeras for all of these different embryos. By so doing, we could then map these cells onto our wild-type reference atlas and really get a finer-grained understanding of what's going on in early development using this combination of single-cell omics and chimera, so intelligent genetic perturbations. I was going to tell you about multi-omics. I don't think I'm going to have time, so I will finish here by acknowledging the people who have worked on this study, and in particular I would acknowledge Blanca, Carolina, and particularly Johnny, my student, who really through a very nice team effort put this work together. There's a quite nice reference atlas, a Shiny app for exploring this if you're interested in these data, for thinking how they might inform your own studies. And as well as Bertie, who I mentioned earlier, Bertie Goodkins, there's also Wolf Reich, Jenny Nichols, and Ben Simons whose labs also contributed to the study. And with that, I'd like to finish, and thank you very much for your attention. Thank you.