Scientist Stories: Peter Kharchenko, Beyond The Cell Atlas

[Speaker] First of all, I want to thank the organizers for inviting me to speak. Let me see. I'll probably stand a little bit to the side so I can see the slides. Yeah, so I'll talk about our efforts to do integration, and mostly in the context of comparative studies that I imagine a lot of people are pursuing and clinical comparisons, so like control versus, let's say, disease cases. And we're focusing on the single-cell transcriptomic data. In general, this is actually single-cell site data, but just kind of to illustrate, I think, the power of these single-cell approaches to look into complex tissues and trace different sort of parts of it, different subpopulations, maybe in dynamics or in static context. And so now these techniques are more mature and a lot of folks are using them for different purposes and there's kinda like two main classes I think that can be easily separated. There's this ongoing, very active tissue mapping efforts at CGI and beyond. Human Cell Atlas is a big collection. HubMap is sort of the NIH effort to get into this kind of mapping initiative, as well as brain. And that's great. They're very active in scaling up. So the other type of efforts are essentially comparative, and a lot of it happens in the context of disease. And so H10, I think, that NIH has started is looking at a human tumor atlas sort of thing. The Lifetime Project that is starting up in Europe that would also longitudinally look at the impact of disease on these tissues, so using some of the same kind of single-cell resolution techniques, but to start to look for the types of perturbations that you see in disease, for instance. And when looking at this type of data, I think initially at least, at first people are trying to figure out the targets to analyze, to pursue, to study in further detail. And a target, the very simple case, can be a cell type, cell population, let's say, that's most affected by the disease. Maybe mediating some of the phenotypes. And obviously since you have deep molecular data on this, you can then go in and try to figure out what specific pathways were perturbed in these cells and then try to intervene somehow, for instance. Yeah, so that's the general logic. And so within this context, I'll show you an example of a study we've been pursuing and the types of tools we had to develop for it, and then kind of discuss where this all leads. And for us, there's still a lot of work to do. Yeah, so in these types of comparative studies, you basically have more complex designs. And particularly in disease, you would have, obviously, usually healthy controls or some sort of controls. So you have multiple patient cohorts. Quite often you want to look at multiple tissues. And if you're looking at longitudinal sampling or maybe effective treatment, then you also have multiple time points there as well. So the example is our study that we're still putting together on the impact of prostate origin metastasis on bone marrow. So just to give you kind of overview of design, so we have multiple patients. We're able to get material from spinal decompression surgeries. And so during the surgery, we can extract actually a sample of the solid tumor. So this is, you know, little t-SNE plots for multiple individuals that we had. But from same individuals, we could get a liquid aspirate of bone marrow as well at the site of the metastasis. So you get, you know, another set of samples that are actually paired between individuals. We can get liquid aspirate or bone marrow at the distal vertebrate site that is supposedly not to be affected by the metastasis. And we could get bone marrow from controls. In our case, controls are actually individuals who have undergone hip replacement surgery, so they still have inflammation, but it's not cancerous. We could get blood over time, obviously, and you could potentially look at other tissues as well. It ends up being kind of this very large collection, and in general, I think the initial questions that we ask, well, what are the compositional changes linked to the presence of the disease, and what is the impact of the disease on the state of individual cell types? So let's say a particular T cell subtype. For the effect of treatment, it's also kind of a comparative study. It might have a temporal aspect to it, so it might get a little bit more complicated, but In general, for effective treatment, you need more samples. We're not there yet. So for these types of questions, to answer these types of questions, the first thing we needed is a way of tracking a particular cell type through this collection. And now, collection can be quite heterogeneous, particularly if you're looking at sort of severe disease cases. It can lead to distortions of tissue composition, appearance of new cell types, disappearance of others. And so there's no guarantee that the same cell type will be actually present in all the samples. And so we wanted to have a kind of a way of tracking a cell type consistently through these potentially heterogeneous collections. And if you simply put together these different samples from different individuals from different tissues and so on, they tend to kind of roughly separate by by patients, so there's a lot of inter-individual variation, sometimes by technical batch. And so what we want is some kind of a more aligned or mixed version like this. And so that's what we've set out to do initially. And so the approach we've developed called CONUS is based on these kind of error-prone pairwise alignments. So we start by looking at a pair of samples and then using some kind of rotation that puts them in approximately informative axis. So it's basically a reduction to either consensus principal component analysis or some generalized SVD and so forth. We've recently added CCA as well that Sera has been using. That's quite useful, I think, for particularly distant alignments like evolutionary alignments and so on. In any case, you could even use— for the most part, we use very simple reductions. We actually use principal component analysis to get the eigenvectors of individual cell types and then individual samples and then project them both to a joint space. And then we establish some kind of a mapping, initial curative mapping between them. We started using these nearest neighbors, and then when John Marioni's group has published their mutual nearest neighbors, we switched to that. So it gives you a period of mapping, let's say, between these red cell types. But in any two pairs of samples, that's likely to be error-prone, so the idea is to do multiple comparisons across all pairs. So if you have a large collection, you have plenty of pairs to work with. And in essence, if this red cell population represented a recurrent cell population within the collection, it would continuously tend to reflect on itself, forming kind of a clique in this large graph. That we can recognize easily using a variety of algorithms, essentially community detection algorithms. So in the end, what we have is this essentially very large graph connecting all the cells. It's a weighted graph collecting all the cells in the collection. We do keep actually some low-weight local edges as sort of a week prior for the cluster seen in only one sample. But in general, this graph allows us to link all the cells and we can do a variety of things with it. So first, we could use it for visualization, and it's essentially just projecting the graph into two dimensions. We could use it for clustering to get these groups that would track across the collection, and we also use it to propagate information between samples and so forth. All right, quick example. So this is looking at bone marrow and core blood controls from Human Cell Atlas. This is the actual projection of the joint graph. This is how different samples mix in there. And this is individual samples basically projected onto this common embedding. And the bone marrow, as you can tell, is consistently similar but consistently different from, let's say, cord blood, right? You have a lot more hematopoiesis actually going on in the bone marrow, at least visible. Now, these are two different tissues, but imagine that there were actually conditions that you're trying to compare, and then visually you can even tell how this type of alignment allows you to to start doing comparisons to see where things are changing. We're looking for some properties of this approach in general for the integration approaches, and particularly we hope that we can get a sort of increasing resolution as the size of the collection increases as well. So it's a little bit challenging because, again, of inter-individual variation, which is nicely hidden here, but if you're sampling folks who are let's say, samples that have differences from one another, you're kind of learning more about population variation, but despite that, you want to be able to distinguish finer and finer, let's say, subsets of cells, let's say T cells, if they recur in the population. All right, oh yeah, since I'm here, I have to show some examples of more kind of diverse stuff. So for instance, the most diverse kind of collection we could think of would be these organism atlases, and this is from telemers from CZI. And so some of the datasets have almost no overlap between themselves, and so even though we reflect them on each other, we're probably not gaining much information from that. We actually went ahead and combined it with another mouse atlas from Gage Goua's group, so it's using a micro-well technique that gives sort of shallow coverage but more cells. And just kind of put it all together. So this is coloring of the joint graph by platform, and blue is micro-well, and then you have 10x and SmartSeq2. In the places where they should align, they do align. So let's say, like, hematopoiesis is nicely laid out. You have progenitors here, neutrophils, let's say B-cell branch, macrophage branch, and so forth. But even from the coloring, you can tell that a lot of things do not align. Line, right? And looking at this carefully, we can tell that, of course, that tissue sampled in one dataset but not another. So despite these tremendous efforts, and it's obviously very valuable data, I think this type of overlap test shows that there's a lot more mouse to sample, essentially. And I know CGI is actually doing a much deeper atlas as well. You can integrate across modalities and Here I'm just showing integration, let's say, of Chromatin Accessibility long data with RNA data from the atlases. And then this is just the RNA view, accessibility view, and this platform view of this integration on the same embedding, essentially. It integrates fine. I mean, you could get reasonable resolution like T cells and K cells. You could tell that the, let's say, epigenetic still, datasets still clustered together a little bit closer than than the expression ones. So in general, these types of integration methods, they don't actually take any kind of instruction currently other than, you know, kind of a magic word to integrate. So the strength of integration and the factors that the integration overcomes, they're poorly defined at this point. And so it's quite natural for, you know, let's say for samples of different modality to still be a little bit closer to each other than than the other modality, right? All right. So the other thing, sort of a word of caution, basically all these cross-modalities comparison depend on linking measures that somehow bring together, let's say in this case accessibility and expression. It can be a latent variable, it can be a translation from one to another, but the quality of that linking measure is essential, right? So if you try to use simpler and simpler linking measures at some point, The integration will break down once the measure stops being informative. And it's very hard to actually know in advance whether you're doing a good job or not. Not a lot of benchmarks here. All right, so back to the comparative studies. So this is kind of an overview of the samples we had from the bone marrow, and this is the integration for this prostate metastasis study. You see hematopoiesis here nicely. These are the actual tumor cells. T cells mature outside of the bone marrow, so they look like a separate cluster. You have a little bit of a stroma here, but mostly immune compartment. Okay, so there are some things you can do, right? And in essence, we split the problem into two aspects that are kind of two sides of the same coin. One is compositional shifts in terms of population frequencies, and the other one will be transcriptional shifts, right, controlling for the compositional difference. So here, just visually, you could easily, for instance, plot cells that are in benign in blue and in all cancer fractions here in gray, and you see that, I would say, cancer fractions are much lower B-cell progenitors and mature B-cell compartments, so it appears to be impacted. At the same time, they have higher, let's say, cytotoxic T-cell and K-cell frequencies, right? So even these types of visual comparisons here, let's say, between the solid tumor and the the liquid bone marrow that we get from the cancer patients. Even visually, it can be reasonably well informative. But of course, we want to do statistics, and the question is, do these types of measurements, at least to the current scales, do they allow you to do that? So the simplest thing you can do, obviously, is to look at the fraction of, let's say, total cells, how that fraction changes, in our case, across different samples. And so, for instance, For the B cells, right, you see in the benign controls, you see, you know, pretty high fraction of the overall sample represented by these B cells and very low in the cancer. Now, we only have, I think, 12 individuals we're screening here, cancer patients, and so it's a relatively modest dataset, but the magnitude of these shifts and their consistency between, you know, from one individual to another is actually such that we can get statistical significance across, you know, many of these cell types in terms of alterations. So obviously these are kind of low-hanging fruits, but they statistically works out reasonably well powered. Now, the other thing that I want to point out that there are many ways of actually looking at this fraction problem. And so the simplest second iteration would be to start looking at subsets. So for instance, here it's fraction of the entire dataset. But in fact, what you probably care about, let's say within the T cell compartment, how did the fractions change within myeloid compartment? So you could do these types of fraction analysis, at least within the cell types recursively going down the tree. The other thing is centering, so if let's say B cells are going up, just because of that, the T cells will come down a little bit in fraction. So there are a variety of approaches that allow you to kind of center it and try to guess which ones went up, which one went down. Which once went down. Given the relatively low number of populations we're actually considering here, it seems like most of these approaches are a little bit underdetermined. So you could actually shift that centering in a variety of ways. But these are kind of considerations to take into account when you're considering compositional shifts. So transcriptional shifts, probably a little bit more interesting. So here we're controlling for compositional variation, and we just want to know let's say for a given cell type, the red one, how far did it shift between conditions? How much was its transcriptional state impacted by, let's say, presence of metastasis? Was it in a consistent manner between different individuals and so forth? So we'll start with kind of an overview. We can do these types of, let's say, expression distance calculations for every cell type between a pair of conditions, and then we can add them up weighting by the size of the populations in general. So then we get kind of an expression distance between samples that controls essentially for the compositional differences. And this is another t-SNE plot except every point here is a sample, right? And we're just looking how they group together. So the benign controls, they group together very nicely. They're much more similar to each other than the cancer individuals. The other thing you see is that these same color kind of points, these liquid bone marrow, for instance, samples from the same individual, so they tend to group together. And then this is essentially inter-individual variation. So we could look at this somewhat quantitatively. So this is expression difference between different individuals for different types of samples. So in the benign controls, so even though these individuals have inflammation, still overall they seem to be quite close to each other in terms of expression states of the cells within them. And as we go sort of move slowly towards the site of the actual metastasis, the variation increases. Interestingly enough, we see this in other diseases as well, that you have substantial increase in amount of state variability as you move towards disease states. You could also ask, you know, which cells are most affected? So here we're just comparing tumor versus liquid bone marrow, and simply every point is a kind of a pair comparison that we've done. In this case, we can control for an individual variation because we have multiple fractions from the same individual, so every point is a— is a sample from a given individual. And here we simply rank cell types to see where the impact is greatest. Again, you could do statistics on this, you know, so like some changes are significant but many of them will not be. You could, in this case, we're doing it on a granularity of these subpopulations which we can interpret, you know, this is kind of the level of let's say clustering at which we decided to work. You could do this regionally by basically looking at neighborhoods, in this case also looking at tumor versus non-involved, trying to screen where the biggest impact occurs in terms of transcriptional shift. And that's somewhere around macrophages. Just quick dive into these subpopulations for this particular study. So we have, I'll say, this macrophage compartment that seems to have several monocyte subpopulations and these, what we will call, tumor-associated macrophages, tumor-inflammatory monocytes, so the strong macrophage signature for this type and monocyte for this one. This is kind of their quick characterization. The macrophages appear to be sort of M2 type and the monocytes show very strong inflammatory signatures. That's why we call them that way. But in just generic analysis, we tend to look at these types of I don't know how well they show up. Essentially density distributions to see how cells shift around, at least within this monocyte population. You can see that there's a huge kind of shift in the tumor population where these two TIM and TAM populations dominate, but also between, let's say, within liquid bone marrow, you have a shift from this kind of more resting monocytes into activated population here in the middle. The T cell compartment, again, we could dive in, get a more detailed clustering there, and again, start looking at these types of shifts. Essentially, there's a shift between the cytotoxic populations here with higher abundance of the CTL2, which we think is an exhausted cytotoxic population. In general, for that exhausted population, we see a nice kind of relationship between the exhaustion and cytotoxicity, so as exhaustion And every point is an individual here, so as exhaustion increases, the cytotoxicity decreases. And we can actually compare across populations. In this case, we could look, for instance, at relationship between abundance of these tumor-associated macrophages and the exhaustion of the T cell compartment in individuals, right? And again, you probably want more individuals, but even with this fairly small collection, we can get reasonable results here. Yeah, so if you have crosstalk between cell types, myeloid and T cells in this case, the key question is how they're actually talking to each other and through what channels. And that's particularly challenging to guess with the current techniques. You could look at the expression of, let's say, cytokines, ligands, receptors, and that's essentially kind of what we started doing. You could do ligand-receptor matching, which is pretty popular. For some signals, you know, the picture looks very simple. For other signals, it will look more complicated or even more complicated. The end result that if you just look at the sort of numbers of potential channels, you know, through which myeloid cells, you know, different myeloid populations could talk to different T cell populations, it's very high. This is like a very relaxed estimate of the number of channels where you just have expression of a ligand, expression of a receptor, which is greater than 0. You can obviously make it more stringent, let's say 1% of cells, 10% of cells. But in any case, if you have, let's say, between TIM and CTL1 you have 57 channels, and between TAM and CTL1 you have 60 channels, it's very hard to argue that somehow one is preferentially talking to T cells rather than the other. So I think the overrepresentation approach hasn't been very informative. Differential context can help, and particularly correlation can help. So in this case, across individuals we can look for correlation between, let's say, ligand and the receptor, which biologically doesn't have to happen, but there are cases like that that are pretty clear. And then finally, you could look at downstream responses as a way of guessing whether the crosstalk happens. So if, let's say, amount of ligand goes up, the downstream response in the T cell population goes up. This one is pretty challenging because you need to have a pretty good model about the downstream effect, so it's hard to screen dozens of channels that way. So that might be a potential approach. So for now, we're basically prioritizing these channels based on correlation, based on abundance change, and so forth. And then with a good consideration of literature and our capabilities, like what mice we have, try to pick the targets to go after. So it's still a manual process. At this point. All right, so quick summary. Basically, these types of comparative analysis require to align first the cells, you know, in your collection so you can compare apples with apples. And there are a variety of methods for doing this now, including the one we've developed. And then in terms of analysis, it seems like the initial view can be broken down easily into this kind of compositional variation, and then for just transcriptional state difference that can be analyzed using these different types of diagnostic plots. In this case, we seem to catch crosstalk between immune cells and then trying to figure out some ways of prioritizing particular channels that might be informative or might be involved in this type of, let's say, sustained, in this case failed immune response. And overall in this particular disease, we see very widespread effects. Most of the populations are affected in state and numbers, so it's a pretty systemic view. Finally, I wanted to thank all my collaborators, the PIs on this bone marrow metastasis effort, as well as folks in the lab who have been helping to develop all these tools. Thank you very much. Thank you, Peter. We have time for maybe two quick questions. I was wondering, you mentioned trying to harmonize like ATAC-seq data and RNA-seq data. And you talked about this linking measure, which was critical. How much of that linking measure is based on having a data set which is basically the same samples analyzed by both techniques, or not? [Speaker:JOSHUA ELKINGTON] I think what's best— and we actually tried to do this in the paper to match up— are the multimodal measurements, where you actually measure both in the same cell. So there's a public data set from Jay Shendure's lab. SkyCar, and also a paper just came out from our collaborators at San Diego called SnareSeq. And these allow you to basically have a barcode mapping, right? So you don't have to guess. And you can align them, forgetting about the barcode, and then check how well you actually align things. The evaluation is— I think this is like the only benchmark in this whole field that I actually believe, to be honest. But the measure of alignment is a little bit tricky because there'll be— I mean, we don't expect the exact same cell to find each other, especially in larger population. So you either have to kind of make it scored by clusters or come up with some kind of a distance measure. So try to do like a rough version of it. But I think these types of approaches would be very useful for both optimizing the linking measures and evaluating which ones work better and in general evaluating which sort of alignment algorithms do a good job. It's exciting research and nice talk. I'm wondering, there's many approaches now for harmonization, data integration. To my eye, not everybody's talking about what do I do with the gene expression measurements on the backend? I've introduced a sort of complicated transformation to them. Oh, yeah. But you still do differential expression. Yeah, I didn't have time to mention it, but basically we do not— in any way use transformed expression. In fact, we don't transform it throughout this entire process. The easy way of thinking about it is that when you're doing, let's say, differential expression, you're trying to figure out, you know, what's different, systematically different in disease versus controls, you very much need the variation that you've seen within disease, within control individuals. And if you have any kind of smoothing that try to equilibrate them either within group or between groups, you're essentially losing signal and inflating your p-values. So we always go back to the original expression values and we consider all kinds of secondary factors that should be taken into account. I don't know, age, platform, and so on. And the way we model it, it essentially looks like, you know, for a given cell type, now let's say T cells, we do kind of essentially in silico purification and then you end up with a setup where it looks like a set of bulk experiments on each side. So you have bulk measurements for, let's say, T cells for disease, for control. And you can use traditional tools like DESeq2 to set up a proper differential expression test.