Scientist Stories: Dana Pe'er, Computational Systems Biology

I'm going to try and be the opening act for a lot of the computational talks that will follow after the break and really try and give a very high-level abstract framing for it. So what do we need? Why do we come together as a community? So first, to— Are we having again this thing stuck and I can't move to the next slide? Well, I can talk without any slides, or maybe someone can— or is this the wrong— There's no clicker, okay. Okay, well, I can use the keyboard. So, as was described earlier today, the Human Cell Atlas will have multiple components, and here I'm not being exhaustive, but really describing the 3 main ones. So we're going to have data-driven cell types, largely driven by the high-throughput molecular profiling methods, such as single-cell RNA-seq, probably single-cell ATAC-seq will shortly follow, and really have our cell types. We're going to have spatial maps using some of these rapidly evolving multidimensional spatial technologies for both RNA and protein. And we're also going to organize these things in lineages and trajectories, both developmental— how cells develop down a developmental lineage— but also in terms of cell transitions and how cells transition from one cell state to another. And this is going to be a deep organization principle. Now, why do we come together? Why don't we all do it in each and everyone's lab? Because as a community, we can reach a scope doing the entire human. We can reach a scale—billions of cells—something that you need computational infrastructure that is certainly beyond what an academic lab can put together. Quality—by putting our minds together, we can really decide on best practices. And most important, compatibility and comparability. So what I collect in my lab and what each of you collect in each of your labs, and also outside in the world, can really be put in one reference frame, compared and learned., and that is the power of the atlas. So, you know, I'm going to be very simple. We can collect this data, even do a really good job, collect high-quality data, and just dump it in some data portal, but really the power of the atlas will be how it's used and how it's accessible. And one of the really great things about the Human Genome Project is way before it was completed, where it was still only sparsely populated, someone could BLAST and say, "Has the community found a gene similar to mine? Was a sequence like this found?" And we need to be able to BLAST cell types, both at the level of a cell type— here's a population, I want to see if this population has been found— and BLAST the cell mixture. Here's my set of cells, where do they map onto the atlas? And even something as simple as that is actually quite complex, and we have to actually think as a community, because we've been hand-waving what is a cell type, and I'll admit I will also hand-wave what is a cell type, but one of the things we have to come together as a community and define it. If we want algorithms, if we want a framework in which we can search, discuss, and compare, we first have to nail a more formal way of what is a cell type. What is a good similarity metric? What does it mean that my data is similar to something in the atlas? We have to define a biologically meaningful similarity matrix. We have to normalize the data. If the data is not properly normalized inside the atlas, we're not going to be able to make meaningful similarities. And finally, with billions of cells, we actually need to develop this search in a scalable manner so the PhD student will be able to graduate before the search is done. So back from the abstract to the very simple normalization. Normalization is a huge but critical challenge, and one of the things I'm going to try and do in this talk, in the community spirit, is I'm only putting one slide from my own lab. This is data from John's lab, and it really demonstrates the importance of normalization. Here are multiple pancreas samples, and they are very different. Multiple cell types that are supposed to be the same ones, colored similarly, they're actually very different in their data. A large part of these differences are actually not the technology, but many of the things that Pam so nicely described, of what happens until the sample hits the lysis buffer, as well as additional technical things once you run the protocol. And once you normalize the data— and this is John's beautiful normalization algorithm— you can realign these cell types together. So things that we know are the same cell type, using our prior knowledge, actually map to the same place. This is based on a k-nearest neighbor method. Which I'll talk to you in a moment, really powerful method to actually capture the meaningful similarities. But John's method aside, as a community, we're going to have to be able to normalize this atlas. And it's actually tremendous, complicated technical challenges which we're going to have to do together. So I'm going to try and talk a little bit— what is a cell type? I'm certainly not going to give you the answer, but I hope we'll engage with that in this meeting, because I really think that that is something that we will have to resolve. I don't think we'll have a final answer until the first draft atlas, where we'll actually use the data itself to begin to understand what a cell type is. But when you collect single-cell data, no two cells are the same. Every cell has a different expression profile. But which ones are meaningful? How do we organize them? And there's been some distinction here between cell type and cell state, and one of them is more static, the other is more transient. I'm actually not going to try and make this distinctions. So when I talk cell type, I mean cell state, because I think we're going to have to define an entire new language. And there's also continuum. So we really have to redefine, based on this data, what do we mean in the Human Cell Atlas when we talk about cell type. We're going to have to drive our definitions by the data we collect, but we'll have to do this in close collaboration and communications with biologists so that these definitions will actually have meaning and impact. So I'm going to make a first crack at some foundations that will help us approach this. In my approach, each cell is a vector in high-dimensional space, and it has features. It has whatever features we manage to measure, whether it's single-cell RNA-seq, whether it's ATAC-seq, whether it's from imaging, where we could get both morphological features, spatial features of where it's located, who its neighbors are, and what it's expressing in the multimodal expression assays. And for each cell, whatever we manage to measure for that cell— and this will evolve, because the technologists are evolving this— we have a vector. And this vector is a point in a high-dimensional space, and once we have many cells, as we will have in the atlas, we actually begin to see these cell densities. One of the advantages of doing this in sort of this mathematical, vectorized way is that you can actually integrate all these different modalities and have one big vector for all the possible measurements one can think of, which can expand with time. Now, as we started looking at such data, we've seen that this data doesn't just randomly diffuse across the high-dimensional space, but cells tend to accumulate in densities, these low-dimensional manifolds. This is the work with Gary, the first Seithoff paper, and you really see that the cells actually form very distinct low-dimensional structures. Neighbor graphs allow us to actually build a mesh to capture and model the structure of this space, and it's, you know, been one of my favorite ways to approach these. And really, there's a lot of tools and math from diffusion maps, from spectral theory, that can really help us now engage with these objects that are neighbor graphs of cells. Generally, why neighbor graphs? Because this actual cell space resides on a lower-dimensional geometry. When we try and look at regular Euclidean space, you can have cells that this— that metric doesn't really capture, and you can have things that are actually rather far away from each other seem close. And by taking small steps in the graph structure between very similar steps, we can actually remain inside the cell space. We could actually remain within the manifold. And so once we have the cells mapped into this graph structure, you can see two components. You can see these densities, which are represented as densely connected components. These are stable cell states, whether they are actually transient in the terms that they can go back and forth. They're stable in the fact that the cells remain there, and there's enough cells there. And there's the transitions, the connectivity between these densities, which either mean transient states that are developmental, how the cells mature, as well as transient steps, states such as cell cycle, how cells cycle from one state to another. And in order to do this well, in order to build this graph well, we really need good feature selection, good similarity metrics, good graph construction, These are abstract concepts, but the devil is in the details, details that we should decide together as a community to have one atlas. And I really want to bring an example of why feature selection matters, and again, I'm going to bring very early work by 4 of my colleagues here in the audience. And this is why self-feature matters. One of my passions is trajectory methods, particularly in the developmental context, and some people write to me and say, "Hey, it doesn't work. I got crap answer by running your algorithm." And the thing is that it matters the feature. Here is a bunch of cells, stem cells, that are undergoing development, and if you just look naively at the cell profiles, you get them completely out of order. And what John, Ollie and Sarah did is they actually tried to decompose the two continuous factors that were at play here. The differentiation process, the one that they were more interested in in this process, and the cell cycle process, which is not noise, it's not artifact, it's a different type of biology, which was less relevant to the questions they were asking. They were doing this using parametric approaches, linear mixed models. You can do this either by parametric approaches or nonparametric approaches, such as those graph structures I showed earlier. The point is that we're actually going to have to decompose and really define features that matter, similarity metrics that matter, and these are going to be dependent on the biological question that we're going to ask. Another group actually might be interested in the cell cycle, but to be able to decompose is the first step. And one of the really exciting things now, as we're trying to build these pseudo-time developmental trajectories, is by having many of these now new genetic approaches to do tracing and tracking, of what actually happened, we can actually have models by which we can test our algorithms so that we can have best practice algorithms, as well as get a prior from model organisms on the human atlas where we're not going to be able to put a lineage tracer in an embryo, at least not ethically. So again, I'm going to try and dig deeper on some of my random thoughts, and this is not clear final thoughts. This is just thoughts for you, and I'm hoping some of these things will crystallize in a meeting. When I'm saying what is a cell type, we have all these measurements, but underneath these measurements, I do think that there's an abstract real state— state that we're trying— cell state that we're trying to get at. And what defines it? Both the past of that cell, its lineage, the environmental exposures that it has seen, and what matters to us. The future. What will it do? What can it do? How will it respond? What is its potential other cell states that it can reach? And this is what we actually want to capture in our unit of cell states, in the new language that we have yet developed together as a community. And one of the advantages of this is that if we can understand this, if we can get a predictive model from looking at a profile or defined cell state now to what a cell could do, to its function, we have gained not only a good model for our atlas, but true understanding. So, as I said, I really believe that cell state connects to past and future, and this is where epigenetics can play a really big role, because epigenetics really does code past history and future potential. Now, this is a slide by my colleagues at Stanford, and here, of course, they did a big cheat and showed how gene expression poorly captures cell state and ataxia captures cell state well. They, of course, used the silliest clustering algorithms for gene expression and didn't optimize it, because they were trying to make a point, and I actually am a big believer in gene expression and think we can do a much better job, But it does go to show that you can't use gene expression naively, and I do think that there is true value encoded in the epigenetic profiles that we should collect for the atlas. And again, I can't reiterate enough, it's not enough just to look at these as vectors. This is why we have to come together as a community, work with the biologists. Feature selection similarity metric is the difference between success and failure in this atlas. I also love imputation. This is the one slide from my lab together with Smita, who has her own lab, and David. And here is the power of imputation. We have a bunch of different cells. We have a lot of dropout. This is Ido Amit's data, who's here. And once we impute the data, we can actually recover the missing values. And this thing is not working. Lovely. Well, you would have seen a beautiful reconstruction of everything recovered had this been my computer and not another computer. The point is, though, that using manifold learning, we can actually impute quite a bit. We can impute not only within single-cell RNA-seq, as I sort of showed here how the animation works, but also across data modalities. And here I stole Gary's slides, as you can notice, but in the speaker before me, I really think that space is a critical component of the atlas. When we first got together in London, we were sort of driven by single-cell RNA-seq, and we were like, yeah, space is important, uh, we certainly think so, but the spatial technologies, they're not quite there. But by bringing it and coming together as a community, we realized the spatial technologies are actually much further along than we thought and are actually progressing at such a rapid pace that we can incorporate them big time into the atlas from the very beginning. And basically, a cell resides in this multi-resolution environment— tissue region, microenvironment, all the way down to its immediate neighbors— and we have to actually begin looking at that as a big component, and as Gary so eloquently showed, understanding that using some— here's just a few technologies that really allow us to begin gauging it at a high-dimensional level. And again, this actually adds another layer of abstract definitions that we're actually going to have to understand. Basically, a key part of the cell's identity is who its neighbors are, and if we go back to what the cell might do, we can get two things from the neighborhood. One, that neighborhood might encode things which we hadn't measured, which if we would have comprehensively profiled everything inside the cell, we wouldn't need to look at its neighbors. And two, because a cell actually interacts with its neighbors, it actually listens to its neighbors when it's responding, if we want to know how a cell will respond, its neighbors have critical input into that thing. So we really need to define a cell in its context in order to understand how it will respond, in order to understand this cell type will do this when given this input. So the initial spatial data, as Gary showed, has clear spatial patterns at multiple resolutions, and we're going to have to develop an entire new formalism, an entire new language on top of molecular profiles and cell types to actually begin talking about these entities that we're trying to build in this atlas. An atlas will never work without a language, without a formalism. We're computational and mathematical scientists, so we'd better get from hand-wavy things, as I'm doing now, to concrete concepts. We will need the data in order to do that. We should discuss it these coming two days, but we won't get there. Spatial mapping allows us to integrate. We can actually put the data, overlay our data in the atlas. Here's work by Raoul when he was in Aviv's lab of actually taking these molecular profiles and mapping them spatially to actually this really low-dimensional imaging analysis. And now we can take these full genomic profiles, map them into spatial coordinates, and have powerful data integration that allows us to look at new cell types, new patterns of gene expression, and networks both within and between cells, given the entire genome profile now mapped onto a spatial coordinate. And this really brings us to the fact that we, as computational scientists— yeah, I'm almost done— as computational scientists, are going to have to map these things inside and iteratively decide the experiments. We're going to have to decide what are the best markers to map in the spatial resolution, because right now it is limited to 50, 100, 300 markers. We're going to have to decide what are the assays, the cell types that we're going to do to maximize our information, imputing potentially missing values, and I really believe that more cells and less depth will be the foundation of the atlas, and particularly because you have so much information to learn across cells. And critically, the computational biologists should work hand in hand with the experimental biologists in designing the data acquisition. Another case study. Again, a disease scenario more complex than "What is my cells like?" Here, we need a multilayer answer. If you want to compare, here's my disease cell, I want to use the atlas to query my disease cell. How does expression differ between matched cell types in the disease and non-disease? Again, we have to define cell types, and then match cell types, and what matters between the disease and not cell state. How does the composition of the cells differ, both new cell types, as Pam so beautifully showed in the context of immunotherapy, as well as differences in their proportions, and how does the tissue architecture change, as Gary so beautifully showed right here. The disease will actually, even though it's not part of the atlas, actually give us an input into understanding our atlas, because when we see things that are not normal, when we see things that are perturbed in disease, disease as a perturbation, we can go back and understand the functional implications. So we have a long and exciting road ahead of us. And we're going to have to do this together. We're going to have to bond together as computational biologists to solve together these big practices and decide what are the best practices. And we're going to have to work together with the experimental biologists side by side so what we do is meaningful biologically and helpful for them. And one final thing about this community effort— and I understood that I, because I was swapped, I'm going to talk about it— and that is the Jamboree Hackathon. So as part of this computational community effort, there's going to be a hackathon where we're going to begin to solve some of these concrete computational problems. There's been 12 computational groups invited, all 12 accepted. It's going to be graduate students and postdocs mixed across labs trying to share our knowledge, our insight, and our information and brainstorm together to actually come up with best solutions for things such as normalization, clustering and handling of this data. The datasets that we derive from this Jamboree, the metrics that we decide and the solutions we've achieved will be posted online into a living Jamboree where anyone around the world can continue to test against these benchmarks and these datasets in an ever-improving computational algorithmic space. Let's do this together.