Scientist Stories: Oliver Stegle, Genomics and Cellular Heterogeneity

So a lot of experimental considerations have already been discussed, but I just want to run through some of these concepts again and sort of, you know, how we could as a community think that the HCA could work and how everybody could come into play, in particular, obviously highlighting the computational aspects that underlie these sort of considerations. So we already heard about the sort of model that the HCA ties together, obviously computational analysis, but also the experimental expertise to really drive experimental efforts. And we have this vision that experimental efforts obviously go back into the ATLAS, and that future experiments will then be again guided by the new data and knowledge that we have acquired in the field collectively in the HTA and beyond. And to sort of fulfill that vision, we need sort of two examples or two aspects. One is computational methods that really guide experimental design choices. The second aspect is obviously standardization collaboration and best practices to really bring together the community efforts and work as a team between computational and experimental labs to make progress in this area. So here are a few considerations for the HCA. You've heard many of these thoughts before, but I think it's just sort of a good idea to go through them again. The one really point to make in the first place is we need to have designs that allow for accounting for different variation we do not care about. I'm talking about inter-individual variation, understanding how the differences between people, how robust are the inferences we're going to make in the HTA, how robustly do we identify cell types and cell states that really are observed not in one human, but really are generalizing across the human population, which I think is a very important aspect. There are tradeoffs between protocols that are high-throughput and can be applied to thousands and millions of cells and very deep analysis for deep dives and really understanding epigenomes and other aspects of the variation in great detail. And there will be interesting design choices to be made around maximizing bang for buck. We want to really get the most out for every dollar we spend as a function of cell types and different biological questions we would like to ask. And then another big important question related to this is when are we done? When have we understood the cell type? When have we understood variation to the extent that it satisfies our criteria and our standards that we collectively as a community want to set? And last not least, it's really how do we do this bootstrapping? How do we use the existing data to drive the next set of experiments? That's going to be an iterative process and an ongoing multi-year effort, so really the question is how do we do this? I'll try to, in the next few slides, highlight a few examples of work that we've been doing and others in the field to sort of try and make the first steps towards addressing these overarching aims that we have to attack. So the first example I want to talk about is interdigital variation and batch. That's a question that's very close to my heart, but I think it really is a general phenomenon and process we have to link in. It also ties the human satellites to human genetic variation and other efforts. And I want to start here with having these sort of three categorizations of what I mean actually by batch. I believe there are sort of 3 different types of batch. On the one hand, there's this simple batch we all know about. It's an experimental covariate we can adjust for, we regress it out, and we are all happy. We do this on a day-by-day basis. Then there's this sort of unknown batch, which is a factor that perturbs our experiments, but we don't really measure it. We don't know its state. We heard about the cell cycle, Dana has mentioned this, some of our old work. There are methods to sort of try and recapitulate or infer these variables from the data that we then can adjust for them in our analysis. And that is obviously an important work forward. But perhaps the toughest batch effect is if batch is confounded with the biology. If you do different experiments in different people, you can't regress out batch because you're actually taking out the signal that you're caring about. And that's actually a really important question. We heard about computational solutions of alignment from the Satija lab and also John Marioni. But there's another question which is of experimental designs. And again, I want to highlight there's other works by others that sort of tackle similar questions, and that's really sort of collectively shown As a community, we're starting to think about these problems, and we have to come together to find solutions. So I want to very briefly here show you one example of how we did that. Actually, that's work with Sarah's lab, where we looked at the reference populations of individuals. These are individuals whose genotypes we know. We have genotypes then, and we want to understand what is the variation between people. So we did a very simple proof-of-concept experiment and took 3 individuals whose genotypes we know. They're different, but they're known. We pooled them, And as a set, we profiled fibroblast cells from all of these individuals here into conditions— control condition and stimulus condition— in this case on a standard 10x run. Now, the question is obviously, how do you then identify which cell comes from which individual? And the approach we're using— and it's actually a very simple idea, but illustrates that we can combine, again, genetic principles and computational methods with experimental designs— is we use the relatedness of these individuals. We call variants and ask the question, Which cell or which donor is the cell we see closest to? This is the population of all the reference donors, and you can see if you pick the right donor, they're much more related. This is a relatedness score that basically says how similar are individuals on a genetic level. And that allows you to really recapitulate who is who. You can do this for one cell. You can do it for all cells. Here I'm showing cells by different sequencing depth, but the upshot really is, if you look at the background here in black, there's a very, very, very good separation between the alignment of donors to those groups of individuals you care about that are in your study and those which are not in the study. And that really allows you to resolve this sort of pooled experiment and disambiguate these very natural barcodes which we all carry in our genomes. Now, you can get more out of that. I just want to show you very two brief vignettes why do we care. The first is you actually find doublets. If individuals are 50-50 related, that suggests that two genomes in that cell, and it's not a single genome. That's very good to know so you can control for those and get that out of the way, which is a very interesting experimental component for our designs and really have good quality controls. The second is you can resolve these mixtures. So here, there's a t-SNE plot on control and stimulus condition. And this is showing you the same thing when I color the cells by donor. So now I know who is who in the cell space. You can see these top populations here mixes very well, the bottom one perhaps less so. We can also now start asking the questions, you know, how robust are these signatures we see? How robust are these cell states we see across different people in the human population? For example, we can break down the variation here, looking at how much variance for every gene is explained by donor, how much by condition, and maybe also other experimental covariates. And it gives you sort of an idea that actually inter-individual variation— we see this in many studies— is not to be neglected. It explains variability. We all are different. Genetic variation or epigenetic factors we carry along, they do drive differences. I think we need to understand how this variation contributes to the findings we have in the HCA. You can even go a step further. This is the same analysis when you don't know their genotypes. It's just basically clustering individuals.. But unless, you know, other PCA plots you may be familiar with, this is not based on transcriptomes. It's based on the genetic variants. So even if you don't know genetic variants a priori, if you don't genotype individuals, you can replicate who is who. You can see the 3 major clusters from 3 donors. There's some mixed population in between, but it's only 6% of the cells that you cannot assign. So you can really very simple, effective designs by pooling individuals together. So this was the first vignette I wanted to give, talking about individual variation. There's a very big, important question, which is next, is when are we done? When have we completed our goals? That depends on our objectives, but I just want to highlight here two pieces of work by colleagues. The first one is really from Rahul Satyendra, who has this very nice online app I'm encouraging you to look at that carries out sort of a very baseline power comparison analysis that asks the question, how many cells do you need to assay in order to identify at least a certain number of cells from each cell type? So in this case, asking, I want 5 cells from every of 10 cell types, and then you can sort of look at what is the number of cells you need to have some confidence and be able to do that. That's the first step, and these methods have limitations, but we have to drive them forward in order to really ask the questions we need to ask to plan our experiments appropriately. The second idea I want to highlight is an example from Dominik Grün together with Alexander von Lünenhardt, who looked at the question of outlying cells. So initially, what the study did is looked at sort of dimension reduction in clusters of cells and asked then, are there cells that are outlying, that don't fit to the clusters we do, that we then could follow up in a second step analysis experiment to really resolve cell types to very high resolution. Now, I just want to highlight you can do this two ways. You can do this to drive new experiments, but you can also use outliers to ask questions about, are we actually done? If you have outliers that suggest there are additional cell states— and these could be early measures to really get an idea of the complexity of cell types and cell states and really ask questions about how comprehensive are we in our profiling efforts and really drive experimental design choices, particularly if you're thinking about redoing experiments in tissues using perhaps samples we have kept and preserved in the next set of analysis. That leads me really to the final and third part of questions I want to highlight, and this is the sort of two-step approaches. We use single-cell transcriptomics primarily because that's cheap and high-throughput to get the first-pass view of how tissues look in humans and how cell states behave. But then there's the question, how can we go deeper? And I just want to really highlight here other multi-omics protocols. This is some of our work in this example where we looked at now methylation and transcriptomes. So you get sort of every cell's sort of two views, and these sort of protocols, they grow and develop. There's now ATAC-seq, we heard about this. There's also now early work with transcriptomes, accessibility, methylation, and that really ties together different aspects of the variability between cells, but these methods have a substantial disadvantage. They're cost-intensive and they're not high-throughput. So, we need to really think about which cells are worthy to assess and profile using such complicated procedures, and for which ones are we really fine with just using our single transcriptomes. And that's really an important question for experimental design. Going forward. So, I just want to end here with a sort of bird's-eye view of some idea of what's the throughput and how many cells here in rows can we profile using RNA-seq, spatial methods— we heard about those— proteomics, and then maybe here at the end, the epigenome variation datasets, which are perhaps least high throughput. And obviously, the scales and the rules about what's the throughput change day by day, so don't quote me on what is high and low throughput here, but the key point is here we need both experimental procedures, for example, preserved samples, but also computational methods to guide the cell selection. And then obviously the analysis of these sort of patchy datasets with lots of missing values. Not every cell will have every information, but several cells will have multiple information to be tied together and integrated. And with this, I just would like to sum up. I talked about some aspects of experimental design, in particular computational aspects and how computation could help us drive those. I mentioned this computational data-driven design principle. You heard this repeatedly today, and I think that's really at the heart of the HCA, that we don't want to do experiments and then look at them. We want to do this together as a team, and we really hope that this will create interactions between experimental and computational groups going forward, that we can decide where to look, how to look, how deep to go, and so forth. I mentioned these sort of pooled experiments, but I think it's quite a nice approach, and maybe a building block. There are many things that could solve— you can jointly disassociate cells and process them, you can even do joint cell culture and grow your cells in populations, and thereby overcome some of these problems we all know about. I mentioned power calculations, which I think is a really growing area for our computational efforts. We'll talk about this more tomorrow. And then the sort of two-step ideas. We first have a broad pass across many cells, and then the deep dives in specific areas of biology to really understand variation to the level we would like to understand them for particular questions. Okay, with this I take questions.

Thank you. Thank you, Ollie. Questions?

No one?

Oh.

Hello.

Hi, great talk. For the individuals, for when you had the 3 individuals, have you scaled out for how many individuals you think that is possible for, given read length and known amount of variation in the genome?

It's a very good question. I think it's important to remember that because every cell is tagged, and we know that every cell has, by definition, to come from any one of those individuals, we don't fall into regimes like, you know, typical on a seq read alignment. It's not like allele-specific expression calling, but rather you just need collectively across the genome have enough reads. We believe you can easily pool tens or even hundreds of individuals at quite shallow sequencing depth. So that's not the limiting factor. The limiting factor is rather experimental. Do you want to pool, you know, do you get the samples at the same time in the same experimental place? I think these are the factors that are really rate-limiting for these sort of approaches.

Any other questions? Okay, since there's a few more minutes left, I'll ask a question. I know that I pressed on you to cut out, be limited. What is the next most important thing you'd want to share with us about experimental design in a minute and a half?

Oh, wow. Thanks, Dana. That's a very nice question to ask. I think what's really, I think, for us important is is quantifying how easy is it to detect a cell and a cell type and a cell state. You know, if you think about experimental design, the question is, how likely is it that we're seeing a particular source of variation? Well, it depends on how many genes, how big are the gene sets, and how specific are they in this variation. And I think getting a handle on this would be very important to actually be able to do the right power calculations, because it obviously depends a lot. If two classes are very distinct, you don't need that many cells, reads, to disambiguate them. If there is more commonalities, then it will be much, much harder. And I think that's actually where we can really start looking at data and sort of making these questions, maybe, you know, improve these tools that we have for doing these sort of power calculations.

Okay, let's thank Ollie again.