Scientist Stories: Aviv Regev, Systems Biology

Yes. So thank you for the opportunity to speak today and to try and pitch to you the Human Cell Atlas. So why cells? Well, cells are our core units of each and every one of us, actually of any living organism on Earth. And our cells come in many different flavors that we like to classify into different types by different types of characteristics. There are function, their shape, more recently their molecular profiles. So based on this, for example, we might distinguish in the immune system dendritic cells that are first line of defense from T cells in adaptive immunity. We can distinguish neurons from glial cells like astrocytes and so on. Now, cells are also a key intermediate in between the genotype and the phenotypes that we care about so dearly. So, for example, a genetic variant manifesting in immune dendritic cells could increase the risk of Crohn's disease. Another genetic variant manifesting in skeletal muscle cells could end up in muscular dystrophy and so on. So it becomes completely obvious that if we actually want to functionally characterize variants and act upon them, we have to know which cells we're dealing with. So what would be incredibly handy is if we had some sort of catalog or atlas of all cells. But what exactly would be in this kind of atlas? Well, obviously the first thing that should be there is a listing of all of the different cell types. So we want to have our neurons and our different immune cells and so on. We also want to be able to distinguish different states for the same cell type. For example, an immune cell that is completely naive compared to the same immune cell type after it had encountered a bacterium. We want to be able to position all of these cells in their three-dimensional location. We want to be able to identify the salient features as these cells undergo dynamic transitions, for example, during differentiation or when they're activated. And finally, in an ideal world, we'll also be able to trace their entire lineage as people can easily do for the— I don't know— the worm, C. elegans. I think it's very fair to say that each and every one of these 5 aspects, our knowledge about it is limited. Some less, some more. It's also very rarely informed by genomic information. It's mostly done by pre-genomic tools. Some of it is, you know, years old. A lot of this knowledge is decades old and some of it is actually, you know, centuries old. But very recently, advances in single-cell genomics all of a sudden make a much more systematic catalog within reach. And this is because, first of all, we have new technologies, both molecular biology and the volume of sequencing that is needed in order to measure the transcriptomes of single cells with great precision. We also have sample preparation methods, including the ability to isolate cells on multi-well plates and handle them with liquid handlers and with microfluidics that lets us process a lot of cells in the right scale and for the right cost. And finally, we have rapidly maturing computational methods for everything from QC to cell type classification to identifying markers. And as a result, even though single-cell genomics is very young, it's already taking on very quickly across the broadest scientific community. And just based on numbers from the Broad, where I'm at, if in May 2012 we had 18 cells profiled, then by now we have profiled about 100,000 cells. Now very importantly, we're also making very quick advances in critical additional capabilities that will bring us the right scale and resolution. So for example, Hopefully this works. Yes. So, for example, using reverse emulsion droplets, which is what we see in this movie, we can handle cells very quickly at the rate of thousands per second. And with clever barcoding, we can bring down the costs of sample prep below 3 cents a cell. In addition, advances from many different groups are showing that we can measure many other variables besides RNA in single cells. And finally, other advances are able to use experimental methods to likely register cells to 2D or 3D positions. Now, these advances are already quickly leading to new insights. So here, for example, we're seeing profiles from 800, roughly 800 dendritic cells isolated from the blood of an Asian American 24-year-old female. The vast majority of these cells allow us to identify completely from scratch all of the 4 known subtypes of dendritic cells in the blood. But there's also a small minority population, under 2% or so, that is quite distinct and can be identified with a unique signature. These cells are only 2% of dendritic cells. They're only 0.06% of peripheral blood mononuclear cells. But they are validated readily in 10 of 10 independent individuals. And looking at these profiles, we can actually infer their position in the lineage and identify them as a likely new blood precursor of human dendritic cells. Now, this is one example of an insight around lineage. We can also identify new neuronal subtypes. We can find new states for well-known types like pathogenic T cells in models of autoimmunity, cells in glioblastoma with different states. We can monitor cells during dynamic transitions in very quick immune responses or in longer differentiation processes in adipocyte cells. And we can even, by coupling single-cell RNA-seq of these aggregated cells from basically a piece of tissue together with clever informatics— and again, let me try and run these for you— we can actually map these cells back— map these cells back to their 3-dimensional position, which is actually what we see in these rotating spheres that represent the zebrafish embryo at about a 10,000-cell stage. So this all suggests that the sequencer may very well become the microscope for the 21st century. And what I want to do in my other 5 minutes is to try and suggest how we can put this microscope for very good use in order to build ourselves a human cell atlas that gives us all of the types, states, transitions, and locations of cells. But first, let me remind you why we would actually want such an atlas. And this is because this kind of atlas will be a reference map for all functional studies. It will provide us foundational, fundamental knowledge across all biological systems. It will allow us to characterize the function of genetic variants, which is what this community cares about deeply, in the relevant cell type. It will allow us to interpret pathology and heterogeneity in disease with respect to an actual reference of normal. And it should eventually allow us to characterize this kind of heterogeneity in individual patients. I also want to claim that this can only be done within the context of one of those large-scale managed consortium-type technology advancing project. All of these underlying terms are actually from Eric Green's presentation earlier today that the NHGRI excels at and really shows the way at compared to other institutes. So first of all, let me say a few words about how. It's a unified project. I probably don't have to sell you so much on the idea that it's large. And because it is so large, I suggest that we should probably first engage in a pilot project in a few complementary systems. Maybe the blood, which is easily accessible but heterogeneous. Maybe the gut, that has a lot of many different types of cells— immune, epithelial, neural, others— all embedded in a three-dimensional tissue. Maybe the liver. Obviously, we need to do it in partnership with the expert communities and the relevant institutes. We want to do it with a standardized, controlled process and by developing and deploying shared analytical tools. And finally, we could set ourselves realistically the goal of driving the cost per cell to about $0.15. And this is already quite realistic with the ability of using these droplets and with the yield of a HiSeq 10x. And if sequencing cost drops even further, then this number can go substantially lower. Why do we want to do it as a big, large consortium-managed project? Well, because in this case there are fundamental scientific and impact-related reasons that for which only a large-scale project makes sense. So first of all, from a scientific perspective, only if we do it in a standardized way from sample acquisition to sample prep to QC to analytics would we ever have a chance of actually creating a meaningful catalog that is not riddled by noise. And our body didn't hear that it actually belongs to different institutes. So almost in any system that you would isolate, you will find cells that belong to more than one. Now scale serves a lot of purposes. One of them is obviously to drive the cost down, and that's very important. But only under the umbrella of scale do you usually actually garner the energy and the focus needed in order to have those technology advancing things. In our case, novel approaches to sample prep and to cell isolation and analytical tools. And these are the things that eventually will serve to our community as resources to the entire community of biomedical researchers and should be commensurate with the clinic. And then finally, because obviously this is a large project, I want to spend half a minute on numbers. So you might not know, but the human adult consists of about 20 trillion cells when you don't count the third of us that is actually red blood cells that fortunately don't have a nucleus, so we don't have to worry about them. That's a big number. With $50 million, that's what I gave myself instead of $100, or at least $50 million that we can spend directly on the cell, that gives us maybe 500 million cells at 10 cents a cell. That's only 10 to the minus 5 of a human. That sounds a little low-bally. But there are some encouraging things, like the fact that we know that there are about 300 major cell types. But then again, if you consider something like the retina, which is one of the best characterized systems for cell types, it might consist of 3 of the 300, but it actually has 100 sub-subtypes inside them. So it's definitely a big project. However, it's also a finite project. And what is very critical is the fact that in order to build an atlas, we don't have to measure every single cell in the same sense that when we chart a map, we actually don't do it in a 1-to-1 scale. We do it in a 1-to-100 or a 1-to-1,000, the level of resolution that is needed. And with simple calculations, we can show that we can sample orders of magnitude less cells and get the right representation for an appropriate survey. So in the example of the retina where there are 150 million neurons in the retina, given its complexity, which is probably unsurpassed in biology, only 40,000 would be required for a sufficient survey. So I posit to you that 10���� of the human body would not be so bad if done judiciously and could give us a very good resolution of an atlas, that the need is enormous, that the tools are in place, and the goal is within reach, so we should do it. Thank you very much. Any questions? Ewan.

Very cool. So in that last bit where you were looking at the kind of the numbers, you didn't put the state expansion, the combinatorics with state. And I mean, my sense is that these spaces are sort of complex and fractal. I mean, in other words, they're sort of boundable, but if one wanted to get more resolution, one could always get more resolution. The question is, is what point are these things useful? And they're useful, I think, very early on. I mean, even before one gets to the scales that you're suggesting. So, anyway, one comment there is about state. And the second comment is combining this with microscopy. You said this is a kind of a new microscope. I buy that. But I know that there are many pushes and drives into imaging and microscopy at a variety of different levels, and especially at the— I mean, you have this on the top left there as well. But I think integrating this kind of process, especially in the model organisms, which are much more tractable, with high-end organismal microscopy would be incredibly powerful.

I will answer both. The first is about state. Yes, they can be a tough one. But state, probably we will be saved by the modularity of the systems that we study. So yes, you could envision an endless number of stimuli that would lead to an endless number of variations. That's what I think you referred to as a fractal. But our experience in those places where people have tried to jiggle biological systems with these types of continuous responses is yes, that there were variations, but things were not endlessly variable in terms of the response. That's maybe because of canalization, maybe because of functionality. In transitions, this could be a lot more difficult, but I would say that they are probably beyond the scope of the first phase of a project like this. I think in states, we could go after some. We couldn't go after everything because we probably don't know all of the necessary stimuli. In terms of the relation to microscopy, I actually think it goes very nicely hand in hand. I think that there is a number of technologies. This one is based on computational mapping using the crudest in situ data out there, straight images out of papers on the fish. There are finer methods that can rely on finer images, and there are also emerging technologies in the single-cell proteomics that actually couples slices straight to readout, and I think would be adapted for the RNA domain as well.

So just to add that I've been really impressed by these light sheet approaches. I don't know if you've seen these light sheet approaches, but they are They are— the way they— the resolution they can sample at is just incredible.

Yeah, there are phenomenal opportunities there. Given the 10 minutes, I just put one cute figure. Ross.

Now that I have a chance to study the slide a little more, I see you've already thought of this very obvious thing, but I want to give you the opportunity to expand on it.

We should—

this should be done in— I love the idea, by the way, and I would do mouse in parallel. I mean, not wait, but go right ahead. I think there's a lot of richness that could come from looking at mouse and in helping us interpret, you know, studies in mouse and how to extrapolate that to human.

I think that's a phenomenal suggestion to do it hand in hand with the probably most prominent model organism of mammalian biology. I think it is completely doable. It would probably also resolve a lot of arguments. That I don't have a strongly held opinion one way or another, but it would at the very least tell us where we can go to the mouse as our model and where we cannot. Thank you. Yes.

I just have a couple of pedantic questions, which is— one is you— I noticed on your slide, you— on your cost estimates, you suggested something like 10 to the 5th reads or something per cell.— and just wanted to get a feeling for you about how thorough you think that would be. And second of all, when you think about the microfluidics solution that you had for lots of cells, you mentioned you've done upwards of 100,000 cells. Is that system ready for primetime yet?

Okay. Let me answer both questions. The first was about the cost and the 100K reads per cell. I think the number of reads per cell actually depends not just on the type of resolution that you want to have, but on the particular system. And we have pretty good ideas from the biological literature when you have to go deep and when you do not. There is a nice paper earlier in Science this year, J.T. Natale, that shows that for broad immune classifications, you can very well do with 1,000 reads per cell. I went higher end because if you talk to a neuroscientist about distinctions, for example, in the retina, which are incredibly important and functionally distinct entities, these neural sub-subtypes, actually the defining characteristics are at a much lower level of expression than the things that typically classify immune cells one relative to the other. The current molecular biology of single-cell RNA-seq, the current conversion rates that actually don't really vary from protocol to protocol, really say that above 500,000 cells, all you're doing is sequencing duplicate reads. So that doesn't help. And it seems to us from our own analyses, and I don't speak for my own lab only, but for the community of people who works with single-cell RNA-seq data, that you get the most bang for your buck from the first 100,000, and then it's diminishing returns. So that's our current knowledge. If our sequencing were a lot— were better, we should go with more reads. But we're just not there. It's not sequencing. If our molecular biology was better. About the drops. No, the 100,000 cells were done by a mixture of some of them with microfluidics, some of them with regular plates followed by automated robot, you know, liquid handling. And some of them were first done, you know, by hand. But nobody does them by hand anymore. The reverse emulsion droplets, I would not put on the slide if we didn't have really encouraging preliminary results. We think we turned the corner on them. But we haven't used them for the 100,000 count.

One more question.

Yes. Is there a strategy to advance the lineage map ahead of the other assessments? Or is the concept that you have to know the states and the location before you can do the lineage map?

No, actually you don't have to know the states nor the locations to do lineaging according to the people who do lineaging. There are several strategies. Most of them are based on judicious genome sequencing, not whole genome sequencing. And there doesn't seem to be an inherent impediment to combining that with RNA readout of the state of the cell, which means that you should be able to infer lineage. Having said that, it would require work. I would say that if I had to say off the top of my head, I would say that the tech dev for that and very early proof of concept would be appropriate for a Phase 1, but not a full application. Thanks very much, Avi. We have to move on.