Scientist Stories: Jonathan Pritchard, Human Genetic Variation and Evolution

Thrilled to get to introduce our first of two keynote speakers, Dr. Jonathan Pritchard. He's just done such amazing work in human population structure and genetics and demography, but I think maybe the best way to summarize that is an email I got when I was a first-year assistant professor. So I had just moved to Arizona State University, I think maybe a year before he had moved to Stanford, And it was from the chair of another department that said, I've just gotten an email from the president of the university asking me about this Jonathan Pritchard guy and read about him in the Wall Street Journal and he's done such great things. Have you heard of him? And I said, oh yeah, yeah, I've heard of him a bit. And then they said, you know, do you think we could recruit him here? And I said, oh, he's just moved to Stanford, probably. I can't bring him to the party in the desert that we have. But I just think it's really amazing that his reputation precedes him across multiple disciplines such that a business— our business major who directs the university just could see the value of the work that you do. And it's just always kind of stuck with me. So with that, I won't take any more of your time because we are on such a short time., and looking forward to your presentation.

Thank you so much for that kind introduction. Can you hear me all right? Great. Can you go back a slide? I guess I can go back a slide. All right, great. Okay, so today I'm going to be telling you about work that we've been doing in my lab in the last few years trying to think about this new type of data that's emerged in human genetics, genome-wide association studies, that in the last 15 years or so, I would argue, have really completely transformed the way that we think about the genetics of a wide variety of different phenotypes. And I just want— before I go on, I just want to highlight key people who have been associated with my lab in the last few years, who have led different aspects of this. So Evan was a PhD student with me and Yang Li, and they started this project, and then Chuan Yao and Yang have continued that, and then most recently, Nason and Sahin in the lab have sort of carried this on. So, I would argue that historically a lot of how we think about genetics has come from studies of simple and single gene traits. I mean, if you go all the way back to the beginning of genetics with Mendel, you know, part of why Mendel was successful was that he had the good fortune to pick traits that are determined in large part in peas by single genes. And then, of course, if we fast forward to human genetics, the early successes in human genetics came from studying monogenic disorders. So, for example, the mapping of the cystic fibrosis gene in 1989. And, of course, none of these early successes would have been possible if these investigators had studied complex traits. What we now know, however, is that most phenotypic variation is enormously complicated. So, if you look around the room, most of the ways in which we differ one from another are genetically complex, meaning that they are influenced by enormous numbers of variants across the genome, as well as usually substantial environmental factors. And so, the question that I want to pose to you today is to think about why it is that so much of the genome affects any given complex trait. So, over the last 10 years or so, I think we've become used to the idea that effect sizes in GWAS are small, there's a lot of loci that matter, but I actually want to pose the question, why is it that so much of the genome has small effects on complex traits? So, in this talk, I'm going to start out by giving you some examples of the architecture of complex traits, and then, in the second half of the talk, I'll tell you a little bit about our model for thinking about the architecture. So, what does this mean? And I think of this, at this stage, really as a hypothesis, and then, at the end, I'll talk very briefly about some next steps in this space. So, the first example that I want to show you is a Manhattan plot here for schizophrenia. So this is a study of about 150,000 individuals looking at schizophrenia. This is the latest meta-analysis from the Psychiatric Genetics Consortium. So, in this Manhattan plot, each dot is showing you the association signal for a single SNP, and these are ordered by genomic position. From chromosome 1 to X, and anything above this red line here is genome-wide significant. And this is a remarkable accomplishment. They identified 108 different loci across the genome that are genome-wide significant. However, you know, I would argue that virtually all of the hits are difficult to interpret if you do analyses like testing for enrichment in geo-categories or pathways, you get some hints of interesting things, but they are really difficult to interpret overall. Secondly, as many people have pointed out, for schizophrenia as well as for many other traits, the hits are responsible for just a small fraction of the expected heritability. So Bogdan Pasaniuc's lab estimated for this dataset that these hits explain about 10% of the expected heritability. And this has been referred to as the mystery of missing heritability, although that's largely been resolved by showing that actually there's an enormous number of additional variants that affect trait risk that are down here in the bottom of the plot. They're not significant, but they're real effects, and most of the heritability is coming from down here. And again, Bogdan's lab has estimated that Together, all the SNPs explain about 80% of the heritability. So, you know, so most of what's going on is an enormous number of small effect variants down here. And another way that we can look at this is to estimate how much of the heritability for schizophrenia comes from each chromosome. And, again, this is the same paper from Persaniuk's lab. So, here, each bar is representing an estimate of the SNP heritability coming from each of the chromosomes, and you can see you can estimate very accurately You can predict very accurately how much heritability is coming from each chromosome simply by its size. And that's what you might expect under a model where there's an enormous number of small effects and they're spread relatively uniformly across the genome. So, just to give you an example of a contrast, this is what the plot looks like for rheumatoid arthritis. So, it turns out that the HLA region plays a very big role in rheumatoid arthritis. That's contributing about 20%. Of the SNP heritability, but again, it's on a polygenic background of small effects across the rest of the genome. And we have argued that— we've made estimates for both schizophrenia and another phenotype we've studied extensively, height, that there may be as many as 100,000 variants spread across the genome that are affecting both of these traits. Okay, now you may say to me that, you know, these are traits that are perhaps incredibly complicated. You know, so for height you could make an argument about how practically any organ system might affect height. So, you know, maybe it's not so surprising. So let's try to think of simpler traits that are not Mendelian, but also not as complicated as something like height. And so I'm going to show you examples for two of these traits today, and we can look at how they're similar or different to these traits I've just shown you. So the first one I'm going to show you is for urate levels. So, urate is a small molecule that circulates in the blood. It's excreted through the urine. It's tightly regulated. If you have too much urate, it crystallizes into these sharp needles, and that produces gout. So, urate levels have been measured in the UK Biobank study in about 430,000 individuals. We've got an enormous sample size, and we can look within population groups. In this case, we're looking within the British ancestry set of that, and so we can minimize population structure as far as possible. So, we know a lot about the biology of urate, which makes it really different from both height and particularly schizophrenia. So, for example, we know what the synthetic pathway for urate is, so this is showing you the production of biochemical production pathways for urate, and we actually know what all the genes are along these arrows. We also know a lot about the control of urate levels in the kidneys. So, urate is exported from the blood into the kidneys, and here are the transporter genes. It's also reuptaken into the blood to keep blood levels at the right level, and, and you can excrete it out here. So we know all of these transporters. So these are two main biological pathways that we should a priori expect to matter. Okay, so now let's do genome-wide association study for urate levels. So how many loci, I'm going to ask you to guess here, how many loci would you expect to contribute to urate? Shout out some numbers. 500. 500, I heard. Other guesses? Excuse me? Did somebody say 40? 42, 1,014. Okay, so you're not such a credulous audience. Sometimes I do this, people say 5. So, the answer is that there are 222 genome-wide significant hits. For urate, and in particular, some of the hits are very, very big. And when we start looking at what the hits are, most of the biggest hits actually make a lot of sense. You can see in green here that all of the biggest hits are these solute channels, solute carriers from the kidneys. We also see a number of transcription factors that are involved in liver and kidney development. I'll show you in a minute that hits in the synthetic pathway are enriched, but they're relatively small part of what's going on. So, here is the transporters labeled now. Green are the hits, and what you can see is that actually out of 10 predefined transporter genes, 8 of these are hits in the genome-wide association study. So the GWAS is really doing an incredible job of pulling out most of the biology of this system. It also pulls out quite a bit of what's going on in the biosynthetic pathway, but it's by no means complete. So, it's pulling out 16 out of 88 genes that we predefined as coming from this biosynthetic pathway. That's a considerable enrichment. It's quite significant, but, you know, it's not overwhelming as it was for the transporters. Okay, so that's one example. The second example I'm going to show you is for testosterone. So, people often measure circulating levels of testosterone in the blood for clinical reasons. And again, this has been phenotyped to the UK Biobank. Again, we know a lot about the biology of testosterone, in particular the steroid biosynthesis pathway. So, here's a cartoon of that. So, we've got molecules like progesterone here. Here's testosterone in the middle. Here are female hormones like estradiol, or I should say hormones that are often thought of as associated with female. We can measure testosterone levels in both males and females, of course. There's about a 10x difference between the levels in males and females. Now, this is a little bit of a tangent on the talk, but I think it's fascinating. It turns out that the the genetics of testosterone levels in males and females are completely unrelated to one another. So, here's a plot showing the effect sizes of hits, of hits that are hits either in males or in females. And so, we're taking all of the SNPs that are hits in either one sex or the other and then looking at what the effect size is in the other sex. So, if these were correlated, if you do this for most traits, you'd see a very high correlation here. Between the two sexes. Instead, what you see is actually that the correlation is essentially zero. And if we look genome-wide across all variants, again, the genetic correlation is zero. And I don't know of any other traits, and we've looked quite a bit, that look like that. And it's a bit of a segue, it's a bit of a tangent, so I won't go into it in too much detail, but one part of what's going on is is that we have several examples of genes where there are actually paralogs that are having male or female-specific effects. So, this is a set of genes called AKR1C, and you can see that there's a female hit here, and then there's a male hit here. So, it seems that these effects in the two sexes are being mediated through different paralogous genes. Okay, so going back to the— biosynthetic pathway. So now, we can ask which of the genes in the pathway are hits. And it turns out that a really large fraction of the genes are hits, and more than half of them are. So we've got female hits in orange, male hits in purple. You can see that most of the female hits are up in this part of the pathway. Most of the male hits are down in this part of the pathway, but there's really very striking enrichment. Of hits for testosterone here. Okay, so although I've argued to you that we see really clear enrichment of signals in these core pathways, both of these traits are really highly polygenic. And so one way we can look at this is to, again, make plots of how the— plots of the heritability sort of as you move along the genome. And so here I'm plotting this slightly differently from how I showed you before. What we're doing is we're computing heritability in small windows along the genome, and then— so this is arranging the chromosomes from 1 to 21 here, and then the Y-axis is an estimate of the cumulative heritability as you go across the genome. And what you can see is that there are some big jumps here corresponding to the transporter genes in orange, but the overall impression you get here is that most of the genome is contributing something. And we get very much the same kind of pattern for testosterone. You can see here that, in fact, the signal's even much more uniform across the genome for testosterone than it was for urate levels. Okay, so that's still a relatively high-level view of— it's still not super high-resolution. Another kind of question we can ask is, like, how many SNPs are actually affecting each trait? And this is actually a bit of a difficult question to answer, so I won't get to a final answer for you for this, but I'll give you a hint of it. It's difficult to answer for two reasons. One is that many of the true effect sizes are very, very small. The second is the effect of linkage disequilibrium. So, if you have a causal variant, that is significant, then its signal will also show up in adjacent SNPs that are in LD with that causal variant. So, you need— so in trying to answer this question, we need to take account of LD. So, I'm going to show you one analysis that gives us a sense of just how much of the genome is contributing to your age. So, what we're going to do is to take every SNP in the genome above a frequency of 5%, and then for that SNP, we're going to first estimate how many LD partners it has. This is called— as measured by something called LD score. And then, for SNPs that are in a bin of having about the same LD score, we're going to estimate what fraction of the tests come from the null and what fraction of the tests come from the alternative. And to do that, I'm going to use a method from Matthew Stevens called ASHR. And you can think of this as being, although the methods are a bit different, it's a little bit analogous to estimating the fraction of true positives in false discovery rate analyses. Okay, so when we take SNPs with the, that are in the very lowest LD bin in the genome, we're estimating from AshR that just a couple, like a few percent of those are associated, or in LD with something that's associated with urate levels. Okay, so now what happens is we sort of step across this plot increasing the LD score. Now you can see that as you increase LD score, the fraction of non-null tests is increasing rapidly. And in fact, by the time that we get to relatively large LD bins in the genome, we're estimating more than half of these have associations with urate levels. But if you take a median SNP, what the sense we get is, from this, is that about about a third of these are in LD with something that affects urate levels. And since the extent of LD, typical extent of LD in the genome is on the order of tens of KB to 100 KB, something like this sort of scale, it's implying that a really large fraction of the genome is close to something that matters for this trait. So, you know, even though we started with a trait that we thought in advance might be relatively simple, in fact, most of the genome matters. For it. The last point that I want to make is that for urate levels, as people see for most complex traits, most of the heritability is coming from regulatory regions. And so I'll just show you one plot to illustrate this. If we look at SNPs that are in active chromatin in the kidneys versus not active chromatin in the kidneys, there's about a 30-fold enrichment for the estimated contribution to heritability per SNP from the active regulatory regions versus the background. And then if we control for what's going on in the kidney, then there's not very much going on in the other tissues, maybe little hints in some of the other tissues. Okay, so now I want to recap what we see for urate. But also for many other traits. So, first of all, I've argued to you that the heritability for typical traits is spread extremely widely across the genome, and usually without very big hits. I've argued to you that genes with trait-relevant functions generally are contributing just a very small fraction of the total disease risk. Although I would say that the biggest hits do often have clearer enrichment in relative gene— in relevant gene sets. And then lastly, the contributing variants are generally highly concentrated in regions that are active chromatin in relevant tissue types, and this, as opposed to, for example, protein coding variants, and this is implying that most of the effects of variants that matter are being mediated through gene regulation. So, the question that I want to pose to you then is how should we think about the molecular links from genetic variation? To complex traits. So, you know, what kind of models can we use to— what kind of conceptual models would produce a result where a large fraction of the genome matters for any given trait? So you can just think about that for a moment, and I'm going to show you in the second half of the talk what we— how we've been thinking about this. So, a couple of years ago in a review paper with Evan and Yang, We started to develop this model that we refer to as an omnigenic model or core gene model, which runs as follows. So, we imagine that we can partition the genes that are expressed in relevant cell types. So, for example, in kidney cells for urate levels into what we refer to as core genes, and these are genes where there's a direct causal arrow from the gene expression to the trait itself. And then peripheral genes, which— and peripheral genes may be relevant, but they don't have direct causal arrows from the gene to the trait. So the only way that peripheral genes can matter in this model is that they're affecting regulation of core genes. And then if you think about SNP variation, because of course we measure SNPs and not genes themselves, SNPs could matter either because they are cis-regulatory variants that affect core core genes, or they could be SNPs that are elsewhere in the genome. They're probably cis to some peripheral gene, and they may have genotype effects that flow through peripheral genes. They may be trans-regulators for the core genes, and thence they're affecting phenotype. And, you know, if you think about these core genes, our notion is that these are sitting embedded in cellular regulatory networks, and so potentially a large fraction of the expressed genes in the cell may wind up mattering, and they're mattering because they're plugged into the same regulatory networks as the cores. So, specifically, we have several hypotheses. First of all, that most of the heritability is flowing indirectly through these weak trans-regulatory effects, and this is to explain why core pathways and genes are actually not contributing very much to the heritability. Secondly, we suggest that perhaps every gene that's expressed in relevant tissues is connected into these regulatory networks and, you know, has the potential to have non-zero effects on the trait. And also that any variants, a regulatory variant for any gene in the relevant tissues might again have effects on the trait. But I note again that for both 2 and 3, that these distributions for most genes are probably centered around 0. With some variance, of course. Okay, so let's think about how, like, how this might work in practice. I'll try to make this a little bit more concrete. So, if we have a core gene, then by the definition I've suggested to you, this means there's a direct causal arrow to the phenotype. Okay, so now how does genetic variation matter, because that, of course, is what we're measuring in the genome-wide association study. So, we could have variants that are regulatory variants in cis. I.e., they're cis-EQTLs. We could have trans-EQTLs, so these are variants that are presumably cis to peripheral genes, and then there are effects that are flowing through the networks. So the first question that we need to address to think about more about this model is how much of the expression variance for a typical gene is coming from cis versus trans effects, because, of course, the effects that are cis to core genes have relatively direct effects on the phenotype., and the effects that are trans to core genes have to flow indirectly through the regulatory networks. So, this is actually not an entirely easy question to answer, because as I'll show you in a minute, trans-EQTL effect sizes tend to be small, but there have been a number of studies that have tackled this question. One of the key studies in this is from Alkes Price, back when he was a postdoc, and he estimated that roughly 37% of the expression variants in blood and 24% in adipose is coming from cis effects. And this was from a design where they're looking at correlation and expression within pedigrees, and they can look at how this is affected by sharing IBD of the cis regions. And in fact, if we look at a variety of studies in the literature, that try to estimate this, we see that overall that there's some degree of agreement. So, on average, we might— I could summarize this by saying that perhaps 70%, 3/4, something like this, of expression variance is controlled by trans effects. Okay, so now we can put some numbers on this. So, we expect that about 30% of the expression heritability is coming from cis effects and the rest from trans. But if you've ever done eQTL mapping, you'll also know that trans-eQTLs are really, really hard to find, and this is because they have tiny effect sizes. So, here's just one example of this, although I think, you know, this is basically everybody's experience. So, we had a study where we ascertained QTLs in one blood expression dataset, and then We measured the effect sizes in a validation set. So these show you the cumulative distributions of effect sizes for cis-eQTLs here, from biggest to smallest, and for trans-eQTLs here. And what you can see, of course, is that the cis-eQTLs are just much, much bigger. And in fact, this is a linear scale, but what actually matters for the variance is this squared, and so that would even greatly exaggerate that difference. So, I've shown you that most of the variance is coming from trans-effects, but at the same time trans-effects tend to be individually extremely weak, and so this implies that a typical gene must have enormous numbers of weak trans-effects on it. So, if we imagine that even a relatively simple trait like urate probably has tens of genes that are core genes, and each of the core genes probably has to make up a number, maybe hundreds of trans variants, then this quickly starts to give us some clue about why it may be that such a large fraction of the genome is able to contribute to any given complex trait. So, the next question, though, is to understand a bit more about why the core genes aren't contributing more to the heritability, because I, you know, have also told you that cis-eQTLs are much bigger than trans-eQTLs. So, you know, you might think, well, you know, a SNP that's in cis to a core gene, you know, sometimes those should have big effects. So, to do this, I'm going to write down a simple model to suggest why this might be. And, you know, I figured that you might want your money back if I had a keynote without any LaTeX in it, so— so I'm going to indulge you here. Okay, so here's the graphical model. So we imagine that we've got a phenotype, Y sub i, it's a phenotype in a particular individual i, and we're going to model this using basically the simplest model that you could come up with for expressing the phenotype in terms of core gene expression. So, we've got the average phenotype Y bar plus the sum of effect of M core genes. So, first of all, we're going to assume that the unit change in expression of core gene J is going to affect the expected phenotype by an amount gamma J. So, it's basically a regression coefficient. And then this is going to be multiplied by the expression level of individual— of gene J in individual I relative to the mean. So, this bit here is just measuring how much the expression of this core score gene in a particular individual is deviating from the population average. And then lastly, we have an error term that's including all of the non-genetic effects. Okay, so next, we need to compute the variance of the phenotype. So, the phenotypic variance, we can calculate this using standard rules of probability. So, there's two parts. So, first of all, we have a sum over variance terms. So, we have a sum for each core gene, we've got gamma squared, so the slope squared, times the variance in expression. And I've told you already that we should expect the variance in expression to be approximately 30% cis and about 70% trans. And there are— if there are m core genes, then there are m of these terms. But, of course, we also have covariance terms, so there's going to be a covariance for every pair of core genes. So, here, we're going to sum over all pairs of core genes, and so there's going to be a product of the slopes times the covariances of each core. And these covariances have to be dominated by trans-effects. And the reason is that most of the time, pairs of core genes are going to be in different parts of the genome, and so a variant that is affecting both of those has to be doing it through trans-effects. And secondly, there are almost m² of these covariance terms. So, if the expression of the core genes tends to be correlated, then you can quickly wind up that these covariance terms wind up dominating the heritability. So, specifically what you need for the core gene— for for the covariances to matter is we need the slope— we need two things. We need that the slope terms generally tend to be in the same direction and that the genes tend to covary. And this is fairly plausible, because you could imagine that if you have a pathway where many core genes are affecting the trait, then that whole pathway may tend to go up or down together, and the genes in the pathway may tend to have shared directions of effect on the phenotype. Okay, and so we think that this is probably actually an important part of what's going on. So if we think about the percent of heritability that's coming from trans-effects, so it's going to depend in this model on how many core genes you have. So if the core genes are completely uncorrelated, then we expect that about 70% of the heritability is coming through these weak trans-effects, but in contrast, If the core genes tend to have strong correlations, then we can easily have a model where virtually all of the heritability is coming from weak trans-effects, and we think actually that in real life that's probably what is happening very much of the time. So, the last point that I want to make is just to think about SNP effect sizes for a moment in this model. So, imagine now that that we've got a SNP, and so that can have— that could be a cis-eQTL for a core gene. And we'll say that the magnitude of effect of the SNP on the core gene is going to be beta, and then this core gene has an effect gamma on the phenotype as before. So now, the effect size on this trait is going to be just the product beta times gamma. Now, we can also have this for a trans-eQTL. And again, the trans effect size is going to— from SNP to gene is going to be beta. But in general, we would expect that if the SNP is just a trans-EQTL for one core gene, then usually beta gamma is going to be quite small because we believe that trans effects generally are very small. Okay. However, of course, there are presumably many core genes, and so we can easily have scenarios where a single SNP is regulating many core genes as a trans-EQTL. And I don't have time to get into it, but there are various examples of this. So, for example, transcription factors that are master regulators for some important process, you could have that a SNP at that transcription factor has effects, presumably, on many core genes that at least show up. If you recall in the urate example, 3 of the top hits were for transcription factors in kidney. So, in this case, we can think about the effect size as being a product of the number of core genes, M, times the average effect size of the SNP on each of the core genes and the core gene effect. On the trait. And then the last case, which is probably what's happening for most SNPs and most regulatory SNPs in the genome, is that this SNP may be having relatively uncorrelated effects on all of the core genes. So, each of the betas will be— each of the betas individually will be small, and then the directions of effects will tend not to be aligned. On average, we might think of these as having an expectation of 0, and there'll be some variance around that mean of 0, so the variance is M times beta— the average of beta gamma squared. So, we think that probably actually most of these effects that are coming from peripheral genes out in the rest of the genome are are going to be small, but when you— but there's an enormous number of them, so basically they add up to explain most of the variants that we see in phenotypic variances. So, to wrap this up, what we've proposed is that, first of all, that essentially all of the genes expressed in relevant cell types have the— at least have the opportunity to affect the function and regulation of core disease-related genes. And secondly, I've argued that most of the heritability is coming from SNPs that are outside of core pathways. And lastly, that most of the heritability we think is coming— is flowing through these very weak trans-effects through regulatory networks. And in the last minute, I just want to tell you a little bit about what I think the main challenges are moving forward. So, first of all, it remains really difficult to get good estimates of how many SNPs there are that matter, how many genes matter for any given trait. And so, I think that that's part of why what I'm showing you is, you know, is still sort of a model and a hypothesis. Secondly, to the extent that this is a good model, I think it's clear that there's a real challenge in thinking about how to use available data to infer what are the core genes, what are the master regulators for any given disease. So, one of the main goals of doing genome-wide association studies is, of course, that we want to be able to identify the core biological pathways to disease. For, you know, for understanding biological processes in disease, we don't really care so much about an enormous number of small effects. Across the rest of the genome, but we need to have good analytical ways of taking GWAS data, taking pathway information, taking all of the other kinds of resources that we have now in, you know, in genomics to be able to tease apart which bits of the signal actually matter for biology and which bits don't. All of that said, I should say that the other major goal of GWAS is prediction of who's at risk, and so polygenic risk scores have been making a lot of progress in the last few years. And for the polygenic prediction problem, I would argue that actually the architecture doesn't really matter very much. So you don't— for polygenic prediction, you don't really care actually whether a SNP is having direct effects or it's some small peripheral effect. All that you really care about is affecting the— is estimating the effects on accurately. And the next thing that I think is a real challenge in the field is getting better measurement of gene regulatory networks. And so, of course, you know, we've had different types of networks around for a long time, but I think it's not really very clear how well these map onto what we actually care about specifically in genome-wide association studies. I think there's a huge opportunity now with very high-throughput perturbation methods, followed by measurements of gene expression. There's a whole bunch of tools in this space. There's a real opportunity to do better at measuring these directly. And then the last thing, which I think is sort of a real challenge or mystery in the field, is when we map cis and trans eQTLs in large studies, so for example GTEx, are these really measuring exactly the same thing as what variation in a complex trait sees? And the reason I bring this up is that I would argue that actually the correspondence between eQTLs that we measure and GWAS signals is not as clear as one would hope for. And so I think that there are various possible explanations for that, but that's a challenge in the field. Field. And so just to end, I want to thank the key people who've been involved in this, so particularly Evan Young and Chuan Yao, as well as NASA and Sahin. And here's a picture of my lab a few years ago. So I'll stop there and take questions.

Thank you. So we have time for just a couple questions and then We'll leave that for the dinner tonight. So we'll give 5 seconds to think of your questions.

You can start here. Yeah. So actually, towards your challenges at the end and perturbation methods, given the amount of redundancy in some of these networks and sort of backups and given the the amount of rewiring that actually happens in a real disease versus the short-time effect of a perturbation effect. How much do you think— how many of the real effects do you think these perturbations will be able to recover? I think we don't really know the answer to that. I mean, we're really at the very beginning of doing this. You know, I think it's— so I think So I should say that we're, you know, we're working quite a lot on perturbation methods for estimating networks. So, you know, I think, you know, we believe this is an important way forward. So maybe that's an answer to this. But I actually think that, you know, the various aspects of the data point to, you know, point to suggest that regulatory networks are, you hugely interconnected that, you know, a lot of the genome must be able to affect any given gene given the trans-EQTL heritability data. And that's just a black box. Like, we really don't know very much about it. We don't know how much of that's going through transcriptional networks, through protein-protein interaction networks, like anything else. I think that there's a whole lot of biology that we just don't really understand very well. The perturbation networks with gene expression readouts are one thing that we can do right now, but I actually think that that's just a very small part of what we need to do. And I hope in the next decade that's going to become a huge focus in biology.

So I'm going to cut off questions now so we can make sure to get through our other keynote and then have time for lots of discussions afterwards.

Thank you very much.