Scientist Stories: Benjamin Neale, Scaling Up Genetic Analysis

I'll probably take a little more, like 30, 35 minutes. So I think it's a great time to be a geneticist. It's really, really exciting, and there are tremendous opportunities for really increasing the scale of genetics and genetic analysis. And I'll talk to you about, you know, some of the things that I think are good about genetics, why we like genetics, and in particular focus on progress in complex traits. I think I've been to ESHG maybe once before., and it's kind of got a little more clinical genetics focus than, say, some of the other meetings that I go to. But I think that kind of relationship between both clinical genetics findings and, you know, almost more Mendelian-style analyses and GWASes is really important. And to kind of start things off, I think one of the reasons that we really like doing genetics is because of what genetics can do to teach us about biology and teach us about novel therapeutics. So there's this really lovely paper by Matt Nelson and colleagues from GSK primarily in Nature Genetics that looks at the fraction of genes, or the fraction of targets in different phases of clinical development that show genetic association evidence, okay? So what you can see is that as you move further and further along the kind of drug development pipeline, so as drugs you know, get to be closer and closer to actually being therapeutics in the clinic, the fraction of genes that have genetic support is increasing. And that's consistent with the idea that support for a target based on genetic information is one of the ways of boosting the chances that we're going to actually make a novel therapeutic. And that really matters, I think, especially for phenotypes that I tend to work on more than any other, and that's in psychiatry where there's been, you know, very little traction in novel therapeutics over the course of the last 50 years. Now, this— these results are split into both OMIM genes here, so these are kind of more traditional Mendelian, strong-acting, single mutations, very clear gene, and then these kinds of GWAS database results, so the things that are in the genome-wide association results list. And what you can see is that both OMIM and GWAS are providing information. They're kind of, as you move further and further through target, you know, kind of pipeline development, you see a higher and higher fraction. But when you combine the evidence, you get really a lot of kind of bang for your buck, a lot of value. And that means that they conclude that, you know, selecting targets that are genetically supported has the chance of potentially doubling the success rate of novel compounds being pushed into the clinic. Now, you know, kind of why genetics in complex traits more than anything else, and this is really consistent evidence across basically every complex trait that we look at, from things like height to diabetes to schizophrenia, there's a clear, you know, an important role for genetics. The earliest indications of this for more complex phenotypes actually came from Francis Galton really founding, in some ways, you know, a lot of the kind of quantitative genetics, maybe not with the formal math that Fisher had, but he made this really elegant plot of looking at mid-parental height and taking individuals that are really short all the way up to individuals that are really tall. And then asking what their children's height were. And what you saw is this phenomenon of regression to the mean. So, the shortest parents had short kids, but the degree to which they were short was not quite as extreme as the extent of the average height of the parents themselves. And so, this regression to the mean is where we get this term regression, but it also shows that there's a similarity between relatives, and that really That similarity between relatives is consistent for something like height, and I think height is a good toy phenotype in some ways. It's easy to measure. It's pretty stable over life. You know, once you get to your adult height, you're pretty much there for the rest of your life, and that kind of consistency of heights, you know, is— it's also like one of those polygenic traits, right? There's not like a single gene that says, "Well, I'm going to be 6'3."' Like, that's not how the genetic contribution to height works. It's really this idea of multiple small genetic effects scattered across the genome. And so, taking these kinds of twin and family studies, you know, there was a kind of entire community focused on this genetic epidemiology kinds of analyses, and Daniela Posma and Tinker Poldeman did this really gorgeous meta-analysis of basically every single trait, every complex trait under the sun, for twin and family studies. And by and large, the general rules of thumb are that pretty much every complex trait that you look at will have some kind of heritability, that the average estimate is around 50% of the variability in the population is being driven by these genetic influences, and the majority, you know, about 70% of twin studies support a kind of additive genetic model, this idea that we have lots of small individual genetic effects that are scattered across the genome that are acting, you know, that can be modeled as acting additively, okay? And there's a— I think, you know, given that this is an educational session, I think there's a really important distinction between what is a— something that can be modeled as behaving additively and true additivity. And it is certainly the case that you can have slight deviations from this additivity. You know, it's not a fully, you know, 0, 1, 2 is the number of doses of an allele and that giving the kind of impact on the phenotype, but that kind of deviation is going to— even if it's not perfectly additive, additivity will work under most scenarios. So, it'll do a pretty good job of capturing what the genetic effect is, even if there's a little bit of non-linearity, which I think a lot of people want to talk about. And a lot of that comes around the fact that things are small, and when effects are small, a linear fit works pretty well. So, picking up this idea of looking at similarity, what's happened over the course of the last 7 or so years, there are, you know, people like Peter Vischer, and Naomi Rae, and Mike Goddard, and Sean Purcell, and others, have been working on this idea of lots of small genetic effects that are scattered across the genome. And so what Peter came up with is this idea of trying to use the degree of similarity across common genetic variation as an index on whether or not there would be phenotypic similarities. So we're just comparing genome-wide how similar your genotype is to a random person in the population. And looking at all of those pairs, looking at all of those kinds of comparisons, you can get estimates of the contribution of common genetic variation to phenotypic variability. And what you see for height, and for BMI, and schizophrenia, and depression, and ADHD, and autism, is that basically all of these traits show, you know, kind of significant common variation contribution to phenotypic variability in the population, okay? So there really are these genetic effects, and we, you know, have looked at this and measured a lot of genotypes, and actually gotten to that kind of perspective. Now, there are kind of two main technologies, and I'll try and talk about both of them, but I'll focus actually more on arrays than on sequencing, because I think there's still an awful lot to learn about arrays, and while sequencing is tremendously exciting, I think there's still a lot of important things to think about and learn from what we've kind of achieved as a community looking at genome-wide association analysis. Okay, so the main reason for arrays is that the sample sizes are the largest, and from a kind of statistical genetics point of view, or quantitative genetics point of view, Sample size is the single biggest indicator of the relative success of the genome-wide association study. As you continue to add samples, you can continue to gain further traction in identifying regions of the genome that you can unequivocally point to as actually being associated, which can then be aligned into, say, new biology. And so, when we talk about arrays, there are a variety of reasons. So the first and foremost is that common variation really is a feature contributing to basically every complex trait that we have assembled sufficient sample sizes to look at, and that cuts across psychiatry, that cuts across autoimmunity, metabolic phenotypes, basically every major domain of phenotypic variation in the population, be it human disease or trait variability. Almost all of these things have a clear and substantial role of common variation in conferring risk. There are a handful of phenotypes for which common variation is not conferring risk. These are Mendelian disorders, are things like Huntington's or cystic fibrosis, but outside of those strong-acting Mendelian phenotypes, these common complex traits are really governed by a non-trivial contribution of common variability— common variation. Now, not only is it the case that common variation contributes to these phenotypes, it is also the case that we can actually detect what this contribution is, so using approaches like GCTA or, or something that I'm going to present called LD score in the next few slides, gives us an opportunity to actually not only, you know, say that it exists, but use it and bend this heritability to start to try and understand how different traits relate. So improving things like epidemiology, and diagnostics, and phenotypic classification, but also learning things about annotation and what functional marks of the genome, and that, you know, people like ENCODE and the Epigenomics Roadmap have really been putting together over the course of the last 5 to 10 years, what of the— which of those marks are actually really relevant? And the more variability in the population that you can characterize, the more power you have to draw conclusions about how different phenotypes relate. And I think that's a kind of important set of questions that genetics can really kind of critically advance our understanding of. Okay, so let's just, you know, give a little bit of a description of, you know, say, one of the large-scale consortium efforts for Common Variation Analysis. This is the PGC that Mick O'Donovan, Stephan Ripka, and Pat Sullivan, and Mark Daly, and many, many others have contributed to over the, you know, since, I guess, 2007, 2008 sort of time. And what the PGC has been doing, first and foremost, is aggregating large-scale genome-wide association datasets and meta-analyzing them, and really trying to put them through uniform QC, clean the data in ways, you know, not too dissimilar from what Teri was talking about in the previous session, ways of trying to remove the hidden unwanted variation. But what has been an absolute, you know, kind of clear trajectory for schizophrenia is that it's a 10,000 cases and 12,000 controls. The first half a dozen genome-wide association results that were, you know, clearly significant started to emerge. And so this is a Manhattan plot where we plotted the minus log 10 p-value of every SNP scattered across the genome. Here's the P-value there, and then, you know, here are all of the SNPs, and this is just thresholded at 10 to the minus 3, and this red line here is 5 times 10 to the minus 8, our kind of estimate of the dimensionality of testing common variation in the genome. Now, as you add sample sizes and get to 25,000 cases, you see 62 loci emerge. As you get to 37,000 cases, this was published in Nature almost 2 years ago now, there are more than 100 distinct loci associated to schizophrenia scattered across the genome. And that's really powerful for a variety of things. You know, each one of these individual genetic variants has the potential to teach us something about biology, about risk for schizophrenia in the population. And while schizophrenia has achieved this kind of dramatic success, it's not the only one that's gone on this kind of arc. And so, as you look at increasing effective sample sizes here on the x-axis, On the Y-axis, you see the number of loci, and this is the case not only for schizophrenia here in purple, but also inflammatory bowel disease, type 1 diabetes, type 2 diabetes, coronary heart disease. All of the kind of major complex traits, these public health burdens for which genetic strategies have had some, but albeit limited, success historically are beginning to yield pretty much uniformly as a consequence of increasing sample sizes. And as common variation continues to be assayed in ever larger sample sizes, we'll see further and further identification of genome-wide significant loci. Now, this activity of identifying loci is all well and good. It's sort of like the first step in a series of downstream activities, some of which I'm going to try and describe in the next few slides. So first and foremost, I think genetics can teach you biology, and there are a ton of examples that are even emerging from common variation, but I think one of the ones that I'm most excited about is some work that Steve McCarroll and Aswin Sekar at the Stanley Center and the Harvard Medical School have really performed. And what they've done is they've taken this most strongly associated region of the PGC schizophrenia meta-analysis, and that happened to land in the HLA. Now, the HLA is one of the hardest parts of the genome to really interrogate. There's widespread correlation, you know, there's kind of lots of LD between the genetic variation in those regions. The association itself did not map cleanly to a classical HLA allele, as is the case for most autoimmune disorders. So if you look at things like, you know, MS, there's— or celiac disease, there's a clear HLA association. But what Steve looked at was— with Aswin— was this C4 gene. And C4 came in various numbers of copies. The C4A, there were, you know, kind of a variety of number of copies of this gene. In the HLA region. So it's this kind of copy number variable gene. And what they did was they lined up basically the implied number of copies that you have of that gene and the degree of schizophrenia risk, and you got a near-perfectly linear relationship, which is like a kind of beautiful dose-response for this imputed C4 expression level. And that's really exciting. And then also what you can do is show that C4 expression is different in, you know, controls versus schizophrenics based on brain tissue analysis, okay? So this is now beginning to build a case that this dose-response kind of association that we see for a specific gene in a region identified from common variant analysis was maybe, you know, relevant to schizophrenia. And then Beth Stevens and Mike Carroll and others picked that ball up and pushed forward and looked at what C4 was doing. And what C4 was really— seemed to be doing in the brain was kind of shaping the extent extent of synaptic pruning, okay? So, the brain goes through this, like, during development a huge number of synapses are formed, and then they get pruned away. And I'll show you a picture of what that looks like from, I think, the kind of late '70s. But what Beth and Mike and others really demonstrated was that the degree of synaptic pruning was less in the C4 knockout, okay? So, that's like there's more kind of multiple overlapping inputs at this kind of critical point in development when C4 is being expressed. Okay, so synaptic pruning now can become like a new, like, hypothesis in a sense, or at least a further supported hypothesis in the context of actually what's driving risk for schizophrenia. And neurobiology is really, really challenging, but this is some work from Hutton-Locker in the late '70s. Showing the degree of synapse formation. So, at birth you have very little synapse, and then as a child there's an explosion in the amount of synaptic kind of development that is generated. And then some of it gets pruned away in adulthood, and that's part of this kind of neurobiological, neurodevelopmental process with respect to synapse. And then second to that, you know, David Lewis did some really nice work looking at postmortem brains showing that there is you know, excess synaptic loss in schizophrenia patients compared to controls. And this is, you know, building a case. It's not to say that synaptic pruning is going to be the sole kind of biological phenomenon that's driving risk for schizophrenia in the population, but now we have genetic evidence that's supporting a specific gene, and that gene has a clear role in synaptic pruning in a mouse model. And then we see this correlation phenotypically in patients versus controls. And for something like schizophrenia, this is tremendously exciting because schizophrenia has, you know, there have been a lot of neurobiological hypotheses, but it's been very difficult to take genetic results and use that to further strengthen specific hypotheses about what may be driving risk for disease in the population. Okay, so let's talk about LD score regression. I think when I look back through my email, this is why I was asked to give this educational session. And this has really been some collaborative work with Alkes and Mark that's been led by Brendan and Hillary and Puru, but really Brendan and Hillary were the main drivers of this particular work. And let's kind of talk about what this kind of genetic— this polygenic inheritance would look like in the population. So, if we think about the genome and we think about the way the genome is organized, there are these regions of high LD, these so-called LD blocks. And this is a heuristic. This is a way of thinking about the way the genome is organized rather than, in truth, the fact that it's, like, discretely blocky. But there are regions of high LD where a lot of these genetic markers are heavily correlated with one another, and then there are kind of lonely SNPs, things that don't show much in the way of genetic— you know, much in the way of correlation between the markers around them. And if we think about what happens when you plunk down causal variants across the genome, and this is little c causal. These are variants that have an influence on phenotype but are not sufficient to cause phenotype. They just are a genetic perturbation, a small increase in risk or a small decrease in risk, depending on which allele you happen to have. And when you plunk down a causal variant in an LD block, you see association not simply to the variant itself, but all of the markers that are correlated with that variant. So it's pulling through the association evidence. In contrast, you can think about a causal variant being a so-called lonely SNP., and in the event that you have such a causal variant as a lonely SNP, you'll only see association to that lonely SNP and that lonely SNP alone. It will not bleed through because there's no correlation between this marker and the surrounding markers. Now, when we have a polygenic scenario, when we have lots and lots of variants that are scattered across the genome, then if we assume that they're just randomly scattered, okay, so if we just say that it's a complete, you know, coin flip as to whether any individual variant is actually truly influencing phenotypic variation. Alternately, we can think about a so-called infinitesimal model, this idea that every single genetic variant contributes a little bit to phenotypic variability. I don't buy that model, but you can think about it. Then what you'll see is that as you have more variation— as you tag more variation as a marker, the chance that you tag a the number of causal variants increases, meaning that you should show more association on average. And so we simulated this. We randomly scattered causal variants across the genome. We created this kind of QQ plot, this distributional look at our p-values, our test statistics. This is really good for interrogating the tail, in contrast to the histogram approach that Teri was talking about that's good at, you know, looking at the distribution across the entire range. Because remember that, you know, 10% of your data lives, like, down here, you know, or 90% of your data lives down here, and this is just 10% of your data, because a p-value of 0.1 is related there. And what we see when we look at our LD score, the amount of genetic variation that is being captured by an individual variant, what its sum of R-squareds is, its kind of SNP-ability, how much it's tagging, as you look at that increasing on the X-axis, On the Y-axis, you can look at the mean chi-square for association, and you see this linear relationship result, okay? So this means that as you go, you know, as you tag more and more genetic variants, you show more and more association, exactly as we kind of set up the toy example. Now we can contrast that with a scenario where population stratification is actually driving the inflation in the distribution of test statistics. That is to say that there's some confounding bias or lurking problem that is really showing most of the association evidence, in contrast to this polygenic model. And under that scenario, there should be no relationship between the degree of LD and the amount of association. And the reason is, is because population stratification is driving— it's being driven primarily by drift, okay? So this is random fluctuation in frequency between populations, and given that it's a random process, it's equally likely that this very large haplotype block is going to drift in frequency between populations as this individual variant, okay? And so that means we've now decoupled this relationship between association, mean association, and the amount of inflation. That's all population genetics. Maybe that's kind of confusing, but we can do the, like, stupid thing of just doing controls versus controls, so we take controls from the UK and compare and compare them to controls from Sweden, and what we see is that we get very little in the way of a relationship between LD score and the mean chi-square. So, we're getting this kind of flat line. Now, this kind of principle, this general idea of the more, say, variation that you tag showing more association is also relevant in meta-analysis. So, if you think about when you perform meta-analysis, hypothesis, things that are true, things that show true effects, as you add more data, your test statistic should continue to increase. And so if I double the sample size, on average, I should double the test statistic, approximately, if I'm a true member of the alternative. And this is a really critical point in terms of how we think about inference and how we think about the confidence of specific observations, and much in the same way that those association test statistics kept getting taller and taller for schizophrenia, that's exactly that same phenomenon. As we add more data, we build a stronger case, and that's really a good way to think about analysis, is how well you're building a case for evidence for association in the population or the relevance of an individual effect. Now, here's a little bit of scary math. I didn't quite do what Teri did and censor all the scary math, but I just want to flag that this we can make an assumption about the distribution of SNP effects. This is a traditional assumption in quantitative genetics, and if we make that assumption about the distribution of SNP effects, then we can actually now create a new estimator of heritability, okay? And that's really what is nice here, is that we can use this idea of heritability and the amount of variability in the population explained by all genetic variation, or a class of genetic variation, to inform on what is important and what is relevant. It is one metric, one arbiter of relevance. Okay, so the first key extension here is really this idea of partitioning heritability, and full credit really goes to Alkes and Sasha Gusev and others that have been working on this idea of trying to take our all common, you know, genetic variation and chunk it up based on things like annotations. So we can do our basic LD score analysis, or we can lay down some annotations like DHS or coding variation. And when we lay down these annotations, we can now create LD scores that are specific to those annotations, okay? So this is how much genetic variation are you tagging that is a DHS, or how much genetic variation are you tagging that is a coding variant, okay? And that now gives you a different way of looking at the contribution to phenotypic variation from classes of variation. So it can tell you what annotation marks are more or less important. And so all we're doing, again, is taking those test statistics and correlating those test statistics that we get from these large-scale genome-wide association meta-analyses with LD scores, the amount of genetic variation being tagged. So that means we need no individual-level data, something that Nick Patterson gives me kind of consistent grief for, because it's kind of perpetuating this idea that just sharing summaries of statistics is okay. He likes individual data-level sharing. But it also means that we can go pretty broad in annotations pretty reasonably. And we can also go pretty broad in terms of phenotypes, right? So we can just take, you know, a whole bunch of different complex traits and see what comes out in the wash. And we can look at this in a couple of different ways. So the first way that we can look at it is to take 17— those 17 traits that I just kind of flashed up across anthropometrics, immunity, psych, et cetera, et cetera, and ask what the kind of average per-SNP heritability is, okay? So if we think about this polygenic model, this idea that there are hundreds or maybe thousands upon thousands of genetic variants scattered across the genome influencing phenotypic variation in the population, an important question now becomes what annotations really matter? How do we actually decide what annotations are important. And so what's plotted here is the degree of enrichment. So you can think about an average per-SNP heritability. That's what was in the equation effectively, is this idea of average per-SNP heritability. And then you can think about enrichment. So decoding variants, say, do explain 2 or 3 times as much heritability on average compared to the baseline SNP. And we can go and look that up. So here's coding variation, and what we get is about 7.5-fold enrichment. Enrichment for coding variation. So, on average, a coding variant will explain 7.5 times as much heritability as a kind of average SNP scattered across the genome. And what's really exciting about this is that, you know, things like conservation— so I see Manolis sitting over there, and so that kind of lovely Lindblad-Thoe... Kellis annotation of the genome does the best in terms of enrichment. It does much better than, say, even coding, which which is, you know, interesting and important. And so just 2.6% of the SNPs in the genome, that class captures about a third of the heritability on average, right? Okay, so whatever evolution is telling us, it's telling us that, you know, these regions are actually very important, and so improvements in things like mammalian conservation can really help with our interpretation of what genetic variants might actually be pathogenic. We can also look at coding, and we see the 7-fold enrichment up here. Or we can look at, you know, epigenetic marks, these H3K4me3 marks, and what we see there is that, you know, you get about a third of the heritability, but that third of the heritability is now in 13% of the SNPs rather than the 2.6% of the SNPs that conservation looks like. We can restrict to the peaks, okay, so this is just trying to narrow down that particular analysis, and that strengthens the enrichment, although it doesn't, you know, explain as much heritability. A smaller fraction of the genetic variation overall, but it does, to me, suggest that there's actually quantitative information in that, you know, peak height that is potentially relevant for interpreting these kinds of epigenetic marks. We can look at specific traits, and I'll actually— I'll probably skip this in the interest of time to move to cell type enrichments, because I think the cell type enrichment is a nice alternate way. So, rather than looking at specific annotation categories, we can ask what cell types and tissue types might be particularly relevant in risk for a given set of phenotypes by aggregating these sorts of epigenetic marks into a meta-annotation that is looking cell type to cell type. And when we do that, we get some, you know, nice results. We get things like CNS for schizophrenia. That looks pretty good. We get things like connective and bone tissue for height. That looks pretty good, too, right? Like, these are confirming the value of these genome-wide association meta-analyses at a very coarse scale, and as we learn more about how the genome operates, we will be able to get much more fine-grained in our understanding of what the biological processes are that are driving these phenotypes in the population. Now, a personal favorite is BMI, right here. So BMI shows the best enrichment for CNS, and that dovetails nicely with some of the mouse models of obesity, where basically it's knocking out genes that make mice compulsively overeat so that they get really, really fat. So, like, you know, the kind of idea that eating behavior is a major driver of BMI in the population, I think, is interesting and something that we can kind of gain support from these kinds of analyses. Okay. Genetics can also reveal how traits relate. Okay, so that's like trying to learn about biology both at a very specific individual variant level, like C4, or looking at broad brushstroke kind of assays that are trying to improve our interpretation of what genetic variants may or may not be relevant, but genetics can also teach us about how different things, you know, relate to one another. Are there shared genetic effects? Is there a genetic correlation between a pair of phenotypes? And when we think about genetic correlation, we can actually think about kind of two flavors of genetic correlation. So we can think about one flavor where there are genetic effects on some trait, maybe we'll call this trait LDL cholesterol, and then that trait could be causally related to a second trait, maybe we'll call that trait, say, heart attack, okay? And I think we're all pretty convinced that LDL cholesterol is a causal risk factor for heart attack, but genetics can help us, you know, understand and interpret that through things like Mendelian randomization that I'm not going to talk about today. Alternately, genetic correlation can emerge from, like, pleiotropy, or influencing some other trait that might influence a pair of traits. So maybe these are genetic effects on eating behavior, maybe genetic effects on BMI, and BMI influences both, say, type 2 diabetes and myocardial infarction, or heart attack, but there's not necessarily a causal relationship between type 2 diabetes and myocardial infarction. There might be, but maybe not, okay? And so, So that distinction in what genetic correlation means is important, but the establishment of genetic correlation suggests that either we can identify genetic effects that live upstream of both things, or we can identify these sorts of causal relationships potentially. Okay, so to draw us back to this idea of LD score regression, where we're plotting the amount of genetic variation that is tagged by an individual genetic variant. So how many SNPs are you correlating with? And that chi-square, that mean association level, this slope estimates heritability, and what we do with genetic correlation is we estimate not one slope, but two slopes for the two traits, and then we take the product of the Z-scores and we plot that line, okay? So what we're doing here is now creating a new sort of fake chi-square, because the chi-square is, you know, the square of the Z. We're creating this fake chi-square, and that fake chi-square here is now what we're going to run this regression on. And what that'll do is it'll pick up the amount of genetic correlation genome-wide, and we can do this between pairs of traits explicitly. It is— it'll work under RG equals 0, so if there's no ordering to the kind of genetic effects, so if the things that increase risk for schizophrenia are equally distributed between increasing risk for obesity and decreasing risk for obesity, then we'll see no genetic correlation, even if that's a consequence of, like, regions of the genome being enriched for functional variation, like, say, around genes or in enhancers or promoters. So, we formed the BrainStorm Project about, you know, 3 years ago now. There's Verneri Anttila, one of my postdocs, and collaborator Aidan Corvin at Trinity College Dublin with a whole bunch of others, and we've taken and tried to aggregate over 200,000, you know, cases across basically every major neurological and psychiatric phenotype that we've looked at. We're calling this the BrainStorm Project, and what we see when we do these kinds of analyses is that we can get an estimate of heritability for pretty much every disease represented here. I'll let you read the diseases rather than talk about them. This scales from about 0.3 to 0.05. 5. Every single major trait, disease, disorder, what have you, that we look at shows clear evidence of influences of common variation on phenotypic variation. I don't— I try and dissuade people from over-interpreting the specific point values of heritability. Heritability is a consequence of where you assume a threshold is for a given phenotype. It's the— there are a whole bunch of other assumptions that kind of fit into into what's going on here, but the kind of take-home message is that we see clear and compelling evidence for the influence of common genetic variation on phenotypic variation really across the board. And then we can do some analyses of genetic correlation. And so what we see over here is that we get, you know, anorexia nervosa and obsessive-compulsive disorder show a high degree of genetic correlation, and that's really interesting because these are two distinct clinical phenotypes as they exist in psychiatry right now. And this is another way that we can start to think about using genetics to try and inform and improve our understanding of things like diagnoses for these very complex, you know, difficult to study kinds of phenotypes. We see not only correlation between those two, but also with bipolar disorder and schizophrenia. Really, schizophrenia is a kind of correlator with almost every phenotype that we look at in major psychopathology, which is particularly interesting. It's also the largest meta-analysis. I mean, it's the phenotype the psychiatrists like working on the most. It seems to be the most debilitating in the scheme of things. But it is also worth mentioning that OCD correlates with Tourette's, but Tourette's does not correlate with anorexia nervosa, okay? So there can be this structure that emerges, and by taking this more multivariate view view, we can start to try and understand and think about reconfiguring how we define these clinical phenotypes. In neurology, things are a lot cleaner, and I think that's probably consistent with the idea that most neurological diseases are, you know, driven by very specific pathologies, you know, histopathology or imaging or, you know, things like plaques and tangles in Alzheimer's disease, or, you know, the loss of dopaminergic neurons in Parkinson's Parkinson's and the substantia nigra, like those kinds of phenotypes are much more specifically defined, or things like stroke or epilepsy, and those phenotypes don't show the same degree of genetic correlation across each other as is the case in, say, psychiatry. We can look across psychiatry and neurology, and we see very little evidence for overlap. You know, there are about as many red squares as there are blue blue squares representing, you know, basically an average estimate of about nearly zero. It's like 0.03 for the average genetic correlation across all pairs of traits across domains. And that supports this idea that there's, you know, kind of reasonable specificity to these deficits, even though patients with Alzheimer's disease, for example, will present with psychosis and psychotic-like symptoms. That psychosis is not being driven by the same phenomena as we see in schizophrenia. You can do this yourself, so Chris Zhang and Dave Evans were like, "This is really fun, LD score and genetic correlation." You can go and upload these results, your own genome-wide association meta-analysis results, and it'll generate genetic correlations across something like 170 or so different GWAS. And so you can go and look at it there. Now, just to kind of touch on sequencing briefly, I got about 10 minutes? 5 minutes? 5 minutes? All right. So I'll just touch on sequencing briefly, just to flag the fact that sequencing really enables this comprehensive capture of genetic variation, and with that comes challenges of scale. So the data that I'll present here is really being driven by this U01 that Nelson Freimer and Arno Peloti are leading, along with many others, and Andrea Ganna, who's a postdoc in my group, has been doing a lot of work on the analysis, and also I'll skip this. And so, when we're talking about doing these kinds of QC and getting rid of the unwanted variation that Teri was talking about, there's a huge amount of, you know, kind of work that goes into looking at the overall data quality. And these methods are largely maturing, but they're also quite computationally expensive. And so, we're working on a new software package called HAIL that'll really leverage, you know, kind of elastic compute and cloud computing environments to really make this go as rapidly as possible. Here's the HAIL team, but what I want to talk to you about is not that, but this primary challenge in whole genome sequencing, this idea that whole genome sequencing is going to be much harder with respect to the analysis of rare variation in complex traits, and that's because it's not very straightforward to figure out what variants you should try and lump together, and the degree of, say, functional enrichment for individual given classes of variation. And so one of the ways that we look at this is on average minor allele frequency, okay? So, from the first 5,000 or so whole genomes, we've started to look at the average allele frequency of variation in different classes. We've got intergenic variation up here. That's most common. And then there's this kind of slow decline in allele frequency with loss-of-function mutations. All the way down here, so these are our nonsense and our frameshifting insertion-deletion polymorphisms, and they're on average the rarest. Next most rare is missense mutation, and then we see that promoters, transcription start sites, and synonymous variation are all about the same. And that's really going to be very challenging in terms of how to interpret what variation is actually functionally relevant phenotypes, and that challenge is also important because when you do rare variation analysis, you have to lump variants together as a class. You can't test the individual variant one at a time. And so our ability to tell what variants are functional is really important. Now, people are already trying to address this through things like EIGEN and CAD, and what you can see is that you can improve from our overall enhancer line here, so that enhancer line's pretty close to synonymous. You can kind of drag it down by focusing on the variants that CAD says are bad, but even those variants that CAD says are bad are not as rare on average as the putative loss-of-function mutations, meaning that we're still kind of somewhere a ways away from being able to tell what a functional non-coding variant is compared to interpreting the coding region more generally. So just to conclude, the genetics of complex traits is working. There's a tremendous opportunity to really dig into these genetic results and start to interpret and learn about disease. And I think that's, like, the most exciting thing about the genetics. I think epidemiology is poised to be completely revolutionized by genetics, and the use of genetics in Mendelian randomization, in particular, for causality, I think, will be extremely powerful. And I think with that, I'll let you, you know, read the rest of the slide and open the floor for questions, and say thank you.

Thank you so much for this exciting talk, and I think, yeah, you've put some interesting issues on the table, and I don't know if somebody has any comments or questions.

Don't be shy, go ahead. Hi, Meghna Prasad, University of Strasbourg. It was a very interesting talk, thanks. I was just— it was probably one thing I just didn't understand. The slide where you showed the fold enrichment of the different kinds of SNPs based on where they were and the percent heritability that could be explained, I found it to be rather counterintuitive that the coding SNPs tended to be more enriched but tended to explain less of the heritability.

How do you explain that? Right, so the degree of enrichment is a concept consequence of the heritability and the number of SNPs in that class. And so coding variants have very few SNPs compared to something like enhancer H3K4me27. You know, those kinds of epigenetic marks have a much larger chunk of the SNP pie. So coding SNPs, I think, are about 1 and maybe 2.5% of the— or 1.5% of the SNPs, whereas, you know, conserved is a little bit more. You know, that— it's a small a small slice of the overall set, and so the number of SNPs in the bin will also influence the total amount of heritability being explained.

Okay. Is that in any which way related to the actual biology, do you think?

Well, I mean, there's only— I guess in a sense it is, insofar as the coding variation, you know, you have to be a coding region, and the coding region itself is only about, you know, 1.5% of the total footprint of of the genome. So enhancers and things that say recruit, you know, kind of transcriptional machinery is a much larger footprint, at least as we kind of deem it now, compared to what the actual coding sequence itself is. And so that's much smaller comparatively. Thanks. Sure.

Another question, yeah. Thank you very much. Burak Karajoran from Turkey. Referring to the GCTA assumptions using uniform priors, instead of this, if we employ biological priors, which is based on maybe prior information where the signals are, do you think that this might be a relevant option, or this might be biased results could create?

Yeah, so it's kind of a difficult question. I mean, the way that functional information I think is best integrated into these kinds of approaches is actually through partitioning. So it's the— so in LD score, for example, we run a general model, and the general model tries to include every functional class and category that we look at, so that when we test for enrichment, we are accounting for as much other kind of lurking components. In terms of actually getting a better estimate of heritability, the estimates are pretty stable. Like, I don't think there's a huge amount of margin in terms of thinning out what markers that you look at versus not. I mean, you probably pay a little bit of penalty for noise in the construction of your, you know, your genetic relatedness matrix, but that price is comparatively small because there are many effects that are really scattered across the genome, and so things are pretty robust.

Any additional questions from the floor?

Oh, Sam Li is going to ask a question then.

I do. Your point number 4 there, annotation is the key for analysis. How do you see— where are we now in terms of the non-coding variation and the annotation? How much should it get better for us to be able to lump together wisely variants?

Well, I think there's, like, I think there are two parts to this. One is maybe more knowable than the other. So the unknowable part is the degree to which genetic variant in a non-coding region is going to have a similar biological or functional impact as, say, a coding variant, okay? So a putative loss-of-function mutation is going to knock down, you know, knock out one of your two copies of the gene. There may be compensatory mechanisms that mean that you transcribe an awful lot more of that gene to compensate for that, and so that may, you know, weaken the effect, and it's not always a guarantee that that will be the most deleterious, but on average it's clearly the most deleterious from a selection point of view. So that's some index on what I would call "effect size," right? So non-coding variants, you know, they may have effect sizes that are comparable to loss-of-function mutations, but I would think that that will tend to be the exception rather than the rule. The question is how well we can call them, right? So, LoFs we do a pretty good job at calling. Like, you can, you know, do a baseline call that's probably a little bit noisy, and then you ratchet up your, you know, thresholds in terms of frequency, in terms of exon inclusion, in terms of whether or not it's, like, at the end and causing nonsense-mediated decay or a truncated variant. Like, that, you know, that space of exploring loss-of-function mutations is pretty well pretty well explored, and we can do a pretty good job of really identifying which ones we really think are going to be functionally impactful. I think we have to get as good for non-coding variation as we are at LoF calling, because I think of the first point, the effect size for these non-coding variants will be weaker than they will be for the LoFs, and so that's the That's the bar that I would hold us to. Like, if we can get as good at predicting the consequences of a genetic variant as we are for calling a predictive LOF, then I think we might stand a shot. I think independent of that, we're going to have an awfully hard time identifying non-coding regions of association from rare variants in particular.

Thanks. I have a question about the SNP. Scaling up, through increasing sample size, through aggregating cases from, you know, different studies. I suppose this is a basic question from a non-statistician, but I'm always concerned about how independent those studies are and whether we are biased by throwing information that comes from, you know, not completely independent studies, meaning some of the patients in one study may be in another study as well.

How do you deal with that? [Speaker:Joshua Elkington] So sample overlap is a big problem. I think in the presence of sample overlap, you're going to get massive bias in your distribution of test statistics. That's really going to kind of hurt those kinds of results, and they're going to, you know, weaken things. And that's, I think, probably the single best argument in favor of widespread data sharing and access to individual-level data is so that you can perform checks looking not only for sample overlap at an identity level, but things like siblings or cousins. I mean, you know, the chance that multiple relatives end up in two disparate studies is maybe non-zero. And so I think if we can, as a community, move more and more towards this idea of individual-level data sharing for the purposes of, you know, genetic analysis, then I think we will be in a better position to ensure the integrity of the kinds of analyses and approaches that we're engaged in doing right now. Okay, thank you.