Scientist Stories: Jonathan Pritchard, Computation to study Genetic Variation

So I'm Jonathan Pritchard, and I'm a professor of biology and genetics at Stanford University.

Can you describe your graduate school experience in Marcus Feldman's group?

So Mark's really had a— unique place in sort of establishing mathematical population genetics in, you know, in sort of the human genetics world. So, he, you know, so, both on the research side as well as in terms of training an enormous number of people. So, on the research side, he's, you know, he kind of developed a lot of mathematical theory, for example, in understanding cultural evolution. As well as his work with Luca Cavalli-Sforza in particular on developing the Human Genome Diversity Panel, and doing a number of really key pioneering studies in understanding patterns of human genetic variation. So, when I came to Stanford as a graduate student in 1994, they they were in the process of putting together the Human Genome Diversity Panel at that time. And the lab tools for surveying large numbers of markers were just coming online at that time. And so, I was interested in understanding, you know, in using genetic variation to study human history. And so, You know, Mark's work in sort of, you know, establishing all of those tools and the mathematical theory was, you know, central to, you know, what I was doing as well as a lot of people in the field. David Goldstein and Magnus Nordborg graduated just as I was leaving. And then I worked closely with Karl Bergstrom, who's at the University of Washington now, Michael Lachmann, David Ardell, Lauren Ansel Myers. So, we had a very interactive group of graduate students. And then, in my last year, Noah Rosenberg joined the lab, and I actually worked very closely with him in my last year. Noah wanted to do a rotation with Mark, but Mark was going out of town, maybe even a sabbatical, in his last— in my last summer, which was basically his first summer in the lab. And we wound up working very closely together on a project that, you know, those last few months which, you know, sort of laid the groundwork for a lot of the work that we did together over the next few years. I considered going to University of Chicago very seriously at that time. So, I would say the main difference between them and Stanford was that University of Chicago kind of had a more— sort of had a stronger evolutionary population genetics with Drosophila kind of flavor. And so, there were really great people there. And I, you know, I almost went there. But at Stanford, there was a heavier focus on human population genetics through Mark and Luca, and then others who I didn't actually know at the time that was making my decision like Neil Resch and David Cox. And, you know, so those were some of the things that drew me to Stanford at that time. So, we actually had an enormous amount of freedom there. And, you know, I would, you know, we'd often work for a couple of months without talking with Mark necessarily. Like, he had his own projects. He traveled a lot. But, you know, when we when we wanted time with him, he was very generous. And, you know, I think as a— like, a lot of his mentoring role was, you know, sort of really high level and, you know, helping us to develop our careers. And, you know, he's been extremely supportive of me over the years, but at the same time, giving us huge amounts of freedom to, you know, kind of explore our own ideas.

What contributions did the Human Genome Diversity Project make?

So, I think the Human Genome Diversity Panel made a number of major contributions. So, one was this idea of collecting samples of individuals who would represent the, you know, who would represent different populations and from which we could learn about, you know, sort of global patterns of diversity. And, you know, that kind of structure has, you know, was in a sense, repeated by the HapMap Project and the 1000 Genomes Project and other sampling projects since then. So, I think that the concept of that kind of sampling was really important. You know, a second major contribution, I think it's fair to say, is the work that Noah Rosenberg and Marc Feldman and I and others did on understanding human population structure using the HGDP. A third kind of area has been work on, you know, sort of specific patterns of variation. So, understanding the structure of linkage disequilibrium and how that relates across different populations, or looking at how genetic diversity declines with distance from Africa. So, that was first done in a paper by Sohini Ramachandran around 2005, which, you know, was a like a really important result that, you know, was only possible because they had gone to this enormous amount of work of creating this panel of individuals.

How did Neil Risch influence your work?

So, Neil was a professor in the genetics department when I was a graduate student. Now, I was in the biology department at that time as a PhD student, but I started taking classes in the genetics department whenever something came up that looked interesting to me. And so, Neil taught a series of classes there. The first one was with David Cox, who was another professor in genetics. And that was really my first introduction to human genetics in, you know, sort of a disease— from a disease point of view, kind of going beyond human history kinds of questions, which I had read about a bit prior to that. And so, then after I took that class, I thought, you know, this is really interesting. So, So, I took Neil's next classes, which were on statistical genetics. And so, he had this pair of classes where he talked about sort of mathematical and statistical techniques in human genetics. You know, so, for example, techniques for computing likelihoods and pedigrees and things like this. Through that, I, you know, really became interested in you know, in kind of the genotype-phenotype question. So, you know, basically in understanding how genetic variation causes disease. And I started to think about how, you know, population genetics tools that I knew a lot about from, you know, from my PhD work could be used in understanding sort of more mainstream human genetics questions. And so, one paper that Neil had written that was influential on the whole field, as well as on myself in particular, was the Rich and Mericangas paper from around 1995 or 1996, that argued that association mapping was going to be much more powerful than linkage mapping. And, you know, it's a very elegant paper. It's two pages.. And, you know, I still sometimes assign it to my students when I teach. It's just like a very, very concise paper that I think really changed the direction of a field. So, anyway, all of those things got me interested in human genetics through Neil's classes. But there was one lecture in particular that had a big impact on me, which was when Neil was teaching the transmission disequilibrium test. The TDT test. And so, he made this comment that, you know, we don't really know how to deal with population structure for case control studies of association mapping. And, you know, instead the, you know, the standard in the field at that time was to use these family-based studies. And so, through my, you know, through my interactions with Luca Cavalli's lab, You know, I knew about the work that they had been doing on using genetic markers to cluster individuals from different populations. And so, they— this is work from Ann Bocock and Joanna Mountain and others. And so, they had shown that using microsatellites that you could basically distinguish people who came from different continents. And so, I thought, well, you know, maybe we could put that idea together with the case control study question and actually use genetic data to solve this problem. And, you know, that would potentially have some big advantages because then you could do case control studies without requiring family members. And, you know, and so that would really simplify studies very much. And so, you know, I went to that one lecture and, you know, it actually changed the course of my research for the next half dozen years or so.

How did you become focused on how to infer population structure from different kinds of data?

I think that was probably about April of the year that I graduated. And so, Noah and I, you know, went back to our office and started talking about this. And so, then we started thinking, well, let's suppose you've got a case-control study. And, you know, let's suppose that you— you know, back then people were thinking in terms of sort of candidate studies of association. So, you'd have some genes that you're particularly interested in because you thought they might be involved. And, you know, of course, that turned out that paradigm didn't really work very well in retrospect. But, you know, at the time what I thought was, well, maybe we could, you know, do genotyping or sequencing of these candidate genes, but also add on a set of additional unlinked markers elsewhere in the genome that would be basically a control set. And so, then the first paper that Noah and I wrote discussed statistics of this. So, if you have a control set of SNPs or other kinds of markers, then how many would you need to convince yourself that there wasn't strong structuring between cases and controls, and that the signals you got at the candidate genes were reliable.

How did you come to work with Peter Donnelly at Oxford?

Mark's approach to population genetics I would characterize as being very mathematical. And by the end of my PhD, I decided that I wanted to become better at statistics, just add a statistical component to my work. And so, Peter Donnelly had published a paper around 1996 that, you know, kind of exemplified the sort of work that I wanted to do. And so, I contacted him and, you know, one thing led to another. And so, I wound up going over to Oxford to study with Peter. But, you know, it was really because I thought that, you know, he sort of had a statistical perspective on population genetics that I could learn from.

How did the STRUCTURE algorithm for studying population structure come about?

Out of all the projects I've done, that's the one that probably unfolded the fastest. So, when I first went to England, there was this several-months program that was going on at the Isaac Newton Institute in Cambridge. And so, I went there, and Peter was spending the autumn quarter there. And so was Matthew Stevens, who was another postdoc in the Donnelly lab, who had been there for a year when I arrived. And so, you know, maybe a week or two went by before I got a, you know, an extended meeting with Peter. And he said, "Well, what do you want to work on?" And so, I told him at that time about the problem that I'd been working on with Noah. And I said, "Well, you know, here's how far we've gotten. You know, we think we know how to detect population structure when it's contaminating a case-control study, but we don't really know how to correct for it. And, you know, we'd been doing a bunch of work on clustering methods, but, you know, we didn't really have anything that we were happy with. And so, you know, Peter liked this question, I think, and he at that point suggested that we could use Bayesian clustering methods. And So, he was at that time, I think, influenced by a paper that had come from the DNA forensics literature. So, Katharine Rooder and Bernie Devlin and others have done some work on this. And so, he kind of suggested this framework. And Matthew Stevens, who became a very long-time collaborator and friend, was with us. And so, we, you know, we kind of went to the board straight away and, you know, started brainstorming models. Matthew knew a lot more Bayesian statistics than I did. And so, you know, so Peter and Matthew started working out ideas and then I was kind of making suggestions to the extent that I could following along. And so, you know, by the end of about 2 hours we had actually sketched out the whole model. And, you know, I think I have this in my notebook somewhere. And so, you know, it was you know, the model was sketched out on the board, essentially wound up being the final structure model. And so, then I went back to my office and programmed that. And it, you know, it took a few days to program it. And I had some data that Lynn Jorde had given me kindly. And, you know, back then there wasn't a lot of human variation data that was suitable. So, but anyway, I took this dataset and I. And I ran it, and, you know, I'm sure there's a little bit of debugging, but, you know, within a week or 10 days or something, I, you know, found that we could actually get, you know, perfect clustering of the individuals from different continental groups from that. And so that example's in the initial paper. And, you know, so, of course, having done that work, you know, which took about 10 days, you know, of course, then, you know, it's a long road from that to actually having a paper that you submit. But, you know, from that point it was immediately clear that we had something that was interesting and that worked really well.

Why is population structure important for understanding human genetics?

So, population structure refers to the idea that populations that are to some extent distinct tend to have different allele frequencies.. And so, you know, all human populations descend from, you know, a single Out of Africa event something like 80,000 years ago. But since then, you know, there's been limited migration between different continental groups. And then when you look within continents, you know, populations have been separated over shorter periods of time. And that partial separation means that allele frequencies can can drift to slightly different frequencies in different populations. So, when you look at individual SNPs, you know, you don't see hard differences between populations. Generally, you know, essentially every allele will be present in all populations aside from, you know, very rare alleles. But the idea with something like, with a method like STRUCTURE is that we can use small allele frequency differences to identify the ancestry of individuals.

What makes the STRUCTURE algorithm successful?

There's been, you know, there's been a lot of work on, you know, on models that are trying to, or statistical methods that are trying to infer low frequency changes over time. You know, if you measure allele frequencies in different populations, then you can, you know, to some extent you can infer what the past allele frequencies may have been. Of course, more recently with the advent of ancient DNA, that's helped with that significantly. I think actually one of the reasons why the STRUCTURE method was so successful was that it actually ignored a lot of those complications. And, you know, by— so, in some sense, it's not really a full population genetics model. It's more of a clustering model that allows us to identify groups of individuals who have similar alleles. And then it doesn't really talk very much about, you know, sort of the historical process that led to that. But by ignoring the historical process, it, you know, kind of allows us to you know, get a lot more power out of, you know, identifying individuals who look similar in genetic terms. And, you know, so, it kind of— I think it kind of raised a new set of things that people could look at. The first motivation that we had was the idea that we would use this for association mapping. And so, we had a a follow-on paper where we talked about how to do this. But essentially the idea was that if you have case and control individuals, then when there is population structure that can cause confounding, if you don't know about it, but our idea was that we would run the STRUCTURE algorithm, learn about the population structure, and basically control for that when we're doing association mapping. And in practice, that kind of concept has you know, has been carried through, although the actual methods that people use, you know, have used more recently are, you know, mainly using different techniques. So, for example, principal components analysis. But they, you know, the concept of using a statistical method with all of the data to learn about population structure up front has borne through. The second kind of thing that we've used it for we and many others have used it for is just kind of understanding how individuals in a sample are related to one another. And so, maybe this is getting ahead of ourselves, but there's the paper that Noah Rosenberg and I published around 2002 or 2003 that looked at global population structure and showed that there's relatively strong structuring and that it follows continental groups fairly clearly. And then, of course, you know, these kinds of methods have, you know, given rise to the, you know, like sort of ancestry estimation and direct-to-consumer testing. So, you know, if you think about like something like the 23andMe ancestry reports or ancestry.com ancestry reports, you know, they're kind of using the way of describing ancestry that we had in that original model, though, you know, the methods now are a little bit different.

How does the STRUCTURE algorithm use genotype information to group samples?

You know, one reason why we wanted to do that at that time was that it was unknown what the structuring of genetic variation would be. So, we talked about this explicitly in the paper that you could You know, up until then, what people did was they would predefine groups and then test for allele frequency differences between those predefined groups. But nobody knew whether those were the right groups or not. And so, the point of structure was that you didn't need to use those predefined labels. In retrospect, it turned out actually that often predefined labels do map onto genetic structure quite well.. But, you know, it wasn't a foregone conclusion prior to doing that. And the main difference between something like STRUCTURE and what the DTC companies do now is that, you know, now they actually do start from the labels because it's known that, you know, that those map to structure pretty well most of the time. And then the last area where the STRUCTURE method is being used a lot is in conservation genetics. So it's had a huge impact. Are there. Because, you know, one of the main goals in conservation genetics is to identify distinct populations that need to be, you know, that need their own protection. And so, you know, under legislation like the Endangered Species Act, the, you know, the government is required to protect genetically distinct groups..

And so, you know, methods like STRUCTURE have had a huge impact there. How does the STRUCTURE algorithm handle genome-scale data?

You know, so, one big difference is that the original programming that I did doesn't really handle genome-scale data very well. So, the most widely used method now for genome-scale data is called admixture, which comes from a UCLA group.. And that has a lot of the elements of the original structure model, but it's got a lot of computational speedups that make it work well for that. And then there are a number of other flavors of models that have jumped off from the original structure model that, you know, have better performance in different— in certain kinds of settings. So, for example, FineStructure from Daniel Fallouche and others. Uses a model of haplotype copying followed by clustering. And so, in some ways the goals are similar to the original STRUCTURE model, but it can actually get much better accuracy in settings of very fine population structure. So, they showed, for example, that when applying this to the British Isles, that they can get really exquisite levels of structure. So, for example, individuals from Cornwall cluster separately from individuals from Devon, even though those are adjacent counties. [Speaker:JOSH] From your perspective, what were some of the key findings of the HapMap project? [Speaker:DR. SIMON] HapMap was incredibly important from the point of view of, you know, giving us the first genome-scale look at variation. So, you know, prior to that, you know, everything was, you know, pick a few, you know, pick a few regions of interest, put some SNPs in there, or do some sequencing. And, you know, we just had a very, very limited view of what genetic variation looked like, both, you know, within populations, between populations. And HapMap was the first time that we actually had a genome-scale view of variation. And, you know, one thing that my lab was particularly interested in, as well as a number of other groups, was identifying selection signals. And, you know, so, you know, for the first time we could ask, you know, what are the biggest signals of selection anywhere in the genome? And so, it was just really a, you know, major change to be able to have that scale of data. You know, like, it gives you a fundamentally different feel to, you know, to know, let's say, you know, by some measures like the lactase signal for selection is the biggest signal in the genome. You know, people believed that lactase was under selection before HapMap came along. But, you know, we didn't know whether that was going to be, you know, one signal alongside 100 other signals that would be of similar strength. And, you know, and actually, you know, lactase really stands out when you have the whole genome. As being an interesting signal. At least, you know, at least in Europeans it stands out. So, oh, the other main contribution of HapMap I think was in terms of driving the technology. So, you know, it really supported SNP genotyping. And, you know, I think that the, you know, as an outsider, my sense is that the investment that was made in genotyping technology to drive HapMap was really a step up towards being able to do genome-wide genotyping for GWAS. And, you know, of course, GWAS— the first GWAS studies came out a couple of years behind HapMap.

What were some of the key accomplishments of the 1000 Genomes Project?

So, by the time the 1000 Genomes started, we were doing quite a lot of work on statistical methods for understanding genome-wide association study data. And so, prior to 1000 Genomes, all of the work was assuming that you've got a framework of SNPs. But, you know, relatively coarse framework of SNPs so that generally the causal variant would be unobserved. And so, we put a lot of work into methods where We would try to infer, you know, the presence of something that you couldn't observe directly. And so, what 1,000 Genomes did was it really transformed the way we could think about the data. So, when you had 1,000 Genomes data, then you suddenly knew that the causal variant, you know, so long as it's reasonably frequent in the population, is actually in your data set. And so, now the way that you think about the data is greatly simplified. And so, you know, that really transformed the way that we thought about all these kinds of problems. The other kind of advance I think that was really important coming from 1000 Genomes was technology. So, you know, obviously they pushed genome-wide sequencing forward a lot. But the other thing I would mention particularly is imputation. So, at the start of the 1000 Genomes Project, there was quite a big debate about whether they should do a small number of individuals at 20 or 30x coverage so that they could get good genotype calls from each individual, or should they do a large number of individuals at something like 4x coverage. And they ended up doing a mixture of both, but it's tilted towards doing a large number of individuals at relatively low coverage combined with imputation methods. And so, you know, one of the things that they showed was that that kind of approach can work.

And, you know, and so that's influenced a lot of subsequent work, that kind of technical innovation. What has been the significance of recent studies on structural variation?

So, I think that over time people's views on the you know, the contribution of structural variation, rare variation to diseases sort of ebbed and flowed over time. So, we were actually involved in one of the early studies showing that deletions in the genome are relatively common. And, you know, so around 2007 or 2008, when it was becoming clear that the, you know, the genome-wide significant hits from GWAS studies didn't explain a lot of the heritability, a lot of people thought that actually maybe structural variants were going to be a big part of the picture. And for complex traits, I think now the answer is that they're, you know, they're a relatively small part of the picture for, you know, for contributing to heritability to, you know, to common complex traits. However, of course, you know, they are important for you know, they're important contributors as major effect loci to, you know, particularly to early onset diseases in some cases. So, you know, structural variation is certainly an important piece of the puzzle. But, you know, I think it's fair to say that in, you know, in terms of the percent of variation that it explains, it's a relatively modest fraction of of everything.

And actually, I think we could say, you know, make similar comments for rare variants.

Can you describe the concept of missing heritability? So, by around 2006, 2007, 2008, there was the first set of genome-wide association studies were, you know, were rolling out. And, you know, these studies were typically finding anywhere from, let's say, 5 genome-wide significant hits to 50 depending on the trait. And so, one of the calculations that various people did was to take the estimated effect sizes of those lead variants. And from that, to try to estimate how much of the expected heritability could be explained by those. And here the heritability refers to basically the similarity between relatives that's estimated to be due to genetic effects. And so, the general observation, which was found across a whole lot of different traits, was that those— was that first batch of genome-wide significant hits was only explaining anywhere from, let's say, 1 to 10% of the expected genetic contribution to these traits. And, you know, these were based on studies with thousands of cases and controls. And so, this raised the question of, you know, if you can only explain, let's say, 10% of the variance and of the expected genetic variance in height based on the first 30 loci, then, you know, where's all the rest of that familial similarity coming from? And so, I think that, you know, the term missing heritability, you know, I'm fine with that term. Like, I think it's a nice evocative term that, you know, kind of summarizes the question. And so at that point, there were a whole bunch of different possible hypotheses. And there's a paper from Terry Manolio and colleagues in Nature, as you know, that basically lists these different hypotheses. And so one hypothesis was that there's a large number of additional common variants that are not genome-wide significant. And these add together to explain the rest of the variants. So, it was clear at that point that there would be additional alleles that were not genome-wide significant. The question, the main question was whether those would add up to a small amount or the rest. And then there were other ideas, like it could be structural variation that wasn't very were well measured, or rare variants, or maybe inherited epigenetic effects, or other kinds of ideas that were being floated at that time. And since then, I think it's fair to say that the consensus that we've landed on is that the majority of the missing heritability is actually due to there being very large numbers of common alleles with small effects. And so, Peter Visscher's lab in particular was responsible for the first analyses that showed that's probably the, you know, responsible for the lion's share of the missing heritability.

How did studies of height illustrate the concept of missing heritability?

Yeah. So, the first paper is from Peter Visscher's lab on height. And for that they showed that the you know, that there's a lot of signal in the rest of the genome that's not genome-wide significant. But if you add all of that together, even though you don't— so, it's basically, it's using these methods where it says that even though you don't know which SNPs are actually responsible, you can estimate how much total signal there is among the non-significant variants. And if you add all of that together, in some sense, it's kind of adding up to explain most of what's missing. You know, and then I think that one thing that that didn't really answer yet was how many of those variants there were. And over the next few years, one thing that started to become clear was that there's actually an enormous number of those alleles that are not genome-wide significant, and that they they span across a large fraction of the genome. So, Alkes Price, for example, had a paper that showed for schizophrenia that if you chop the genome up into 1-megabase blocks that, in their estimate, anywhere from 70% to 100% of the 1-megabase blocks contain non-zero amounts of heritability. So, the heritability is sort of spread across the whole genome. There's, you know, there are no loci in the genome that are contributing a lot of the heritability, and most of the genome is contributing something. Can you describe the omnigenic model of genetics? So, starting maybe 3 years ago, we were working with a number of different genome-wide association studies datasets. And one thing that was striking to me was by various measures, just, you know, what a large— how such a large fraction of SNPs or genomic windows show signal of association. And as I alluded to a moment before, there are, you know, other people have also observed similar sorts of things. So, for example, Alkes Price's work also points to that. And so, actually, this kind of started as a thought experiment, you know. What would it mean if our null hypothesis actually is that, you know, maybe instead of there being a relatively small fraction of genes that can have effects on the phenotype you're looking at, what if, you know, what if essentially every gene that's expressed in a relevant tissue can matter? And so, that's basically the, you know, this omnigenic model that we that we proposed. So, essentially what we said, you know, what that hypothesis says is that if you take the set of genes that are expressed in the cell types that are relevant for driving a particular disease, then, you know, then potentially any of those genes can have non-zero effects. And, you know, so then we start to think about what, you know, what kinds of models could could fit that. So, like, if, you know, what kind of mechanistic process could give you that sort of result. And so, you know, we've done a fair bit of modeling work on this. And so, the kind of model that we like is that there's a set of genes that are functionally close to the phenotype that you're interested in. So, we refer to these as core genes. And so, specifically, we define core genes as being the set of genes where there's kind of a direct causal arrow from the genotype or expression level of that gene that goes to the disease phenotype that's not flowing through gene regulatory networks inside a cell. And then all of the other genes that are expressed in in those disease-relevant cells. They don't have direct causal arrows to the disease. We refer to those as peripheral genes. And because of the way we set up that model, that implies that the only way that they can matter is that their effects, for example, their gene expression, is perturbing regulatory networks in such a way that it perturbs the behavior of core genes. And. And so, what we believe is that, based on some, based on both observation and mathematical modeling, is that although the biggest signals probably often come from core genes, the core genes themselves actually only contribute a small fraction of the heritability. So that it depends a little bit on the structure of the network, but you can actually for quite reasonable models, it turns out you can have a very small fraction of the total heritability coming from these core genes, even though in some sense biologically they're the most important. And so, you know, this kind of connects to the missing heritability argument, because it's, you know, in the, you know, the sort of preconceived notion that we had before GWAS studies was that there'd be, you know, some limited number of genes that are relatively close to the phenotype. And so, when you do mapping, the hits should concentrate in those or in a subset of those. But instead, what we think is happening is that there's actually, you know, a huge number of genes that are contributing.

And the genes that are biologically proximate to the phenotype only contribute a relatively small amount to heritability. Does the omnigenic model work for simple and complex disorders?

So, we think that this kind of pattern is quite wide-ranging. So, we have some unpublished work where we've been looking at small molecules or other biomarkers from the UK Biobank. And so, these are often things like urate or testosterone as well as lipid levels. So, these are kind of intermediate phenotypes where a lot of the biology is known. For these. So, we know often what are the metabolic pathways that produce these molecules, or what are the key transporters. And so, that seems like a setting where, you know, perhaps one would expect that actually the, you know, those genes that we know should play a role, you know, maybe you think those should explain a lot of the heritability. Well, actually what happens is very interesting. So, we find that the biggest hits are very heavily concentrated on the genes that matter. But they still only explain a modest fraction of the heritability. And there's a genome-wide background that's, you know, again, a highly polygenic signal. It's probably not as polygenic as it would be for something like height or schizophrenia. But the qualitative pattern is somewhat similar. So, you get you get a sharper signal on the genes that we know should matter, but still on this polygenic background. And I think what probably happens with more complex traits is that they're composed of many different independent pieces that are all kind of flowing through towards the same phenotype. And each of those independent pieces has this kind of structure. And then when you, you know, when you map the end phenotype, like like height, then you're basically superimposing this large number of intermediate kinds of signals. And, you know, and that quickly begins to implicate a large fraction of the genome. We actually think that it's really important to think about conceptual models for, you know, for why so much of the genome is involved in any given complex trait. So, you know, obviously there's been a lot of focus on trying to understand what are the top hits for typical traits. And, you know, that's rightly so. But, you know, I think that it is, you know, it is a surprising finding of the last 15 years that, you know, probably tens of thousands of variants contribute to a typical complex trait spread thinly across widely across the genome. And so, what kinds of conceptual models would produce a signal like that? And so, you know, our conceptual framework is one attempt to explain that. And, you know, I think that's something that can, you know, we can debate the merits of that model. But, you know, it is important for the field to try to come up with, you know, conceptual frameworks that can explain that, you know, whether or not our particular model winds up being exactly correct. Or not. I mean, the history of human genetics was people systematically, you know, overestimating their power. So, you know, as you know, in the 1990s people did association studies with a few hundred individuals. And, you know, and almost nothing came out of any of that work. And then, you know, in the, you know, by the mid-2000s, studies had gotten up to a few thousand cases and controls. And that was enough to, you know, usually to detect a handful of the biggest hits, you know. And now, with, you know, the biggest studies are, you know, meta-analyses of on the order of a million individuals or more, you know, we're getting many hundreds of hits. But it's still, you know, it's still detecting confidently only a small fraction of the true set of loci.

What are your thoughts on the development of polygenic risk scores?

I mean, it's still a little bit complicated and nobody's fully dissected exactly what's going on. But, you know, we have reasonable— some reasonable idea about that. But, you know, if it's the case that actually the, you know, the accuracy of PRS scores within a population depend very strongly on all these environmental covariates, then, you know, that that does suggest that the implementation of polygenic scores is going to be complicated. Like, we need to think harder in the long run about what exactly environment means and how the phenotype is, you know, how the phenotype is being produced by some combination of environmental factors and genetic factors. And, you know, I'm optimistic actually that polygenic scores will become useful in the clinic. But I think that there are a lot of issues that need to be worked out around this. And some of those issues relate to porting across populations, but some of them relate to things that we don't really understand at all. And actually, I mean, it's worth noting, of course, that the results showing low portability across populations you know, while in some part those are surely due to genetic differences between the populations, you know, there's very often major environmental differences. Nobody really has a good measurement of, you know, the relative contribution of allele frequency differences and environmental differences.

And I'm sure that environmental differences are important. [Speaker:JOSHUA_ELKINGTON] What are the important contributions of the Human Genome Project?

You know, I think it's just stunning, the, you know, the rate of progress that we've seen in this science during the last 20 years. So, I like to tell my students that, you know, during my postdoc, you know, I never encountered a dataset that was too big for me to open it up on the screen and scroll top to bottom in the dataset. And, you know, that's just unimaginable to them today. And, you know, that's in part a technology revolution that's allowed us to, you know, measure, you know, measure sequence data, genotype data, unfathomable, you know, scales that were unfathomable 20 years ago. But at the same time, you know, that has also completely transformed the way that we understand genetics. And, you know, I think that the, You know, these questions that we were talking about around polygenicity or the missing heritability of the omnigenic model. You know, I think that those, you know, those are examples of something where, you know, in the 1990s before we really had data on this, people kind of imagined that, you know, maybe complex traits would be driven by half a dozen loci.. And, you know, the fact that they're probably driven by tens of thousands of loci is just, you know, completely, you know, eye-opening and baffling if you look at that from a 1990s viewpoint. And, you know, that's just one example of, you know, 100 different ways that you know, the Human Genome Project and everything that's followed after it, you know, it's completely transformed our view of biology. You know, there's this paper from— another paper from Neil Risch that I like to cite, which is from 1999, where they had a sib pair study of autism. And they, you know, initially they thought there would be power to find loci that affected autism. In fact, they didn't find anything. But the beauty of the study is that they didn't stop there. Neil had these very careful calculations to look at how many loci there would have to be in order for them to have done a study of this size and in fact not find anything. And, you know, and he estimates it's, you know, it's about 10 or more. And, you know, that was actually a big deal. I remember reading that paper for the first time in 1999 and thinking, "Wow, you know, like, that's really surprising. There's a lot of loci." And, you know, It just seems ridiculously low now. So, and to your point about technology, you know, I think, you know, I mean, I draw on one example that, you know, relates closely to what we've been working on. But, you know, of course, you can tell a similar story about almost any aspect of biology that relates to genetics and genomics where, you know, the you know, the new findings that have been enabled by technology that, you know, came from the genome or the $1,000 genome sequencing and all of the related assays. You know, all of those areas have been completely transformed by that. Yeah, and you can take the, you know, pocket sequencers out into the field and, you know, in remote locations and sequence on site. You can you know, potentially do, you know, clinical sequencing of individuals in remote locations.

I mean, it's very exciting.

What areas of research are you excited to pursue in the future? You know, I continue to be really interested in this, you know, this sort of omnigenic question. You know, like, why does so much of the genome matter? And so, you know, we have You know, as I said, we have this hypothesis that, you know, most of the heritability is flowing through gene regulatory networks. But, you know, these are largely unknown or unmeasured. You know, until now we just really haven't had tools for measuring those. And so, one kind of thing that I'm very interested in is, you know, is using new technology. So, for example, CRISPR perturbations followed by some kind of measurement of gene expression to, you know, start mapping these out. And I think there's a whole bunch of different technologies that are in this kind of space that are starting to allow us to get a picture of the sort of the regulatory structure of cells. And, you know, how are, you know, how are different gene products interacting with each other to produce cellular phenotypes and ultimately organismal phenotypes. And there's really an enormous black box from genetic variant to end phenotype. And so, that's one of the things I'm really excited in for the next 5+ years is to start seeing how that piece begins to get filled in.