Okay, I'm going to share my screen. Great, so interrupt me if you want to ask a question. I'm prepared for a talk of about 40 or 50 minutes, and then we can have questions as needed. But I'm going to minimize my screen with a picture of people, so you'll have to talk to catch my attention. Okay, so this work was almost entirely led by a UCSD PhD graduate, Ali Akbari, who worked with Vagheesh Narasimhan and has been working in my research group first as a postdoc and then as a staff scientist on this project for a number of years. And the background for this work is related to what Jening just said, which is the ability to generate large amounts of ancient DNA data, and this work is powered by that. That's why Ali came to our group, to be able to take advantage of that information, which makes it possible to study phenotypic change over time with genetic data.
So the background for this is ancient DNA, which is a technology that came online as a really serious genomic technology beginning in 2010. It starts with an ancient bone. This is an ear bone of the skull that's a particularly rich source of DNA, and in a clean room like the one we have at our institution, we try to get beneath the surface of the bone in a way that protects the bone from being contaminated from the people who handle it. So there's lots of measures that are used to do this, and people get suited up, and people use gloves, and there's positive pressure in the rooms, and there's HEPA filtering, and so on. We try to get a particularly DNA-rich part of the bone. We make some powder from it, a very, very tiny amount of powder, and then we purify the powder — we extract the DNA in a solution that removes the protein and mineral content, and then we sequence it in a next-generation sequencer. This is an old one now; we use a much more powerful and less expensive sequencing technology now. Often we use an in-solution enrichment step where we take the DNA sequences, the libraries we make, and we hybridize them to sets of synthesized targets that are relevant to the parts of the genome that we use most for our analyses, because often more than 90% of the sequences from the libraries are microbial — from bacteria and fungi, for example — that colonize the individual after they died. And so we want to enrich for the parts of the genome that are informative for our analysis.
Because of the great increase in efficiency in data generation, there's been an explosion in the amount of data that's available. And so from the first five whole genome sequences in 2010, we now have more than 20,000 individuals with reported ancient DNA data. And this has been a really rapid growth in the last years, and it makes it possible to ask and answer questions that were not possible to address before. So we've been trying to distribute this collective data through what we call the Allen Ancient DNA Resource, and you can download a relatively recent update of it from there. So the focus of this talk is not about human population history, which most of my papers are about. Most of my papers are about using ancient DNA to study how people are related to each other and to people today in the past, but it's about: can we study adaptation with this data? And that's what Ali Akbari came to my research group to do. And the premise for my presentation is the kind of orthodoxy or starting point in the field, the one that I've learned through reading many papers in this area over the last several decades, which is that natural selection has been quiescent in our species in the last 10,000 or even hundreds of thousands of years.
So here's an example of an important paper that argues this. So this is a paper by Ryan Hernandez and colleagues, and what it does is it looks around non-synonymous substitutions in the genome and in variant sites, or substitutions, new mutations that are seen in modern humans but not in chimpanzees, for example. And it looks to see if there's evidence of clearance of genetic variation surrounding these substitutions, as you would expect if there was a classic selective sweep on a newly arising non-synonymous mutation that brought it to very high frequency over a relatively short period of time. And if there's a reduction in variation near these that's not seen in synonymous sites, and based on the absence of seeing these, whereas in Drosophila they're seen a lot, these authors argued that there's no evidence for substantial numbers of classic selective sweeps over the last couple of hundred thousand years. And about a year ago, Sasha Gusev at my institution summarized this evidence and said, what do we make of the fact that there are so few loci under selection in the past thousands of years?
Indeed, one of the things that Sasha focused on was the ancient DNA evidence. So beginning in 2015, in a paper I got to be involved in, we started screening the genome in ancient DNA data, taking advantage of the incredible promise of time series data to look to see if selection had occurred. So previously, when people looked at selection — and I've been involved in this for a long time now, and was involved in developing some of the tests that have been used for this — previously when people looked for natural selection, what we would look for is scars in the present-day genomes. So you look at patterns of genetic variation, and you look at scars, distortions of genetic variation compared to what you'd expect by chance that could only arise in a non-random way due to natural selection. But we all know the power of time series experiments, like watching bacteria or fruit flies evolve in the lab over time. And we can now do that with human data. We're not just trapped in two dimensions. We have the present of, you know, latitude and longitude. We can also go back in time.
So when we ran this study in 2015, we only had a couple of hundred individuals. They were all in Europe, and they were the three main sources of ancestry for most Europeans today: European hunter-gatherers who lived more than 8,000 years ago, farmers who lived after 8,500 years ago, and steppe pastoralists who moved into Europe and contribute much of their ancestry after 5,000 years ago. And we looked to see whether frequencies of lots of genetic variants in the genome today are different from what you expect from the mix of those ancestral populations that we know occurred in real life. And we found 12 locations that looked genome-wide significant after this analysis. Some of them were known and some of them were not known, and that was very exciting, and we thought maybe as we gather more data we'll find a lot more. But last year in 2024, the largest such study only found 21 such locations in the genome. And it felt really like the power was asymptoting, and that we might not really find a lot using this approach we were so excited about.
So what we did, and what Ali did, was to work on a variety of approaches that were going to increase power and theory and reduce false positives from these scans. So the first was a new statistical method, at least new to studies of natural selection in humans. There's no explicit modeling of demographic history. So in our paper from 2015, for example, we knew what the right mixtures of these ancestral populations were to form present-day Europeans, and we mixed them in those proportions and saw whether today people deviate from it. We simply ask a simple question, which is: we use the data itself to teach us how closely related each sample is to each other. So we have, say, 16,000 people in our data set — 16,000 ancient and 6,000 modern people in our data set — and we have 10 million positions we're analyzing, and we multiply this table of 10 million positions times 10 million positions to see how closely related each individual is to every other individual, and we get a genetic relatedness matrix which measures how closely related every individual is to every other individual. And we don't use labels of time. We don't use labels of space. That's all we use: the genetic data. And then we try to predict the genotypes — you know, adenine-adenine, adenine-cytosine, or cytosine-cytosine — at a particular point based on the genetic relatedness matrix and the degree of relatedness of that person to all other people in the data set. And so that's the game we're playing: we're trying to predict genotypes based on the relationship of each sample to every other, and we try to make the best linear prediction that we can.
And then we ask the question: if we give the model an additional piece of information — a single point estimate of the time of the sample — does that predict the genotypes a little bit better, significantly better? And basically what we're asking is: does natural selection keep blowing the frequencies again and again over time, in each time and place, always in the same direction by the same amount? And so we're simply asking: does that model, where time matters and has been changing the frequency at a constant rate, is that a better fit to the model than not needing to use the time, after correcting for the extra degree of freedom that that adds?
And so we implemented this approach. Another thing that we did is we massively increased the sample size. So the sample size has increased 14-fold compared to the largest previous study. Almost 16,000 ancient people with super high-quality data are included in the study, and more than 10,000 individuals with newly reported data are released in this study. And then the data quality is also improved through a series of data cleaning steps. When we applied this to the data, we got this incredibly weird result that we already had several years ago, and we were just very incredulous about it. So the result was that instead of 21 genome-wide significant locations, we found hundreds and hundreds of them. And we thought this must be wrong. And so we spent several years trying to make this go away. But we believe it's correct. And in fact, these are really just the tip of the iceberg. So these are sites that we think have a 99% posterior probability of being real or more. But if you look at the top 7,000 positions, we think half of them or more are real.
So one reason that we think they're real is the following. So if you take each of these approximately 10 million positions in the genome, we have a score for selection. That's on the x-axis. We call the X score. And then we can also ask for each of these variants: what is the probability that they're associated to some trait in an association study, a genome-wide association study? So you probably most of you know that genome-wide association studies, what they do is they take millions of genetic positions and they see if these positions modulate a trait. For example, in the UK Biobank, which has studied half a million British people for a variety of traits — hundreds and hundreds of traits. So what we did is we took the entire UK Biobank. We took 452 traits that had been scanned in the UK Biobank and that were well-curated traits, and then we looked at the full list of positions that were significant in at least one of the 452 scans. So that was, I think, 15 or 20% of all of the 10 million positions in the genome were associated to one of these traits, and very convincingly so.
And so then we ask the question: as we slide the X statistic upward, is there a change in the fraction of these sites that are associated to some phenotype, like height, for example, or body mass index, or schizophrenia, or blood pressure, or something like this? And what you see as you slide it upward is a heavy enrichment, until instead of 15 or 20%, it's something like 70% of the positions are associated to some GWAS trait, and then after about five or so it plateaus and flatlines. And our interpretation of that, our hypothesis, is that what's going on is that above this point everything is real, and therefore increasing the significance does not cause further enrichment, because all of these positions are genuine signals of selection, and we're using a phenotypic impact, an independent data set in genome-wide association studies, as a readout of function.
So we were incredulous about these results and thought that maybe some other phenomenon might be contributing to this. Maybe, for example, signals of selection might be easier to detect in functionally important regions, and functionally important regions might also be enriched in GWAS signals. So we simulated that. We carried out semi-realistic simulations of a population history sort of inspired by that of Europe and the Middle East. We simulated what's called background selection or purifying selection, where there's selection to remove the rain of slightly bad mutations or very bad mutations that rain down on the genome. And you know, it occurs. We know it occurs. We also simulated stabilizing selection, which is selection to remove very polymorphic positions that affect traits, because those will create more variance in a trait which is ideally less optimal. And none of those types of selection can generate even a hint of the observation that we see. And when you control for proximity to genes in various ways, you can't even see a hint of the rise. But then when you simulate directional selection, you see exactly the type of signal that we're observing. The fact that it goes 40-fold in our directional selection simulations rather than five-fold is just related to the nature of the simulation: a smaller fraction of the simulation were real hits.
We also used an approach called the haplotype score, which was developed by Ali and his colleagues in Vagheesh's lab when he was a grad student, which is a statistic that is sensitive to natural selection. So if there's a partial selective sweep, or a sweep that is in the process of becoming a complete sweep, there is a period of time where the squared frequency of genetic variance around the position of the sweep will go up relative to what it is otherwise. And in all other non-directionally selected scenarios, it's expected to go down. So in our data, we see that when we slide up the haplotype score, there's a rise also around five. And this suggests that this is potentially a meaningful threshold of significance. And again, when we simulate, we see only in the case of directional selection do we see this haplotype score rise. So we have these two independent lines of evidence that this is a real signal.
So now going into the interpretation of the results, the data show overwhelmingly that not only are there hundreds of positions of the genome that are affected by directional selection in this period, but there has been an intensification of natural selection over this period in Europe and the Middle East. And I'll show you the evidence of this. So the first thing that we did is we looked at about 107 genome-wide association studies, and we looked to see whether the variants that were involved in these studies had an enhanced probability beyond background of being associated to a particular trait. And we see a huge five-fold or so enrichment for blood, immune, and inflammatory traits, indicating immune traits are unusually selected, and almost no signal at all for behavioral and mental psychiatric traits. So this is actually a third line of evidence that this is real, because it suggests that these variants are very differently affected by selection in their interaction with complex traits depending on what type of variant that they are.
You can also do something where we compare the intensity of selection in different time periods: the period earlier than 5,000 years ago, the Bronze Age from 6,000 to 2,000 years ago, and the historical era less than 3,000 years ago. And what we find is that there's more intense selection at these variants, with an increase in the magnitude of selection coefficients from the Mesolithic-Neolithic period to the last 5,000 years, a period when people began to live at much higher densities and much closer to animals and potentially more at risk for infectious diseases coming from animals. This is of course speculative, but also very different economies and cultures and ways of living. And again, if we look at behavioral traits or mental psychiatric traits, we don't see evidence in this analysis of an increase.
There's also evidence not just that it's intensified, but that selection has fluctuated over time. So if you look at the approximately 479 positions that are genome-wide significant, and you plot the age of the origin of these mutations based on patterns of linkage disequilibrium around them — how large the haplotypes are around them, which provides a proxy for when they arose in the genome. This is an estimate based on a genealogy reconstruction software called Relate. You can then plot that against the age of the variant if it had arisen as a new mutation and was subject to the selection coefficient that we estimate over the whole history since it arose. And if it's been under a partial sweep and constantly selected, you expect the variants to cluster on the x-y axis like these. And you see a lot of those, but it's only about a third of them. The remaining two-thirds fall off the line, and in fact many of them are ancestral variants that are favored, which means that at some point these variants — the new mutation is not under positive selection, the ancestral one is — and so there is reversal of selection on these traits.
So not only is there an acceleration of selection, there are changing selection coefficients. And probably the reason why you don't see — or potentially the reason why we don't see when we look at scars in modern human genomes evidence for lots of selective sweeps — is that some of these variants have been going up in frequency and going down in frequency over time under the pressure of selection. But for example, we don't see 100% differences between Europeans and East Asians very much under the pressure of selection, simply because they're going up and down.
So here's some characteristics of the variants that we've identified. So again, here's these 479 positions plotted based on their frequency of the minor variant versus the selection coefficient of that variant. And what we see is that the ones that had been identified before by the earlier scans are these labeled ones, and they have huge effects. So very strong selection coefficients. That's why it was possible to detect them using smaller sample sizes and different methods. And they don't seem to particularly have a strong relationship between frequency and between the strength of selection. There's very strong variants that are like KITLG, which is related to pigmentation, over at this site. There's other ones that are strong that are at 50% frequency. So there doesn't seem to be a strong relationship to frequency, and maybe these are cases of balancing selection or partial sweeps or sweeps in process. The important point here is that they're very atypical.
When you have more power, you start to see this motherload of genetic variation, these many gray dots, and they have this very strong relationship to frequency: fewer strong effects with high frequency, many more strong effects with lower frequency. This is exactly what you see with variants that contribute to genome-wide association studies. It's what you see if there's stabilizing selection to remove variation, and suggests that we might be able to learn something about these selective variants, about the way that selection happens, about the way that variants that affect traits evolve.
So now I'm going to talk about some individual loci. So as you remember, our test was for a constant selection coefficient over time, but we can also apply this in intervals. And so here is a variant that is the strongest known risk factor for multiple sclerosis. A previous study by William Barrie and colleagues had found evidence that it had rapidly increased in frequency after 6,000 years ago. Here's European hunter-gatherers, here's the farmers, here's the steppe pastoralists. But it kept going up. But they didn't notice that we can see with this much higher resolution that it crashes after this point. There's a reversal of selection. So something happened potentially immunologically that meant this made this advantageous before 2,000 years ago and disadvantageous after 2,000 years ago. We also see when we break it up into different parts of Europe that it's super advantageous in Northern Europe, with a really truly massive selection coefficient, even stronger than lactase persistence, and a very, very strong reversal. In southwestern Europe, Iberia for example, there's a much weaker selection coefficient, and it doesn't even clearly crash at all.
Here's two more cases of fluctuating selection. So here's the TICAM2 risk factor for severe tuberculosis. People who have two copies in some context will have an 80% lifetime risk of clinically severe tuberculosis. A previous study by Gasparyan and colleagues had found that it had decreased in frequency based on the ancient DNA available at the time in the last couple of thousand years, and we replicate that, but we detect also a rise before that time, before two or three thousand years ago. So again, there seems to have been very clear positive selection pressure to increase its frequency, then a reversal. Also here's a very common risk factor for hemochromatosis, and again evidence of a reversal. So what you can see is fluctuating patterns even in these variants.
Here's examples of variants that are under consistent positive selection: two risk factors for celiac disease, gluten insensitivity, which rocket up in frequency, both of them after around 4,000 years ago. You might think that variants that make a person susceptible to celiac disease might be deleterious, might be bad for the organism, but both of these variants were positively selected, presumably because they may have had some beneficial effect.
So I wanted to summarize a little bit what we learned so far from single-variant analysis. There we have evidence of strong directional selection in Europe and the Middle East acting on hundreds of genetic variants in the last 10,000 years. Natural selection has intensified since the advent of agriculture, mostly immune-related. Selection has fluctuated over time. You can go to this website which we call AGES, and you can type in your favorite SNP variant and look at the trajectory of selection over time at this position, and see whether there is evidence of selection at this variant, as there is at this one shown here, for example. You can also, in the same browser, correlate it to the effects on this trait and all of the 452 traits we analyzed in the UK Biobank, and look at the colocalization of selection and GWAS traits.
So now I'm going to move on to complex traits, traits that are underpinned not by a single variant but by many variants, and see what we can see with this type of data. So I'm going to introduce the concept of the polygenic score, which some of you are probably familiar with, but for that I will nevertheless repeat this so that you can have a sense of what it's about. So a polygenic score is a combination of genetic variants that are predictive of a trait. So a very good and clear case of a polygenic score is the one that's been built for height. Loic Yengo and colleagues a few years ago published a genome-wide association study based on data from more than five million people of European ancestry whose genomes had been scanned. And in this study they found more than 12,000 independent positions that independently modulated height, many of them by very sub-millimeter effects, but together were highly predictive of height and predicted almost all the common genetic variation associated with height.
And so what we do with a polygenic score is simply do a sum over all of the positions that are genome-wide significant, or you could actually sometimes use non-genome-wide significant ones as well. And you look at the genotype a person has: zero, one, or two copies of the variant that increases height, multiplied by the effect size of that variant on height. You know, increasing by 0.2 millimeters. So having one copy of the variant will be 0.2 millimeters higher prediction, and two copies will be 0.4. And you just add it up over all 12,000 variants. So that's the polygenic score. And the test for selection we want to carry out is: the change in polygenic score over time more than could be due to chance? That's the question we want to ask. And we do this in a variety of ways. One of them is we do an analysis where we randomize the effect direction for each variant, and we do it many times. So sometimes it's positive, sometimes it's negative, and that will cause a cancellation of the signal in the random traits, and then we see if the results that we have are more than we would expect by chance.
Another thing we do is we don't even use the effect size; we just use the sign. And the last is we just look for whether there's a significant correlation coefficient between the effect sizes as measured by GWAS and the effect sizes as measured by our selection coefficient estimate from our scan. So those are the three tests, and we have out of the 563 genome-wide association studies we analyzed — which were all the UK Biobank ones plus about a hundred additional ones that we picked to especially focus on dichotomous traits like type 2 diabetes or schizophrenia or bipolar disease, for example — we had 100 traits that were highly significant, that had a polygenic score change over time that was too large to be affected by chance.
So almost like a kind of control for this analysis is pigmentation. So lots of evidence from single variants, actually, as well as makes sense evolutionarily, has shown that Europeans have depigmented over the last 10,000 years. So if you look at European hunter-gatherers, they have, for example, from western Europe, the predicted skin pigmentation based on known predictors of skin pigmentation is extremely dark, almost sub-Saharan African. And the early farmers, it's lighter, but still much darker than today. And the steppe pastoralists, again, lighter, but still darker than today.
If you look at the polygenic score for skin pigmentation, it's dominated by a few very strong effects, but there's actually hundreds of weaker effects, and there's a consistent trend over time decreasing by about two standard deviations over this time scale. So you can also pull a sliding window of say 2,000 years and estimate what the selection coefficient is over time in each sliding window for pigmentation. And you can see that at the beginning of the period, from like 6,000 to 4,000 years ago, there is selection for depigmentation at a modest level — well, pretty strong, but still a modest level — at a level of about one to two standard deviations on the scale of modern pigmentation over a 10,000-year period. And then between 4,000 and 2,000 years ago, there is much, much more intense selection for depigmentation after controlling for population structure. And in the most recent last couple of thousand years, the selection for depigmentation is still very strong, but less than in the previous couple of thousand years.
We also see very powerful evidence with approximately one standard deviation effects for a set of traits related to...
Type two diabetes risk. So this includes body fat percentages, waist circumference, waist to hip ratio, body mass index, indeed type two diabetes itself. And what you see is a consistent downward trend in the risk factors for type two diabetes. You can again do this trick of pulling a sliding window through the data and asking is there a change in the selection coefficient over time? And there's no clear evidence over time in the strength of change. Maybe over here there's less strong selection for waist to hip ratio in the most recent period than in some of the earlier periods.
One thing that's important to keep in mind, and it will come up again and again in the next couple of slides, is that genetics here is pushing potentially in the opposite direction to phenotype at least recently. So, it's not the case that people in Europe today are less obese or less prone to type 2 diabetes than they were thousands of years ago, even if they're genetically less prone because there's also environmental risk factors. So, for example, people have very strong access to food that is likely to contribute risk for type two diabetes or to obesity.
So next one I wanted to show you about were clear evidence of selection for variants that decrease risk for psychosis. We get very strong signals for bipolar disease and schizophrenia again decreasing by close to a standard deviation over 10,000 years on the variation scale of Europeans today. And these are correlated traits in terms of their polygenic score. They're correlated traits in terms of their phenotypes, and we see a correlated selection signal to decrease genetic risk for it. Again, this doesn't mean that there is less schizophrenia and diabetes now than in the past. There might be more because possibly we live more in environments that might make people susceptible to these things.
Then there's selection for variants that increase health span. So, measurements of overall health decline and walking pace are also strongly selected. And here you see very strong fluctuations. For example, the strongest signals are for four to 5,000 years ago for these traits and weaker in the last couple of thousand years. And then finally there's selection on variants that affect these social science measures of cognitive performance. So people's scores on IQ tests, their household income or their years of schooling. Now, these are ridiculous phenotypes because they were unmeasurable and would not have existed in the past. So, if you look 5,000 years ago, there's no school, there's no IQ test, there's no income. Yet, nevertheless, the genetic variants that modulate these traits today were systematically being selected. So, it's probably not, certainly not those traits that were being selected. Certainly not schooling, for example, or household income. But it was something else that's correlated to them.
You also see that there's huge fluctuation over time. So if you look at the polygenic score that predicts performance on IQ tests, there's very, very strong positive selection from about 5,000 years ago to 2,000 years ago. And then there's almost no selection at all in the last 2,000 years. If you asked sort of sociologists or other people when is there going to be the most strong selection for this trait, I don't know what they would say, but I don't know if people would have argued that it was the last 2,000 years that there was the least selection for this trait. So a caveat as I said is there's no IQ test, wealth or school in the deep past, what was selected is unknown. Out of the hundred traits where we detect signals like this, there is many other traits that are interesting here. Some of them are for example food preferences like cereal intake or the type that you choose, you know, do you like Cheerios or frosted flakes or whatever, cheese intake. There's also other behavioral cognitive traits like ADHD, age at which people did various things such as first sexual intercourse, and also a series of traits related to lung behavior and to smoking behavior. So it's a rich data set that's interesting to query.
We were very, very skeptical about the cognitive performance results but they replicate in East Asians and I'll show you the evidence for this. So we had 31 traits where we could perform an analysis because we had access to genome-wide association studies also in East Asians, either in Chinese or Japanese people. And what we did is we did the following test. We measured the effect sizes of these 10 million variants on the traits in East Asians and then we correlated them to the trajectories in Europe, which has an essentially uncorrelated population history to East Asia over this time period. So there's essentially no way for these variants to be correlated unless this is really a real signal. So if you look at the 31 traits, what you see is here is the correlation between selection and GWAS in Europeans and here's the correlation between selection and GWAS with the East Asian measurement. And you see for this number of years of schooling trait, there's huge signals, five standard deviations or six standard deviations whether you measure the effect sizes in Europeans and East Asians. We don't see how that could be a false positive. And so it suggests that somehow these variants whether you measure them in Europeans or East Asians are being selected. We also measured these in a genome-wide association study based on sibling pairs where there's two siblings that have exactly the same family environment that's immune to population structure confounders and even though it's only a p value of 0.04 or a two or three replication, it's also a replication.
So what phenotypes were adaptive in the past is unclear. It's very confusing what's actually being selected. So if you look at for example 12 traits that we analyze that replicate powerfully across all three tests of selection for polygenic score that we look at, here is the correlation of their polygenic scores. And what you see is that pigmentation is uncorrelated to anything else. It's being selected on its own trajectory. It's not related to any of the other traits. Bipolar disease and schizophrenia, they're not correlated to anything else but each other where they're strongly correlated. And then all the other traits are in a big block of strong correlation. And they're quite different traits like smoker status, score on IQ tests, walking pace, body fat percentage, health decline rates. You know, it's not clear if what you're measuring is really IQ or is like a type two diabetes related trait. These are all correlated to each other in one way or another. So one almost wonders whether there's some kind of metatrait that is being selected and toggled back and forth and advantageous or disadvantageous at different times. Maybe something related to delaying gratification or executive function or something.
The other thing that we've been doing recently, and this work is not quite as advanced but is we're very excited about, is we're trying to repeat this analysis in people of East Asian ancestry. The ancient DNA data from East Asians is much poorer than that of Europeans because the ancient DNA field started in Europe, but it's rapidly progressing now. And most people of East Asian ancestry for which there's ancient DNA data actually don't come from China or Japan. They actually come from Central Asia and even Europe because of East Asian Avars ancestry, Avars who appear in Europe. Our current version of the scan actually cuts out the data set west of this point just so that we don't have any overlap geographically with our West Eurasian scan. And so what we did in these 2300 ancient people is we looked to see whether there were changes over time. So we found in this version of the analysis 26 locations with genome-wide significant evidence of natural selection, many fewer than in West Eurasia but the study is much poorer in terms of sample size and also in terms of coherence of population structure. So there's every reason to think that as sample size increases one will get more powerful inference. So the thing that's most interesting to us is the correlation between the West Eurasian and the East Eurasian scans. So 26 is very interesting. Almost all of these are new and not known before and we're trying to think about them and understand them. But what's most interesting to us is the correlation to West Eurasians. If you take these 26 positions that are genome-wide significant in East Asians, so that's all of these positions and here's their selection coefficient estimates in East Asians. And then you look at their selection coefficients in West Eurasians in our independent scan. You see there's a clear correlation with an R coefficient of 0.7.
So that indicates that variants that are subject to selection in East Asia tend with exceptions like this one to be also subject to selection in West Eurasia. And if you reverse the experiment and look at the 479 positions in West Eurasia, again there's a strong correlation to East Asians, again about 0.7. So it looks like there's correlation of selection coefficients in West and East Eurasia over this time period despite the population histories being completely independent. So here's a completely crazy result. So we did the same analysis looking at the 107 traits and seeing whether these selection coefficients are enriched in certain traits rather than others. And what we see is it's exactly the same pattern. Very strong enrichment for immune and inflammatory traits. No evidence at all of enrichment for mental, psychiatric and behavioral traits. And it looks like a shared architecture in some sense. If you look at for example body mass index and the trajectory of the polygenic score over time, it goes down over time by a lot in East Asia. It goes down over time by a lot in West Eurasia. You see similar trajectories for these polygenic traits.
Here is the 452 traits from the UK Biobank. So caveat is that these are measured in Europeans, and we're looking at the correlation of the score for our test for polygenic selection based on signs not based on effect sizes in Europeans and East Asians and there's this huge correlation between the polygenic directions between Europeans and East Asians, not for for example pigmentation where there doesn't seem to be a strong effect correlation and in fact there's not very strong evidence for depigmentation being important in East Asians in the last 10,000 years. So perhaps that makes sense. But here's body mass index which I just showed you an individual plot of, fat mass which are related to body mass index traits, and here are some of these social science traits that are again correlated. So this is like a crazy result. So the result is that there seems to be correlated selection processes in both extremes of Eurasia, West and East Eurasia in the last 10,000 years as these two groups of people were exposed to parallel processes such as the advent of agriculture, increased density of people, urbanization and so on. And so it's as if there's an ancestral distribution of genetic variance that due to this set of cultural and environmental and lifestyle changes tend to get pushed in the same direction by this profound impact of living differently that's happened to many people from our species in the last 10,000 years.
So thank you again. This work is overwhelmingly led by Ali Akbari. Allison Barton has been leading the East Asian work along with Ali Akbari. Annabelle Perry carried out many of the computer simulations. Nadine Roland runs as the director of our wet laboratory which generated this more than 10,000 newly reported ancient DNA samples which will hopefully be available publicly within the next couple of months. And Sha Malik is our bioinformatics director who's been responsible for organizing the processing of these data and many other wonderful colleagues who we've worked with on this project, archaeologists and anthropologists whose new data are being reported in parallel publications, and the ancient and modern people whose genetic data we analyze. This work was supported especially by the Howard Hughes Medical Institute, by the NIH, by the John Templeton Foundation in a generous grant that allowed us to collect a lot of this data set, and also from the Allen Discovery Center for brain evolution that supported this work too over the last years. Thank you.
