463: Why brain traits leave faint, common genetic signals

[00:00:20] Welcome to Base by Base, the papercast that brings genomics to you wherever you are. Thanks for listening and don't forget to follow and rate us in your podcast app. [00:00:30] Base by Bass is now on YouTube too, at base by bass, where every episode gets a video with chapters and the full description. Come subscribe. [00:00:44] Imagine standing in a crowded hall, trying to pick out individual voices. In one room, a few people are shouting. You hear them clearly one by one, and you can point to each of them. In another room, hundreds of people are talking, but every one of them is barely above a whisper. You know they are there. You can count them, but each voice only just rises above the noise. [00:01:08] Geneticists have been standing in both of those rooms for about 20 years, and the strange thing is that the second room keeps turning out to be about the brain. [00:01:20] Here is what that looks like in practice. A genome Wide association study scans the genomes of many thousands of people looking for variants that are more common in people with a trait. Each signal that clears a strict statistical bar is called a hit. For a trait like LDL cholesterol, the hits are loud. Some of them clear the bar by an enormous margin. For schizophrenia, studies of similar power have found a similar number of hits, yet almost every one of them clears the bar by only a hair. [00:01:55] There's a second oddity hiding in the same data. The schizophrenia hits tend to be common variants carried by a large share of the population. [00:02:04] Why would that be? A disorder that lowers the chance of having children should push its risk variants toward rarity, not toward being common. So if Is the pattern a statistical artifact of how these studies are run, or is it telling us something real about how evolution has shaped the genetics of the brain? [00:02:23] That is the question behind today's episode. [00:02:31] Today we celebrate the work of Hui Xiang Tzu, Yuval Simons, Jeffrey Spence, Guy Sella, and Jonathan Pritchard from Stanford University, the University of Chicago, the University of California, San Francisco, and Columbia University, who have advanced our understanding of how natural selection shapes the genetics of brain related traits. Their paper, Genetic Architectures of Brain Related Traits are Shaped by Strong Selective Constraints, was published in the Proceedings of the National Academy of Sciences in September 2026. [00:03:08] Let's start with a word that will carry the whole story Architecture. [00:03:13] The genetic architecture of a trait means how many variants contribute to it, how common they are, and how big their effects are. [00:03:21] Some traits have a few variants with large effects. Others have thousands of variants, each nudging the trait by a tiny amount. [00:03:29] Over two decades, genome wide studies have found tens of thousands of robust associations across a wide array of traits. [00:03:37] And the architectures they reveal vary a lot from one trait to another. [00:03:44] Why does architecture matter? From a practical angle, it decides how well a study can map a trait and how accurately we can predict that trait from someone's DNA. From a deeper angle, it is a record of evolution. The number of variants, their frequencies and their effect sizes all reflect the four forces that have acted on a trait over thousands of generations. [00:04:06] Read the architecture carefully and you can learn something about the selection that shaped it. [00:04:14] Recent work by Simons and colleagues gave researchers a way to read that record. Their model assumes what is called pleiotropic stabilizing selection. Many traits have an optimal value and fitness drops as you move away from it in either direction. [00:04:29] And most variants touch many traits at once. [00:04:32] Under that model, a new mutation is selected against whether it raises or lowers a given trait. The model fit the architectures of 95 quantitative traits better than earlier models did. But it was built for quantitative traits like height or cholesterol, not for diseases. [00:04:51] Meanwhile, psychiatric genetics kept running into the same puzzle. [00:04:56] Several studies had already suggested that traits linked to the brain are unusually polygenic, spread across an enormous number of variants, with effect sizes much smaller than for other complex traits. But there was a real worry hanging over all of it. Disease studies and trait studies are designed differently. A case control study of a disorder has different statistical power than a study measuring a continuous trait. [00:05:21] Could the strange pattern simply be a side effect of study design? [00:05:30] The first step was to decide which traits count as brain related without guessing. [00:05:36] The team used a method called stratified LD score regression. Put simply, it asks where in the body a trait's heritability is concentrated. By checking whether the genetic signal population piles up in the active regulatory regions of particular cell types. [00:05:52] They tested 220 cell types grouped into 10 categories. [00:05:57] A trait whose signal concentrated in cells of the central nervous system was labeled brain related. [00:06:03] The study used brain related as shorthand for exactly that kind of enrichment. [00:06:11] They applied this to 151 quantitative traits from the UK Biobank, each with at least 20 roughly independent hits, and to 13 complex diseases with public data. [00:06:23] 50 of the quantitative traits came out as brain related, including seven behavioral and cognitive traits such as fluid intelligence and the age at first sexual intercourse. [00:06:33] Among the diseases, four were brain related and all four were ADHD, bipolar disorder, major depression and schizophrenia. Schizophrenia. The other 101 quantitative traits and nine diseases formed the comparison group. [00:06:51] Next came the design problem. How do you compare a disorder with a continuous trait? Fairly, the team leaned on an old idea called the liability threshold model. Think of a hidden dial for every person. When the dial passes a threshold, the disease appears. [00:07:08] Using that idea, they derived how much power a case control study has compared with a study of the dial itself. [00:07:15] Then they tested the math directly. They took LDL cholesterol, a continuous trait, and turned it into fake diseases by calling the top slice of people cases. [00:07:28] They did this at several thresholds, from 20% of people down to 1%. [00:07:34] Converting the trait into a disease weakened the signals, and so did simply shrinking the sample. In both cases, the hits lost strength along the lines the formulas predicted. That gave them a tested way to put every study on the same footing. They could now shrink a large study down to the effective size of a smaller one and ask whether the differences in architecture survive. [00:07:59] The last tool was the evolutionary model itself, extended to handle diseases as well as continuous traits. Here is the each new mutation gets a selection coefficient, a number for how strongly evolution pushes against it. The model also has a mutational target size, the number of sites in the genome where a mutation would affect the trait. Selection decides how common a variant can become. [00:08:25] Then a simulated study decides whether that variant is detected. [00:08:29] Fit the model to the real hits, and you can estimate both quantities. [00:08:38] So what did they find? Start with three traits that each produced a similar number of LDL cholesterol, coronary artery disease and schizophrenia. All three studies found roughly the same same count of independent signals. On a Manhattan plot, which shows every variant's strength along the genome, LDL and coronary artery disease have towering peaks. Schizophrenia has none. Its signals form a low, even ridge. Almost all of them narrowly exceed the significance threshold, and very few go far beyond it. [00:09:14] The difference shows up in the numbers, too. The schizophrenia hits have a narrow range of strength, so their curve climbs steeply and stops early. And when you plot how common the variants are, a new pattern appears. Schizophrenia hits are generally higher in frequency. They are common variants, while the LDL hits include many rarer ones. Weak and common. That is the signature the team set out to explain. [00:09:43] Is it just schizophrenia? No. The same signature appears across the board. The other psychiatric disorders, adhd, bipolar disorder and major depression share it. So do the behavioral and cognitive traits measured in the UK Biobank. Their hits span a narrow range of significance and skew common, while the other quantitative traits look different. [00:10:06] Interestingly, neurological diseases such as Alzheimer's, Parkinson's and multiple sclerosis were not classified as brain related here. And they did not share this architecture. [00:10:18] Put every trait on one map with the typical frequency of its hits on one axis and their typical strength on the other. The brain related traits gather in one corner. High frequency, low strength traits enriched in other tissues such as the adrenal gland, the kidney or skeletal muscle also cluster, but none as sharply apart. [00:10:39] Body composition traits, which are shaped partly by the brain through appetite, land in between. [00:10:45] That middle position is exactly what you would expect if the tissue behind a trait shapes its genetics. [00:10:54] Now the crucial could study design explain all of this. [00:10:59] The team took the LDL and coronary artery disease studies and shrank them on paper to the effective size of the schizophrenia study. About 192,273 people on the liability scale. After that, power matching the gap remained. Schizophrenia hits were still weaker and still more common. [00:11:19] They also checked that the pattern was not driven by population structure and found the strongest signals behaved alike in of front family based studies. Study design was not the answer. [00:11:32] So the team turned to evolution. They fitted the model separately to brain related and other traits. For the brain related traits, the inferred distribution of selection shifted towards stronger selection. Variants affecting these traits are on average pushed against harder by evolution. And there was a second difference. The mutational target was far larger. The median target for brain related traits was about 1.32% of the genome, compared with 0.27% for other traits. Roughly five times as many places where a mutation matters. [00:12:08] How large is that? About 10% of the genome is thought to be under selection and putatively functional. So a target of 1.32% is a big slice of everything that matters. [00:12:19] Here is the striking. The heritability itself did not differ much between the two groups. What differed was how that heritability was spread out. Brain related traits spread the same total heritability across a much larger number of sites. So each site carries less of it. That alone makes each signal harder to detect. [00:12:41] But that still leaves the puzzle of frequency. Why would the stronger selection produce common hits? Simulations gave the answer and it is counterintuitive. Without any study filter, stronger selection makes variants rarer, as you would expect. But a genome wide study only sees variants with effects large enough to clear the bar. Under strong selection, only variants with small effects can drift up to high frequency. Those few common ones are the ones assigned study can catch. Stronger selection filtered through detection leaves common hits behind. [00:13:19] There is an obvious objection here. All of this was inferred from the sane hits that were being explained. So the team looked for independent evidence at the level of genes they used rare loss of function variants, mutations that break a gene together with an existing measure of how strong strongly each gene is constrained by selection. [00:13:40] Then they asked how strongly breaking a gene shifts brain related traits compared with other traits. [00:13:48] The answer lined up with the model for the most constrained genes. Breaking them had larger effects on brain related traits than on other traits. For the least constrained genes, it was the other way around. The two curves crossed if the brain if brain related traits were simply more heritable, one curve would sit above the other everywhere. Instead, the crossing says that what matters most for the brain is tied most tightly to fitness. Selection shows up at both levels variance and genes. [00:14:22] These evolutionary numbers also explain the history of the field. Take major depression. Early studies found few or no significant hits. And then recently the number of discoveries shot up. The model reproduces that trajectory. Because each site carries so little heritability, a study needs a large sample just to start crossing the threshold. Once it does, the huge mutational target means discoveries rise steeply. The authors expect future hits for brain related traits to grow faster than for other traits as studies get bigger. [00:14:58] Why would the brain be under stronger selection? The authors point to several clues. Genes expressed in the brain tend to be longer and more complex. Disrupting the machinery that regulates gene activity often produces developmental delay or intellectual disability. And 83% of chromatin modifying disorders involve intellectual disability. [00:15:19] Among highly constrained genes, 30% are highly expressed in the brain, compared with 14% in lymphocytes, the next tissue. The brain seems especially sensitive to small changes in how genes are turned up or down. [00:15:35] Now the limits and they matter. The brain related traits studied here are best seen as proxies under the model. Most selection on a variant comes from its effects on other traits. So the study cannot say which specific traits evolution is actually acting on. In particular, it does not mean that intelligence or any behavior is itself under stronger selection. The authors are explicit about that the traits tag the variants in genes involved in essential brain processes. [00:16:07] There are other caveats. The analyses used cohorts of European ancestry because the demographic model was built built for that population. [00:16:15] Family based studies give the cleanest estimates of direct genetic effects. But they are still underpowered, so confounding could not be completely ruled out. [00:16:24] The rare variant burden tests covered only quantitative traits, not the disorders. And the model assumes that variants act through many traits at once. A high dimensional space that the authors tested in simulation but cannot observe directly. [00:16:44] So here is where this leaves us. Psychiatric disorders and other brain related traits share a distinct genetic signature. Many hits, each barely significant and unusually common. That signature survives a fair comparison of study power, and it is best explained by evolution. Traits that were run through the brain have enormous mutational targets, and the variants and genes behind them face stronger selection. [00:17:12] Back in our crowded hall, the whispers are not a fault of the microphones. They are how the brain's genetics was built. [00:17:22] More broadly, the tissue through which a trait works seems to shape its genetic architecture. That idea reaches well beyond psychiatry into every complex trait we try to map and predict. What does this mean for the future of genetic studies of the brain, and for how many whispers we will need to hear before the picture becomes clear? [00:17:48] This episode was based on an Open Access article under the CC BY 4.0 license. You can find a direct link to the paper and the license in our episode description. [00:17:59] If you enjoyed this, follow or subscribe in your podcast app and leave a five star rating. If you'd like to support our work, use the donation link in the description. Thanks for listening and join us next time as we explore more science base by base.