After the first somewhat successful genome wide association study (GWAS) of height identified ~20 genetic variants, which collectively explained ~3% of individual differences in stature (Weedon et al. 2008), David Goldstein wrote perspective in New England Journal of Medicine (Goldstein 2009). He predicted some of the trouble ahead for GWAS. He outlined that the effect sizes found for individual genetic variants in GWAS were small, and while he takes care to note this doesn’t mean the genes implicated have small effects, it does make following up on them experimentally harder as sample sizes and the number of identified variants grow. He goes on to state: “if effect sizes were so small as to require a large chunk of the genome to explain the genetic component of a disorder, then no guidance would be provided: in pointing at everything, genetics would point at nothing.“. Based on the state of the art in GWAS in 2009 he predicts that ~93.000 variants would be needed to explain 80% of the variation in height, a guess that was frighteningly close to the true answer. Goldstein suggested larger GWAS then customary at the time were definitely worth pursuing, but that they would at some point become uninteresting, and that we should pivot to rare variant analysis (GWAS only targeting more common genetic variants). In hindsight Goldstein seemed to have a laser-sharp vision of what was to come, now 16 years later we are at that point in time: The beginning of the end, the first of the last GWASs, GWAS: endgame [queue dramatic music].
GWAS: endgame
GWAS: endgame is be a series of posts I began writing in the transition betwene jobs, to help me think about the fure of my field. It once was one 15.000 word mega draft I am now frantically re-organizing into multiple posts so bear with me over the coming weeks as I drip feed you parts of the story. In each of the upcoming posts ill be covering a GWAS that is at, or nearly at, the sample size where we expect to have found all we can.
The next episode in this series will cover height, or stature, a trait of moderate or even minimal medical relevance but one that is frequently measured as a covariate or just as background demographic info. Given its neat distributional properties (height is nearly normally distributed within a region/cohort), and how easy it is to measure it is frequently seen a a sort of “model" trait, a trait to try new statistical genetic techniques on even. It was also the first trait for which GWAS got so big that a polygenic score based on the GWAS, could predict height nearly as well in a holdout sample of individuals as one could every hope to. The variance explained by that score tops out at about ~45%, which is at or near the SNP heritability (50%), a measure of the total signal in the GWAS (Yengo et al. 2022). When it came to heigth the standard additive genetic model appeared to be right, many additive but very small effects collectively influence stature, not only trough biological paths, or even height specific paths, but their effects added up “explains” approximately the portion of height we expected it to. GWAS worked, regardless of whether it will eventually proof useful. Now GWAS of various other key traits are closing in on similar sample sizes, and its time to see what we can learn from those.
As other GWAS near or reach that point where most, or all, common variant signal should be absorbed we see mixed results. For major depressive disorder (MDD, colloquially “depression”, episode three in this series) each new GWAS brought with it a smaller estimate of SNP heritability, which is the total variance in the outcome captured by the GWAS. For educational attainment (episode 4 in this series) the GWAS appeared to be, and perhaps in some ways are successful, the predictive accuracy of the polygenic score based on that GWAS did near the SNP heritability of educational attainment. The SNP heritability in turn didn’t shrink dramatically with each new bigger GWAS. However, within family GWAS, which controls for environmental confounding of almost any kind, revealed the SNP heritability and polygenic score results in the education GWAS were inflated due to gene-environment correlation. Finally bigger and bigger schizophrenia GWAS (episode 5 in this series) seem more like height than like depression or education. However, the ability to compare between the common variant and rare variant analyses of schizophrenia (rare variant analysis for schizophrenia being widely available) suggest that due to the strong negative natural selection on schizophrenia, rare variant associations point to a smaller more biologically relevant set of genes, and it remains to be seen what the added value of the broader set of genes derived from GWAS will be. In addition, there is evidence of assortative mating for schizophrenia where people who have kids together are correlated in their risk for schizophrenia, which has consequences for how we model genetic effects.
Over the course of 4-8 weeks (potentially an Elon Musk like prediction that Ill never be able to live up to…) we’ll go trough each of these “endgame” GWASs, basking in the powers of hindsight. Leaning into hindsight risks annoying some of the O.G. authors reading along, so let’s be mindful of the gap between what we know now and what they knew then (turns out they saw a lot f it coming early, leaving a trail of aamzing theoretical work). I’ll jump ahead to the conclusion for the impatient: each of these massive GWAS of Height, MDD, Schizophrenia and education in some new way surprised the field. While there are those that after these GWAS came out claimed (on social media usually) “they told us so”, no one predicted the different ways in which each of these GWAS played out. Precisely because each of these GWAS has new unexpected results at scale, it’s probably worth letting the GWAS endgame play out for all traits and medical outcomes we are interested in. There are general lessons we can take from one GWAS to the next GWAS, there are also spectacular new maybe even unique ways in which things unfolded at scale. As these GWAS came out many have begun to formulate answers to that core question: “the GWAS is sort of done, now what?”, because these GWAS played out in different ways, the answers to that core question also differ.
From a deeply personal perspective on the GWAS saga, there are larger implications to what is essentially a failure. This is particularly true for “behavioral” traits. It can all be looked at as a scientific pursuit, now, but that ignores the fact that these studies were hyped and used as a justification for, for example, a biocentric view of my own field of psychiatry, wherein my entire career diverged into a “prescriber” model of treating presumed biological diseases, while all other avenues were closed off to me and little debate could be had, without being regarded as a crank. Moreover, the grasping of the fading results for so-called “educational attainment,” continue to fuel “race scientists,” and racially motivated individuals.
The point here is not to do an “I told you so,” but to consider whether replacing GWAS with something else will give me a third “I told you so” (lest we forget about candidate gene era). In any case, it is my view that whatever direction you pursue will not be good science unless the field of behavioral genetics seriously considers in their exploration what I think should be obvious, at this point: behavioral traits are, at best, minimally influenced by nonpathogenic genetic variation in human beings.