,
20 tweets,
5 min read
Read on Twitter
(1) Here's an explainer thread on our new paper on the difficulties of interpreting polygenic score differences between populations. Paper w/ Noah Rosenberg, @jkpritch, and Marc Feldman, but speaking for myself here: academic.oup.com/emph/advance-a…
(2) With so much genome-wide association study (GWAS) data becoming available, it's become easy to compute mean polygenic scores for human populations for many traits, which are mean values of phenotypes predicted from genetic information.
(3) The most straightforward way to compute a mean polygenic score is to take a weighted sum of allele frequencies, where the weights are effect size estimates from GWAS.
(4) (Most polygenic scores aren't very accurate as trait predictions right now, but the accuracy will go up as GWAS progresses.)
(5) As it's easy to compute mean polygenic scores, You can imagine people finding that group A has a higher mean polygenic score for, say, height, than group B, and then saying things like "group A is taller than group B for genetic reasons" or "x% of the height gap is genetic".
(6) These kinds of inferences are slipshod for a lot of reasons, and our goal was to collect those reasons and put them in one commentary. Our sense is that people working in the field know these things, but we wanted to go on record before misinterpretations proliferate.
(8) So why are these inferences so slippery? A lot of the reasons have to do with how genes and environments interact to shape phenotypic variation, and are actually basically the same reasons that have been discussed widely since at least the 1970s by Lewontin, etc.
(9) In short, genetics doesn't necessarily work in a simple additive way that's constant across populations---there can be gene-gene interaction (epistasis), gene-environment interaction, and plain old environmental differences that swamp or even counter any genetic differences.
(10) These are all glosses for quite general classes of genotype-phenotype relationships: it can be really complicated.
(11) Beyond these conceptual issues, there are a lot of reasons to think that polygenic scores don't translate well between populations. @genetisaur's work has been a major touchpoint for a lot of us here, for example figures 2 and 3 here: biorxiv.org/content/biorxi…
(12) Why do phenotype predictions developed from European populations perform less well in other places? We don't really know, but there are at least 5 possible reasons.
(13) i) correlations between genetic sites (i.e. LD patterns) vary across populations, meaning that a particular GWAS SNP may be correlated with different sets of causal variants in different populations. (e.g. figure 2 in the @genetisaur paper linked earlier).
(14) ii) Some variants that affect a phenotype in one population may not be variable in another population, making them impossible to map by GWAS in the population where they're fixed (i.e., not variable). science.sciencemag.org/content/early/…
(15) iii) The original GWAS may have issues with uncorrected population stratification, which seems to have driven exaggerated signals of selection on height. biorxiv.org/content/early/… and biorxiv.org/content/early/…
(16) Differences in either (iv) genetic background or (v) environment may cause the effect sizes measured in GWAS to differ in different populations (epistasis and GxE again), as in the apparent epistasis in this ApoE4 story: journals.plos.org/plosgenetics/a…
(17) So all told, we just don't understand genotype-phenotype maps well enough to leap from current data to explanations of group differences.
(18) Further, even if a group difference is "genetic," that doesn't mean it's immutable---it might disappear in another environment. (Imagine a "genetic" difference in lipid levels that's erased in a world where everybody takes statins.)
(19) We also talk a bit about attempts to explain observed trait differences in terms of drift vs. selection. There are ways to test for selection on polygenic scores, but the tests are vulnerable to stratification, and there are further difficulties in interpretation.
(20) Finally, we spend some time working through a sort of back-of-the envelope calculation regarding health disparities to give an example of the kinds of things one has to consider when approaching these questions. (end)
