Emil O W Kirkegaard

Loads of things that could go wrong when testing a polygenic score using one population as the discovery & training sample on another population. Effect sizes, effect directions, interactions of SNPs can vary across populations;

Some SNPs that cause variation in the test sample may not vary in the discovery sample and therefore wouldn't be included in the score; there might not be enough correction for unrelated pop genetic diffs e.g.https://twitter.com/salonium/status/1087950206304436226 …

There are also other technical problems related to identifying snps, like LD decay, where the snps tagged by the analyses capture the causal variant less often bc of more recombination as populations differ more

And then there are problems when we ignore rare variants that affect the phenotype which are quite different across populations

In general, polygenic scores are useful for prediction rather than causal explanation, and especially when they're being tested on populations that weren't properly represented in the discovery sample

Why would that be useful? Going back to this ex https://twitter.com/salonium/status/1087950206304436226 … 1) the correlation b/e pgs & height in N Koreans might be lower than the correlation in S Koreans, e.g. bc even w a higher pgs, envt deprivation is so large that genetic propensity doesn't help N Koreans much