Finally, we addressed our hypothesis that IMPACT annotations improve the trans-ancestry portability of PRS (Fig. 1d-(iii)). For each of the 21 previously analyzed traits, we built a PRS using effect size estimates from EUR summary statistics and applied it to predict phenotypes of EAS individuals from BioBank Japan (BBJ) (Fig. 5a). Here, we compare two PRS models, both blind to any EAS genetic or functional information and removing SNPs with LD r2> 0.2, according to European individuals from phase 3 of 1000 Genomes53: (i) standard P+T PRS and (ii) functionally informed P+T PRS using a subset of SNPs prioritized by the lead EUR IMPACT annotation (Methods). In functionally informed PRS models, for each trait separately, we selected a priori the subset of top-ranked IMPACT sNps (top 1%, 5%, 10% or 50%) that explained the closest to 50% of common SNP heritability (Methods). This ensures that functional prioritization captures approximately the majority of trait-relevant genetic variation and the cumulative genetic signal among functionally prioritized variants was consistent across traits, allowing for varying degrees of polygenicity. For all PRS models, we report results from the most accurate model across nine EUR GWAS P value thresholds.
