For example, in height, EUR–EAS effect size estimates of SNPs in the top 5% partition are 2.1-fold more similar (Pearson r = 0.29, Fig. 4a) than those in the bottom 95% partition (r = 0.14, Fig. 4b), and 1.6-fold more similar than the set of all SNPs (r = 0.18). For each of 17 GWAS P value thresholds, the marginal trans-ancestry effect size correlation among the top 5% of IMPACT SNPs tended to be greater than the set of all SNPs genome wide across 21 traits (all 17 one-tailed paired Wilcoxon P < 6.9 × 10−4) (Fig. 4c,d). Furthermore, this observation was consistent across individual traits (Supplementary Fig. 12) and was comparable to using alternative functional annotations (Supplementary Note). Since allele frequency greatly affects disease predictive power, we next analyzed the trans-ancestry concordance of allelic heterozygosity and population divergence (Fst). We found that neither increased concordance of heterozygosity nor substantial difference in Fst is a consequence of IMPACT prioritization (Extended Data Figs. 6 and 7 and Supplementary Note). Overall, our results suggest that we might anticipate improved trans-ancestry portability of PRS models by prioritizing SNPs in key functional annotations by decreasing the likelihood of selecting SNPs associated solely by linkage.
