lead IMPACT annotations from S-LDSC results using GWAS summary statistics, as done above, on the partition of the BBJ cohort excluding the 5,000 PRS test individuals. We defined improvement as the percent increase in R2 from standard to functionally informed PRS; therefore, differences in PRS performance due to intrinsic factors, such as GWAS power or genotyping platform, cancel out. In both scenarios, we observed substantial positive improvements: averaged across the 21 traits in the trans-ancestry setting (mean percent increase in R2=47.3%, s.e.m. = 8.1%, one-tailed z-test P< 2.7 × 10−9) and in the within-population setting (mean percentage increase in R2 = 20.9%, s.e.m. = 6.6%, one-tailed z-test P< 7.5 × 10−4). Indeed, this revealed a significantly greater improvement in the trans-ancestry application than in the within-population application across the 21 traits (one-tailed paired Wilcoxon P < 0.012, Fig. 5e). Moreover, the disease predictive power of our PRS was not driven by a few loci of large effect nor the scale of our effect size estimates (Extended Data Fig. 10 and Supplementary Note). Overall, our results reveal that functional prioritization of SNPs using IMPACT improves both trans-ancestry and within-population PRS models, but is especially advantageous for the trans-ancestry application of PRS.
