We designed PRS models using two strategies: standard PRS and functionally informed PRS. For standard PRS-EUR, we performed conventional LD clumping to acquire sets of independent SNPs using EUR LD reference panels from phase 3 of 1000 Genomes. Similarly for PRS-EAS, we utilized EAS LD reference panels from phase 3 of 1000 Genomes. We used PLINK (v.1.90b3)63 to remove variants in LD with r2 > 0.2 with a significance threshold for index SNPs of P = 0.5. For functionally informed PRS, we restricted the analysis to variants with high IMPACT score according to the lead IMPACT annotation before conducting LD clumping. As before, we define the lead annotation as the one with the largest τ* estimate that was significantly greater than 0. When we designed PRS-EUR, we utilized the lead IMPACT annotation in EUR GWAS summary statistics (EAS summary statistics were not taken into account to avoid overfitting). Similarly, when we design PRS-EAS, we utilized the lead IMPACT annotation in EAS GWAS summary statistics for which 5,000 EAS individuals for PRS analysis were removed to avoid overfitting. We performed LD clumping using variants within a predefined top percentage of IMPACT scores.
