After selecting approximately independent GWS SNPs using LD clumping, we then generated PGS in each target population by adding up the product of minor allele counts times effect sizes of GWS SNPs estimated from GWAS summary statistics. The prediction accuracy was then estimated using the squared correlation (R2) between the true phenotypes and the PGS. For diseases, we calculated the liability-scale R2 as described in Lee et al.31. We estimated the disease prevalence as the proportion of cases in each ancestry given that UK Biobank is a population-based study. We combined the two target populations of EUR ancestry (N = 19,979) to calculate the corresponding R2 for traits/diseases in UKB. The empirical relative accuracies (RA) were calculated as the ratio of the R2 in UKB participants of non-European ancestry over the same R2 estimated in ~20,000 independent UKB participants of European ancestry.
