to RApred3 in Fig. 1) strongly overestimated the RA in non-European ancestries. The average overestimation relative to the observed RA ranged from +4.3% (i.e. (96.3%−92.3%)/92.3% × 100%)) in SAS up to +103.0% in AFR ancestry. This result suggests that population differences in LD between causal variants and GWS SNPs contribute a larger fraction to the decreased RA than allele frequencies differences at GWS SNPs only. It is also worth noting that our predictions were nearly insensitive to using either whole-genome sequence (WGS) data from the 1KGP or imputed genotypes of UKB participants as reference panels (Supplementary Fig. 5), which is reflected by the highly correlated allele frequencies (Supplementary Fig. 6) and LD scores (Supplementary Fig. 7) between WGS and imputed data (Supplementary Note 3). However, LD reference panels with larger sample sizes are still recommended to achieve more accurate estimates of LD correlations. Finally, we assessed the robustness of our results by varying the size of the discovery GWAS between 100,000 and 300,000 participants (Supplementary Note 4). We found the accuracy of PGS to increase proportionally in all ancestries, such that the RA remained constant and thus independent of sample size (Supplementary Fig. 8).
