This study has several limitations. First, we limited to concordant variants in both AA and EA, thus, variants that may have discordant but true effects were excluded, reducing the performance of PRSgene. Second, although most intergenic concordant variants did not contribute to the PRS signal as shown in the analyses of PRSintergenic and extended gene boundaries with different window sizes, some of them are truly AUD related and contribute to the risk of AUD, and excluding them leading to a further reduction in the performance of PRSgene. Third, we used posterior effects estimated from the meta-analysis of AA-MVP and EA-PAU. As EA-PAU had a much larger sample size (>7 times of sample size of AA-MVP), more weight was put on effects estimated from the EA samples. Therefore, for those variants that had different sizes of effects between AA and EA, effects from the meta-analysis were biased toward the EA GWAS. Fourth, studies have shown that using functional annotations can improve the performance of PRS and increase the transferability of PRS between different populations [11, 54–56]. However, most of the available
