48] or multi-ethnic meta-GWAS [7, 44], and PRS evaluated in this work using more sophisticated PRS construction methods such as PRS-CS and LDpred2, our trans-ancestry PRS constructed by PRS-CSx demonstrated improved prediction accuracy and transferability across ancestral groups, reflecting the combined effect of methodological advances and increased sample sizes in the training GWAS. We note that the T2D GWAS of African ancestry from the MEDIA Consortium was substantially underpowered relative to the GWAS in European and East Asian populations, which limited the performance of PRS-CSx. As the diversity and scale of discovery GWAS continue to expand, principled computational models that can appropriately integrate multi-ancestry genome-wide data are expected to further improve the prediction in non-European populations and demonstrate bigger advantages over single-population methods. In addition, while HapMap3 variants represent a good balance between computational cost and prediction accuracy for PRS construction in European populations, they may tag a lower proportion of genetic variation in non-European populations, limiting the transferability of PRS. Future work is needed to develop computationally efficient algorithms that can incorporate genome-wide genetic variants into PRS calculation.
