Our study has several limitations. Compared with GRS for other common conditions, such as CAD21, the performance of our metaGRS for stroke is limited. A likely reason is that stroke is more heterogeneous and that GWAS sample sizes for mechanistically defined stroke subtypes are still limited in comparison with other diseases. As stroke GWAS progress, GRS will become more powerful39,40. We did not observe substantial advantage from incorporating GRSs based on GWAS summary statistics for specific IS subtypes (LAS, CES, SVS) over that of IS as a whole; however, we were not able to examine subtype-specific outcomes in the UKB due to the lack of such detailed information. There may still be benefit from developing subtype-specific scores that take advantage of the unique genetic architecture of each subtype41. The number of older individuals (>75 years) in UKB is limited at this stage, reducing our ability to model stroke risk in the age strata where the majority of events occur. Furthermore, the duration of follow-up in UKB is relatively limited and, because of the limited number of assessments, we could not
