As our cohort comprises ethnically diverse participants, reflective of the diversity among those serving in the US military, we examined the degree to which genetic ancestry affects PRS prediction performance. In order to define more genetically homogeneous subgroups of our cohort, instead of using self-identified ethnicity/race, hierarchical clustering was performed with the genetic data in an unsupervised manner (i.e., without making use of self-identified ethnicity/race label), see Methods. Having defined four genetic ancestry clusters (Fig. S2), we examined the performance of the PRS in the four clusters separately. Not surprisingly, the EA PRS has the most predictive value for clusters 1 and 2 (Table S1). In contrast, its performance on clusters 3 and 4 are near one representing a random classifier. It should be noted that limitations of PRS to predict across ancestral groups has been reported for other traits as well33,34.
