Dimensionality reduction reveals fine-scale structure in the Japanese population with consequences for polygenic risk prediction

The diversity in our genome is crucial to understanding the demographic history of worldwide populations. However, we have yet to know whether subtle genetic differences within a population can be disentangled, or whether they have an impact on complex traits. Here we apply dimensionality reduction methods (PCA, t-SNE, PCA-t-SNE, UMAP, and PCA-UMAP) to biobank-derived genomic data of a Japanese population (n = 169,719). Dimensionality reduction reveals fine-scale population structure, conspicuously differentiating adjacent insular subpopulations. We further enluciate the demographic landscape of these Japanese subpopulations using population genetics analyses. Finally, we perform phenome-wide polygenic risk score (PRS) analyses on 67 complex traits. Differences in PRS between the deconvoluted subpopulations are not always concordant with those in the observed phenotypes, suggesting that the PRS differences might reflect biases from the uncorrected structure, in a trait-dependent manner. This study suggests that such an uncorrected structure can be a potential pitfall in the clinical application of PRS.