Deep learning-based polygenic scores enhance generalizability of psychiatric disorders prediction.

Polygenic scores (PGSs) have emerged as promising tools for predicting complex traits from genetic data, however, their predictive performance for psychiatric disorders remains limited and the added value of deep learning (DL) over linear models is underexplored. In this study, we compared our DL model, Genome-Local-Net (GLN), with the linear model bigstatsr in predicting five psychiatric disorders-ADHD, ASD, BIP, MDD, and SCZ-using individual-level genotype data. We further assessed whether combining these internal (individual-based) PGSs with external (GWAS-derived) PGSs and family genetic risk scores (FGRSs) could improve prediction additively or synergistically. While GLN and bigstatsr performed similarly in-sample, GLN showed better generalization on an out-of-sample replication set for ADHD, ASD, and MDD, with an average AUROC gain of 0.026. Integrating internal, external, and family-based scores significantly improved ADHD prediction, though DL-based integration provided no consistent advantage over logistic models. These findings suggest that while DL may enhance generalizability for specific psychiatric traits, linear models remain competitive and effective for genetic risk prediction.