DeepU-PRS: nonlinear and uncertainty-aware SNP heritability modeling for annotation-informed polygenic risk scoring.

Polygenic risk scores (PRSs) are widely used to quantify genetic predisposition to complex traits, and their predictive accuracy can be enhanced by incorporating functional annotations into single-nucleotide polymorphism (SNP) effect-size estimation. Existing approaches typically assume linear relationships between annotations and SNP effects, neglecting predictive aleatoric variance and potential nonlinear regulatory interactions. To evaluate the gains of nonlinear and uncertainty-aware modeling, we developed DeepU-PRS, a deep ensemble framework that extends the linear BLD-LDAK model with a neural network trained on functional annotations to jointly estimate SNP heritability and its aleatoric uncertainty, which are then integrated into a Bayesian PRS model. Across 28 continuous and 8 disease phenotypes in the UK Biobank, DeepU-PRS improved the mean continuous R2 by approximately 10% over linear BLD-LDAK, with the largest gains for traits whose association signals were concentrated, whereas the more diffuse disease traits favored BLD-LDAK. These patterns were reproduced in individual-level data from All of Us. Against a panel of established methods (C + T, Lassosum, PRS-CS, LDpred-funct, and SBayesRC), DeepU-PRS was among the strongest single methods on continuous traits, and ensembles combining DeepU-PRS with the linear baseline and SBayesRC further improved accuracy, indicating complementary predictive signal. These findings highlight the value of modeling nonlinear and uncertain regulatory effects to advance PRS prediction.