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

DeepU-PRS enhances SNP heritability modeling by incorporating nonlinear and uncertainty-aware methods, improving prediction accuracy by approximately 10% in UK Biobank traits.

• Why it matters: Accurate polygenic risk scores are crucial for understanding genetic predisposition to complex traits, but current linear models overlook nonlinear effects and aleatoric uncertainty, limiting their predictive power.
• What they did: The authors developed DeepU-PRS, a deep ensemble framework that extends the linear BLD-LDAK model with neural networks trained on functional annotations to jointly estimate SNP heritability and its uncertainty across 36 phenotypes.
• The result: DeepU-PRS outperformed linear models for traits with concentrated association signals, and ensemble approaches combining it with other methods further improved prediction accuracy, emphasizing the benefit of modeling nonlinear and uncertain regulatory effects.