Poster #40 - Junhyeong Lee

DeepU-PRS: Integrating Uncertainty-Aware Deep-ensemble SNP Heritability Estimation to Improve PRS / Bioinformatics methods development.

Junhyeong Lee, MS¹,²; Dokyoon Kim, PhD²; Seunggeun Lee, PhD¹ Affiliations ¹ Graduate School of Data Science, Seoul National University, Seoul, Republic of Korea ² Department of Biostatistics, Epidemiology and Informatics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA

To enhance the predictive performance of polygenic risk scores (PRS), incorporating functional annotations of genetic variants is important. Typically, this involves estimating SNP-level heritability informed by functional annotations. However, existing methods often rely on linear annotation-informed priors, overlooking both annotation uncertainty and potential nonlinear interactions in regulatory elements. Here, we introduce a deep ensemble-based framework, DeepU-PRS. By leveraging a neural network trained on functional annotations, our method simultaneously estimates SNP heritability and quantifies associated uncertainties, subsequently integrating these estimates within a Bayesian PRS model. Specifically, DeepU-PRS employs an ensemble of five feed-forward neural networks trained on 205 SNP annotations-including epigenomic marks, QTL indicators, and sequence embeddings-to estimate per-SNP heritability along with uncertainty estimates. SNPs with high uncertainty were subsequently down-weighted or filtered within an SBayesR-based PRS pipeline, enhancing prediction robustness and accuracy. We evaluated DeepU-PRS on 31 continuous traits from UK Biobank and externally validated in an independent All of Us cohort. DeepU-PRS yielded mean relative R² improvements averaging 9-10% compared to a linear annotation-based model (BLD-LDAK) across both UK Biobank and All of Us. Traits with highly concentrated genetic signals, identified by higher Herfindahl-Hirschman Index (HHI) values and lower Effective Number of Signals (EFF), showed the largest performance improvements. Compared to models without uncertainty modeling, incorporating uncertainty improved the average R² by 15%, demonstrating the clear advantage of modeling annotation uncertainty. Simulations varying trait heritability (h² = 0.1-0.5) and causal SNP counts (10k, 50k) also demonstrated the improved performance of our method. DeepU-PRS provides a flexible framework to integrate functional annotations and quantify uncertainty, enabling more accurate and biologically informed genetic risk prediction.