Policy gradient-guided ensemble learning for enhanced polygenic risk prediction in ultra-high-dimensional genomics

Polygenic diseases challenge genetic risk prediction due to extreme dimensionality, low per-variant effect sizes, and non-additive interactions. Conventional marginal P-value-based methods potentially overlook subtle signals and complex dependencies, while inefficient random sampling in ensembles misses sparse signals. We introduce ELAG, an ensemble learning framework that advances feature bagging by reformulating variant selection as an approximate reinforcement learning problem. Leveraging policy gradients, a parameterized policy optimizes adaptive sampling--evolving beyond uniform strategies like random forests through importance-distribution guidance--to direct non-linear classifiers toward synergistic variants. This enables scalable navigation of millions of loci without pre-filtering, handling intricate architectures. In high-polygenicity, low-heritability simulations, ELAG boosted predictive accuracy ({Delta}AUROC = 0.0644). For neuro-immune diseases like rheumatoid arthritis, it enhanced AUROC from 0.6866 to 0.7354 and polygenic scores (e.g., Lassosum AUROC 0.7186 [->] 0.7543), outperforming mainstream methods. ELAG is robust to missing data, integrable with covariates, and yields interpretable variant sets enriched for pathways and networks. It can replace random sampling with learned guidance, advancing machine learning for ultra-high-dimensional data such as genetic risk prediction.