Genome-Wide Association Studies Using Deep Neural Networks

Abstract

Genome-wide association studies have identified thousands of genetic variants linked to complex traits and diseases, yet standard linear regression methods capture only additive effects and miss the non-linear interactions and epistatic relationships that deep neural networks can model. We evaluated 210 DNN-based GWAS programmes conducted across genomics centres in France and Spain between 2016 and 2021, spanning five modelling approaches: convolutional variant-effect networks, graph neural networks for variant interaction, variational autoencoders for population structure, attention-based polygenic risk scoring, and multi-task transfer learning across phenotypes. A DNN-GWAS Effectiveness Index (DGEI) was constructed from five sub-scores -- phenotype prediction improvement over linear baselines, biological interpretability of learned features, replication across independent cohorts, computational tractability at biobank scale, and clinical utility of polygenic scores -- with weights from regression against adoption into active genomics research pipelines. DGEI correlated with adoption at r = +0.83 and discriminated adopted from non-adopted methods with an AUC of 0.880. Attention-based polygenic risk scoring led at mean DGEI 0.818, while variational autoencoders trailed at 0.592. Only 34.3 percent of programmes exceeded the 0.75 threshold. Phenotype prediction improvement carried the largest weight (beta = +0.278), followed by replication breadth (beta = +0.230)