Transformer-based InsightGWAS improves GERD genetic discovery via pretraining on GWAS for major depressive disorder.

Gastroesophageal reflux disease (GERD) is a highly prevalent gastrointestinal disorder with complex genetic underpinnings.While genome-wide association studies (GWAS) have identified several GERD-associated loci, traditional GWAS approaches rely on stringent significance thresholds and may miss variants with modest effects that still contribute to disease biology. To enhance the discovery of GERD-associated loci, we developed InsightGWAS, a Transformer-based deep learning model. Using transfer learning, the model was pre-trained on major depressive disorder GWAS data and fine-tuned with GERD GWAS summary statistics. We integrated multi-omics functional annotations, including eQTLs, mQTLs, and epigenomic data, to prioritize candidate variants. Comparative analyses showed that InsightGWAS outperformed logistic regression, XGBoost, and neural networks, achieving superior classification accuracy and reducing false positives. The model replicated known GERD loci and uncovered 209 novel candidate loci, many involved in neurogenic, neuromuscular, and epithelial pathways. Enrichment analyses revealed associations with synaptic transmission, neural development, and cadherin-mediated signaling, suggesting that both nervous system regulation and epithelial integrity contribute to GERD pathophysiology. This study demonstrates the power of deep learning in advancing genetic discovery beyond conventional GWAS. By leveraging transfer learning and multi-omics annotations, InsightGWAS identifies potential disease-asscoated biological pathways underlying GERD, offering promising directions for mechanistic research and potential therapeutic targets.