DeepDisSNP: Predicting disease-associated SNPs by representation learning on disease and SNP linkage disequilibrium networks.

Genome-wide association studies (GWAS) have identified numerous disease-associated single nucleotide polymorphisms (SNPs), yet many potential disease-SNP associations remain undiscovered due to the high dimensionality and sparsity of genomic data. Computational approaches that integrate biological network information can complement existing GWAS resources by prioritizing candidate disease-associated SNPs for downstream investigation. In this study, we propose DeepDisSNP, a deep learning framework for disease-SNP association prediction that integrates disease similarity networks and chromosome-specific SNP linkage disequilibrium (LD) networks through graph attention network (GAT)-based representation learning. Disease similarity networks were constructed from MeSH-based disease relationships, while SNP LD networks were generated from Phase 1 and Phase 3 datasets of the 1000 Genomes Project under multiple LD thresholds. DeepDisSNP independently learns disease and SNP embeddings using weighted GAT encoders and subsequently predicts disease-SNP associations using a multilayer perceptron classifier. Extensive experiments across all 22 autosomal chromosomes demonstrated that DeepDisSNP consistently outperformed the state-of-the-art DisSNPNet framework under multiple experimental settings. Under the best-performing configuration, DeepDisSNP achieved AUROC and AUPRC values of approximately 0.96 and 0.95, respectively. Additional analyses demonstrated robustness across LD thresholds, 1000 Genomes Project phases, and increasingly imbalanced negative sampling settings. External GWAS resources, including NHGRI-EBI GWAS Catalog, PhenoScanner, and OpenGWAS, provided supportive biological evidence for many highly ranked predicted associations. Functional enrichment analyses further suggested biological relevance of the predicted SNP-associated genes. Overall, DeepDisSNP provides an effective network-based framework for large-scale disease-SNP association prioritization and may facilitate downstream genomic and translational research.