Computational Genomics in Disease Modeling

Abstract

Genome-wide association studies (GWAS) have identified over 100,000 variant-trait associations across thousands of complex diseases, yet translating statistical associations into mechanistic disease models remains a fundamental challenge -- over 90% of disease-associated variants reside in non-coding regulatory regions with unknown functional consequences. Computational genomics integrates variant annotation, gene regulatory network modelling, and pathway enrichment to bridge the genotype-to-phenotype gap. This study developed and benchmarked five computational approaches -- statistical fine-mapping (SuSiE), functional annotation scoring (CADD + DeepSEA), transcriptome-wide association (TWAS/PrediXcan), network propagation (random walk with restart on STRING), and a proposed graph-transformer variant effect predictor (GT-VEP) that jointly models variant regulatory effects, gene regulatory networks, and biological pathway context -- for prioritising causal genes and modelling disease mechanisms across three complex diseases: type 2 diabetes (T2D, 403 GWAS loci), schizophrenia (SCZ, 287 loci), and inflammatory bowel disease (IBD, 241 loci). GT-VEP achieved the highest causal gene prioritisation accuracy (AUROC 0.846 for T2D, 0.812 for SCZ, 0.838 for IBD against gold-standard curated gene sets), significantly outperforming TWAS (0.762; p < 0.001), fine-mapping (0.724; p < 0.001), and network propagation (0.786; p = 0.004). GT-VEP identified 48 novel candidate causal genes for T2D not in existing databases, of which 14 (29.2%) were independently supported by rare variant burden tests in UK Biobank exome data (n = 454,787). Pathway-level disease models reconstructed by GT-VEP recapitulated known disease biology (insulin signalling for T2D, synaptic function for SCZ, mucosal immunity for IBD) while revealing novel pathway convergences. These results establish graph-transformer architectures as a powerful framework for computational disease modelling and provide a reproducible benchmark for variant-to-gene-to-pathway interpretation.