Genome-wide association studies (GWAS) identify single nucleotide polymorphisms (SNPs) that are correlated with complex traits by computing the marginal effect size of each SNP on the trait. Subsequently, by aggregating the individual SNP contributions into a single genetic risk estimate per individual, we can compute the polygenic risk score (PRS), serving a clinical utility especially for early disease screening.
Current PRS models explain only a fraction of trait heritability and provide limited mechanistic insight. While GWAS identifies SNP-trait associations, linking these signals to causal genes and pathways typically requires separate fine-mapping. Standard PRS approaches inherit this limitation by aggregating SNP effects without modelling relationships between variants and genes, resulting in a lack of integrated interpretability. In contrast, the omnigenic model suggests that genetic effects propagate through gene regulatory networks, where peripheral genes influence core disease pathways.
Here we introduce OmniGRS, a graph neural network that models PRS under the omnigenic hypothesis. OmniGRS constructs a per-individual bipartite graph connecting SNPs to genes and extends it with protein-protein interaction networks (STRING) to propagate regulatory context. We incorporate functional annotations (CADD) and SNP-gene relationships (e.g., genomic proximity or eQTLs) as node features and edge weights, enabling meaningful variant prioritisation.
OmniGRS outperforms baselines PRS models in both classification (AUC) and regression (Pearson correlation) tasks. Attention-weighted pooling over SNP and gene nodes yields interpretable importance scores that accurately recover causal variants in simulation.
Overall, OmniGRS provides a biologically informed framework for modelling non-additive genetic effects, improving risk prediction and interpretability in complex traits.

Co-authors: Sergey Vilov, Carsten Marr, Matthias Heinig