Cross-Species Graph Neural Network for Translating Animal Disease Resistance to Human Drug Targets

Therapeutic target discovery has traditionally relied on data from human patients and mouse models, which primarily capture disease states rather than mechanisms of resistance or repair. Animals with evolved disease-protective adaptations offer complementary insights that, when properly combined with human data, can drive more discoveries. Here, we introduce CentaurGNN, a cross-species graph neural network for novel target discovery. It leverages species-specific message passing to integrate animal data with large-scale human data. In particular, CentaurGNN incorporates data from animal disease resistance (ADR) adaptations to conditions like obesity, heart disease, kidney disease, and neurodegeneration, providing complementary biological signals to human-only datasets. We demonstrate that CentaurGNN has up to 2 times stronger generalizability (via ROC) to unseen gene-indication associations than state-of-the-art baselines. We also find that integrating animal data to a model trained only on human data improves predictive performance by 12% (in ROC). Additionally, feature attribution analysis confirms the significant contributions of animal data toward CentaurGNN's predictions, with animal-derived features having at least an order of magnitude higher average importance than human-derived features. Finally we validate our findings using gene-based burden tests in the UK Biobank (~400,000 exomes). Our analysis shows that CentaurGNN, when enhanced with ADR data, nominates an average of 178 more genetically-validated novel targets than our human-only baseline, indicating its translational potential for drug discovery.