Hyperbolic Graph Embeddings Reveal the Host-Pathogen Interactome

Infections depend on interactions between pathogen and host proteins, but comprehensively mapping these interactions is challenging and labor intensive. Many biological networks have hierarchical, scale-free structure, so we developed a deep learning framework, ApexPPI, that represents protein networks in hyperbolic Riemannian space to capture these features. Our model integrates multimodal biological data (protein sequences, gene perturbation experiments, and complementary interaction networks) to predict likely interactions between pathogen and host proteins through multi-task hyperbolic graph neural networks. Mapping protein features into hyperbolic space led to much higher accuracy than previous methods in predicting host-pathogen interactions. From tens of millions of possible protein pairs, our model identified thousands of high-confidence interactions, including many involving human G-protein-coupled receptors (GPCRs). We validated dozens of these predicted complexes using AlphaFold 3 structural modeling, supporting the accuracy of our predictions. This comprehensive map of host-pathogen protein interactions provides a resource for discovering new treatments and illustrates how advanced AI can unravel complex biological systems.