Predicting phage-host interaction via hyperbolic Poincaré graph embedding and large-scale protein language technique

Bacteriophages (phages) are increasingly viewed as a promising alternative for the treatment of antibiotic-resistant bacterial infections. However, the diversity of host ranges complicates the identification of target phages. Existing computational tools often fail to accurately identify phages across different bacterial species. In this study, we present GE-PHI, a machine-learning-based model for predicting phage-host interactions (PHIs) by integrating knowledge graph embedding algorithm with a large-scale protein language model. First, a phage-host heterogeneous association network (PHAN) was constructed that incorporated phage-phage and host-host similarity networks. Then, the multi-relational Poincaré graph embedding (MuRP) was used to extract topological patterns. Additionally, we employed the ESM-2 protein language model to capture evolutionary information from phage tail proteins and host-receptor-binding proteins. GE-PHI achieved a cross-validation area under the curve (AUC) of up to 0.9453 in silico and maintains this performance in case studies. This study provides insights into machine-learning-guided phage therapeutics and diagnostics in microbial engineering.