A Multi-Modal Framework for Phage-Host Interaction Prediction Using Multi-View Contrastive Learning

Phage therapy is emerging as a promising strategy to combat the growing challenge of antibiotic resistance. The efficacy of this therapeutic approach critically depends on the precise prediction of phage-host interactions (PHI). However, existing computational methods exhibit notable limitations in multi-source bioinformatics data integration and heterogeneous network representation. To address these limitations, we propose MCL-PHI, a novel framework that achieves high-precision prediction by integrating multi-modal features with multi-view contrastive learning. Specifically, a Large Language Model (LLM) utilizing Chain-of-Thought reasoning is employed to generate structured knowledge descriptions. These descriptions are subsequently encoded by BioBERT to yield domain-enhanced semantic representations, which are then fused with genomic features to construct a comprehensive heterogeneous network. Furthermore, the core of our framework is a multi-view graph neural network that captures both local topology and high-order meta-path semantics, reinforced by a contrastive learning strategy to maximize representational consistency. Finally, an adaptive multi-head attention mechanism fuses these representations for link prediction. Extensive experiments demonstrate that MCL-PHI significantly outperforms six state-of-the-art baseline methods. Ablation studies and case analyses further validate the efficacy of individual modules and the framework's potential in discovering novel PHIs.