ARCHIVES

Understanding the molecular mechanisms underlying host–pathogen interactions (HPIs) is vital for developing effective therapeutics and vaccines against infectious diseases. However, existing computational methods for HPI prediction often rely on incomplete annotations, suffer from poor cross-species generalization, and lack biological interpretability. To overcome these limitations, this paper introduces an AI-Assisted Host–Pathogen Interaction (AI-HPI) Network Mining Framework, which integrates relational graph neural networks (R-GAT/R-GCN) with graph contrastive pretraining and positive–unlabeled (PU) learning for label-efficient and interpretable HPI discovery. The proposed framework constructs a heterogeneous biological knowledge graph combining host and pathogen proteins, Gene Ontology (GO) terms, KEGG/Reactome pathways, and known protein–protein interactions (PPIs). A contrastive learning objective (InfoNCE) enhances feature alignment across graph augmentations, while a DistMult decoder performs link prediction between host and pathogen nodes. The inclusion of PU-risk correction and temperature scaling ensures robust, uncertainty-aware inference under data sparsity. Experimental results on multi-species datasets—Homo sapiens–Plasmodium falciparum, H. sapiens–Mycobacterium tuberculosis, and Arabidopsis–Pseudomonas—show superior predictive accuracy (AUROC = 0.943, AUPR = 0.902), improved calibration (ECE = 0.038), and strong zero-shot transferability (>90% AUPR retention). Biological interpretation via attention-based attribution and pathway enrichment identifies key host immune regulators (NF-κB, MAPK, and TNF signaling) and potential druggable effectors in Plasmodium and Mycobacterium.