LLM-H2G: biomedical semantic-enhanced hypergraph contrastive learning for herb–disease association prediction

Abstract Background Herb–disease association prediction is central to computational traditional medicine, but existing graph and hypergraph methods mainly rely on observed topology and underuse biomedical textual semantics, especially in heterogeneous or sparse association networks. Results We propose LLM-H2G, a biomedical semantic-enhanced hypergraph contrastive learning framework that integrates herbs, compounds, proteins and diseases through an enhanced heterogeneous hypergraph. Entity names are encoded as large language model-derived semantic representations, injected into compound and protein propagation, and aligned with herb and disease structural readouts through an InfoNCE objective. Across TCM-suite and Ethnobotany benchmarks, LLM-H2G consistently outperforms representative network embedding, graph neural network, graph contrastive learning and hypergraph baselines, with the largest performance gaps observed for sparsely connected herbs. Ablation and parameter analyses show that both semantic enrichment and enhanced hypergraph construction are necessary for the observed gains. Case studies further show two complementary uses of the model: attribution and docking recover plausible compound–protein mediators for known associations, while literature-supported de novo predictions reveal candidate unannotated therapeutic paths. Conclusions LLM-H2G provides an effective framework for integrating biomedical semantics with heterogeneous hypergraph structure and may support interpretable discovery of biologically plausible herb–disease associations. As a representative new path, the Ginkgo biloba –Radiation Injuries prediction is supported by a coherent radiation-related MeSH cluster and PubMed-indexed experimental evidence.