MHGNN: Multiplex Hypergraph Neural Networks for Predicting Herb--Symptom Interactions

Herb-symptom interaction (HSI) prediction is crucial for understanding the multi-target mechanisms of herbal therapies and enabling data-driven precision traditional Chinese medicine (TCM). Existing computational approaches mainly employ graph neural networks (GNNs) to model herb-symptom relationships in biological networks. However, these methods are limited to low-order pairwise associations and fail to capture herbs' high-order and indirect regulatory effects. In addition, random negative sampling from unobserved herb-symptom pairs often results in false negatives due to incomplete annotations. To address these limitations, this study proposes MHGNN, a novel hypergraph learning framework that models HSIs as a multiplex hypergraph integrating protein-protein interactions (PPIs) along with high-order herb-protein and symptom-protein relationships. MHGNN employs hierarchical message passing to aggregate high-order features across multiple edge types and adopts a network-based negative sampling strategy, which selects herb-symptom pairs with larger proximity distances on the human protein interactome as negative instances to reduce false-negative bias. Comprehensive experiments on two public TCM datasets show that MHGNN consistently achieves superior predictive performance compared with thirteen state-of-the-art baselines, highlighting its potential to advance mechanism-aware modeling and enable precision herbal medicine.