Interpretable Herb-Disease Association Prediction with Molecular-Aware Multi-View Representation Learning

Predicting herb-disease association plays a crucial role in accelerating herb repositioning and expanding clinical applications of Traditional Chinese Medicine (TCM). However, most of the existing methods neglect the fine-grained biomedical information, such as the molecular structures, which leads to inaccurate and incomplete representations of herbs and diseases. Since TCM involves multiple components and modulates various targets, existing methods lack an effective approach to represent these relations, resulting in the omission of critical information. In addition, the lack of interpretability in existing methods limits their credibility and acceptance. To tackle the above problems, we propose a herb-disease association prediction framework, named MVGP, which incorporates Multi-View and multi-Granularity representation learning as well as meta-Path-based relational modeling to accurately capture the relationships between herbs and diseases. Specifically, we design a multi-view learning frame-work that integrates latent information from both TCM and western medicine views. We incorporate fine-grained molecular structures of herbal compounds and target proteins to better characterize features. We design a hypergraph convolutional network to model high-order relations among herbs and compounds, diseases and target proteins. Moreover, we employ meta-path-based relational modeling to make the results interpretable. Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed model.