Revealing Herb-Symptom Associations and Mechanisms of Action in Protein Networks Using Subgraph Matching Learning.

Abstract

In traditional Chinese medicine, deciphering herb-symptom associations (HSAs) and revealing their mechanisms of action are crucial for bridging traditional knowledge and modern biomedicine. While previous studies have investigated HSAs using protein-protein interaction (PPI)-based network medicine method, they often treat all proteins equally, failing to capture the heterogeneous contributions of individual proteins to HSAs. This limitation hinders their capacity to reveal the mechanisms of action. To address this challenge, we propose a subgraph matching learning method, GraphHSA, for HSA prediction. GraphHSA maps herbs and symptoms onto the PPI network to construct subgraphs. Then, GraphHSA utilizes an attention mechanism to compute the importance of each protein on the subgraph, and weighted aggregate protein information to generate herb/symptom embeddings. Subsequently, these embeddings are combined to model the matching relationship between herb and symptom subgraphs, enabling association prediction. Additionally, a dual-contrastive learning strategy is introduced to generate discriminative representations to enhance prediction. Experiments indicate that GraphHSA not only applies to individual herbs but also extends to compound formulations composed of multiple herbs. By capturing the dynamic interactions among their components, GraphHSA enables the identification of key biological targets and the elucidation of the mechanisms underlying their therapeutic efficacy.