Drug-drug interaction prediction of traditional Chinese medicine based on graph attention networks.

Predicting drug-drug interactions (DDI) is crucial for preventing adverse reactions in patients and plays a vital role in drug design and development. However, traditional Chinese medicine (TCM) formulations, typically composed of multiple herbal ingredients with diverse bioactive compounds, present a unique challenge in comprehensively assessing potential adverse interactions among their components. To address this challenge, we propose a novel Dual Graph Attention Network (DGAT) designed to predict TCM drug-drug interactions (TCMDDI) by extracting key structural features of active molecules within the herbal ingredients. Our approach leverages graph-based representations of chemical molecules and employs attention mechanism to extract deep structural features, enabling the effective prediction of TCMDDI by capturing spatial structural relationships among different compounds. Furthermore, we construct a comprehensive dataset encompassing three different categories of herbal ingredients, informed by traditional TCM principles. Experimental results reveal that the proposed DGAT method significantly outperforms currently advanced deep learning techniques, including Graph Convolutional Networks, Weave, and Message Passing Neural Networks. Compared to traditional rule-based two-dimensional molecular descriptors, DGAT more effectively captures the spatial structural information of molecules. Notably, DGAT exhibits robust performance and strong generalizability on unseen samples, providing valuable insights for future research on TCMDDI prediction and advancing the integration of artificial intelligence in TCM studies.