目的 针对中药复方成分复杂、相互作用关系不明导致配伍机制难以解析的问题，以成分间相互作用预测为切入点，构建基于深度学习的分子相互作用预测模型，为识别中药成分间潜在相互作用关系及后续方剂配伍机制研究提供方法学参考。方法 提出基于分子3D结构特征与扩展连通性指纹的分子相互作用预测模型（3D structure and extended connectivity fingerprints molecule-molecule interaction，3DFPMMI）。模型利用SchNet网络提取分子3D图结构的深层特征，同步获取扩展连通性指纹，以对比学习策略实现多源特征的有效融合；模型在DrugBank药物相互作用数据集上完成训练，并通过跨域泛化实现中药成分间相互作用的预测。结果 3DFPMMI模型在DrugBank数据集上准确率达到0.921，精确率-召回率曲线下面积为0.972，均优于其他基准模型。消融实验证实，分子3D结构特征、扩展连通性指纹特征及对比学习特征融合策略均为模型性能提升的核心有效模块。将该模型应用于经典方剂补阳还五汤和半夏泻心汤的成分相互作用预测，其部分预测结果与已有的实验研究结论高度吻合。结论 构建了一种中药成分相互作用预测的新方法，并初步验证了其应用潜力，为从分子层面解读中药方剂配伍机制提供了新的技术手段与方法学借鉴。

Objective Given the complexity of traditional Chinese medicine (TCM) formula constituents and their unclear interactions, a deep learning-based molecular interaction prediction model is constructed taking the prediction of interactions among constituents as the entry point, providing methodological support for identifying potential constituent interactions and subsequent studies on formula compatibility mechanisms. Methods We propose a molecule-molecule interaction prediction model based on molecular 3D structural features and Extended Connectivity Fingerprints, designated as 3D structure and extended connectivity fingerprints molecule-molecule interaction (3DFPMMI). The model employs the SchNet network to extract deep features from the 3D graph structure of molecules while simultaneously acquiring Extended Connectivity Fingerprints, and achieves effective fusion of multi-source features through a contrastive learning strategy. The model is trained on the DrugBank drug-drug interaction dataset, and cross&8209;domain generalization is implemented to realize the prediction of interactions among TCM constituents. Results The 3DFPMMI model achieves an accuracy of 0.921 and an area under the precision-recall curve of 0.972 on the DrugBank dataset, both outperforming other benchmark models. Ablation studies confirm that molecular 3D structural features, extended connectivity fingerprint features, and the contrastive learning-based feature fusion strategy are all core effective modules for improving model performance. When applied to predict component interactions in the classic TCM formulas Buyang Huanwu Decoction and Banxia Xiexin Decoction, some of the prediction results are highly consistent with the conclusions from existing experimental studies. Conclusion This study establishes a novel method for predicting interactions among TCM ingredients and preliminarily validates its application potential. This work offers new technical tools and methodological references for interpreting the compatibility mechanisms of TCM formulas at the molecular level.

TP18;R284

国家中医药管理局中医药创新团队及人才支持计划项目（ZYYCXTD-D-202408）；国家自然科学基金项目（61702164）；河南省卫生健康委国家中医临床研究基地科研专项课题（2021JDZX2135）