Unexpected drug–drug interactions (DDIs) are important issues for both pharmaceutical research and clinical applications due to the high risk of causing severe adverse drug reactions or drug withdrawals. Many deep learning models have achieved high performance in DDI prediction, but model interpretability to reveal the underlying causes of DDIs has not been extensively explored. Here we propose MeTDDI—a deep learning framework with local–global self-attention and co-attention to learn motif-based graphs for DDI prediction. MeTDDI achieved competitive performance compared with state-of-the-art models. Regarding interpretability, we conducted extensive assessments on 73 drugs with 13,786 DDIs and MeTDDI can precisely explain the structural mechanisms for 5,602 DDIs involving 58 drugs. Besides, MeTDDI shows potential to explain complex DDI mechanisms and mitigate DDI risks. To summarize, MeTDDI provides a new perspective on exploring DDI mechanisms, which will benefit both drug discovery and polypharmacy for safer therapies for patients.

意外的药物相互作用（DDI）是药物研究和临床应用的重要问题，因为引起严重药物不良反应或停药的风险很高。许多深度学习模型在 DDI 预测方面取得了高性能，但揭示 DDI 根本原因的模型可解释性尚未得到广泛探索。在这里，我们提出了 MeTDDI——一种具有局部-全局自注意力和共同注意力的深度学习框架，用于学习基于主题的图以进行 DDI 预测。与最先进的模型相比，MeTDDI 取得了具有竞争力的性能。在可解释性方面，我们对 73 种药物的 13,786 个 DDI 进行了广泛的评估，MeTDDI 可以准确解释涉及 58 种药物的 5,602 个 DDI 的结构机制。此外，MeTDDI 显示出解释复杂 DDI 机制并降低 DDI 风险的潜力。总而言之，MeTDDI 为探索 DDI 机制提供了新的视角，这将有利于药物发现和多药治疗，为患者提供更安全的治疗。