基于多层次结构建模的药物相互作用类型预测  

Drug-drug interaction prediction via hierarchical structure modeling

作　　者：吴怀瑾 卜家梓 杨念祖 孙垚 宋海涛[1,3] 柳宁 严骏驰[1,4] Huajin WU;Jiazi BU;Nianzu YANG;Yao SUN;Haitao SONG;Ning LIU;Junchi YAN(MoE Key Lab of Artificial Intelligence,Shanghai Jiao Tong University,Shanghai 200240,China;Zhejiang Lab,Hangzhou 311121,China;Shanghai Artificial Intelligence Research Institute,Shanghai 200240,China;School of Artificial Intelligence,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学人工智能教育部重点实验室,上海200240 [2]之江实验室,杭州311121 [3]上海人工智能研究院,上海200240 [4]上海交通大学人工智能学院,上海200240

出　　处：《中国科学:信息科学》2025年第2期356-371,共16页Scientia Sinica(Informationis)

基　　金：国家自然科学基金重大研究计划(批准号:92370201);国家自然科学基金优秀青年基金(批准号:62222607)资助项目。

摘　　要：药物相互作用(drug-drug interactions,DDIs)通常在多种药物共同使用时发生.某些药物之间的相互作用可以带来积极的治疗效果,而某些药物相互作用引起的反应可能会带来严重的影响,甚至危及生命,这使得预测某两种药物之间是否存在相互作用乃至何种类型变得至关重要.而实验研究DDIs的需求和成本不断增长,凸显了采用计算方法,尤其是机器学习方法以预测DDIs的价值.然而,此前基于机器学习的方法均未能全面建模DDIs问题中包括药物自身分子子结构、全结构和药物相互作用网络结构在内的3个层级信息,部分方法还依赖于大规模生物知识图谱,增加了模型的推理成本.本文提出HierDDI,它仅通过数据集中的DDIs即能构建一个多层次图结构,而无需依赖附加知识图谱.更具体地说,HierDDI通过建模药物自身分子子结构、全结构、药物相互作用网络结构这3个层面,来预测药物相互作用.由此不仅捕获了药物自身的结构信息,还将药物间的相互作用考虑在内,并通过端到端的训练,建立模型对DDIs的全面理解.本文所有模块的设计都非常简洁,广泛的实验展示了本文提出的方法明显优于目前药物相互作用领域的领先方法,成为目前表现最佳的模型.在DrugBank和Twosides等公开数据集上,本文提出的方法在直推式(transductive)场景中的准确率比当前领先方法提高了0.58%和2.02%,而在归纳式(inductive)场景中则更是提高了4.20%和3.52%.源代码发布在https://github.com/wuhuaijin/HierDDI.Drug-drug interactions(DDIs)often occur when a drug is co-administered with other drugs.Some interactions between drugs can result in positive therapeutic effects,while reactions caused by others can have serious consequences,even life-threatening,making it crucial to predict whether the interaction or what type exists between two drugs.The escalating demand and cost of experimental DDI studies underscore the requirement for automated computational tools,especially machine learning methods,to predict DDIs.However,previous machine learning methods have failed to comprehensively model multiple levels of information,including the substructure level,molecule level,and DDI network level.Some methods also rely on large-scale biomedical knowledge graphs,increasing the inference cost.In this paper,we propose HierDDI,which constructs a hierarchical graph structure solely from known DDIs,without relying on external knowledge graphs.Specifically,HierDDI predicts DDIs through modeling the aforementioned three levels,capturing both inter-drug and intra-drug interactions to establish a comprehensive machine understanding of DDIs.While all modules were designed with simplicity,extensive experiments showcase that our method outperforms baselines by a notable margin.Particularly,on DrugBank and Twosides public benchmarks,our method achieves 0.58%and 2.02%improvement in accuracy over the strongest baselines in the transductive setting,while these two values reach up to 4.20%and 3.52%in the inductive setting.The source code is publicly available at https://github.com/wuhuaijin/HierDDI.

关 键 词：机器学习 药物相互作用 多层次 图神经网络 端到端 

分 类 号：R9[医药卫生—药学] TP181[自动化与计算机技术—控制理论与控制工程]