基于磁性符号图神经网络的药物互作用关系预测  

Predicting Drug-drug Interactions via Magnetic Signed Graph Neural Network

作　　者：陈明 姚斌 胡子涛 马华 CHEN Ming;YAO Bin;HU Zi-tao;MA Hua(School of Information Science and Engineering,Hunan Normal University,Changsha 410081,China)

机构地区：[1]湖南师范大学信息科学与工程学院,湖南长沙410081

出　　处：《计算机技术与发展》2024年第6期103-109,共7页Computer Technology and Development

基　　金：国家自然科学基金项目(62077014);湖南省自然科学基金(2023JJ30411);长沙市自然科学基金(kq2202248);湖南省教育厅优秀青年项目(22B0097)。

摘　　要：药物互作用(Drug-Drug Interactions,DDIs)是病人同时或在一定时间内先后服用两种或两种以上药物后产生的复合效应,它可使药效加强或减弱,甚至产生不应有的毒副作用。图神经网络(Graph Neural Networks,GNNs)是当前流行的一种DDIs预测框架,它已在单类型和多类型的DDIs场景中取得了很好的效果。然而,大多数现有方法忽视了DDIs的语义描述细节,例如,药理学变化中的符号和角色信息。针对DDIs数据中药理学变化符号和角色的共存现象,构建了带节点属性的符号有向图,并将DDIs关系预测定义为符号与方向的联合预测任务。引入磁性符号图神经网络(Magnetic Signed GNN,MSGNN),提出了MSGNN-DDI预测方法,用于同时挖掘DDIs网络的符号和方向信息。该方法不仅可用于预测DDIs关系中的符号或方向,而且适用于它们的联合预测任务。从DrugBank和PubChem数据库中提取药物节点属性与DDIs,所得实验结果表明该方法能有效应对药理学变化的符号与方向联合预测,且在符号、方向单任务上与基线方法具有可比性。Drug-drug interactions(DDIs)refer to the complex effects that occur when patients take two or more drugs simultaneously or within a certain period,potentially leading to an enhancive or degressive effect on drug efficacy or even the occurrence of undesired toxic side effects.Graph neural networks(GNNs)are a popular framework for predicting DDIs,which has achieved promising results in the scenarios of mono-type and multi-type DDIs.However,most studies tend to overlook semantic details of DDIs,like the signs and roles information in pharmacological variation.Aiming at the coexistence of pharmacological variation signs and roles in DDIs data,the signed directed graphs with node attributes are built,and DDIs prediction is defined as a joint task involving signs and directions.By introducing the magnetic signed GNN(MSGNN),a MSGNN-DDI prediction approach is proposed to simultaneously extract sign and direction information within the DDIs network.This method can be used not only to predict the signs or directions in DDIs relationships but also for joint prediction tasks.Experiments based on DrugBank and PubChem databases demonstrate that MSGNN-DDI not only effectively handles joint prediction of pharmacological signs and directions but also exhibits comparability with baseline methods in the single task of sign or direction prediction.

关 键 词：药物互作用 符号有向图 谱图神经网络 磁性符号拉普拉斯 符号与方向联合预测 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]