基于图注意力网络的抗癌药物组合协同性预测方法  

A prediction method for anti-cancer drug combinations synergy based on graph attention network

作　　者：秦伟琦 包欣 陈晓[3,4] 邱建龙 王东琳 QIN Weiqi;BAO Xin;CHEN Xiao;QIU Jianlong;WANG Donglin(Affiliated Hospital of Shandong Medical College(Linyi Geriatric Hospital),Linyi 276004,China;School of Automation and Electrical Engineering,Linyi University,Linyi 276005,China;Key Laboratory of Complex Systems and Intelligent Computing in Higher Education Institutions of Shandong Province,Linyi University,Linyi 276005,China;School of Information Science and Engineering,Linyi University,Linyi 276005,China)

机构地区：[1]山东医学高等专科学校附属医院(临沂市老年病医院),山东临沂276004 [2]临沂大学自动化与电气工程学院,山东临沂276005 [3]临沂大学山东省高等学校复杂系统与智能计算重点实验室,山东临沂276005 [4]临沂大学信息科学与工程学院,山东临沂276005

出　　处：《南通大学学报(自然科学版)》2025年第1期10-17,共8页Journal of Nantong University(Natural Science Edition)

基　　金：国家自然科学基金面上基金项目(62173175,61877033)。

摘　　要：抗癌药物组合的协同性筛选对于临床治疗具有重要意义,但随着药物组合数量的爆炸式增长,传统检测方法存在耗时长、成本高等问题,难以有效发现新的协同药物组合。针对上述问题,提出一种基于图注意力网络的抗癌药物组合协同性预测模型(multi-scale feature fusion model based on graph attention network for anticancer synergistic drug combination prediction,MFGSynergy)来辅助抗癌药物组合筛选。首先,该模型将药物简化分子线性输入规范(simplified molecular input line entry system,SMILES)编码为分子图及分子指纹数据,并对癌细胞系数据进行预处理;然后,通过图注意力网络(graph attention network,GAT)和多层感知机(multilayer perceptron,MLP)对药物数据及癌细胞系数据进行特征提取,并将提取到的多种药物特征和癌细胞系特征进行特征融合用于预测抗癌药物组合的协同性;最后,基于公开数据集将MFGSynergy与Deep DDS、DeepSynergy及6种机器学习方法进行对比实验,实验结果表明,MFGSynergy在五折交叉验证上的ROC曲线下的面积(receiver operating characteristic area under the curve,ROC AUC)、PR曲线下的面积(area under the precision-recall curve,AUPR)、准确性(accuracy,ACC)、精准度(precision,PREC)、真阳性率(true positive rate,TPR)和F1分别达到了0.94、0.94、0.86、0.87、0.86、0.86,均高于其他对比模型,这说明MFGSynergy的预测性能优于其他对比模型。此外,独立测试实验表明,对于未知的药物组合,MFGSynergy仍具有良好的预测性能,这证明模型具有良好的泛化性。Screening for synergistic anticancer drug combinations is essential for clinical treatment.However,the exponential rise in potential combinations renders traditional methods time-intensive and expensive,impeding the discovery of novel synergies.To overcome this,multi-scale feature fusion model based on graph attention network for anticancer synergistic drug combination prediction(MFGSynergy)is introduced,a graph attention network-based model to streamline anticancer drug combination screening.Initially,the model converts drug simplified molecular input line entry system(SMILES)into molecular graphs and fingerprint data while preprocessing cancer cell line data.It then employs a graph attention network(GAT)and multilayer perceptron(MLP)to extract features from both drug and cell line data,fusing these multi-source features to predict combination synergy.Evaluated on a public dataset,MFGSyne-rgy outperforms Deep DDS,DeepSynergy,and six machine learning methods,achieving receiver operating characteri-stic area under the curve(ROC AUC),area under the precision-recall curve(PR AUC),accuracy(ACC),precision(PREC),true positive rate(TPR),and F1 scores of 0.94,0.94,0.86,0.87,0.86,and 0.86,respectively,in five-fold cross-validation.Moreover,independent tests on unknown combinations validate its robust predictive power,underscoring MFGSynergy′s superior generalization.

关 键 词：抗癌药物联合治疗 分子指纹 图注意力神经网络 深度学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] R9[自动化与计算机技术—控制科学与工程]