基于多任务消息传递神经网络的药物不良反应预测

Prediction of adverse drug reaction based on multi-task message passing neural network

对已知药物不良反应和药物蛋白质链路分别构建二分图,并分别基于消息传递神经网络(MPNN)和TransE模型进行建模,同时以交叉压缩单元(CCU)作为共享单元,联结药物不良反应预测和药物蛋白质链路预测,构建多任务MPNN(MT-MPNN)模型,提高对未知药物不良反应的预测效果。对SIDER公开数据集的89855例样本数据和DrugBank的5928例数据进行五折交叉验证,实验结果显示,在测试集上其平均受试者工作特征曲线下面积(AUROC)和平均F1值分别为0.9469和0.8753,表明本研究提出的MT-MPNN模型可以辅助临床有效挖掘潜在未知的药物不良反应。

In this study,bipartite graphs of known adverse drug reactions and drug-protein links were constructed respectively,which were modeled by message-passing neural network(MPNN)and TransE models separately,while the multi-task MPNN(MT-MPNN)model was constructed by integrating the prediction of adverse drug reactions and drugprotein links using cross-compress units(CCU)as a shared unit to improve the prediction performance of unknown adverse drug reactions.A 5-fold cross-validation was performed on a dataset containing 89,855 samples from open SIDER and 5,928 samples from open DrugBank,the results showed that the average area under the receiver operating characteristic curve(AUROC)and the average F1-score on the test set were 0.9469 and 0.8753,respectively,indicating that the MT-MPNN model proposed in this study can assist the clinical practice to effectively explore potential unknown adverse drug reactions.