基于深度学习与多层次信息融合的药物靶标亲和力预测*

药物研发是非常重要但也十分耗费人力物力的过程。利用计算机辅助预测药物与蛋白质亲和力的方法可以极大地加快药物研发过程。药物靶标亲和力预测的关键在于对药物和蛋白质进行准确详细地信息表征。提出一种基于深度学习与多层次信息融合的药物靶标亲和力的预测模型,试图通过综合药物与蛋白质的多层次信息,来获得更好的预测表现。首先将药物表述成分子图和扩展连接指纹两种形式,分别利用图卷积神经网络模块和全连接层进行学习;其次将蛋白质序列和蛋白质K-mer特征分别输入卷积神经网络模块和全连接层来学习蛋白质潜在特征;随后将4个通道学习到的特征进行融合,再利用全连接层进行预测。在两个基准药物靶标亲和力数据集上验证了所提方法的有效性,并与其他已有模型作对比研究。结果说明提出的模型相比基准模型能得到更好的预测性能,表明提出的综合药物与蛋白质多层次信息的药物靶标亲和力预测策略是有效的。

Drug-target Affinity Prediction Based on Deep Learning and Multi-layered Information Fusion

Drug discovery is a very important and costly process. Computer-assisted methods for predicting drug-protein affinity can greatly speed up the process of drug discovery. The key to the prediction of drug target affinity lies in the accurate and detailed characterization of drug and protein information. In this paper, a prediction model for drug target affinity based on deep learning and multi-layered information fusion is proposed, in an attempt to obtain better prediction performance by integrating multi-layered information of drugs and proteins. Firstly, the drug is expressed as molecular graph and ECFP, GCN module and fully connected(FC) layer are used for learning, respectively. Secondly, protein sequence and K-mer feature of protein are input into CNN module and FC layer, respectively to learn potential protein features. Finally, the features learned from the four channels are concatenated and the FC layer is used for prediction. In this study, the availability of the proposed method is verified on the two benchmark datasets of drug-targets affinity and compared with other existing models. The results show that the proposed model can obtain better prediction performance than the baseline model, which indicates that the proposed strategy for predicting drug target affinity based on multi-layered information fusion of drug and protein is effective.