Drug-target Relationship Prediction Based On Joint Learning And Interpretability Research

The process of new drug development includes target confirmation,model construction,lead compound discovery and selection and lead compound optimization.The three stages of target confirmation,model construction and lead compound discovery and selection are mainly to screen and confirm new compounds that can interact with the target.Effectively identifying new indications that may exist from approved drugs or more mature clinical drugs,that is,predicting the potential relationship between drugs and targets,can save a lot of labor and time costs in the early stages of new drug development.With more and more data sources related to drugs/targets,these heterogeneous data sources provide rich information for drugtarget relationship prediction from multiple different perspectives,which can improve the accuracy of drug-target relationship prediction to a certain extent.sex.The current drug-target relationship prediction methods based on heterogeneous networks have problems of insufficient interpretability and inability to perform relationship prediction for samples outside the training dataset.In response to these problems,this paper conducts the following joint learning-based drug-target relationship prediction and interpretability research.(1)Aiming at the problem of poor interpretability of methods based on heterogeneous networks,a drug-target relationship prediction method LUNAR based on graph convolutional network and attention mechanism is proposed.The method is based on heterogeneous graph convolutional neural network to learn the embedding representation of nodes in heterogeneous relational networks,and combines the attention mechanism to reflect the importance of different types of relational networks to node embeddings,so as to obtain interpretable node embedding representations.The edge-weight relationship in the heterogeneous relational network is reconstructed according to the learned node embedding,and the relationship is predicted by the edgeweight strength between drug targets.The experimental results show that LUNAR solves the problem of poor interpretability of prediction methods based on heterogeneous networks,and can reflect the degree of influence of different relational networks on node embedding representation.The AUC metric and AUC-PR metric of LUNAR on public datasets are 0.949 and 0.866,respectively.LUNAR outperforms other methods on the most important metric,AUC-PR,in which unknown edges are treated as negative samples.(2)Aiming at the problem that methods based on heterogeneous networks cannot predict samples outside the training dataset,a multimodal and sequence-based drugtarget relationship prediction method Multi DTI is proposed.Methods based on joint representation learning in multimodality combine interaction or association information in heterogeneous networks with sequence information of drugs/targets to map drug,target,side effect and disease nodes in heterogeneous networks into a common space,the drug-target potential relationship is predicted according to the distance between the drug and the target in this common space.Multi DTI can map the sequence information of new samples outside the training data set to the common space after training,and predict the relationship of new samples according to the sequence similarity relationship and network association relationship in the common space,thus solving the problem that existing methods cannot The problem of predicting samples outside the training dataset.The experimental results show that Multi DTI supports sample prediction outside the training dataset,and the AUC metric and AUC-PR metric on the public dataset are 0.961 and 0.947,respectively.Multi DTI outperformed other methods on the most important metric,AUC-PR.