A Drug-Target Interaction Prediction Based on GCN Learning

Abstract

In recent years, the use of deep learning methods for drug-target interaction (DTI) prediction has become the mainstream research direction. Drugs, targets, and other related biological and chemical properties have constructed a very complex network structure. How to effectively extract network features and predict target has become a big challenge. Graph Convolutional Neural Network (GCN) is one of the effective deep learning methods for complex networks. It extends the convolution operation from traditional European space to non-Euclidean space, and can simultaneously perform end-to-end node attribute information and structural information. End-to-end learning, its core idea is to learn a function mapping, through which nodes in the mapping graph can aggregate their own features and its neighbor features to generate a new representation of the node. In this study, we introduce the GCN link prediction method decagon for DTI prediction research. The experimental data comes from the DTI-net model. By combining the drug-drug interaction relationship matrix, the target-target interaction relationship matrix and the drug-target interaction relationship matrix provided by DTI-net, the drug characteristics and target characteristics are expressed as the attribute characteristics of the network nodes, thereby obtaining DTI Heterogeneous Network. In order to improve the ability to predict the drug-target relationship, this article has done a lot of tuning experiments in parameter selection and optimization strategies, and analyzed and compared the prediction results. The best predicted AUC is 0.919, and the best AUPR is 0.922. In terms of traditional drug-target prediction methods, the GCN method can effectively extract the features contained in heterogeneous networks, which proves the feasibility of this method in predicting drug-target interactions.