Prediction of Drug-Disease Relationship on Heterogeneous Networks Based on Graph Convolution

Abstract

Drug-disease association prediction is essential in drug development and repositioning. At present, the proposed drug-disease association prediction models based on graph convolution usually learn the characterization of the entire drug-disease heterogeneous network. However, the obtained characterization information come more from the characteristics of neighboring nodes in the homogeneous network, it lacks attribute information of nodes in the heterogeneous network, thus affecting the model's predictive performance. In this paper, an end-to-end model named DAHNGC based on graph convolutional neural networks is proposed to predict drug-disease association, which divides the characteristic learning of drugs and disease nodes into two parts. The proposed model uses the graph convolutional network to learn the attribute characteristics of drugs and disease nodes in the homogeneous network. Based on the known relationship between drugs and diseases, we design a method to automatically learn the characteristic information of drugs and disease nodes in heterogeneous networks. Subsequently, the drug-disease association matrix is reconstructed using a bilinear decoder to obtain a potential drug-disease association. In addition, we also adopt the DropEdge method to alleviate the over-smoothing problem of graph convolution. The experimental results show that the average AUC of the DAHNGC is 0.9113 through five-fold cross-verification, which is superior to that of the comparative method.