Drug Repositioning Based On The Enhanced Message Passing And Hypergraph Convolutional Networks

Drug repositioning,an important method of drug development,is utilized to discover investigational drugs beyond the original approved indications,expand the application scope of drugs,and reduce the cost of drug development.With the emergence of increasingly drug-disease related biological networks,the challenge still remains to effectively fuse biological entity data and accurately achieve drug-disease repositioning.Therefore,this paper proposes a drug repositioning method based on message passing enhancement and Hypergraph convolutional network,the specific content is as follows:(1)Firstly,it constructs homogeneous multi-view information with multiple drug similarity features,and then extracts the feature information of each view through the Hypergraph convolutional network and uses the channel attention mechanism to score each view to obtain drug embeddings corresponding to each view,and finally the drug embeddings of all views are combined to obtain the intra-domain embedding of the drug.Similarly,the intra-domain embedding of the disease is also extracted by the Hypergraph convolutional network.(2)Secondly,it extracts the inter-domain information of known drug-disease associations by constructing a graph convolutional network that combines node and edge embeddings,and enhances the message passing between drug and disease domains by constructing a heterogeneous network consisting of association information of drug,protein and disease with graph attention mechanism.Then,the inter-domain information of the known drug-disease association is fused with the enhanced interdomain information to obtain the inter-domain embeddings of drugs and diseases respectively.(3)Finally,the intra-domain and inter-domain embeddings of drugs and diseases are integrated to form the final embedding,and the drug embedding matrix is multiplied by the disease embedding matrix to obtain the final predicted drug-disease association matrix.Each value in the matrix represents the likelihood of a drug-disease association.Through the 10-fold cross-validation on some bench-mark datasets,we find that the AUPR of our approach reaches 0.593(T1)and 0.526(T2)respectively,and the AUC achieves 0.887(T1)and 0.961(T2)respectively,which shows that our approach has an advantage over other state-of-the-art prediction methods.In the case study,our approach also achieves satisfactory results in real drug repositioning for breast carcinoma and Parkinson’s disease,providing a reliable reference for subsequent drug development.