The Research Of Drug Repositioning Based On Graph Convolutional Network

Traditional drug development is a time-consuming,expensive and high-risk process.Drug repositioning,also known as in silico reuse of old drugs,has gradually become an important research topic in the field of computational biology due to the known drugs with greatly reduced risk of use,which can shorten the drug development timelines and reduce the investment cost.In order to discover candidate drugs to cure diseases,researchers have proposed many drug-disease association prediction models based on graph convolutional neural networks.However,the graph structure ignores the differences of nodes in different domains.Considering that the drug node and the disease node belong to different biological domains,this paper proposes a drug repositioning algorithm based on heterogeneous information fusion graph convolution network(DRHGCN).DRHGCN uses the chemical structure of drugs,phenotypic information of diseases and confirmed drug-disease associations to construct drug similarity subnetwork,disease similarity subnetwork and drug-disease association subnetwork respectively.Based on the assumption of guilt-by-association,the drug-disease association subnetwork is used as a bridge to connect the above three subnetworks into a unified network,and the drug repositioning problem is modeled as a link prediction problem.First,the intra domain feature extraction module and inter domain feature extraction module based on graph convolution neural network are designed respectively.The intra domain feature extraction module extracts the internal features of the sub network,and the inter domain feature extraction module emphasizes the commonness between the sub networks and desalinates the inconsistent information.Then layer attention mechanism and residual connection are used to enhance the expressive ability of features.Finally,the matrix factorization decoder is used to decode the features to obtain the drugdisease association probability matrix.Aiming at the decision bias caused by the extremely sparse known associations in the training samples,this paper uses a weighted binary cross-entropy loss function to strengthen the influence of positive samples.The experimental results show that DRHGCN has excellent performance compared with six state-of-the-art association prediction algorithms on four drug repositioning benchmark datasets.In the case study of Alzheimer’s disease,Parkinson’s disease and COVID-19,according to the results of external data sources and molecular docking experiments,DRHGCN can effectively predict potential associations and provide theoretical guidance for biologists to effectively narrow down the search space of candidate drugs.