Research On The Prediction Method Of Drug-disease Association Based On Deep Learning

Developing new therapeutic effects for existing drugs can help reduce the cost of drug development.However,previous prediction methods often only consider singlesource drug and disease data,do not effectively integrate multiple information,and do not consider the sparsity of drug-disease association data.Therefore,it is necessary to develop a drug-disease association prediction method based on a variety of bioinformatic data.We propose two prediction methods which integrate a variety of biological information,one is based on random walk prediction method,the other is based on depth learning prediction method.(1)A prediction method of drug-disease association based on the random walkWith the accumulation of bioinformatics data,more and more data can be applied to drug-disease prediction.How to efficiently integrate these data and use this data to predict drug candidate diseases is an important issue.In this part of the study,we calculated three different similarities between drugs by combining the chemical substructure of the drug,the target protein domain of the drug,and the gene ontology of the drug.Three kinds of drug similarity networks from different perspectives were constructed.Combined with drug similarities,disease similarities and drug-disease associations,a multi-layer heterogeneous network is constructed,which contains a variety of drug similarities and disease information.A novel method for predicting the drug-disease association based on the random walk is proposed(MultiNRW).We weight different network layers to balance the impact of each network layer,and construct the transition matrix of heterogeneous networks.At the same time,the restart mechanism of random walk is added to control the range of walk and prevent the introduction of too much noise data.We compare the MultiNRW method with the other four prediction methods.The comparison results show that the MultiNRW method has better prediction ability than other prediction methods.In addition,the case studies of 50 candidate diseases of 5 drugs further proved that the MultiNRW method has the ability to discover potentially diseases.(2)A prediction method of drug-disease association based on dual convolutional neural networkThere are many complex and non-linear relationships between drugs and diseases.Traditional prediction methods are shallow models,and it is difficult to capture these associations.We propose a novel drug-disease association prediction method based on dual convolution neural network(DCPreRD).Combined with drug-associated disease information,the fourth similarity was constructed.By combining the four drug similarities,disease similarities and drug-disease associations,we established a deep learning model to predict drug-related diseases.The left part of the model learns the original representation of drugs and diseases from the original characteristics of drugs and diseases.The right part learns the neighbor representation from the neighbor node information of the drug and disease.We compare the DCPreRD method and the MultiNRW method with several other prediction methods.Among the three evaluation methods of AUC,AUPR and Top k,the DCPreRD method achieves the better prediction performance.The case studies of 50 candidate diseases show that DCPreRD method can identify potential relevant disease candidates,which provides a very reliable basis for further clinical trials of biologists.