Research On Candidate Prediction Methods For Drug-related Diseases Based On Deep Learning

Traditional drug research and development not only requires a lot of investment,but also takes a long time to complete the development of a novel drug.Thence,drug repositioning can reduce the cost and the period during drug development.However,most of the previous methods are based on shallow learning methods.These methods fail to take into account the sparsity of drug-disease associations and cannot fuse the multiple information between the drugs and diseases.Therefore,it is important to develop a prediction method that can deeply represent and fuse the information of drugs and diseases.We propose two methods in this work.The first method is random walk and convolutional neural network through similarity fusion,and the second method is based on convolutional neural network and gate recurrent unit for model fusion.(1)Prediction method for the association between drugs and diseases based on convolutional neural network and two-way random walkWith the development of bioinformatics,more and more information were discovered by researchers.How to select and integrate these information becomes a big problem.In this part,we calculated two types of drug similarities and two types of diseases similarities.Drug similarities include chemical substructure similarity and dichotomy network projection similarity.Disease similarities consist of disease phenotype similarity and dichotomy network projection similarity.We calculated the similarities from two views,and then constructed a drug-disease heterogeneous network.First,the network transfer matrix is constructed to propose a two-way restart random walk to learn the network topology information.Second,deep learning and feature fusion are performed through the convolutional neural network to obtain the CNNRW prediction model candidate diseases for drugs.Comparing CNNRW with the other four advanced prediction methods,the results show that CNNRW is superior to the other methods.In addition,a case study on 50 candidate diseases of 5 common drugs shows that CNNRW has the ability to discover the candidate diseases of drug.(2)Association between drugs and diseases prediction method based on convolutional neural network and gated circulation unitBecause most of the previous drug disease prediction models are focused on shallow learning method,and there are many nonlinear relationships between the drugs and diseases,we propose a deep learning-based model,called CGAPred,to capture the complicated relationship between the drugs and the diseases.The left side of this model builds a feature matrix based on the drug disease heterogeneous network and learns the node information of drug and disease through the convolutional neural network,and the right side of the model obtains the path information of the drug to the disease based on the drug disease network and adds attention weights to different paths.Finally,through the fusion of left and right information as the final score of association of the drug and disease prediction.By comparing the CGAPred method with the other four advanced methods,CGAPred is indeed superior to other methods.In our three evaluation indicators(AUC,AUPR,TOP K),CGAPred also achieved the best results.In the case study on 50 candidate diseases of the five common drugs,CGAPred has the ability to find candidate diseases of drugs,providing a reliable reference for subsequent drug development.