Prediction Of MiRNA-disease Associations Based On Simplified Graph Convolutional Networks

There are a variety of non-coding RNAs in the living organisms,miRNA is widely present in eukaryotes and plays an important role in gene expression by interfering with post-transcriptional regulation.In recent studies,a large number of miRNAs are associated with the development of human diseases.Prediction of miRNA-disease associations has important implications for explaining the causes of human diseases and promoting the prognosis and treatment of related disorders.Traditional experimental methods for validating miRNA-disease associations are time-consuming,laborious,very expensive,and have a high failure rate.Furthermore,Discovery of miRNA-disease associations by computational methods has attracted extensive attention in the scientific community.A series of bioinformatics methods have been developed to identify possible miRNA-disease associations,such as models based on scoring functions,complex network algorithms,machine learning,and various biological information.The successful application of previous models to identify possible miRNA-disease associations is based on the biological premise that similar miRNAs are more likely to be closely related to similar diseases.However,the previous methods also have shortcomings that cannot be ignored,such as not fully considering the high-order link information between miRNAs and diseases,the implementation process is complicated,the performance of the method can be further optimized,the application of local information leads to high false positives,and so on.In this paper,we propose a novel method for identifying miRNA-disease associations(LGCNMDA)based on a simplified graph convolutional network,on the basis of existing miRNA-disease associations,the main work is to construct the expression modeling of high-order connectivity between miRNAs and diseases through a simple graph neural network,and integrate and update the existing miRNA-disease similarity network by combining the Gaussian kernel function,matrix completion is performed by integrating disease similarity network,miRNA similarity network and disease-miRNA association information to update the miRNA-disease association matrix,the resulting data is the prediction score for the new miRNA-disease association.The performance of the method is compared with existing methods by leave-one-out cross-validation.The AUC score of LGCNMDA,IMCMDA,RLSMDA,TLHNMDA and WBSMDA were 0.9012,0.8378,0.8193,0.8795 and 0.8010,respectively.Then,according to the predicted scores of the unknown associations,the potential actual associations are found out by comparing with the database HMDD v3.2,miR2Disease and dbDEMC to demonstrate the performance of the method.We selected three common human diseases colon tumor,kidney tumor and breast tumor as case studies,the predicted top 10 miRNAs were all obtained in the comparison database,and 48,48 and 47 of the top 50 miRNAs were obtained in the comparison database,respectively.In conclusion,the LGCNMDA model has reliable performance for predicting miRNA-disease associations.