Research On Prediction Of MiRNA-Disease Association Based On Mixed Neighborhood Graph Convolution

MiRNA is an endogenous noncoding RNA that plays a key role in a variety of biological processes.Identifying potential associations between miRNAs and diseases is helpful for the diagnosis and treatment of diseases.However,there are disadvantages such as high cost and long cycles to discover the associations between them through biological experiments.Therefore,some researchers in the field of bioinformatics try to build suitable computational models to predict potential miRNA-disease associations based on the existing miRNA-disease associations and related information.This provides guidance information for subsequent biometric verification.Graph convolutional networks have a strong representation learning ability for graph-structured data,and have been successfully applied to the prediction task of miRNA-disease associations.In the learning process of existing methods,only direct neighbor node information is used to update node embeddings of miRNAs and diseases,which ignores the valuable information contained in higher-order neighbors.To learn the node embeddings of miRNAs and diseases more effectively from high-order neighbors and improve the prediction effect of miRNA-disease associations,this thesis proposes a miRNA-disease association prediction model based on mixed neighborhood graph convolution(MINIMDA).The main work is as follows:(1)Building a multi-view network using multi-source data on miRNAs and diseases.The network includes miRNA similarity network,disease similarity network and miRNA-disease association network,so that the MINIMDA model can synthesize the complex interaction between miRNA and disease in the process of learning node embedding,and obtain a more effective node representation.(2)Aggregating mixed neighborhood information and updating node embeddings by iteration.MINIMDA uses a mixed neighborhood graph convolutional layer to expand the aggregation range of node neighborhoods,so that the learned miRNA and disease node embeddings can explicitly fuse neighbor information at different orders.(3)In this thesis,we compared 13 mainstream miRNA-disease association prediction models on two benchmark datasets,and the experimental results verified the superiority of MINIMDA.In addition,this thesis presents case studies of esophageal cancer,colon cancer and lung cancer.Among the 50 most relevant miRNAs predicted by MINIMDA,48,49,and 47 miRNAs have been verified by biomedical experiments.