MiRNA-Disease Association Prediction Based On Heterogeneous Network Representation Learning

MiRNA is a type of endogenous non-coding single-stranded small RNA composed of 21～25 nucleotides.A large number of studies have shown that the abnormal expression of miRNA is closely related to the occurrence and development of human diseases.The discovery of potential miRNA-disease-related information is of great significance for miRNA function research,pathogenic mechanism analysis and disease treatment.Traditional biological experiment methods have the problems of long experiment period and high resource consumption.In recent years,under the guidance of bioinformatics technology,many computational model methods have emerged.These methods design computational models based on similarity features on verified biological data sets and predict potential miRNA-disease associations on a large scale.The task has achieved remarkable results.However,most of these computational models are designed based on co-acting target genes,ignoring the influence of target gene interaction on research tasks.Studies have shown that miRNAs mainly regulate the abnormal expression of target genes to cause the occurrence and development of diseases.Therefore,we believe that when predicting the relationship between miRNAs and diseases,it is necessary to consider the influence of target genes and the interaction between target genes.By collecting and sorting out related biological data such as miRNA-gene association network,disease-gene association network,miRNA-disease association network and protein interaction network,this thesis constructs a three-layer heterogeneous biological network and proposes two computing model to predict potential miRNA-disease associations,meta-path random walk-based prediction method HEMDA and graph attention network prediction method GANMDA.Both methods use network representation learning technology to mine heterogeneous network structure information on the basis of heterogeneous network modeling.Among them,the HEMDA model uses a meta-path random walk-based algorithm to learn the embedded representation of the structural features of the heterogeneous network on the three-layer heterogeneous biological network,and then combines the matrix factorization algorithm to predict the miRNA-disease correlation score.The GANMDA model integrates miRNA functional similarity and gaussian nuclear similarity features,disease functional similarity and gaussian nuclear similarity features,uses graph attention network to learn the vector representation of node features,and then uses matrix completion algorithm to analyze potential miRNA-disease associations are predicted.In the experimental part,we use the cross-validation method to test the validity of the model.In addition,we conducted case analysis on a variety of diseases,and verified the miRNA candidate set with high prediction scores on external data.The experimental results show that the two computational methods proposed in this article fully mine heterogeneous network structure information,and can predict potential miRNA-disease associations more efficiently,and provide a more reliable association candidate set for biological experiments.