Prediction Of MicroRNA-disease Association Based On Graph Convolutional Network

Recently,an increasing number of studies have demonstrated that miRNAs are involved in human diseases,indicating that miRNAs might be a potential pathogenic factor for various diseases.Therefore,figuring out the relationship between miRNAs and diseases plays a critical role in not only the development of new drugs,but also the formulation of individualized diagnosis and treatment.As the prediction of miRNA-disease association via biological experiments is expensive and time-consuming,computational methods have a positive effect on revealing the association.In this thesis,we fully consider the spatial characteristics of the datasets,for the the relationship between miRNAs and diseases.Here are three major contributions:(1)In the data extraction,considering the important role of genes in the association of miRNA and disease,gene network is introduced between miRNA and disease networks to form a three-layer heterogeneous network structure(disease,gene,miRNA);In addition,to reflect similar information between nodes,we uses hierarchical directed acyclic graph(DAG)and MISIM to generate disease semantic similarity and miRNA functional similarity,respectively,which update to the disease and miRNA layer,and form a new three-layer similarity heterogeneous network.(2)This thesis constructs a new network framework model GCSENet to predict miRNA disease association.The model first captures features(including disease-gene,miRNA-gene)by Graph Convolutional Network(GCN)based on a similarity heterogeneous graph including diseases,genes and miRNAs.A feature weight is set according to the degree of gene node,and the combination of different features is added as the new feature component input for the convolution operation in CNN,the squeeze and excitation blocks of SENet were applied to determine the importance of each feature channel and enhance useful features by means of the attention mechanism,thus achieving a satisfactory prediction of miRNA-disease association.For the optimization of the model,the cross-entropy loss function with regularization term and cross-validation are used to obtain the tuning model.(3)In order to illustrate the performance of the model framework,firstly,verifying the effect of the four model settings.Then,the proposed method was compared against other state-of-the-art methods(WBSMDA?PBMDA?MDACNN?SAEMDA ? NIMCGCN).The results show that the average AUROC,AUPR,Precision,Recall,F1-score indicator in this paper can reach 95.02%,95.55%,87.95%,84.94% and 86.42%,respectively,which are higher than the comparison method.we set a different proportion of positive and negative examples to test the performance of our GCSENet model.In addition,the robustness of the model is proved on miRNA phenotype prediction.Finally,we carried out case studies on four important diseases(lung neoplasm,heart failure,breast cancer and glioblastoma),which were used to evaluate the performance of our model regarding to the associations with experimental evidences in literature.Each of the above experimental results and analysis are data visualization,which can show the good performance of the model more intuitively.Finally,this thesis summarizes the research work of miRNA disease prediction and prospects for future work.