| MicroRNA(miRNA)is a single-stranded RNA that can regulate gene expression.Studies have shown that abnormal expression of miRNA can cause a variety of complex diseases.Uncovering the interaction between miRNA and disease is conducive to understanding the pathogenesis of the disease and has great significance for disease diagnosis and drug development.It is helpful to construct a reliable miRNA-disease prediction model which can provide the high probability potential miRNA-disease association for verification in traditional biological experiments,thereby reducing the number of biological experiments and reducing time and capital expenditure.Based on miRNA and disease-related data,we combined the Bayesian matrix decomposition algorithm and graph regularization transduction regression algorithm to construct two miRNA-disease association prediction models.Compared to traditional machine learning methods,Bayesian matrix decomposition has higher accuracy and significantly reduces the computational complexity.Heterogeneous network-based graph regularized transduction regression method can efficiently extract information on sparse networks and improve the prediction accuracy of the model.The first model we proposed was the association prediction model based on the Bayesian matrix decomposition algorithm(KBMFMDA).First,we used miRNA similarity and disease similarity to project miRNA and disease into a unified subspace.Then,the conjugated Bayesian probability formula was used to infer the target output of the projected subspace.Finally,the product of the two target outputs of the subspace was used as the miRNA-disease association prediction score.The AUCs of KBMFMDA in global and local leave-one cross-validation and five-fold crossvalidation are 0.9132,0.8708 and 0.9008+/-0.0044,which are better than many previous models.Three case studies of colon,lymphoma and esophageal tumors show that there are 44,47 and 48 of the top 50 disease-related miRNAs predicted by KBMFMDA were validated by the database,respectively.The experimental results show that the KBMFMDA has reliable and highly accurate prediction performance.The second model we proposed was a graph-regularized transduction regression model based on heterogeneous networks to predict miRNA-disease association(GRTRMDA).We first constructed a three-layer heterogeneous network about miRNA,lncRNA and disease.We then made a preliminary estimate of the association scores of miRNA and lncRNA without known related disease to the disease.The final miRNA-disease association score could be obtained by performing transduction regression method in the heterogeneous network.The AUCs of GRTRMDA in global and local leave-one cross-validation and five-fold cross-validation are 0.9057,0.8372 and 0.9033+/-0.0008 which are better than many previous models.Three case studies of the three diseases(lymphoma,breast tumor and esophageal tumor)show that there are 44,44 and 50 of the top 50 disease-related miRNAs predicted by GRTRMDA were validated by the database,respectively.The experimental results show that the GRTRMDA can reliably predict the potential associations between miRNA and disease. |
