A MiRNA-disease Associations Prediction Method Based On Higher-order Proximity In Heterogeneous Networks

The prediction of the associations between miRNAs and diseases plays a very important role in identifying miRNAs related to human diseases.This method not only helps to develop personalized diagnosis plans but also helps to understand the pathogenesis of the disease.It is very time-consuming and expensive that use biological methods to identify the associations between miRNAs and diseases.Using some calculation methods is an effective supplementary method to identify the associations between miRNAs and diseases.Therefore,it is great to research significance to develop relevant calculation methods and accurately and effectively predict the associations between miRNAs and diseases.The main contents of this thesis are:(1)First,we use the latest miRNAs and diseases association data set(HMDDv3.0)to generate a heterogeneous network of miRNA and disease for it.This data increases the number of miRNAs and diseases and the associations between miRNAs and diseases compared to previous data.This provides the basis for future methods.(2)A higherorder proximity calculation method(HOP_MDA)is proposed,which fully exploits the higherorder interaction between miRNAs and diseases,and uses supervised learning to tune higher-order weights.This is a solid step for further research.(3)In this paper,a neural induction matrix completion method(NIMCHOP_MDA)with higher-order proximity is proposed by combining the high order proximity calculation method and the induction matrix completion method with better performance.This method no longer uses the bilinear evaluation model,but improves the projection matrix in the inductive matrix completion method to a nonlinear fully connected layer.Then the similarity features of miRNA and disease are generated by the higher-order proximity calculation method and consider as the input of the nonlinear fully connected layer.(4)In order to improve the operating efficiency of the higher-order similarity method,a higher-order approximation method is proposed in this paper.This method can approximate the higher-order matrix decomposition results using the lower-order matrix decomposition results.This method reduces the time complexity to a linear level.This approximation method is applied to the HOP_MDA and NIMCHOP_MDA methods,respectively.The HOPA_MDA and NIMCHOPA_MDA method are proposed.This thesis uses a variety of verification methods to test the proposed method on three data sets.These include 5-fold-cross-validation,leave-one-cross-validation,and case studies.In terms of data,the two most common data from previous researches and the latest data sets were used.The results of AUCs obtained by the method of five-fold cross-validation on the latest data set are HOP_MDA: 0.9163 +/-0.0014,NIMCHOP_MDA: 0.9221+/-0.0023,HOPA_MDA(With higher-order approximation): 0.9072+/-0.0012,NIMCHOPA_MDA(With higher-order approximation):0.9137+/-0.0014.The case study method is to list the top 50 most diseaserelated miRNAs output from the prediction model and query the database one by one to determine whether there is a true correlation.For the diseases lung cancer and breast cancer,the accuracy of the HOP_MDA method is 96%,100% on the latest data set.The accuracy of the NIMCHOP_MDA method is 96%,100%.All in all,compared with the existing research on three different data sets,the methods proposed in this paper are superior to other methods.