Study On Prediction Algorithms Of Complex Disease Association Network Based On Data Fusion

In recent years,microRNAs(miRNAs),as diagnostic biomarkers and therapeutic targets of diseases,have been confirmed to be involved in many important biological processes and closely associated with kinds of human complex diseases.Therefore,developing effective computational methods to identify potential disease-miRNA associations will provide a new perspective for disease treatment.Currently,targeting miRNA related to diseases with drug small molecules can treat a variety of human complex diseases.Similarly,identification of miRNAs related to small molecules is of great significance for the treatment of diseases and the clinical application of drugs.Furthermore,studies have shown that in addition to the disorder of miRNAs can cause diseases,the disorder of microbes can also cause diseases.Therefore,predicting potential microbes related to diseases can help people understand the pathogenesis of diseases,which play an important role in the prevention,diagnosis,monitoring,prognosis and treatment of diseases.The purpose of this thesis is to construct heterogeneous network based on a variety of biological data and establish predictive models to predict potential disease-miRNA associations,small molecule-miRNA associations and disease-microbe associations.The main contents of this thesis are as follows:(1)Disease-miRNA association prediction model of TLHNMDA based on triple layer heterogeneous network was proposed.We integrated known disease-miRNA associations,known long noncoding RNA(lncRNA)-miRNA interactions,disease similarity,miRNA similarity and lncRNA similarity into a triple layer heterogeneous network,and proposed iterative update algorithm based on global network to predict potential disease-miRNA associations.In the model,we introduced known lncRNA-miRNA interactions and Gaussian interaction profile kernel similarity for lncRNAs,and treated lncRNAs as a bridge between diseases and miRNAs to construct an iterative update algorithm that can fully exploit network information to predict potential disease-miRNA associations.Meanwhile,TLHNMDA can also predict potential lncRNA-miRNA interactions.The results of cross validation and case studies showed that TLHNMDA can predict potential disease-miRNA associations effectively.(2)Disease-miRNA association prediction model of MDLPMDA based on matrix decomposition and label propagation was proposed.We first used sparse learning method(SLM)to decompose known disease-miRNA association matrix into a new disease-miRNA association matrix.Then,known disease-miRNA associations and new disease-miRNA associations were incorporated into disease similarity network and miRNA similarity network,respectively.At last,label propagation algorithm(LPA)was used to predict potential disease-miRNA associations.In MDLPMDA,noise can be removed from known disease-miRNA association matrix to get a new disease-miRNA association matrix by using SLM,which could improve the accuracy of the prediction model.In addition,MDLPMDA predicted potential disease-miRNA association score based on disease similarity networks and miRNA similarity networks,respectively.The final association score can be obtained by integrating the two association scores,which could make the prediction result more reliable.The results of cross validation and case studies showed that the model has better prediction performance.Moreover,the model can also predict potential miRNAs associated with new diseases.(3)Small molecule-miRNA association prediction model of TLHNSMMA based on triple layer heterogeneous network was proposed.We integrated small molecule similarity,miRNA similarity,disease similarity,known small molecule-miRNA associations and known disease-miRNA associations into a triple layer heterogeneous network,and proposed iterative update algorithm based on global network to predict potential small molecule-miRNA associations.In TLHNSMMA,we introduced multi-source biological data and treated diseases as a link to construct an iterative update algorithm that could fully mining the topology information of three-layer heterogeneous network to predict potential small molecule-miRNA associations.The results of cross validation and case studies showed that TLHNSMMA has reliable prediction performance.(4)Small molecule-miRNA association prediction model of HSSMMA based on HeteSim was proposed.We first combined known small molecule-miRNA associations,miRNA similarity and small molecule similarity into a two-layer heterogeneous network.Then,by considering all search paths from small molecule to disease with length less than 4,the measure method of HeteSim was used to calculate potential small molecule-miRNA association score base on each path.Finally,the final small molecule-miRNA association score was obtained by average integrating association scores of all paths.In HSSMMA,we not only integrated multi-source biological data to build heterogeneous network,but also constructed paths with appropriate length and introduced path-based measurement method of HeteSim to fully mine the path information from small molecules to miRNA.The results of cross validation and case studies showed that HSSMMA can be as a useful tool for predicting potential small molecule-miRNA associations.In addition,the model can predict potential miRNAs associated with new small molecules.(5)Disease-microbe association prediction model of MDLPHMDA based on matrix decomposition and label propagation was proposed.Based on the known disease-microbe associations,disease similarity and microbe similarity,we first used SLM decompose known disease-microbe association matrix to obtain a new disease-microbe association matrix.Then,based on the known disease-microbe associations and the new disease-microbe associations,LPA was used to predict potential disease-microbe associations from the perspectives of diseases and microbes,respectively.Finally,the final disease-microbe association score was obtained by integrating the two predicted association scores.In the model,SLM can decompose known disease-microbe association matrix into the liner combination of low rank matrix containing its real information structure and the sparse matrix containing noise.The new disease-microbe association matrix obtained based on the low rank matrix can effectively improve the prediction accuracy of the model.In addition,LPA can predict the potential disease-microbe associations from the perspective of disease and microbe,respectively,which can increase the reliability of the prediction results.The results of cross validation and case studies showed that MDLPHMDA can predict the potential disease-microbe associations effectively.There are 24 figures,21 tables and 255 references in this thesis.