Association Prediction Of Biological Networks Based On Regularized Self-Representative Model

Association analysis in biological networks is critical to the discovery of pathogenic factors and potential drug targets,which is of great value in the diagnosis,treatment and prevention of human diseases.The prediction of associations in biological networks has drawn much research attention in the fields of biology and medical.The biological networks can be classified into two categories,namely homogeneous network and heterogeneous network.In this dissertation,we focus on the lncRNA-disease association prediction in heterogeneous network and synthetic lethality association in human cancers prediction in homogeneous network.Regularized self-representative matrix factorization-based models are proposed to predict the various associations in biological networks.The main contributions of this dissertation can be summarized as follows:(1)lncRNA-lncRNA,disease-disease and gene-gene similarity matrices are constructed to promote the performance of the prediction models.The verified lncRNA-disease and synthetic lethality associations in human cancers are collected from some relevant databases.After data preprocessing,the experimental datasets are constructed.Disease classification information collected from the Medical Subject Heading(MeSH)database is used to construct disease semantic similarity and lncRNA function similarity.Genetic semantic information collected from the Gene Ontology(GO)database is used to calculate disease semantic similarity.(2)A sparse regularized self-representative matrix factorization method is proposed for lncRNA-disease association prediction.This model focuses on learning two non-negative sparse representation matrices which capture the self-representation of diseases and the selfrepresentation of lncRNAs,respectively.Moreover,this model could also draw support from the intra-associations among disease and lncRNAs that derived from other public databases to generate more accurate estimation of the representation matrices.After ranking the candidate lncRNA-disease pairs based on the predicted propensities in descending order,the top-ranked pairs are considered to be the most likely true association pairs.This model is evaluated by10-fold cross-validation experiments,case analysis and parameter sensitivity analysis on three datasets of LncRNADisease,MNDR and Lnc2 Cancer,and the experimental results demonstrate the effectiveness of this model.(3)A graph regularized self-representative matrix factorization method is proposed for synthetic lethality prediction in human cancers.This model focuses on self-representative and learn affinity matrix from known synthetic lethality matrix.It also uses biological knowledge derived from Gene Ontology(GO)similarity to promote the novel prediction of synthetic lethality.The probability of each gene pair being synthetic lethality gene pair is calculated by this model.The results in 5-fold cross-validation experiments and the parameter sensitivity analysis on SynLethDB dataset fully demonstrate the effectiveness of this model in predicting Synthetic lethality in human cancers.In this study,the sparse regularized self-representative matrix factorization model and graph regularized self-representative matrix factorization model are proposed for the association prediction of heterogeneous network and homogeneous network,respectively.The experimental results demonstrate these two proposed computational models are effective in predicting associations in biological networks.It is anticipated that our work could provide a new perspective for subsequent related research.