Study On Association Prediction Algorithms Of Biomedical Networks Based On Matrix Factorization

With the fast development of biomedical and computer technology,and the postgenome era coming,identifying relationships between biomedical entities(such as proteins,drugs,diseases,etc.)from complex networks and how to develop effective computer-aided prediction models have become hotspots in computational biology.Potential associations could be confirmed through biological experiments when we provide promising candidate lists by models.It is of great importance for understanding the biological process and molecular mechanism,the diagnosis and treatment of human diseases,and development of drugs.Matrix factorization algorithms received widespread attention in solving biomedical problems due to their excellent performance.This degree dissertation mainly used the matrix factorization theory and studied the interaction problems of the molecular level in proteomics and transcriptomics,and association prediction problems of disease-associated human microbes,and the data sparsity problem and the unification problem of association prediction models of biomedical networks.The specific research work and innovations are as follows:(1)For the interaction problem of the same molecular type that can obtain molecular sequences,we proposed two methods for numerical characterization to encode proteins based on the graph energy in chapter 2.It can comprehensively consider the physical and chemical properties as well as the contact and location information of amino acids.We utilized the pseudo amino acid composition method to extract composition information.After multi-information fusion,we designed a machine learning model for predicting protein-protein interaction.We demonstrated the effectiveness of the numerical characterization method and model.(2)For the interaction problem of the different molecular types that can obtain molecular sequences,we proposed a novel joint regularized nonnegative matrix factorization(?)hm and the gradient descent is utilized to solve the optimization pro(?)chapter 3.There are not just nonnegative matrix factorizations of the adjacency matrix but also symmetric nonnegative matrix factorizations of the similarity matrix in the optimization objective function.We constructed computer-aided prediction models to identify the lncRNA-protein interaction by integrating internal information of molecular sequence and external information such as network topology.We demonstrated the effectiveness of the algorithm and model.(3)For the interaction problem of the different molecular types that can not obtain molecular sequences,we developed a microbe-disease association prediction algorithm based on the Kronecker regularized least squares by introducing the Kronecker product and combining regularized least squares method in chapter 4.It used a vector operator to stack the elements of the association matrix into a vector for prediction.Especially,we improved the solution method through the eigenvalue decomposition in the process of solving algorithms.It can improve operational efficiency.Experimental results demonstrated the effectiveness of the model.(4)For the data sparsity problem of biomedical networks and the unification problem of model association prediction,we proposed a collaborative weighted nonnegative matrix factorization algorithm and utilized two optimization methods to solve the optimization problem in chapter 5.It is a generalized model for solving a class of problems,not a particular problem.We used the proposed collaborative weighted nonnegative matrix factorization algorithm and ran extensive experiments on different kinds of datasets.Experiments adequately demonstrated the effectiveness and applicability of the algorithm.We designed a similarity calculation method founded on the network to fuse more information and alleviate the impact of missing information.Models are designed to solve practical problems based on the matrix factorization theory while it can promote the application of matrix factorization in the biomedical field.Computer-aided prediction models are proposed for interaction prediction at the molecular level in proteomics and transcriptomics,which is of great significance for understanding complex biological systems from a network perspective and the occurrence and progression of diseases.Algorithms of microbe-disease association prediction can help identify potential associations and provide new directions for the diagnosis and treatment of diseases.Especially,collaborative weighted nonnegative matrix factorization algorithm can effectively alleviate data sparsity problems and assist in resolving a class of problems.It can play an important role in biomedical network association prediction and provide more assistance for biomedical researchers.