Study On Modeling Gene Regulation Network Based On Bayesian Theory

Microarray technology provids a data base for gene regulatory networks, construction of gene regulatory networks is an important research topic of functional genomics. Bayesian graphical models based on Bayesian theory have a solid theoretical foundation and flexible reasoning mechanism, they can simplely represent knowledge, so they become powerful tools to construct gene regulatory networks. Based on Bayesian graphical models, for the current problems in curent study of genetic regulatory networks, the contribution of this paper is as follows:To outcome short problem of sample points in single gene expression dataset, a distributed MFD-GRN algorithm combines the multiple distribution of static gene expression data sets to construct gene regulatory networks, each distributed data sets contain the same genes, different numbers of sample points. MFD-GRN algorithm is divided into two processes: local learning and global learning. In the local study, based on each gene expression data sets, using a search scoring method, learning each independent local structure; in the global study, the method based on correlation analysis to fusion local structures, considers the mutual information and conditional mutual information as a variable, taking their mathematical expectation, as a global learning evaluation criteria. This simply passes the local mutual information and conditional mutual information to the fusion center node, without direct access to local individual data, the effective protection of individual privacy.For time series gene expression data, we propose TSMI-GRN algorithm to construct the gene regulatory networks.TSMI-GRN algorithm defines a time series mutual information between genes, calculated using covariance matrix time series mutual information, compared with traditional mutual information, an increase of time features, and more consistent time series data.To solve varied structured problem of gene regulation network with time varied, we propose a VS-GRN algorithm to construct varied structure network. The algorithm has three phases. The first phase: cutting steady time slices from time series.we propose a cutting algorithm. The second phase: Learning Bayesian network on each steady time slice. We propose a P-BIC score to fuse protein-protein interaction data. The third phase: Learning transaction network between two steady time slices.For different problems, we consider threee algorithms. We test these algorithms on true gene expression datasets, and compare with present algorithms, prove our algorithms are valid.