Reconstruction Of Gene Regulatory Networks Based On Mutual Information

Systems biology is an interdisciplinary field that focuses on complex interactionsand mechanisms of the elements (genes, mRNAs, proteins, small molecules etc.) withinbiological systems from systematic perspective. A gene regulatory network (GRN) rep-resents the map of gene-gene interactions that are important in genetics, development,evolution, and disease, etc. The reconstruction or inverse engineering of GRNs, whichaims to infer the interactions among genes, remains the bottleneck in the pipeline todiscover biological mechanisms from high throughout data. In this thesis, the weakness-es and strengths of current network inference methods were addressed and some novelmethods were proposed for inferring GRNs based on information theory. The developedmethods improve the accuracy of GRNs inference by overcoming the disadvantage ofcurrent methods. The follows are the main works in this thesis.(1) A novel GRN inference method was developed based on conditional mutual infor-mation (CMI) and path consistency algorithm (PCA). In the proposed method, theregulatory strengths are firstly estimated by mutual information, which can quanti-fy the nonlinear correlation between gene pairs, and then the indirect interactionsare deleted recursively using PCA algorithm based on the conditional independencemeasured by CMI. With above processes, the network achieved is as sparse as pos-sible. With the general hypothesis of Gaussian distribution underlying gene expres-sion data, CMI between a pair of genes is computed by a concise formula involvingthe covariance matrices of the related gene expression profiles. The experiments onthe simulation data and the widely used SOS data in Escherichia coli showed thegood performance of the method. Besides its high accuracy, our method is able todistinguish indirect interactions from direct associations.(2) A novel approach was developed to improve the accuracy of GRN inference com-bining recursive optimization (RO) and mutual information (MI). In the proposedalgorithm, the noisy regulations with low pair-wise correlations are first removedutilizing MI, and the redundant regulations like indirect regulators are further ex-cluded by RO to improve the accuracy of inferred GRNs. The regulatory strengthsare determined with the integration of linear and nonlinear correlations betweenregulators and targets. The accuracy of the proposed method was proved throughthe analysis on the simulation data and the genome-wide gene expression data inEscherichia coli. The method not only helps to determine regulatory directionswithout prior knowledge of regulators, but also can detect the nonlinear correla-tions.(3) A novel GRNs inference method was proposed to accurately quantify causalstrength between gene pairs. In the proposed method, the causal strength or cor-relation between two variables can be accurately quantified by tuning conditional mutual information. For the computation of regulatory strength, a concise formu-la involving the covariance matrices was given. The proposed method can not onlyaccurately quantify causal strength by the tuning CMI but also reconstruct the topol-ogy sparseness of biological networks. The experiments on the simulation data andthe widely used SOS data in Escherichia coli showed the good performance of themethod. The developed method greatly improves the accuracy of GRNs inferencethrough accurately quantifying the causal strength between two variables.