Some Results Of Modeling And Learning Genetic Regulatory Networks

The genes of higher eukaryotic organisms often exhibit different expression patterns at different developmental stages, in different tissues and environments, This is attributed to the underlying genetic regulatory networks, which direct and regulates the expression of each gene in a systematic and cooperative manner according to the functional demand of each cell, and controls accurately the quantity of each type of protein, consequently achieving the growth and development of the organisms in a scheduled sequence, and keeping the normal physiological function of the cell in some limited environmental conditions. A genetic regulatory network functions as the nerve center or brain in terms of the cell level, it conducts the space-time specific expression of each gene. Generally speaking, a genetic regulatory network is a complex, dynamic feedback system with multiple regulatory levels. Construction of genetic regulatory network is an important but challenging task in the post-genome era, and has raised a research upsurge in the systems biology and computational biology fields. This thesis addresses the genetic regulatory networks via probabilistic modeling from microarray data, major contributions are as follows:1. A new algorithm for learning Bayesian networks (BNs) is proposed and applied to infer the genome-wide genetic regulatory network of yeast. The proposed method employs a divide and conquer scheme, which first decomposes the task of learning a large BN into learning some relatively small BNs, and then builds the final BN by reunifying these learnt small BNs. Our method can greatly reduce the scale of the network and computational complexity through the decomposition of network, and effectively solves the problem that the number of genes is very large but the number of samples is small.. We apply the proposed method to building genetic regulatory network from yeast microarray data and compare the learning result with that of biochemical experiments, which shows that our method can learn high-quality genetic regulatory networks.2. A new method for building quantitative transcriptional regulatory network is developed. Most existing methods focus on identifying the qualitative regulatory relationship among genes, while ignoring the quantitative aspects of the transcription regulation, which are critical to understand the function of regulators. By assuming that the transcription rate of a gene and the degeneration of its mRNA reach an equilibrium state, this paper proposes a generative model that integrates the concentration of TFs, binding energy between TFs and binging sites, and other kinetic parameters into a general regulation function, and thus has strong expressive ability. The experimental results on the yeast microarray data show that our model can accurately learn the concentrations of transcriptional factors and predict whether a transcription factor is a activator or a repressor.3. A novel approach for predicting the target genes of microRNAs is introduced. Considering that there are increasing number of evidences suggesting that microRNAs(miRNAs) are not only a type of post-transcriptional regulators as important as transcriptional factors, but also the indispensable members of the genetic regulatory network, it is great significance to building miRNAs-mediated genetic regulatory network. However, identifying the bona fide target genes of miRNAs is the stepping-stone of the goal mentioned above. This thesis proposes two methods for identification of miRNA targets, appMirTar and superMir-Tar. appMirTar first shows that miRNAs mainly mediate two types regulatory circuits, and then proposes a variant of affinity propagation algorithm customized for bipartite graph, and combines the results of sequence-based prediction algorithms and microarray data of both miRNAs and mRNAs to predict the target miRNAs. Experiments on human dataset demonstrate that our method greatly improves the prediction accuracy; superMirTar tries to identify miRNA targets in terms of bipartite graph learning, which can be further reformulated as the problem of kernel-based supervised distance metric learning. Taking the experimentally supported miRNA-mRNA interactions as training set, our experimental results show that superMirTar achieves better performance than most existing methods.The four methods proposed in this thesis has been successively used to build qualitative, quantitative and miRNAs- mediated genetic regulatory networks. They form a set of systematically analytic methods that provide powerful tools for investigating genetic regulatory networks.