Item

Abstract

Reconstructing gene regulatory networks (GRNs) from expression data is a challenging task that has become essential to the understanding of complex regulatory mechanisms in cells. The major issues are the usually very high ratio of the number of genes to the sample size, and the noise in the available data. In this thesis we investigate the effect of the number of samples and noise on the performance of statistical methods. The results indicate that in the case of not having many samples and/or in facing high amount of noise like the case for gene expression data, the performance of all methods decreased significantly compared to the well behaved case (many samples and no noise). Integrating biological prior knowledge to the learning process is a natural and promising way to partially compensate for the lack of reliable expression data and to increase the accuracy of network reconstruction algorithms. In this thesis, we present PriorPC, a new algorithm based on the PC algorithm that uses prior knowledge. Despite being one of the most popular methods for Bayesian network reconstruction, the PC algorithm is known to depend strongly on the order in which nodes are presented, especially for large networks. PriorPC exploits this flaw to include prior knowledge. We show on both synthetic and real data that the structural accuracy of networks obtained with PriorPC is greatly improved compared to the PC algorithm. Furthermore, PriorPC is fast and scales well for large networks which is important for its applicability to experimental data. Another challenge in GRN reconstruction is to detect (direct) nonlinear interactions between genes. A recently proposed association measure named distance correlation is a powerful method to find nonlinear relationships. In this thesis, we propose a novel approach to estimate partial distance correlation, the generalization of distance correlation which accounts for the influence of other variables and therefore it can detect direct nonlinear relationships.