Bayesian Models for Gene Regulatory Networks Applied to Cancer Tissues

Cellular behavior is controlled through multivariate interactions between various biological molecules such as proteins and DNA. Various methods have previously been proposed to model such interactions. However many of these methods require large volumes of data to effectively estimate the associated unknown parameters. In this work we explore the use of Bayesian methods to exploit the prior knowledge about pathway information in combination with collected data in order to make accurate and useful inferences about tissue level behavior. These predictions would in turn help in the discovery of better therapeutic strategies such as the development of better combination therapies involving kinase inhibiting drugs. Various problems of modeling cancerous and healthy tissues from a Bayesian perspective have been addressed in this work. We give a short description of these problems here in this section. An important problem in the study of cancer is the understanding of the heterogeneous nature of the cell population. The clonal evolution of the tumor cells results in the tumors being composed of multiple sub-populations. Each sub-population reacts differently to any given therapy. This calls for the development of novel (regulatory network) models, which can accommodate heterogeneity in cancerous tissues. Here we present a new approach to model heterogeneity in cancer. We model heterogeneity as an ensemble of deterministic Boolean networks based on prior pathway knowledge. We develop the model considering the use of qPCR data. By observing gene expressions when the tissue is subjected to various stimuli, the compositional breakup of the tissue under study can be determined. We demonstrate the viability of this approach by using our model on synthetic data, and real world data collected from fibroblasts. Another problem which is addressed in this work is the determination of locations of dysregulations in a Boolean network used to model signal transduction networks. Knowledge about which proteins/genes are dysregulated in a regulatory network, such as in the Mitogen Activated Protein Kinase (MAPK) Network, can be used not only to decide upon which therapy to use for a particular case of cancer, but also help in discovering effective targets for new drugs. The posterior inference problem is solved using a version of the message passing algorithm. We have done simulation experiments on synthetic data to verify the efficacy of the algorithm as compared to the results from the much more computationally intensive Markov Chain Monte-Carlo methods. We also applied the model to analyze data collected from fibroblasts, thereby demonstrating how this model can be used on real world data. Another important issue in Bayesian computation is that the processing of the collected data must be done as efficiently as possible in terms of computational speed and memory requirements. The use of Markov Chain Monte Carlo methods is time consuming and hence other methods need to be used for the analysis. The use of conjugate exponential models is investigated in the modeling of the heterogeneity of cancerous tissues where variational methods could be used in a straightforward manner. Variational algorithms, which allow for the fast computations of posterior probability distributions of variables of interest, have been used in the inference of the compositional breakup of the heterogeneous tissue under study. The efficacy of these methods has been demonstrated by comparing them with other methods such as Markov chain Monte Carlo and Expectation maximization