Evaluation of causal Bayesian network search algorithms using simulated mesotheliomas gene expression data
Chan, F., Marinova, D. and Anderssen, R.S. (eds) MODSIM2011, 19th International Congress on Modelling and Simulation. · 2011
To understand the physiology of a complex disease, such as mesotheliomas, it is necessary to learn how the genes that are involved in developing the disease interact with the environment.To this end, statistical methods that can detect these gene-environment interactions will help scientists in detecting causal relationships among genes.These predicted causal relationships among genes can then be later verified through actual laboratory experiments.In this paper, we have developed a novel causal discovery system that incorporates recent advances in Bayesian network search methods.We introduce a novel algorithm called Equivalence Checking Local Implicit latent variable scoring Method with Markov Chain Monte Carlo (EquLIM-MCMC) search algorithm that extends existing causal Bayesian network discovery algorithms, EquLIM and the Local Implicit latent variable scoring Method (LIM).Markov Chain Monte Carlo (MCMC) search has been shown to be very useful especially in analyzing datasets where the number of input variables greatly exceeds the number of cases that are collected (Friedman and Koller 2000;Hageman, Leduc et al. 2011).More and more datasets that are collected for gene expression studies have thousands of genes' expression levels (input variables) that are measured from tens or hundreds of subjects (cases).Datasets collected in gene-environment interactions studies will show similar trends. We use LIM with MCMC (LIM-MCMC) and EquLIM-MCMC to analyze purely observational simulated gene expression datasets.To test these algorithms' abilities to detect causal relationships from realistic data, we generate datasets from a gene regulation pathway model of malignant mesothelioma formation proposed by an expert.Using the metrics of Area Under Receiver Operating Characteristic (AUROC) curve, Positive Predictive Value (PPV), and Shannon Entropy, we show that EquLIM-MCMC exhibit clear advantages over LIM-MCMC with causal relationship predictions.EquLIM-MCMC therefore improves over LIM-MCMC's ability in detecting causal relationships in gene networks and gene-environment interactions from presently available observational gene expression data.