Learning and validating Bayesian network models of gene networks

We propose a framework for learning from data and validating Bayesian network models of gene networks. The learning phase selects multiple locally optimal models of the data and reports the best of them. The validation phase assesses the confidence in the model reported by studying the different locally optimal models obtained in the learning phase. We prove that our framework is asymptotically correct under the faithfulness assumption. Experiments with real data (320 samples of the expression levels of 32 genes involved in Saccharomyces cerevisiae, i.e. baker's yeast, pheromone response) show that our framework is reliable.