Abstract:
Gaussian Graphical Models provide an alternative to Bayesian Networks for inferring regulatory network relations from gene expression data. We introduce a Monte Carlo algorithm based on Metropolis sampling with a suitable guidance function to estimate a network with a minimal number of links to describe observed correlation functions. A central role plays the elimination of indirect correlations by a Markov Chain Monte Carlo process that iteratively prunes and extends the network. This process is enhanced by an empirical guidance function which selects candidate links for addition and removal. The resulting network (or set of networks) describes a minimal set of correlative relations between quantities. The method is evaluated both on synthetic data as well as a biological data set.