An information theoretic method for reconstructing local regulatory network modules from polymorphic samples.

Statistical relations between genome-wide mRNA transcript levels have been successfully used to infer regulatory relations among the genes, however the most successful methods have relied on additional data and focused on small sub-networks of genes. Along these lines, we recently demonstrated a model for simultaneously incorporating micro-array expression data with whole genome genotype marker data to identify causal pairwise relationships among genes. In this paper we extend this methodology to the principled construction of networks describing local regulatory modules. Our method is a two-step process: starting with a seed gene of interest, a Markov Blanket over genotype and gene expression observations is inferred according to differential entropy estimation; a Bayes Net is then constructed from the resulting variables with important biological constraints yielding causally correct relationships. We tested our method by simulating a regulatory network within the background of of a real data set. We found that 45% of the genes in a regulatory module can be identified and the relations among the genes can be recovered with moderately high accuracy (> 70%). Since sample size is a practical and economic limitation, we considered the impact of increasing the number of samples and found that recovery of true gene-gene relationships only doubled with ten times the number of samples, suggesting that useful networks can be achieved with current experimental designs, but that significant improvements are not expected without major increases in the number of samples. When we applied this method to an actual data set of 111 back-crossed mice we were able to recover local gene regulatory networks supported by the biological literature.