Parallel information theory based construction of gene regulatory networks

We present a parallel method for construction of gene regulatorynetworks from large-scale gene expression data. Our method integratesmutual information, data processing inequality and statisticaltesting to detect significant dependencies between genes, and efficientlyexploits parallelism inherent in such computations. We present a novelmethod to carry out permutation testing for assessing statistical significancewhile reducing its computational complexity by a factor of Θ(n2),where n is the number of genes. Using both synthetic and known regulatorynetworks, we show that our method produces networks of qualitysimilar to ARACNE, a widely used mutual information based method.We present a parallelization of the algorithm that, for the first time, allowsconstruction of whole genome networks from thousands of microarrayexperiments using rigorous mutual information based methodology.We report the construction of a 15,147 gene network of the plant Arabidopsisthaliana from 2,996 microarray experiments on a 2,048-CPUBlue Gene/L in 45 minutes, thus addressing a grand challenge problemin the NSF Arabidopsis 2010 initiative.