A Parallel Algorithm for Estimating Genome-Wide Gene Networks using Nonparametric Bayesian Networks

We present a novel algorithm for estimating genome-wide gene networks using nonparametric Bayesian network models [3]. The algorithm, which is called the Neighbor Node Sampling & Repeat (NNSR) algorithm, is capable of searching a Bayesian network structure consisting of more than 20 000 nodes, which is fitted to given gene expression data. To realize the large scale Bayesian network structure search, the algorithm is designed to run on massively parallel computers where a massive amount of independent computation nodes are linked by fast connection. Such a computer system is also known as a distributed-memory architecture supercomputer. A Bayesian network is widely used as a gene network model [2]. Learning of a Bayesian network structure from gene expression data, however, is known as an NP-hard problem, and therefore the optimal network structure can be estimated only up to less than 30 genes or fewer. Thus, for the larger network, a heuristics algorithm such as the greedy hill-climbing (HC) algorithm is often used. The HC algorithm is applicable to networks with up to 1000 genes. A gene network with 1000 genes, however, involves fewer than 5% of all human genes. Therefore, the current Bayesian network estimation technology is far from a genome-wide scale analysis for most species.