Inferring large-scale gene regulatory networks using a low-order constraint-based algorithm

Recently, simplified graphical modeling approaches based on low-order conditional (in-) dependence calculations have received attention because of their potential to model gene regulatory networks. Such methods are able to reconstruct large-scale gene networks with a small number of experimental measurements, at minimal computational cost. However, unlike Bayesian networks, current low-order graphical models provide no means to distinguish between cause and effect in gene regulatory relationships. To address this problem, we developed a low-order constraint-based algorithm for gene regulatory network inference. The method is capable of inferring causal directions using limited-order conditional independence tests and provides a computationally-feasible way to analyze high-dimensional datasets while maintaining high reliability. To assess the performance of our algorithm, we compared it to several existing graphical models: relevance networks; graphical Gaussian models; ARACNE; Bayesian networks; and the classical constraint-based algorithm, using realistic synthetic datasets. Furthermore, we applied our algorithm to real microarray data from Escherichia coli Affymetrix arrays and validated the results by comparison to known regulatory interactions collected in RegulonDB. The algorithm was found to be both effective and efficient at reconstructing gene regulatory networks from microarray data.