DoGeNetS: using optimisation to discriminate regulatory network topologies based on gene expression data

Gene regulatory networks (GRNs) determine the dynamics of gene expression. Interest often focuses on the topological structure of a GRN while numerical parameters (e.g. decay rates) are unknown and less important. For larger GRNs, inference of structure from gene expression data is prohibitively difficult. Models are often proposed based on integrative interpretation of multiple sources of information. We have developed DoGeNetS (Discrimination of Gene Network Structures), a method to directly assess candidate models of GRN structure against a target gene expression data set. The transsys language serves to model GRN structures. Numeric parameters are optimised to approximate the target data. Multiple restarts of optimisation yield score sets that provide a basis to statistically discriminate candidate models according to their potential to explain the target data. We demonstrate discrimination power of the DoGeNetS method by relating structural divergence to divergence between gene expression data sets. Known models are used to generate target expression data, and a set of candidate models with a defined structural divergence to the true model is produced. Structural divergence and divergence of expression profiles after optimisation are strongly correlated. We further show that discrimination is possible at noise levels exceeding those typical of contemporary microarray data. DoGeNetS is capable of discriminating the best GRN structure from among a small number of candidates. p values indicate whether differences in divergence of expression are significant. Although this study uses single gene knockouts, the DoGeNetS method can be adapted to simulate a virtually unlimited range of experimental conditions. [Includes supplementary material].