Background: Network reconstruction methods that rely on covariance of expression of transcription regulators and their targets ignore the fact that transcription of regulators and their targets can be controlled differently and/or independently. Such oversight would result in many erroneous predictions. However, accurate prediction of gene regulatory interactions can be made possible through modeling and estimation of transcriptional activity of groups of co-regulated genes. Results: Incomplete regulatory connectivity and expression data are used here to construct a consensus network of transcriptional regulation in Escherichia coli (E. coli). The network is updated via a covariance model describing the activity of gene sets controlled by common regulators. The proposed model-selection algorithm was used to annotate the likeliest regulatory interactions in E. coli on the basis of two independent sets of expression data, each containing many microarray experiments under a variety of conditions. The key regulatory predictions have been verified by an experiment and literature survey. In addition, the estimated activity profiles of transcription factors were used to describe their responses to environmental and genetic perturbations as well as drug treatments. Conclusion: Information about transcriptional activity of documented co-regulated genes (a core regulon) should be sufficient for discovering new target genes, whose transcriptional activities significantly co-vary with the activity of the core regulon members. Our ability to derive a highly significant consensus network by applying the regulon-based approach to two very different data sets demonstrated the efficiency of this strategy. We believe that this approach can be used to reconstruct gene regulatory networks of other organisms for which partial sets of known interactions are available.