Abstract Background Little is known about bacterial transcriptional regulatory networks (TRNs). In Escherichia coli, which is the organism with the largest wet-lab validated TRN, its set of interactions involves only ~50% of the repertoire of transcription factors currently known, and ~25% of its genes. Of those, only a small proportion describes the regulation of processes that are clinically relevant, such as drug resistance mechanisms. Results We designed feed-forward (FF) and bi-fan (BF) motif predictors for E. coli using multi-layer perceptron artificial neural networks (ANNs). The motif predictors were trained using a large dataset of gene expression data; the collection of motifs was extracted from the E. coli TRN. Each network motif was mapped to a vector of correlations which were computed using the gene expression profile of the elements in the motif. Thus, by combining network structural information with transcriptome data, FF and BF predictors were able to classify with a high precision of 83% and 96%, respectively, and with a high recall of 86% and 97%, respectively. These results were found when motifs were represented using different types of correlations together, i.e., Pearson, Spearman, Kendall, and partial correlation. We then applied the best predictors to hypothesize new regulations for 16 operons involved with multidrug resistance (MDR) efflux pumps, which are considered as a major bacterial mechanism to fight antimicrobial agents. As a result, the motif predictors assigned new transcription factors for these MDR proteins, turning them into high-quality candidates to be experimentally tested. Conclusion The motif predictors presented herein can be used to identify novel regulatory interactions by using microarray data. The presentation of an example motif to predictors will make them categorize whether or not the example motif is a BF, or whether or not it is an FF. This approach is useful to find new "pieces" of the TRN, when inspecting the regulation of a small set of operons. Furthermore, it shows that correlations of expression data can be used to discriminate between elements that are arranged in structural motifs and those in random sets of transcripts.