(210d) Optimization-Based Approach for Transcriptional Regulatory Network Refinement

Integrated constraint-based metabolic and regulatory network models are effective in predicting cellular growth phenotypes arising from genetic and environmental perturbations. A challenging task in the construction of regulatory network models is identifying accurate interactions between transcription factors and their corresponding gene targets. We have developed a mixed integer linear programming (MILP) algorithm that identifies which TF-target interactions are inconsistent with experimental data, and can aide in the refinement of transcriptional regulatory models. In addition, the algorithm can be used to identify genes that if allowed to violate their regulatory constraints (such as by over-expression) can rescue lethal phenotypes.

The optimization algorithm allows an integrated metabolic and transcriptional regulatory model to choose and violate specific regulatory rules in order to achieve non-zero growth phenotypes under different environmental conditions. We applied this algorithm to an updated regulatory network model for E. coli (iMC1010) to identify Boolean regulatory rules which caused incorrect growth phenotype predictions. Out of 33,146 knockout mutant/environmental cases evaluated, 3,019 (9.1%) were incorrectly predicted by the existing model to be unable to grow. The algorithm was applied to these 3,019 cases to identify the genes that are incorrectly down-regulated in the regulatory model resulting in an incorrect no-growth phenotype prediction. Nearly a third of these cases correspond to growth on N-acetyl-glucosamine, N-acetyl-mannosamine, and N-acetylneuraminate and many of these can be corrected by relaxing the regulatory rule for glmU (whose gene product is required for peptidoglycan and lipopolysaccharide biosynthesis). Allowing for constitutive expression in the Boolean rule for glmU expression would result in substantially higher agreement between the regulatory model predictions and growth phenotyping data.

In an additional 2,076 cases (6.3%) the existing transcriptional regulatory model correctly predicted knockout mutants would be unable to grow even though the metabolic network has other enzymes which would allow for growth. These cases correspond to lethal strains that could be rescued by the over expression of other gene products. On average, most lethal mutant strains would require 2.72 genes to be over-expressed so that they could grow on a given condition. Growth on guanine as a nitrogen source required the most number of genes (seven) to be over-expressed. We have subsequently applied model to an updated metabolic model of E. coli and found further improvements and rescue lethals.

The developed approach allows for the automated evaluation of experimental data for both wildtype and knockout mutant strains in order to identify what transcriptional regulatory interactions are consistent and inconsistent with experimental observations. The approach can also be used to hypothesize strategies to rescue lethal mutants which if confirmed would help validate metabolic and regulatory network models.