Benchmarking kinetic models of Escherichia coli metabolism

Predicting phenotype from genotype is the holy grail of quantitative systems biology. Kinetic models of metabolism are among the most mechanistically detailed tools for phenotype prediction. Kinetic models describe changes in metabolite concentrations as a function of enzyme concentration, reaction rates, and concentrations of metabolic effectors uniquely enabling integration of multiple omics data types in a unifying mechanistic framework. While development of such models for Escherichia coli has been going on for almost twenty years, multiple separate models have been established and systematic independent benchmarking studies have not been performed on the full set of models available. In this study we compared systematically all recently published kinetic models of the central carbon metabolism of Escherichia coli. We assess the ease of use of the models, their ability to include omics data as input, and the accuracy of prediction of central carbon metabolic flux phenotypes. We conclude that there is no clear winner among the models when considering the resulting tradeoffs in performance and applicability to various scenarios. This study can help to guide further development of kinetic models, and to demonstrate how to apply such models in real-world setting, ultimately enabling the design of efficient cell factories. Author summaryKinetic modeling is a promising method to predict cell metabolism. Such models provide mechanistic description of how concentrations of metabolites change in the cell as a function of time, cellular environment and the genotype of the cell. In the past years there have been several kinetic models published for various organisms. We want to assess how reliably models of Escherichia coli metabolism could predict cellular metabolic state upon genetic or environmental perturbations. We test selected models in the ways that represent common metabolic engineering practices including deletion and overexpression of genes. Our results suggest that all published models have tradeoffs and the model to use should be chosen depending on the specific application. We show in which cases users could expect the best performance from published models. Our benchmarking study should help users to make a better informed choice and also provides systematic training and testing dataset for model developers.