High Throughput Metabolic Ensemble Modeling Assists Strain Engineering

Dynamic metabolic modeling has been a desirable and yet challenging goal because of the lack of detailed enzyme kinetics. As such, perturbations at the enzyme level cannot be properly represented, and strain optimization cannot be guided by data generated from typical strain evaluation experiments. Metabolic ensemble modeling (MEM) framework utilizes ensemble approach to bypass the need of detailed enzyme kinetics. The basic idea of MEM is to construct an ensemble of models that are anchored to the same steady-state fluxes, but span all the kinetic space allowed by the known enzyme mechanistic structure and thermodynamics. These models give different responses to enzyme overexpression or knockouts (typical strain engineering efforts). The different responses provide a basis for screening models against experimental data commonly generated in metabolic engineering efforts. We have shown that the models would converge to a small set with only a few rounds of screening and become increasingly predictive.;The steady-state anchor is the most unique and important property of MEM. We have demonstrated two benefits of anchoring ensemble of models to the same steady-state flux distribution. First, by requiring that all dynamic models yield the same steady-state flux distribution, the allowable range of possible kinetic parameters is significantly reduced, which allows meaningful sampling schemes to explore the dynamic behavior of the model. Second, by reducing the kinetic parameter space, further screening of models based on limited data (steady-state fluxes or transient metabolite profiles upon perturbations) becomes possible. Without such anchoring constraints, the parameter space is too large to be sampled effectively. Thus, MEM differs from the traditional ensemble approaches in that it utilizes the available flux data and constrains the sampling space to a realistic size.;In addition to implementing numerical approach to solve for steady state, we have also improved the ordinary differential equation solving algorithm to solve for reasonable initial conditions more efficiently. Thereby, we realize a high throughput model screening strategy based on strain construction experiments without the need for designated kinetic experiments. The concept mimics the high throughput screening of biological activities, and the parameters generated are portable across different metabolic engineering efforts based on the same organism. By extracting the kinetic information from strain engineering data, the retained models are not only able to describe these data but also possess predictive capability for future modification. MEM has been applied to model abundant isobutanol strain engineering data generated in our lab. The trained parameter spaces are further utilized for 2-methyl-1-butanol (2MB) model construction. Using this approach, we achieved improved 2MB production in engineered Escherichia coli strains.