Linking intra- and extra-cellular metabolic domains via neural-network surrogates for dynamic metabolic control

In this study, we aim to optimize biotechnological production by manipulating intracellular metabolic fluxes in microbial cell factories. Model-based dynamic optimization is proposed to determine the optimal dynamic trajectories of the manipulatable intracellular fluxes. A challenge emerges as existing models are often oversimplified, lacking insights into intracellular metabolism, or are excessively complex, leading to numerical and implementation challenges in optimal control (e.g., related to bilevel optimizations). We propose a solution involving a machine-learning surrogate derived from steady-state constraint-based metabolic modeling. This surrogate bridges the gap between manipulatable intracellular fluxes and process exchange rates. By integrating the surrogate model with simple macro-kinetic dynamic models, we can develop hybrid machine-learning-supported dynamic models. Conveniently, the manipulatable intracellular fluxes in these augmented models can be exploited as dynamic optimization degrees of freedom. We apply this modeling and optimization strategy to a representative metabolic network that showcases common challenges in dynamic metabolic control. We also present an example of cybernetic control to counteract system uncertainties. Our approach facilitates the in silico evaluation of dynamic metabolic interventions and can aid in the selection of suitable control and actuation strategies.