Soft-sensor development for real-time state estimation as an enabler for robust processing with Escherichia coli

The demand for biotechnological products has grown steadily in recent years. To facilitatethe increased production needs, advanced bioprocessing strategies are required. Theseare often realized via model-based monitoring and control approaches. In this work,a mechanistic model was developed for Escherichia coli cultivations using the alkalinephosphatase (phoA) expression system for the expression of a recombinant fragmentantigen binding. The goal was to establish a robust predictive framework for real-timestate estimation in fed-batch cultivations.First, a mechanistic model was developed to describe the cultivation trajectory. Thereby,higher-order Moser-Kinetics were used to accurately describe the phosphate uptake ofthe cells over the whole cultivation period. With a variable phosphate to biomass yield,the shift in phosphate uptake once limitation occurred was accurately represented. Toensure stability and reliability of the established model structure, parameter identifiabilityand sensitivity analyses were performed. The model was able the describe the observedcultivations with an NRMSE below 10 %. Further model improvements were tested byincluding a maintenance term. However, this measure was rejected due to the risk ofoverfitting and reduced model stability.Second, the optimal harvest time point for a fed-batch cultivation was determined usingthe created model via a Monte Carlo Simulation. Here, the time points for the maximumproduct titer and glucose accumulation threshold were simulated. The obtained timepoints were compared to experimental measurements for validation. This successfullydemonstrated the model’s predictive capability and applicability for process optimization.Subsequently, a soft sensor was implemented using a square-root unscented Kalman filter.The description of the cultivation trajectory was further enhanced by including on-linemeasurements of base consumption and carbon dioxide off-gas content. As a result, the RMSE was reduced to up to one-third compared to a ’mechanistic model only’ approach.Overall, this work presents a comprehensive modeling and soft-sensing framework forEscherichia coli cultivations with real-time process monitoring applications. The developedpredictive framework provides valuable insights into phosphate-dependent regulation,enables the prediction of optimal harvest time points, and serves as a basis for model-basedprocess control strategies.