From process variables to metabolic states: Hybrid constraint-based analysis for upstream bioprocesses

Upstream bioprocess optimization and control strategies rely predominantly on extracellular measurements, yet the underlying intracellular metabolic states that drive process performance remain largely inaccessible for real-time monitoring and intervention. In this study, we address this challenge by developing a hybrid modeling framework that integrates machine learning with first-principles models to predict specific reaction rates and bridge reactor-scale observations with genome-scale metabolic analysis for real-time monitoring of intracellular states. With a monoclonal antibody producing Chinese Hamster Ovary (CHO) cell upstream bioprocess as a case study, we apply the hybrid modeling strategy to predict the viable cell density, product titer, and metabolite concentration profiles throughout the process. When compared to purely data-driven and purely mechanistic methods, the hybrid model displayed better predictive performance for both interpolative and out-of-domain process conditions. Predictions from the hybrid model were used to constrain a genome-scale reconstruction of the CHO cell, and the solution space was sampled uniformly without assumptions of a metabolic objective. Using this methodology, we identified distinct metabolic capabilities in desirable high-yielding processes, including flux reversal in lower glycolysis and TCA cycle coordination during transitional periods leading to enhanced energy generation efficiency. This approach represents a fusion of hybrid modeling and genome-scale metabolic models for simultaneous predictive modeling and intracellular insight associated with the prediction.