Metabolism-Informed Bioreactor Model for Transferring Knowledge from Fed-batch to Continuous Biomanufacturing

Chinese hamster ovary (CHO) cells are the primary platform for producing monoclonal antibodies, and the industry is transitioning from fed-batch to continuous manufacturing to meet increasing demand. Continuous cultures offer sustained high cell densities and productivity, but they require reliable real-time monitoring, prediction, and control. However, limited observability of intracellular processes and sparse extracellular measurements create a gap between understanding the metabolism and deploying it in practice.

The authors present a metabolism-informed bioreactor model for mAb production that incorporates reactor- and intracellular-level mass balances within a learning architecture. They avoid predefined kinetic rate laws and steady-state assumptions to create a structure that generalizes across operating modes. This model functions as a soft sensor, inferring unmeasured intracellular fluxes and forecasting extracellular concentrations for process monitoring and control.

The performance of the model is consistent regardless of network size, and the exchange flux fidelity is independent of the solver choice, while network size only affects intracellular interpretability.

The authors demonstrate that transferring the model from fed-batch to continuous operation through fine-tuning outperforms other transfer learning methods. This indicates that the metabolic structure remains invariant to the mode of operation. The predicted fluxes reveal that growth rate, rather than operating mode, is the dominant axis of metabolic variation in fed-batch.

The study also identifies potential regime-shift signals to guide experimental design for characterizing continuous culture, ultimately paving the way towards digital models grounded in first-principles knowledge yet adaptive for regime shifts.