Metabolism-Informed Bioreactor Model for Transferring Knowledge from Fed-batch to Continuous Biomanufacturing
Chinese hamster ovary (CHO) cells remain the dominant platform for monoclonal antibody (mAb) manufacturing, and rising demand is pushing the industry from fed-batch toward continuous operation. Continuous cultures promise sustained high cell densities and volumetric productivities, but their operation depends on reliable real-time monitoring, prediction and control. Limited intracellular observability and sparse extracellular measurements hinder this, leaving a persistent gap between mechanistic understanding and predictive deployment. Here we describe a metabolism-informed model for mAb production that embeds reactor- and intracellular-level mass balances directly within a learning architecture, while deliberately avoiding predefined kinetic rate laws and explicit steady-state assumptions. Because conservation relationships remain invariant across operating modes, this formulation provides a principled structural prior for cross-regime generalisation, positioning the model as a soft sensor for continuous bioprocesses: inferring unmeasured intracellular fluxes and one-step ahead forecasting extracellular concentrations to support process monitoring and control. Benchmarking two metabolic network sizes and three metabolic optimisation strategies, concentration predictions are indistinguishable across network sizes. Exchange flux fidelity is agnostic to solver choice, with network size mattering only for intracellular interpretability. Transferring the developed model from fed-batch to continuous operation with fine-tuning outperforming other transfer learning variants. This is consistent with the metabolic structure being invariant to mode of operation. Predicted fluxes reveal growth rate in fed-batch, rather than operating mode, as the dominant axis of metabolic variation, and identify potential regime-shift signals to guide experimental design for continuous culture characterisation. These results chart a route towards digital models sufficiently grounded in first-principles knowledge, yet adaptive for deployment under regime shifts.