Media Optimization of CHO Cell Culture using a Hybrid Dynamic Flux Balance Analysis Model

Abstract

The manufacturing of pharmaceuticals relies heavily on upstream cell-culture processes that must achieve high, reproducible productivity and quality under tight timelines. Mathematical modeling is an essential tool for bioprocess prediction and optimization. Classical kinetic models provide mechanistic detail but require high parameterization, thus making them prone to overfitting and prediction inaccuracy outside the domain of conditions studied for model calibration. On the other hand, purely data-driven (black-box) models are easy to train but extrapolate poorly due to their lack of physical constraints. This thesis addresses that gap by developing, validating, and experimentally applying a hybrid modeling framework for Chinese hamster ovary (CHO) cell culture that couples dynamic flux balance analysis (dFBA) with partial least squares (PLS) regression models to describe concentration-dependent kinetic constraints. A key challenge in formulating dynamic metabolic models without overfitting is defining a minimal set of parameter-dependent constraints that is sufficient for accurate data fitting. In practice, the dominant drivers of growth and productivity include both abundant extracellular metabolites (e.g., glucose, glutamine) that can be tracked over time and minor components (vitamins, growth factors) that are often unknown due to confidentiality or only known at inoculation and are not routinely measured during the run. Direct optimization over all media components is therefore impractical. This work addresses that limitation by optimizing proportions of commercially available basal media and feeds rather than individual trace constituents. Kinetic bounds embedded in the hybrid model are then expressed as explicit functions of the media proportions, allowing the indirect, but operationally meaningful, optimization of media without time-resolved measurements of each species. Like many empirical and hybrid models, accuracy is strongest near the calibration domain, which can bias an optimizer if the predicted optimum lies outside the data support. To mitigate this, the thesis implements a run-to-run (batch-to-batch) optimization strategy in which each iteration consists of (i) model identification using newly collected data and (ii) model-based optimization to recommend the next experiment. The recommendation is executed, new trajectories are acquired, and parameters are updated for the subsequent iteration, thereby guiding the model and the process toward the true optimum through successive refinements. In this study, the hybrid dFBA–PLS model is integrated with experiments on an Ambr15® microbioreactor platform and enables efficient exploration of the media-blend simplex under consistent operating conditions. The availability of multiple parallel runs allows a design-of-experiments strategy to be layered onto the run-to-run loop, accelerating convergence to high-performing blends while quantifying variability across batches. In particular, the experimental study demonstrates the key importance of matching gradients between experiments and model predictions, an intermediate step in our methodology, to drive the process close to an optimum. Without such matching of gradients, it is shown that the optimization is not meaningful. The overall optimization goal is to improve culture performance by identifying media blend compositions, encoded here by inoculum and feed fractions of commercial media, that maximize monoclonal antibody (mAb) production while maintaining target viability. Systematically varying these fractions tunes both major nutrients and traces in a controlled, scalable manner. Within the hybrid model, a piecewise PLS layer maps measured states (e.g., extracellular concentrations, viable cell density) and media proportions to metabolite uptake/production rates; these rates are transformed into upper–lower kinetic bounds for selected exchange and lumped reactions, which the dFBA layer enforces alongside intracellular stoichiometry and mass balances. In this way, the model links media composition to feasible flux distributions and, in turn, to dynamic trajectories of biomass and mAb. A key novel contribution of the modeling approach is the use of uncertainty bounds for the regression models describing the constraints. It is shown that relaxing or tightening these bounds for the regression models provides several advantages: i- it addresses the multiplicity of solutions of dFBA by limiting the solution space, ii- it reduces overfitting by widening some bounds, thus making them less sensitive to the corresponding constraint, and iii- the relaxation of bounds for a particular constraint reduces sensitivity with respect to this constraint without the need of completely eliminating the constraint that would require expensive mixed integer optimizations. The specific contributions of the thesis are 1. Development of a novel hybrid CHO model that combines a dFBA core with PLS defined, concentration- and media-proportion-dependent kinetic bounds, using a minimal set of tunable uncertainty parameters to avoid overfitting. 2. Implementation and validation of the hybrid model for CHO cultures conducted on Ambr15® cultures under diverse inoculum and feeding formulations, demonstrating the ability to reproduce key metabolic behaviors (e.g., lactate and ammonia dynamics) and product formation profiles. To our knowledge, this is the first CHO model of the dFBA type that explicitly accounts for mixtures of media. 3. Integration of the hybrid model into a run-to-run optimization procedure that recommends next-batch media blends to maximize mAb titer at target viability, using parallel experiments to update parameters, assess variability, and improve recommendations iteratively. This is the first application of the Ambr15® in the context of a run-to-run model-based optimization approach. The application of this methodology led, after 3 iterations, to an almost 30 percent improvement in the value of an objective function consisting of the specific productivity at 80 percent viability. Together, these elements yield practical modeling and model-based optimization frameworks that respect physicochemical constraints, leverage data efficiently, and directly support media-blend design. By expressing kinetic limits as functions of media proportions, the approach enables optimization over both major and minor components without requiring time-resolved measurements of every trace species. Embedding the model in a run-to-run loop further aligns the model to the plant response as the search advances toward the true optimum. The resulting dFBA–PLS methodology provides accurate, interpretable predictions and actionable guidance for upstream process development in CHO cell culture.