A hybrid modeling framework to optimize Chinese hamster ovary cell cultures for monoclonal antibody (mAb) production reduces the number of modeling parameters needed while returning results that compare favorably to the best kinetic models.
The framework designed by researchers at Sartorius and the University of Waterloo in Canada integrates hybrid dynamic flux balance analysis with piecewise partial least squares regression (dFBA-PLS). Results, they report, were highly predictive, with a “normalized mean squared error of 0.15 for most metabolites and interpretable flux profiles.” This approach, they continue, “successfully simulate[s] the temporal evolution of key metabolites, biomass, and mAb production across diverse CHO-fed-batch experiments.
“Multivariate analysis…revealed specific feed-inoculum combinations drove shifts in glycolysis, TCA cycle flux, and nitrogen metabolism,” they point out, which suggests this framework could be valuable in CHO cell culture optimization.
Because of significant metabolic variation among CHO cells and complex culture media, biomanufacturers typically expend considerable effort to optimize the growing environment. Often, these involve combinations of empirical and mechanistic models, such as metabolic flux analysis, flux balance analysis, or kinetic modeling.
The hybrid mode developed by first author Zahara Negahban, a PhD candidate in the lab of Hector Budman, PhD, and colleagues uses a “dFBA framework with kinetic rate constraints described by PLS-based regression functions.” This combination can model media mixing ratios directly, without needing the formulation’s precise composition and while accounting for various blends of cell culture media. It can predict the dynamic evolution of amino acids, glucose by-products, biomass, and titer, as well as the dynamic evolution of intracellular fluxes and, thus, correlations among specific media and key pathways.
Across the 18 runs the researchers performed, the hybrid dFBA-PLS model identified major trends and more accurately predicted reaction rates than did the study’s initial use of PLS followed by mass-balance equations. The dFBA-PLS model “uses the metabolic network to enforce mass balance and feasible fluxes, which lets it infer the right rate changes even when some variables are missing,” the researchers noted.
Overall, the model accurately predicted minute details of the cell culture behavior despite continuous formulation changes. That capability, they suggest, makes this hybrid modeling method valuable when designing and optimizing media blends by reducing the number of necessary experiments. Other potential benefits include the ability to evaluate designs in silico, accelerate media development, reduce process variability, and extend models.
This hybrid dFBA-PLS model for CHO cell culture optimization is scalable and can be ported to other CHO DG44-based cell lines, the researchers say.
