Methods

Cell culture experiments

Two production fed batch processes were used, Fed batch 1 and Fed batch 2. Both fed batch processes used chemically defined media and feeds over the 12-day cell culture. Fed batch 1 used a glucose restricted fed batch process called HiPDOG(Gagnon et al., 2011). Glucose concentration is kept low during the initial phase of the process, Day 2-7, through intermittent addition of feed medium containing glucose at the high end of pH dead-band and then glucose was maintained above 1.5 g/L thereafter, restricting lactate production without compromising the proliferative capability of cells. In Fed batch 2 a conventional cell culture process was used where glucose was maintained above 1.5 g/L throughout the process.
For both process conditions, bioreactor vessels were inoculated at 2 x 106 viable cells/mL. The following bioprocess characteristics were quantified daily using a NOVA Flex BioProfile Analyzer (Nova Biomedical, Waltham, MA): viable cell density, average live cell diameter and concentrations of glucose, lactate, glutamate, and glutamine. Viable cell density data were converted to growth rates by following equation to be compared to model-predicted growth rates.
\(\text{Growth}\ \text{rate}=\ \frac{1}{\text{vcd}}\ \frac{\Delta\text{vcd}}{\Delta\text{time}}\)
Flash-frozen cell pellets (106 cells) and supernatant (1 mL) were collected from bioreactor runs for each sampling day. Collected samples were sent to Metabolon (Metabolon Inc, Morrisville, NC) for metabolomics analyses. Metabolomics measurements were used as input data to the model by converting their units to model units of mmol per gram of dry weight of cell per hour.

Metabolic network modeling

We used a previously described metabolic network model that is tailored to the investigated CHO clones(Schinn et al., 2020). Experimental measurements for clone and culture day were used to constrain model reactions for biomass production, monoclonal antibody secretion and consumption of glucose, lactate, glutamate, and glutamine. Then, we computed distributions of likely amino acid consumption rates by stochastically sampling 5000 points within the model’s solution space via a Markov chain Monte Carlo sampling algorithm, as described previously(Megchelenbrink et al., 2014; Nam et al., 2012), usingoptGpSampler (Megchelenbrink et al., 2014) and COBRApy(Ebrahim et al., 2013). Upon completion, the sampled distributions’ statistical features were noted – that is, their mean, median, standard deviation, 25 percentile, and 75 percentile values.

Statistical methods

For each amino acid, the mean of the sample distribution was interpreted as the likely consumption rates. These predicted consumption rates deviated from experimental observations by a consistent fold amount. Fold change error was also correlated with culture day, as the model predicted the exponential growth phase better than the subsequent stationary phase. Therefore, the model predictions were refined by a regression model as follows, with growth rate and the predictions themselves as explanatory variables.
\(\text{Corrected\ prediction}=\ \beta_{0}+\beta_{1}\bullet prediction+\beta_{2}\bullet growth\ rate\ \)
The time-course amino acid consumption profiles were described mathematically by the Monod equation, as follows:
\(\text{Consumption\ rate}=\ \beta_{0}\bullet\frac{\text{time}}{\beta_{1}+time}\)
Here, β0 represents the minimum consumption rate which the cells asymptotically approach during later stationary phase. The variable β1 is the half-velocity constant, or the time point at which the consumption rate reaches half of β0. These analyses were carried out and visualized using COBRA Toolbox 2.0(Schellenberger et al., 2011) in MATLAB R2018b (MathWorks; Natick, Massachusetts, USA)