(634c) Leveraging Prior Knowledge and General Cell Culture Models to Configure and Validate Bioreactor Digital Twins across Scales and Cell Lines

Mammalian cell cultures have the ability to express large complex recombinant proteins, which makes them a crucial platform for the production of biopharmaceuticals such as monoclonal antibodies, viral vaccines, and immuno-regulators. High-fidelity models with a good predictive capability of the cell culture behaviour in different culture conditions have great potential to accelerate and derisk process scale-up and technology transfer. However, generating accurate bioreactor models is not a trivial task, and it is well accepted that, for cell culture processes, model calibration and external validation is often a complex and time-consuming activity.

The most essential and challenging aspects in the development of such models are finding the correct model structure, identification of model parameters from experimental data, and consequent model validation^[1]. Typically, cell culture models are validated across different process conditions but for a single scale and a single cell line only^[1,2]. This means that almost the same model validation effort is required each time a new clone is developed or the scale is changed.

In this work, we used experimental data from a fed-batch cell culture for the production of a monoclonal antibody to configure, calibrate and validate a bioreactor model across different scales and test it with different cell lines. Approaches to leverage prior knowledge to develop a general cell culture model with elements transferable across different scales and clones were tested. The approach aims to minimise data requirements for new assets, and extend the usual range of model validation and testing.

The effect of critical process parameters on monoclonal antibody (mAb) production in Chinese hamster ovary (CHO) cells was investigated. Cells were grown with a fixed feeding strategy in three different bioreactor systems: i.e., 1L Erlenmeyer orbitally shaken flasks (flask), 250 mL mini bioreactors (Dasbox) and 5 L stirred tank reactors (STR). In flask scale, a 3-level full factorial design was used, with the three factors being: cell line (GS-CHO producing IgG₄, GS-CHO Null, CHO-S producing trastuzumab), osmolality shift (no shift, shift to 380 mOsm∙kg^-1, shift to 440 mOsm∙kg^-1) and temperature shift (no shift, shift to 32^oC, shift to 34.5^oC). In Dasbox scale, a one-factor-at-a-time method was employed where the critical factors included pH shift (no shift, shift to 6.8, shift to 7.5). The osmolality shift (no shift, shift to 450 mOsm∙kg^-1, shift to 520 mOsm∙kg^-1) and temperature shift (no shift, shift to 32^oC, shift to 34.5^oC) were also tested to assess the process performance in a controlled environment (pH, temperature and dissolved oxygen) compared to the uncontrolled environment in shake flasks. These experiments were used to evaluate the possible impact of shear stress and sparging on the cell culture dynamics. Findings were then confirmed at the larger scale in the STR system. Osmolality and temperature shifts occurred in the mid-to-late exponential phase when cells transitioned to hyperosmotic and/or hypothermic conditions until the end of the culture. Both STR and Dasbox experiments were performed using GS-CHO producing IgG₄.

The mathematical model described in this work was implemented in gPROMS FormulatedProducts® 2023.1 (Siemens, UK), using the inbuilt numerical solution methods for simulation and parameter estimation. The bioreactor model, included in the upstream bioprocessing libraries, is a predominantly mechanistic model that describes the phenomena associated with cell culture dynamics (i.e., cell growth, death and lysis, nutrient consumption, metabolite and product secretion) as well as gas-liquid mass transfer phenomena and extracellular reactions. The rate of phenomena associated with cell culture was modelled using well-established kinetic expressions. Cell growth was represented by simple exponential growth kinetic and Moser kinetics, which include terms for limiting nutrients and inhibitory metabolites^[3]. Similarly, toxic and depleting metabolites were accounted for in the rate of cell death, as expressed through Moser kinetics. The cell lysis rate was calculated with a zero-order expression. Regarding metabolite components in the medium, diverse kinetic terms were set for nutrient consumption and metabolite secretion (among growth-dependent, Moser, zero order, etc) ^[3]. Customized kinetics expressions where adopted when the out-of-the-box options available in the libraries could not represent the actual experimental trends. Finally, a hybrid approach, combining mechanistic and data-driven modelling, was used to describe product secretion, as the mechanistic understanding of the impact of process conditions on productivity was not good enough to reproduce the data. The bioreactor model validated in this work also simulates gas-liquid mass transfer of oxygen and carbon dioxide. Differences in mixing and aeration conditions were accounted for in the model.

Three fed-batch model flowsheets were configured for the different equipment types: flask, Dasbox and Stirred Tank Reactor (STR). All three flowsheets, represented in Figure 1, include a bioreactor, feed models providing nutrients to the bioreactor, and a sampler. Sodium chloride feed is used to promote osmolality shifts. The bioreactor contains an initial medium seeded with cells, and the temperature is controlled over culture time. In the flask flowsheet, the dissolved gases are not controlled. The Dasbox and STR flowsheets include two independent control systems, one for pH and another for dissolved oxygen. The pH control system comprises a pH sensor and a controller connected to CO₂ and base sources that feed the bioreactor to maintain the pH at the desired level. The oxygen control system includes a dissolved oxygen sensor and a controller connected to air, O₂ and N₂ sources to keep a stable dissolved oxygen value. Both bioreactor models account for gas-liquid mass transfer.

Following the configuration step, the calibration activity started by using the experimental data from flask experiments, as this data set presents the most varied process conditions among the different data sets. The data set is for a single cell line (GS-CHO) and includes multiple conditions of osmolality and temperature shift. Since the model structure is not known a priori, we adopted a systematic parameter estimation strategy that allows the definition of the model structure and the identification of a good set of initial guesses for model parameters. Then refining the parameter estimation through a series of sequential steps allows us to obtain a model suitable for relevant modelling activities, such as the design space characterisation and process optimisation. The workflow is shown in Figure 2 and the steps described below.

1. Analysis of available data, selection of measurements to use in parameter estimation.
2. Estimation of the exponential growth rate parameters.*
3. The exponential growth rate is fixed and the stationary phase is included in the model. Cell growth limiting nutrients and inhibitory metabolites are selected.
4. Estimation of model parameters for stationary phase. Cell growth saturation and inhibitory constants estimated along with consumption/secretion rates of limiting/inhibitory metabolites.*
5. The parameters estimated for the exponential and stationary phases are fixed, and the model introduces the decline phase. Cell death and lysis models are defined.
6. Estimation of model parameters of the decline phase. Cell death saturation constants and secretion rate of toxic metabolites are estimated.*
7. All previously estimated parameters are fixed, and the model of product secretion is defined, as well as a model for lactate and any other relevant metabolites.
8. Estimation of model parameters related to product secretion and additional relevant metabolites.*

*In steps 2,4,6,8: when model validation failed, model structure was modified and experimental data sets used were reconsidered.

Using the flask data of GS-CHO cell line and the model validation procedure elaborated above, we successfully estimated the kinetic parameters for cell growth, cell death, cell lysis, nutrient consumption, metabolite secretion and product secretion. The calibrated model was further validated with the flask data for the other cell lines (GS-CHO Null and CHO-S). Data from CHO-S, a cell line with divergent behaviour, was used to test the extent of model transferability and identification of model expressions and parameters that need to be modified to build a platform model. As expected, different cell growth-limiting nutrients and different rates of consumption and secretion of some key metabolites were identified for different cell lines.

Analogous to the validation done for different cell lines, the model was validated for different scales, Dasbox and STR. It was also expanded to include the impact of a pH shift based on the data from the Dasbox experiments.

Ultimately, we developed a strategy to use data from different scales to validate the cell culture dynamics model. The evaluation of model behaviour with different cell lines and at different scales increases the usability range of the model and gives insights into bottlenecks for building a platform model. This work presents a fundamental step in the development of a platform model for bioprocessing, a general model that can be easily adapted and used for diverse products.

[1] Kontoravdi C., Pistikopoulos, E.N., Mantalaris, A., (2011), Systematic development of predictive mathematical models for animal cell cultures. Comput. Chem. Eng., 34(8):1192-1198. doi:10.1016/j.compchemeng.2010.03.012

[2] Kiparissides, A., Pistikopoulos, E.N., Mantalaris A. (2014), On the model-based optimization of secreting mammalian cell (GS-NS0) cultures. Biotechnol. Bioeng., 112(3):536-548. doi:10.1002/bit.25457

[3] Pereira, S., Kildegaard, H.F. and Andersen, M.R. (2018), Impact of CHO Metabolism on Cell Growth and Protein Production: An Overview of Toxic and Inhibiting Metabolites and Nutrients. Biotechnol. J., 13: 1700499. doi:10.1002/biot.201700499