In the cultivation process for mAb production, impurities (e.g., host cell proteins (HCP), DNA, and charge variants) are produced as well as mAb, i.e., the final product. These impurities have a significant impact on quality in the time- and resource-intensive cultivation process1, making it a major factor influencing the overall production cost, time, and quality. Since these impurities and mAb are affected by many cultivation conditions, it is essential to select conditions appropriately. Therefore, mathematical models that describe the relationship between quality attributes and cultivation conditions are useful in the design of this critical cultivation process without relying on trial-and-error experiments.
In the field of process systems engineering, several studies have worked on developing models toward improving the cultivation process, e.g., hybrid modeling to improve predictions of impurities2 and dynamic simulations based on models for elucidating optimal feeding, bleeding, and harvest strategies3. However, existing models had limitations in quality attributes considered and the range of applicable cultivation conditions, posing challenges for application in practical design.
This work presents a hybrid-modeling approach to optimize the cultivation process. We first developed a hybrid model using six cultivation conditions (e.g., agitation rate and dissolved oxygen (DO)) as inputs to describe not only mAb but also impurities (HCP, DNA, and charge variants). To obtain experimental data for developing the model applicable to a wide range of cultivation conditions, automated cultivation equipment with twelve 250 mL bioreactors was set up which enabled reducing experimental burden by minimizing manual operations. Four cycles of fed-batch cultivation were performed while varying the six cultivation conditions which were used as model inputs. In the experiments, a newly developed CHO-MK cell line4 producing an IgG1 mAb was used, and multiple items including viable cell density, mAb, metabolites, and impurities were measured as time series data.
To validate the model, three additional experimental cycles were conducted under 15 different conditions. High prediction accuracy was confirmed in twelve conditions, while discrepancies were observed in the remaining three. The discrepancies were considered to result from high agitation rate and DO, which hindered cells from controlling environmental factors, leading to unexpected cell behaviors. By using the model, simulations were conducted within the range where high prediction accuracy was confirmed, and the optimal operating condition was found to maximize mAb production while keeping impurities below given thresholds. Finally, the validity of the optimal condition was confirmed through experiments. The developed hybrid model and the identified optimal point are expected to contribute to improving the cultivation process without relying on trial-and-error experiments . In the ongoing work, we are working on developing an operational support approach to address unexpected cell behaviors as observed in the validation experiments.
References