Hybrid Dynamic Modelling With Gaussian Process Regression for Intensified Fed-Batch CHO Cell Culture Processes.

The increasing complexity of cell culture processes and sustained drive towards efficient, scalable bio-manufacture accentuate the need for advanced process modelling techniques. Hybrid models, which integrate the flexibility of machine learning algorithms within the structure of material balances, offer a promising route towards digital bioprocess development. In this study, a hybrid dynamic model based on Gaussian Process Regression (GPR) is proposed for prediction of cell growth and metabolism, and is subsequently evaluated across two contrasting intensified fed-batch CHO cell culture processes. Predictive performance of the hybrid model is compared against an alternative Monod-based kinetic modelling approach. In relation to the hybrid strategy, the study incorporates data from both development and manufacturing scale to explore how the incremental nature of data availability and impact of data quality affect the predictive capacity of GPR-hybrid models. Presented results highlight the value of hybrid approaches over mechanistic methods, emphasise the importance of data quality, and indicate the need for considerate experimental design to ensure model reliability and performance. Challenges are presented regarding transient bolus feeding effects, indicating a need for further investigation into the GPR-hybrid model structure. Overall, this study demonstrates the capability of GPR-hybrid models to capture the dynamics of intensified CHO cell culture processes across development and manufacturing scales, highlighting both the promise and current limitations in the context of data-driven bioprocess development.