(588bp) Application of Hybrid Dynamic Modelling for Fed-Batch Cell Culture Processes Using Gaussian Process Regression

The development of dynamic process models provides an indispensable approach for capturing the intricate, time-dependent behaviour of biological processes, particularly in dynamic systems such as fed-batch Chinese Hamster Ovary (CHO) cell cultures [1-4]. Dynamic models help bridge the gap between experimental observations and process understanding, allowing for improved design and performance predictions during process development. However, accurately modelling the complexities of cellular processes, including cell growth, death, and metabolism, remains a significant challenge due to the non-linear nature of these systems and the inherent variability in bioprocess experimental data [5].

This study presents a hybrid dynamic process model that integrates Gaussian Process Regression (GPR) for estimating cell-specific rates, coupled with material balance equations to capture the underlying dynamics of CHO fed-batch cell culture processes. The principal aim of this research is to investigate how specific properties of the GPR model, such as data smoothing, kernel function selection, and input feature selection, affect GPR performance, and how these properties influence the predictive capability when integrated into a larger dynamic process model. By focusing on these GPR-specific attributes, this work aims to advance current understanding of how model development choices impact the overall performance of hybrid dynamic models within bioprocessing applications.

The exploration of key GPR model properties in this study is driven by the need to maximise the predictive capacity of dynamic models while accounting for the complexity and variability of biological data. Firstly, the role of data smoothing techniques prior to rate estimation is investigated, to assess potential impact on model accuracy. Secondly, the study explores the relevance of kernel selection, through comprehensive comparison of Squared Exponential (SE), Matern, and Rational Quadratic (RQ) kernel options, to understand the influence on model performance and generalisability. Finally, the study evaluates the impact of input feature selection, particularly the importance of leveraging domain-specific knowledge to identify critical model features that contribute towards improved prediction.

To complement primary investigations, the study employs global sensitivity analysis (GSA) to provide an understanding on how different input features contribute to model variance, thus providing deeper insight into the most influential factors on cell-specific rate estimates. While a strategy for applying GSA to advanced regression models, such as GPR, has not been published to date, the results indicate that variance-based approaches can lead to enhanced process insight through characterising the distribution of variance in model predictions across input features.

Ultimately, the work presented aims to offer valuable insight into the development of accurate and efficient hybrid dynamic models for bioprocess applications. By carefully examining the impact of GPR model properties on predictive performance, this work highlights the importance of balancing model complexity and predictive accuracy. The findings contribute to the growing body of knowledge in dynamic bioprocess modelling, particularly in the context of fed-batch cell culture processes, while providing a strong foundation for future research in dynamic cell culture modelling and the deployment of bioprocess digital twins.

References

1. Pinto, J., J.R.C. Ramos, R.S. Costa, S. Rossell, P. Dumas, and R. Oliveira, Hybrid deep modeling of a CHO-K1 fed-batch process: combining first-principles with deep neural networks. Front Bioeng Biotechnol, 2023. 11: p. 1237963.

2. Yatipanthalawa, B.S., S.E. Wallace Fitzsimons, T. Horning, Y.Y. Lee, and S.L. Gras, Development and validation of a hybrid model for prediction of viable cell density, titer and cumulative glucose consumption in a mammalian cell culture system. Computers & Chemical Engineering, 2024. 184.

3. Doyle, K., A. Tsopanoglou, A. Fejér, B. Glennon, and I.J. del Val, Automated assembly of hybrid dynamic models for CHO cell culture processes. Biochemical Engineering Journal, 2022.

4. Luo, Y., V. Kurian, L. Song, E.A. Wells, A.S. Robinson, and B.A. Ogunnaike, Model‐based control of titer and glycosylation in fed‐batch mAb production: Modeling and control system development. AIChE Journal, 2023. 69(4).

5. Tsopanoglou, A. and I. Jiménez del Val, Moving towards an era of hybrid modelling: advantages and challenges of coupling mechanistic and data-driven models for upstream pharmaceutical bioprocesses. Current Opinion in Chemical Engineering, 2021. 32.