(432f) A Kinetic Model-Based Transfer Learning Approach to Predicting Cell Line-Specific Metabolic Behavior in Biomanufacturing

Metabolism is the most important driver of environmental changes in cell culture processes and exerts profound effects on productivity and product quality [1]. A cell line’s metabolic behavior is dynamic, varying with growth rate and chemical environment, and the dynamics even change with different production cell lines. A kinetic metabolic model that integrates the effect of growth control, metabolic regulations, and the dynamics of the cell’s metabolic machinery can greatly facilitate the development of predictive tools for process enhancement [2]. A major challenge in the construction of such predictive models is the large amount of data needed for parameter estimation, which is usually available only to well-established processes but not for new production cell lines.

In this work, we develop a transfer learning approach [3] that facilitates the data-driven construction of kinetic metabolic models for cell lines with limited data using knowledge obtained for cell lines for which richer datasets are available. To this end, we collect a comprehensive set of process and transcriptome data from different CHO cell lines at different metabolic states over time in culture; these cell lines have different but related metabolism. The parameter estimation problem is formulated as a nonlinear programming program with a tailored regularization term in the loss function. The proposed regularization effectively allows for the transfer of certain learned parameters across cell lines while keeping subtle differences in the models that account for varying cell-line behavior. The resulting cell line-specific kinetic models incorporating dynamic transcriptome recover the metabolic shifts observed in the culture, indicating the significance of transcriptome variability over time and across cell lines to the cell metabolism. In addition, the growth rate-dependent models reclaim the importance of the growth regulations to the control of the glycolytic states in the production processes. Finally, we perform model simulations under different culture conditions to identify factors that impact the metabolic shifts the most, providing insights into potential cell line improvement.

[1] Le, H., Kabbur, S., Pollastrini, L., Sun, Z., Mills, K., Johnson, K., Karypis, G., & Hu, W.S., 2012. Multivariate analysis of cell culture bioprocess data—Lactate consumption as process indicator. J. Biotechnol. 162, 210–223.

[2] O’Brien, C.M., Zhang, Q., Daoutidis, P., & Hu, W.S., 2021. A hybrid mechanistic-empirical model for in silico mammalian cell bioprocess simulation. Metab. Eng. 66, 31–40.

[3] Pan, S.J., & Yang, Q., 2010. A survey on transfer learning. IEEE Trans. Knowl. Data Eng. 22, 1345–1359.