Advanced computational approaches for the modeling and optimization of mammalian cell metabolism.

Abstract

Over the past decades, biologics have transformed the modern treatment of complex human diseases. Mammalian cell culture forms the basis for manufacturing high-value biologics due to its capacity for high cell density and proper post-translational modifications. Despite the significant strides made in cell culture process development, it remains time-consuming and costly, which motivates the development of computational tools that can help accelerate the timelines and improve process efficiency. In this thesis, we develop advanced computational approaches for the modeling and optimization of cell metabolism, addressing challenges in cell culture processes for biologics production. In the first portion of the thesis, we explore the use of both mechanistic and data-driven modeling techniques to better understand the underlying mechanism of the complicated metabolic shifts that affect the performance of cell culture processes. We first developed mechanistic model with additional cellular machinery to capture metabolic characteristics across multiple CHO cell lines, using a new parameter estimation approach to construct models with small datasets. Next, we assess the capability of recurrent neural networks for the forecasting of cell culture profiles using a large biomanufacturing dataset. The results show that the input structure accounting for fed-batch operations is crucial for accurately modeling metabolic shifts, and the process history can strongly affect cell culture dynamics at the final production scale. In the subsequent part, we present a new solution framework for kinetic-model-based pathway optimization and apply it to study reverse glycolysis in mammalian cell lines, potentially reducing the need for glucose feed in cell culture. The key findings show good agreement with the existing mechanistic understanding of gluconeogenesis in liver and kidney metabolism, emphasizing the importance of metabolic kinetics and regulation machinery when studying metabolism in mathematical models. In the last part of the thesis, we develop a novel algorithm to solve inverse optimization problems using Bayesian optimization with a statistical analysis that offers uncertainty quantification. The algorithm is first applied to learn unknown cellular objectives in flux balance analysis. Additional case studies in system engineering problems demonstrate the efficiency and robustness of the algorithm in solving various classes of inverse optimization problems. Overall, the advancements in computational tools documented in this thesis extend our understanding of cell metabolism and provide insights into mitigations of undesired metabolic shifts, and can potentially help accelerate the development cycles and improve the efficiency and robustness of biomanufacturing processes.