A computational framework for predicting CHO cell culture performance in bioreactors

Abstract

Breaking the trade-off between speed and productivity is a key milestone across industries. In particular, in the biopharmaceutical industry this trade-off is exacerbated by a highly regulated environment, which hinders continuous improvement and fixes future manufacturing costs. Given the complexity of living organisms and the improvement in quality of life offered by the product - which demand agile development - the industry has traditionally taken phenomenological approaches to process development, generally sacrificing costs. Nonetheless, technological developments and lower entry barriers make the biopharmaceutical industry far more competitive than in its origins, demanding efficient and reliable processes. Developing efficient manufacturing processes for new products while being agile to market is a key differentiating capability of Amgen's process development organization.

In collaboration with the process development team at Amgen, a computational framework for in-silico upstream bioprocess development has been developed, allowing for faster, more robust, and more optimal process development. Specifically, a mechanistic model of a bioreactor has been designed, implemented, and applied to an Amgen product. The project was divided into three major components: The first was a survey of internal Amgen capabilities and the state of the art in external industrial and academic models to identify the algorithms and design the signal flow required to support the range of expected process engineering applications. The second consisted of implementing a modular, extensible software platform with the architecture and interfaces dictated by the first component. The third part consisted of applying the software to an actual product development problem capturing the primary process variables.

A constraint-based model of a metabolic network consisting of 35 reactions of the main carbon-nitrogen metabolism relevant in energy and redox balance was adapted from literature (Nolan & Lee, 2011). The metabolic network was coupled with glucose, glutamine and asparagine kinetics with temperature, dissolved oxygen, pH and osmolarity dependence. Stress induced by temperature shifts was modeled as a first-order step response coupled to a non-growth associated ATP of maintenance. The cellular model was coupled with a well-mixed bioreactor model consisting of mass balance equations. We solved the model using dynamic Flux Balance Analysis (dFBA). We first calibrated the model with experimental process characterization data for a product in development. We used a Non-dominated Sorted Genetic Algorithm (NSGA-II) to solve the calibration problem, minimizing the error in metabolite concentrations to yield estimates of 13 strain-specific parameters.

We then assessed the calibrated model's predictions of biomass growth and metabolite concentrations against a second experiment run with different process settings. Finally, I developed a graphical user interface for subject-matter-experts to simulate experiments and test hypotheses using the model. We applied the tool to three process-relevant case studies, and analyzed the in-silico results. The calibrated model can predict biomass and titer from process settings, potentially reducing experimental time from 20 days to 30 seconds, in addition to reducing the experimental cost.