UC Davis

In this dissertation we focus on the application of several design-of-experiments (DOE) methods to cell culture media development in order to sequentially learn optimal media formulations. These sequential DOE methods use data collected from an experimental system to simultaneously improve the model with additional suggested experiments while at the same time learning the optimal conditions of the experimental system. The purpose of this media is for applications in the cellular agriculture industry, where animal cells are grown for consumption. Starting with a hybrid scheme utilizing radial basis functions with a genetic algorithm and coordinate search method, we discovered that long-term cell growth is not fully correlated with the short-term chemical assays typically used in cell culture. We solved this by successfully deploying a Bayesian model that correlates long and short-term growth assays. We could then predict the information value of new experiments and assays jointly, reducing the overall number of experiments needed to solve the optimization problem. This improved Bayesian methodology focuses long-term experiments only on the most promising areas of the design space while allowing simpler short-term growth experiments to fully explore the design space. Using this new approach, we designed a medium with 181% more cell growth than a common commercial formulation with a similar economic cost, while doing so in 38% fewer experiments than an efficient DOE method using a desirability function to parameterize the outcome space. This medium even managed to maintain robust cell growth over four passages. Next, we used a hypervolume function to design experiments to sequentially learn the trade-off between cell growth and media cost in a serum-free system. We found a medium with a 184% improvement in growth over the control at a 71% increase in cost that maintained a high level of cell growth over five passages. Both optimal formulations resulted in robust long-term proliferation of cells, indicating the success of our multi-assay Bayesian approach to optimizing media. Future work could tie imaging software, bio-marker quantification, and techno-economic analysis to improve the accuracy and usefulness of predictions and experimental designs.