(283c) COSMIC-dFBA: A Novel Dynamic Flux Balance Analysis Framework to Interface Reactor Conditions with Metabolism Using Machine Learning Methods

Accurately modeling a bioprocess is critical to process optimization and real-time control. The inability of traditional dynamic flux balance analysis-based methods to accommodate changes in cell state and metabolism limit their application to bioprocess modeling. Here, we present COSMIC-dFBA as a novel data-driven approach to model cellular metabolism in a bioreactor. COSMIC-dFBA leverages machine learning principles to predict cell state and metabolic objectives based on metabolite concentrations in the bioreactor to dynamically update the uptake and secretion rates for all tracked metabolites. COSMIC-dFBA assumes that the net instantaneous uptake and secretion rates describing the observed concentration profiles in a bioreactor can be decomposed into contributions from individual cell states weighted based on the fraction of the cell population in each state. We describe the pre-processing of the genome-scale metabolic model, identification of state-specific cellular objectives, and feature selection to train the phase classifier model. The phase classifier model constructed using a combination of dimensional reduction and logistic regression had a Matthews correlation coefficient of 0.45, indicating superior performance compared to a random classifier. Integrating the phase classifier with the traditional dFBA framework enabled COSMIC-dFBA to seamlessly transition between the growth and production phases in a bioprocess. We validated COSMIC-dFBA using metabolite concentration profiles from seven different growth media containing varying quantities of glucose, amino acids, and oxygen. Using COSMIC-dFBA, we demonstrate the role of cell state transition in the observed lactate concentration spike in a perfusion bioreactor and the impact of switching from a glucose-centric metabolism to an amino acid-centric metabolism on cell growth and antibody productivity in a perfusion bioreactor. Thus, we show the utility of machine learning models in modulating cellular metabolism and expand the applicability of dFBA to model the dynamic conditions in a bioreactor.