(253a) Multi-Scale Modeling of a CHO Production Process Using Integrated Machine Learning Models and Genome-Scale Metabolic Models | AIChE

(253a) Multi-Scale Modeling of a CHO Production Process Using Integrated Machine Learning Models and Genome-Scale Metabolic Models


Gopalakrishnan, S. - Presenter, University of California San Diego
Tat, J., University of California San Diego
Schlegel, F., Amgen Inc
Rolandi, P. A., Amgen Inc.
Johnson, W., Amgen
Kontoravdi, C., Imperial College London
Lewis, N., University of California
The goal of pharmaceutical bioprocess optimization is the maximization of antibody titer at harvest. Common strategies to achieve this goal include cell line optimization for improved growth and antibody specific productivity, and process optimization for optimal media feeding and timing of hypothermic shift to induce antibody production. Mathematical models that can accurately predict cellular responses and antibody production under changing bioreactor conditions in the bioreactor are a valuable tool that inform interventions and strategies to accelerate cell line and process development. This talk presents a novel dynamic flux balance analysis (dFBA)-based framework to predict concentration profiles of cells, glucose, lactate, amino acids, and therapeutics in a perfusion bioreactor. Because process performance depends on the changing metabolic states over the course of the process, we first characterize the metabolic features of five CHO cell lines engineered to produce three types of antibodies. In four of five cell lines, we find that high producer clones typically grow to higher cell densities than the unselected clones, and that the final titer positively correlates with the final cell density at the end of the growth phase. Interestingly, the final titer negatively correlated with cell density for the fifth cell line, implying that this clone follows a different pathway to achieve optimal antibody titer. Exo-metabolomic characterization of the cell lines revealed two distinct metabolic states corresponding to the growth phase and the production phase, respectively. Upon performing a principal component analysis on the computed fluxes, we found that a shift from growth to production phase was marked by a reduced growth, reduced lactate secretion, and increased antibody production. Hierarchical clustering of gene expression data measured for each cell line revealed that differences in cell state were primarily based on process phase (i.e., growth vs production phase) and that the cell state varied minimally across different cell lines. Upon extracting context-specific models using mCADRE, we found that the growth phase models preferentially activated anabolic pathways such as lipid and nucleotide biosynthesis, whereas amino acid degradation pathways were preferentially activated in the production phase models. Next, we developed a novel framework to simulate the concentration profiles of all metabolites and cells in a perfusion bioreactor using the knowledge gained from the characterization of the five cell lines. The framework expands on the traditional dFBA framework by (i) incorporating nutrient uptake using kinetic rate laws, (ii) constraining fluxes through key metabolic tasks based on process phase, and (iii) limiting scope of the metabolic model based on the process phase. A machine-learning model is trained to predict the process phase based on reactor conditions. The concentration profiles predicted by this multiscale dFBA model are validated using experimental data generated across different bioreactor operating conditions.