The control of nutrient availability is critical to large‐scale manufacturing of biotherapeutics. However, the quantification of proteinogenic amino acids is time‐consuming and thus is difficult to implement for real‐time in situ bioprocess control. Genome‐scale metabolic models describe the metabolic conversion from media nutrients to proliferation and recombinant protein production, and therefore are a promising platform for in silico monitoring and prediction of amino acid concentrations. This potential has not been realized due to unresolved challenges: (1) the models assume an optimal and highly efficient metabolism, and therefore tend to underestimate amino acid consumption, and (2) the models assume a steady state, and therefore have a short forecast range. We address these challenges by integrating machine learning with the metabolic models. Through this we demonstrate accurate and time‐course dependent prediction of individual amino acid concentration in culture medium throughout the production process. Thus, these models can be deployed to control nutrient feeding to avoid premature nutrient depletion or provide early predictions of failed bioreactor runs.

Genome-scale metabolic models describe cellular metabolism with mechanistic detail.Given their high complexity, such models need to be parameterized correctly to yield accurate predictions and avoid overfitting. Effective parameterization has been wellstudied for microbial models, but it remains unclear for higher eukaryotes, including mammalian cells. To address this, we enumerated model parameters that describe key features of cultured mammalian cells -including cellular composition, bioprocess performance metrics, mammalian-specific pathways, and biological assumptions behind model formulation approaches. We tested these parameters by building thousands of metabolic models and evaluating their ability to predict the growth rates of a panel of phenotypically diverse Chinese Hamster Ovary cell clones. We found the following considerations to be most critical for accurate parameterization: (1) cells limit metabolic activity to maintain homeostasis, (2) cell morphology and viability change dynamically during a growth curve, and (3) cellular biomass has a particular macromolecular composition. Depending on parameterization, models predicted different metabolic phenotypes, including contrasting mechanisms of nutrient utilization and energy generation, leading to varying accuracies of growth rate predictions. Notably, accurate parameter values broadly agreed with experimental measurements. These insights will guide future investigations of mammalian metabolism.

10Genome-scale metabolic models describe cellular metabolism with mechanistic detail. 11Given their high complexity, such models need to be parameterized correctly to yield 12 accurate predictions and avoid overfitting. Effective parameterization has been well-13 studied for microbial models, but it remains unclear for higher eukaryotes, including 14 mammalian cells. To address this, we enumerated model parameters that describe key 15 features of cultured mammalian cells -including cellular composition, bioprocess 16 performance metrics, mammalian-specific pathways, and biological assumptions behind 17 model formulation approaches. We tested these parameters by building thousands of 18 metabolic models and evaluating their ability to predict the growth rates of a panel of 19 phenotypically diverse Chinese Hamster Ovary cell clones. We found the following 20 considerations to be most critical for accurate parameterization: (1) cells limit metabolic 21 activity to maintain homeostasis, (2) cell morphology and viability change dynamically 22 during a growth curve, and (3) cellular biomass has a particular macromolecular 23 composition. Depending on parameterization, models predicted different metabolic 24 phenotypes, including contrasting mechanisms of nutrient utilization and energy 25 generation, leading to varying accuracies of growth rate predictions. Notably, accurate 26 parameter values broadly agreed with experimental measurements. These insights will 27 guide future investigations of mammalian metabolism. 28

NH4)2S04, 1.6 g of KH2PO4, 0.20 g of CaCl2, 0.20 g of MgSO4-7H2O, 0.80 mg of H3BO3, 0.80 mg of ZnSO4 7H2O, 0.80 mg of MnCl24H2O, 0.40 mg of FeCl3 6H20, 80 ,ug of CuSO45H20, 80 ,ug of KI, 2.4 mg of inositol, 0.80 mg of f3-alanine, 80 j±g of nicotinic acid, 16 ,ug of pyridoxine hydrochloride, 16 iLg 864 on August 11, 2020 by guest

The control of nutrient availability is critical to large-scale manufacturing of biotherapeutics. However, the quantification of proteinogenic amino acids is time-consuming and thus is difficult to implement for real-time in situ bioprocess control. Genome-scale metabolic models describe the metabolic conversion from media nutrients to proliferation and recombinant protein production, and therefore are a promising platform for in silico monitoring and prediction of amino acid concentrations. This potential has not been realized due to unresolved challenges: (1) the models assume an optimal and highly efficient metabolism, and therefore tend to underestimate amino acid consumption, and (2) the models assume a steady state, and therefore have a short forecast range. We address these challenges by integrating machine learning with the metabolic models. Through this we demonstrate accurate and time-course dependent prediction of individual amino acid concentration in culture medium throughout the production process. Thus, these models can be deployed to control nutrient feeding to avoid premature nutrient depletion or provide early predictions of failed bioreactor runs.