Introduction
Improved robustness in production of monoclonal antibodies (mAbs) from Chinese Hamster Ovary (CHO) cell culture processes depends on gaining better understanding of biological changes, interactions, and impact to product quality and yield. Advances in systems biology have led to the development of computational tools to compliment emerging technologies that can perform data-rich experimentation (Schaub, Clemens, Kaufmann, & Shulz, 2011; Sharma, Tripathi, & Mukherjee, 2016). As a result, the employment of computational tools such as empirical or mathematical modeling have gained acceptance as a potential path to provide a deeper understanding of cellular mechanisms to inform process decisions that result in more efficient experimental practices during pre-commercial biologics development (Insaidoo, Banerjee, Roush, & Cramer, 2017; Roush et al., 2020; Smiatek, Jung, & Bluhmki, 2020). For instance, computational efforts have shown great potential in optimizing nutrient concentrations within chemically defined (CD) media for CHO cells by modeling routine cell culture and metabolomic data to highlight key limiting nutrients (Cuperlovic-Culf & Culf, 2016). Optimizing base and feed media compositions is a critical step of process development as components in these media directly impact critical quality attributes and cellular productivity of mAbs (Li, Vijayasankaran, Shen, Kiss, & Amanullah, 2010). However, despite long-term benefits of raw material control and cost reduction, development of proprietary CD media is a resource-and time-intensive process which requires a significant number of experiments (Kelley, 2009). By leveraging historical data in computational algorithms, process development researchers are able to identify key components in media that can significantly reduce the complexity of large experimental designs and reduce the development time to optimize media (Torkashvand et al., 2015; Xing et al., 2011).
In this study, we explored the utilization of multivariate data analytics (MVDA) modeling to optimize amino acid concentrations in CD feed media with aims to increase cell growth and mAb productivity. While other components in CD media such as vitamins, trace elements, and co-factors are required for various cellular functions, amino acids are considered one of the most important component group in CD media representing about 74% of CHO cell mass (Carrillo-Cocom et al., 2015; Fan et al., 2015; Pan, Streefland, Dalm, Wijffels, & Martens, 2017; Templeton, Xu, Roush, & Chen, 2017a). Balancing and optimizing amino acid concentrations therefore is critical to ensure CD media can sustain cell growth and mAb productivity. Thus far, majority of laboratory-based experimental and computational approaches for optimizing amino acid levels in CD media have been built on supplementing the amino acids that were being consumed in excess (Dietmair et al., 2012; Duarte et al., 2014; Robitaille, Chen, & Jolicoeur, 2015a). A challenge with such an approach is that over supplementation and consumption of certain amino acids can lead to the production of harmful byproducts. For instance, varying the amount of essential amino acid concentrations can lead to the differential regulation of amino acid metabolic pathways that could ultimately result in increased formation of harmful byproducts such as lactate or ammonia (Alden et al., 2020; Morris et al., 2020; Park, Reimonn, Agarabi, Brorson, & Yoon, 2018; Pereira, Kildegaard, & Andersen, 2018). Alternatively, supplementing amino acids in a time-dependent manner based on cell demand would ensure optimal consumption and cellular metabolism presumably resulting in better cell growth and mAb productivity.
In aims to understand the dynamics of amino acid metabolism, Xie and Wang described the relationship between the theoretical demand and the actual consumption of amino acids as their stoichiometric balances (Xie & Wang, 1994b; Xie & Wang, 1994c). According to their work, the theoretical demand of an amino acid is the amount needed to maintain both biomass and mAb productivity per cell, and with a steady state approximation, the amount consumed is balanced stoichiometrically with the theoretical demand. However, recent work on CHO cell amino acid requirements showed that most cells do not behave in a steady state and that the production process by which nutrient feeds are provided can greatly modulate the dynamics of consumption (Kyriakopoulos, Polizzi, & Kontoravdi, 2013; Luo et al., 2020). Templeton et al described the existence of two possible metabolic states. Within the first state, the actual consumption of an amino acid exceeds the theoretical demand resulting in a catabolic state in which the cell can breakdown the excess nutrients for energy or biomass. In contrast, within the second state, the theoretical demand exceeds the actual consumption of an amino acid resulting in an anabolic state in which the cells must biosynthesize or acquire the required nutrients (Table 1) (Templeton, Xu, Roush, & Chen, 2017b). Time-dependent dynamics of amino acid stoichiometric balances can therefore highlight concentration constraints and metabolic requirements for different amino acids throughout the culture (Ahn & Antoniewicz, 2011; Pan, Dalm, Wijffels, & Martens, 2017a). Although a stoichiometric-based model can provide a systematic and mathematical approach to balancing component composition in CD media, it does not eliminate the resources, and time needed for experimental work (Kumar, Bhalla, & Rathore, 2014). Accordingly, computational MVDA approaches have shown potential in utilizing historical cell culture data and nutrient consumption rates to minimize the resource demand of lab-based experimental designs and execution (Suarez-Zuluaga, Borchert, Driessen, Bakker, & Thomassen, 2019).
MVDA methods rely on a dimensional reduction strategy in which information from a group of variables can be consolidated and summarized into latent variables that in turn, can help describe the overall distribution of experimental observations (i.e. culture time points, cell culture batches) (Mevik & Wehrens, 2007). Each latent variable is based on the individual contribution of variance from all of the variables within the group, and variable ranking statistics can therefore be used to highlight and sub-select a set of key variables towards specific response parameters (Prieto, Eriksson, & Tyrgg, 2014). As a result, MVDA approaches can be used to analyze batch-to-batch variability, test the impact of culture performance from process variations, and identify key nutrients from cell culture media (Nitish & Rathore, 2011). In addition, statistical MVDA models in bioprocess rely on the “black box” approach in correlating extracellular metabolites and fluxes to cellular phenotypical behavior without mathematically modeling the intracellular complexities of cellular metabolism (Kanehisa & Goto, 2000; Noll & Henkel, 2020). Accordingly, MVDA models are easier to develop, do not require systematic constraints or curation of biological parameters, and are computationally efficient in determining key correlations (Rouiller et al., 2013).
Several MVDA algorithms that are routinely used within bioprocess development include Principal Component Analysis (PCA), Partial Least Squares (PLS) Regression, and variations of PLS such as Orthogonal Partial Least Squares (OPLS) (Mevik & Wehrens, 2007; Yamamoto et al., 2009). For instance, Kirdar et al utilized PCA to diagnose similarities and differences between small-scale and large-scale manufacturing batches for biopharmaceutical products (Kirdar, Conner, Baclaski, & Rathore, 2007). Dietmair et al utilized PLS to determine a range of metabolites in the culture medium that correlate with the cell growth rate (Dietmair et al., 2012). In most cases, MVDA models have provided support in upstream bioprocess development, process characterization, platform development, and more recently, media development that drive specific responses (Behrouz et al., 2020; MKS Umetrics AB, 2013; Pybus et al., 2014; Worley & Powers, 2013). Therefore, we hypothesized that amino acid stoichiometric balances (SB) used in combination with MVDA models, could not only highlight key amino acids, but also provide information on cell culture performance characteristics. While SB may not be able to identify specific intracellular pathways or enzyme activation due to amino acid changes, it may however serve as an effective surrogate to extrapolate the intracellular “black box” activities that result in the detected culture outcomes. Furthermore, we postulate that supplementing stoichiometrically desired amino acids that are positively correlated to cell growth or mAb productivity would help improve the desired cell culture performance. The amino acid enrichment ‘cocktails’ can then be designed and established based on SB and MVDA for a defined or training set of culture runs and may be used as feed supplements for future or subsequent culture runs.
In support of our hypothesis, we trained an MVDA model using a 25-batch study in which cells were supplemented with various amounts of a high nutrient feed in aims to identify metabolite stoichiometric balances predictive of cell growth or mAb productivity. Model-derived key stoichiometric balances were translated into amino acid cocktail feeds and supplemented at model-informed time points. In addition to the feed media used in the training set, the model-generated amino acid cocktail feeds were also tested as supplements for two other chemically-defined feeds at lower nutrient concentrations, referred to as moderate and low nutrient feeds. Moderate and low nutrient feeds contained 40% and 17% of the carbon concentration relative to the high nutrient feed level. Experimental validation of our platform and model-directed amino acid supplementation showed about a 20% increase in cumulative viable cells (CVC) and a 30% increase in mAb titer within the supplemented high nutrient condition, about a 30% increase in CVC and 80% increase in titer within the supplemented moderate nutrient condition, and about a 55% increase in CVC and 80% increase in titer within the supplemented low nutrient condition. Results from our study provide evidence towards a novel method to help improve and streamline CD media development for bioprocess.