(468f) Development of a Computational Framework Incorporating Parameter Uncertainty for the Dynamic Simulation of Protein N-Linked Glycosylation to Guide Glycoengineering

1. Background

Protein N-linked glycosylation is a complex post-translational modification occurring in the endoplasmic reticulum (ER) and the Golgi apparatus of eukaryotic cells. This process involves the enzyme-mediated attachment of oligosaccharides, known as glycans, to proteins and plays a vital role in modulating their structure, stability, and function. Therapeutic proteins, such as monoclonal antibodies, undergo N-linked glycosylation, and the resulting glycan heterogeneity is crucial to determining their pharmacological properties. Cell line development plays an important role in the production of therapeutic proteins, with glycoengineering strategies employed to modify the N-linked glycosylation profile for improved therapeutic efficacy and safety. By driving the N-glycan heterogeneity towards desired, non-immunogenic structures, biotherapeutic protein production can be tailored to more targeted and effective treatments. Despite the scientific and industrial interest in glycoengineering strategies within cell line development, challenges persist in large part due to the complexity of N-glycan biosynthesis, which occurs in a non-template driven manner and involves multi-step concerted action of competing glycosidases and glycosyltransferases and co-substrates called nucleotide sugar donors (NSDs). Kinetic models of N-linked glycosylation can be valuable for aiding the model-guided design of glycoengineering strategies by providing insight into N-glycan biosynthetic machinery through first-principles. Although several research efforts have aimed to develop such models and predict the desired relative abundance of N-glycans in Chinese Hamster Ovary (CHO) cell cultures, most N-linked glycosylation models focus only on IgG-based antibodies and do not consider uncertainty in the estimation of key parameters, such as Golgi-resident enzyme concentration levels. Expanding the predictive capabilities of kinetic models of N-linked glycosylation to include the simultaneous prediction of the glycoprofiles of multiple proteins, while accounting for parameter uncertainty, would be an important step forward towards facilitating the model-guided design of glycoengineering strategies for CHO cells.

2. Scope

This work aims to address the aforementioned limitations of existing kinetic N-linked glycosylation models by developing an extensive model parameterization strategy, which enables the integration of prior biological information about model parameters in order to match experimental observations of protein-specific glycoprofiles in CHO cell cultures.

Methods

A previously published mathematical model of Golgi N-glycan processing, developed in our group, has been adapted to run dynamically and, importantly, to include protein-specific N-linked glycosylation pathways not limited to IgG, but also to intracellular host cell proteins (HCPs). This inclusion enables the discernment of cellular glycosylation effects on recombinant protein glycoprofiles, thus facilitating a comprehensive overview of N-glycan biosynthesis in CHO cells. The model requires inputs regarding protein-related information, such as molecular weight, and glycosylation machinery information, including enzymes and NSDs concentrations, and outputs the desired N-linked glycoprofiles as experimentally measured at the end of the culture.

The methodology followed in this work comprises 3 steps. The first step involves extracting pathway information, such as the stoichiometric matrix, from an underlying N-linked glycosylation reaction network generated based on enzyme reaction rules. Next, the pathway information is used to assemble the model equations, which describe the non-linear dynamics of protein-specific N-glycan biosynthesis, resulting in a system of differential and algebraic equations (DAEs) assuming Michaelis-Menten and Sequential Bi-Bi kinetics. Finally, the inference of key model parameters, such as enzyme concentration levels, is achieved via Bayesian inference, particularly by implementing Approximate Bayesian Computation (ABC), which incorporates prior knowledge about enzyme concentrations in the Golgi and updates it by comparing simulated data with observed data. In this case, the latter corresponds to experimentally observed IgG and HCPs glycoprofiles from a given CHO culture. This methodology enables the identification of enzyme concentration ranges that can explain observed cellular and recombinant protein glycoprofiles, thus quantifying parameter uncertainty and providing more insight into the biosynthetic machinery. All computational analysis is carried out using the Python programming language.

Results & Implications

The results of this work indicate a considerable effect of HCPs glycosylation on recombinant IgG glycoprofiles because the N-glycans on both protein components serve as competitive substrates for the same set of glycosidases and glycosyltransferases and draw NSDs from the same intracellular pool. This model-based investigation of this enzymatic competition enables the generation of experimentally-testable hypotheses to assess whether blocking specific HCPs glycosylation pathways can lead to more desirable IgG glycoprofiles. The HCPs N-linked glycosylation network can be used as a proxy for the glycan structures on more complex glycoprotein therapeutics and the model, therefore, considers the full potential complexity of the N-glycan biosynthetic pathways. Additionally, using experimentally measured glycoprofiles from CHO fed-batch cultures, it is demonstrated that enzyme concentration values required to match predicted and experimental glycoprofiles are distributed in an enzyme-specific manner. This inference approach provides more reliable estimates of true Golgi enzyme protein levels, which, in turn, can be used to determine whether alterations in the observed protein N-glycome in CHO cell cultures under different metabolic and physiological states can be attributed to differences in these distributions. These insights facilitate better-informed decision in the targeted manipulation of glycosylation pathways towards desired, non-immunogenic glycoproteins for effective treatments.