1 Computational tools for predicting and controlling the glycosylation of
3 Highlights
4 Glycosylation as a Critical Quality Attribute of biopharmaceuticals.
5 Review of the sixteen mathematical models for protein glycosylation that have been
6 published since 2014.
7 Use of glycosylation models in bioprocess simulation, glycoengineering and glycoprofiling
8 data analysis.
9 Unified strategies for modelling protein glycosylation applied to biopharmaceutical
10 quality control and optimisation.
11 Abstract
12 Glycosylation is a critical quality attribute of biopharmaceuticals because it is a major source of
13 structural variability that influences the in vivo safety and therapeutic efficacy of these products.
14 Manufacturing process conditions are known to influence the monosaccharide composition and
15 relative abundance of the complex carbohydrates bound to therapeutic proteins. Multiple
16 computational tools have been developed to describe these process/product quality
17 relationships in order to control and optimise the glycosylation of biopharmaceuticals. This
18 review will provide a summary highlighting the strengths and weaknesses of each modelling
19 strategy in their application towards cellular glycoengineering or bioprocess design and control.
20 To conclude, potential unified glycosylation modelling approaches for biopharmaceutical quality
21 assurance are proposed.
22 Introduction
23 Twenty of the fifty top-selling pharmaceuticals are glycosylated recombinant proteins that
24 achieved worldwide revenues of over US$90 billion in 2017 [1]. Eighteen of these
25 biopharmaceuticals are monoclonal antibodies (mAbs), which contain two consensus
1 of 20 26 asparagine-linked (N-linked) complex carbohydrates (glycans) on their constant fragment, Fc
27 (Figure 1A). The remaining two blockbuster products, Enbrel® and Eylea®, are heavily
28 glycosylated Fc fusion proteins, which contain up to six N-linked and twenty-six
29 Serine/Threonine-linked (O-linked) glycans [2] (Figure 1B). Many other therapeutic proteins are
30 also glycosylated, with tissue plasminogen activator (tPA), interferon gamma (IFN-γ) and human
31 recombinant erythropoietin (rHuEPO) (Figure 1C) being key examples [3].
32 The glycosylation of therapeutic glycoproteins (TGPs) is highly variable and heavily influences
33 the safety and therapeutic efficacy of these products. Presence or absence of glycans
34 (macroheterogeneity) on TGPs affects their serum half-life in patients [4,5], while the glycosidic
35 linkages and monosaccharide composition (microheterogeneity) are widely reported to impact
36 the safety, pharmacokinetics and pharmacodynamics of TGPs [6]. Microheterogeneity arises from
37 varying degrees of mannosylation, antennarity, core fucosylation, galactosylation and sialylation
38 (Figure 1D through G) [3].
39 All TGPs are produced through large-scale culture of mammalian cells, in particular of Chinese
40 Hamster Ovary (CHO) cells, to ensure compatibility for administration in humans. Importantly,
41 the conditions under which mammalian cells are cultured heavily influence the glycosylation
42 profiles of biopharmaceuticals [3].
43 Herein, we provide an overview of recent glycosylation models in the context of glycoprotein
44 quality assurance for biopharmaceutical manufacturing and discuss their advantages and
45 disadvantages. The review concludes with perspectives for potential unified modelling strategies
46 to control the manufacture of biopharmaceuticals with optimal and consistent glycosylation
47 patterns.
48 Glycosylation as a critical quality attribute of biopharmaceuticals
49 Based on the definition of Critical Quality Attributes (CQAs) within the Quality by Design (QbD)
50 framework [7], industry and regulatory agencies consider glycosylation a CQA of TGPs because it
2 of 20 51 is a property that must be controlled within an appropriate range or distribution to ensure
52 product safety and therapeutic efficacy [3]. The influence glycosylation has on the safety,
53 pharmacokinetics and pharmacodynamics of TGPs is summarised in Table 1.
54 Table 1. Known impacts of glycosylation on TGP quality
Attribute Effect
Aglycosylation Reduced serum half-life [4] (macroheterogeneity) α-1,3 tandem Anaphylaxis (e.g. Erbitux®) [8] galactose (αGal) Immunogenicity [9] Presence of Neu5Gc Reduced serum half-life [9] High-mannose glycans Reduced serum half-life [10] Absence of galactose Reduced serum half-life [11]
Presence of Neu5Ac Increased serum half-life [12] sialic acid residues Enhanced immune modulation [13] Enhanced antibody-dependent cellular cytotoxicity (ADCC) [14] Man5 residues Reduced complement-dependent cellular cytotoxicity (CDC) [14] G0 glycoforms Enhanced CDC [15] Absence of core fucose Significantly increased ADCC [16] Increased ADCC [17] Presence of galactose Increased CDC [18] 55
56 Glycosylation is widely acknowledged as a major source of variability and one of the most difficult
57 to control CQAs because even modest changes in manufacturing process conditions can influence
58 TGP glycan distributions [19]. Despite the available regulatory guidelines for the assessment and
59 control of TGP glycosylation-associated quality [3], substantial variations have been reported
60 across different production lots of marketed products [20]. Glycosylation remains a key challenge
61 for manufacturers and regulators alike and highlights the need for strategies that mechanistically
62 link bioprocess conditions with TGP glycan distributions.
63 Protein glycosylation in mammalian cells
3 of 20 64 Glycosylation is a non-template driven processes which is thought to have evolved to confer
65 glycoconjugates with additional levels of variability and enhanced functional adaptability [21].
66 The mammalian N-linked glycosylation process, which is summarised in Figure 2 [22], begins in
67 the endoplasmic reticulum and concludes in the Golgi apparatus. Throughout the process, the
68 glycans are modified by multiple enzyme-catalysed monosaccharide removal and addition
69 reactions, many of which may occur in parallel. Nineteen enzymes can give rise to over 50,000
70 different glycan structures [23].
71 All monosaccharide addition reactions require nucleotide sugar donors (NSDs) as co-substrates.
72 NSDs are synthesised in the cytosol or nucleus of cells from simple metabolites, such as glucose,
73 galactose and glutamine, and are then transported into Golgi via bespoke transport proteins.
74 Nutrient depletion during culture can impact both the macro and microheterogeneity of TGPs by
75 limiting the intracellular pool of NSDs [24-26]. In order to curb these effects, cell culture media is
76 commonly supplemented with precursors of NSD biosynthesis [27,28].
77 The inherent variability of the glycosylation process is compounded by the fact that multiple
78 manufacturing environment conditions are known to impact TGP glycosylation (recently
79 reviewed in [3]). These can be split into two broad categories based on whether they influence
80 glycosylation machinery (e.g. glycoenzyme and NSD transporter localisation, activity and
81 availability) or are the result of drifts in NSD metabolism.
82 Models for protein glycosylation (2014-Present)
83 The first ever effort to model glycosylation aimed to predict site occupancy (macroheterogeneity)
84 [29]. Efforts to model microheterogeneity began with a reduced reaction network [30], followed
85 by the seminal work of Krambeck and Betenbaugh [31], who formalised the entire network of
86 possible N-linked glycosylation reactions in the Golgi apparatus and associated kinetics. The Golgi
87 apparatus has been simulated as a series of four continuous stirred tank or plug flow reactors
88 (PFRs), or a single PFR, to capture secretory pathway dynamics [32,33].
4 of 20 89 Building on the above work [29-35], sixteen computational studies on the glycosylation process
90 have been performed since 2014 (Table 2). Herein, we review recent efforts in three key areas:
91 (i) modelling of the manufacturing process, (ii) predicting the effect of cellular and metabolic
92 glycoengineering interventions and (iii) analysing TGP glycoprofiling data. Depending on their
93 formulation and solution strategy, models for TGP glycosylation can be categorised as kinetic,
94 flux-based or statistical. Each approach has distinct advantages and disadvantages that make
95 them better suited for specific applications.
96 Briefly, kinetic models describe time-dependent variations within the system, which makes them
97 more appropriate for process modelling and control applications when considering the
98 inherently dynamic nature of cell culture processes. Kinetic models present two key drawbacks:
99 (i) they require a significant amount of information-rich data for parameterisation and (ii) they
100 cannot make de novo predictions of enzyme regulation in response to environmental changes.
101 Flux-based approaches have been developed in the form of either low-parameter Markov chain
102 models [36,37] or parameter-free genome-scale metabolic reconstructions [38,39]. Such models
103 are mainly used for the prediction of genetic intervention outcomes. The key disadvantage of flux-
104 based models is that their solution assumes steady state, which makes them unsuitable for
105 describing the dynamic shifts in TGP glycosylation during cell culture processes. A notable
106 exception is recent work where a parameter representing the residence time of TGPs in Golgi was
107 introduced to capture secretory dynamics within a flux-based framework [39].
108 A different approach that is gaining ground in industry is the use of statistical models. Advantages
109 are that statistical models can be deployed as black-box software packages that require little or
110 no end-user expertise and that, due to their nature, they align closely with the QbD framework
111 [43]. The intrinsic drawback of statistical models is that they are purely data-driven and, thus,
112 cannot provide mechanistic bases for the bioprocess input/output relationships they identify.
5 of 20 113 Table 2. Mathematical models for protein glycosylation (2014-Present)
Study Model type Scope & Novelty Process models Links extracellular conditions to NSD metabolism and recombinant protein antibody glycosylation in Jedrzejewski et al., 2014 [45] Kinetic hybridoma cells. Predicts full process, from cell growth, metabolism and NSD synthesis to antibody glycosylation. Confirmed experimentally at each stage. Analyses interplay between availability of glycosylation machinery and protein secretory capacity. del Val et al., 2016 [46] Kinetic Mechanistic description of how specific productivity of recombinant proteins affects their glycoform distribution. Mechanistic modelling used to link pH, ammonia, galactose, and manganese chloride supplementation with Villiger et al., 2016 [47] Kinetic NSD availability and mAb glycosylation. Includes the effect of pH on glycoform distribution. Multivariate analysis for media design to tailor product quality, incl. glycosylation. Methodology applied to Sokolov et al., 2017 – I [40] Statistical early process development data for a mAb biosimilar. Toolset of multivariate methods to support decision‐making at every stage of process development. Sokolov et al., 2017 – II [41] Statistical Statistical analysis of process data at different scales, then used to predict mAb quality attributes, incl. glycoform distribution. Mechanistic model used to analyse culture data from two different temperatures and identifies underlying Sou et al., 2017 [41] Kinetic intracellular differences. Use of modelling for data interpretation and biological hypothesis generation. Combination of Semi-empirical model that identifies correlations between culture parameters and recombinant protein Aghamohseni et al., 2017 [48] flux-based and glycosylation at physiological temperature and under mild hypothermia. Use of metabolic flux analysis to kinetic reduce burden of model parameterisation. Comparison of statistical methods and mechanistic modelling for guiding process design for perfusion Kinetic and Karst et al., 2017 [44] reactors with the aim of increasing the production of complex mAb glycoforms. Application of modelling statistical for process design in continuous CHO cell systems. Development of glycosylation flux analysis model for predicting intracellular production and consumption Hutter et al., 2017 [39] Flux-based rates of glycoforms. Analysis used to decipher individual effects of enzymatic perturbations versus specific mAb productivity. Includes a parameter that represents dynamic variations in the process. Principal component analysis, partial least square regression and genetic algorithm used to predict Sokolov et al., 2018 [42] Statistical product quality, incl. glycosylation, based on process information. Statistical workflow that reduces data complexity and identifies main correlations between process data and product quality attributes.
6 of 20 114 Table 2 (Cont.). Mathematical models for protein glycosylation (2014-Present)
Study Model type Scope & Novelty Cellular models Glycan flux analysis was used to identify the rate-limiting processing steps at individual glycosites within a Hang et al., 2015 [49] Flux-based model glycoprotein. First implementation of glycan flux analysis; was used to confirm the importance protein structure has on site-specific N-glycan processing. Theoretical framework for calculating the metabolic demand towards host cell protein and lipid del Val et al., 2016 [50] Kinetic glycosylation. Enables the consideration of partitioning of metabolic resources between cellular and recombinant protein glycosylation in CHO. Theoretical framework to reconstruct the reaction network for O-linked glycan biosynthesis. Represents Network up to 98% of human and glycoengineered CHO cell O-glycans. Only known computational framework for O- McDonald et al., 2016 [51] reconstruction linked glycosylation. Uses a pattern-matching algorithm to generate the reaction network. Identifies the key enzymes that drive O-glycan heterogeneity. Used a time-discrete Markov chain to describe glycosylation process and predict the effect of genetic Spahn et al., 2016 [36] Flux-based glycoengineering. Low-parameter approach that does not require kinetic data to generate predictions for genetic interventions. Markov chain flux modelling for cell line selection and cell engineering strategies to achieve glycosylation Spahn et al. 2017 [37] Flux-based biosimilarity of a mAb and rHuEPO. Predicts the cell lines and genetic interventions that are most likely to achieve glycosylation biosimilarity of two TGPs. Genome-scale metabolic reconstruction of pathways relevant to protein glycosylation. Novel Discretized Kremkow and Lee, 2018 [38] Flux-based Reaction Network Modelling using Fuzzy Parameters (DReaM-zyP) that integrates all glycosylation genes with central carbon metabolism to predict which TGP glycoforms are produced after genetic interventions. Data analysis Analysis of mRNA microarray and mass spectrometric glycosylation data using mathematical model of N- Bennun et al., 2013 [52] Kinetic linked glycosylation. Integration of different –omics techniques and data interpretation. Mechanistic model of N-linked glycosylation is used to analyse mass spectrometric data and predict the Krambeck et al., 2017 [23] Kinetic concentrations of glycosylation enzymes and the direct and indirect effects of genetic glycoengineering strategies. 115
7 of 20 116 Process modelling
117 The primary objective of these modelling studies is to calculate the TGP glycoform distribution
118 from process-level data. The first holistic kinetic modelling efforts towards this goal came from
119 Ohadi et al. [35] for CHO cells and Jedrzejewski et al. for hybridoma cells [45]. In addition to
120 integrating the effect of cell culture dynamics and nutrient availability to better describe the
121 glycosylation process, abiotic process variables, such as pH and temperature, have been included
122 in more recent kinetic models.
123 Changes in culture pH have been linked with shifts in glycoenzyme localisation along Golgi [53]
124 as well as reduced glycoenzyme and NSD transporter activity [25,47]. Culture pH was
125 quantitatively linked with mAb glycoform distribution in CHO cells by Villiger et al. [47], who
126 considered two aspects: the effect of pH on (i) specific mAb productivity and on (ii) glycoenzyme
127 activity. Model validity was confirmed across a range of process conditions including the dynamic
128 addition of galactosylation precursors.
129 Temperature shifts to mild hypothermic conditions are also commonly employed in mammalian
130 bioprocessing [24]. Sou et al. used a kinetic model of cell growth, metabolism and glycosylation
131 to identify the root causes of reduced mAb galactosylation under mild hypothermia [54]. The
132 model correctly predicted that, in conjunction, reduced NSD availability and β4GalT
133 activity/concentration caused the observed results.
134 In parallel to mechanistic modelling efforts, Sokolov and co-workers have proposed a set of
135 statistical tools for multivariate analysis that have been applied to tailor various product quality
136 attributes, including glycosylation, in a mAb biosimilar development campaign [40]. The
137 statistical analysis was further coupled with a genetic algorithm and applied to process data at
138 different scales to identify the main process-product correlations, thus significantly reducing
139 problem dimensionality [41,42]. Interestingly, a comparison of statistical response surface and
140 mechanistic modelling of CHO cell perfusion bioreactors showcased the superiority of the latter
141 for process design [44].
8 of 20 142 Cellular & metabolic glycoengineering
143 A key target for glycosylation models has been to identify cellular glycoengineering strategies
144 that achieve target TGP glycoform distributions. Spahn et al. [36] used a flux-based Markov Chain
145 model to successfully predict the glycoform distributions of a mAb, EPO and the CHO secretome
146 that would be produced upon the knockout of α6FucT and GnTIV. This group subsequently used
147 their modelling strategy to guide cell line selection and identify the number and extent of genetic
148 engineering interventions required to achieve glycosylation biosimilarity in mAb and rHuEPO
149 products [37]. del Val and collaborators [46] used a kinetic model to estimate the level of GnTI
150 overexpression required to minimise Man5 mAb glycoform secretion by CHO cells cultured under
151 mild hypothermia. More recently, Kremkow and Lee [38] demonstrated that their Glyco-Mapper
152 model can predict which glycans are produced upon altered expression of glycoenzymes, NSD
153 transporters and NSD biosynthetic enzymes.
154 Model-derived cellular glycoengineering strategies have yet to be quantitatively validated with
155 independent experimentation. In most cases, model outputs for glycoengineering are
156 quantitative (e.g. β4GalT should be overexpressed 2.2-fold to achieve biosimilarity), and it would
157 be ideal for bioprocess development, optimisation and control if they could be implemented in
158 practice. However, regulation of gene expression with the required degree of quantitative
159 precision is, to our knowledge, currently infeasible in mammalian cells. Because of this, the
160 implementation of glycoengineering strategies remains limited to glycogene knockouts.
161 Interestingly, only a statistical response surface model that correlates cell surface and mAb
162 galactosylation has been used to metabolically control TGP glycosylation [28]. The impact of NSD
163 precursor feeding strategies on cell growth and productivity along with uncertainty on how much
164 of the precursors are destined to central carbon metabolism and cellular glycosylation [50] have
165 limited the use of mechanistic models for the design of such strategies. Further model
166 development has the potential of rationally designing NSD precursor feeding strategies that
167 control TGP glycosylation.
9 of 20 168 Data analysis
169 Analysing glycan profiling data is extremely challenging because the distribution of protein-
170 bound carbohydrates results from a non-template driven process where multiple components
171 (e.g., glycoenzymes, NSD transporters, NSD metabolism and secretory pathway dynamics)
172 interact. In this context, mechanistic models have been recently proposed as potential enabling
173 platforms for the integrated analysis of multiple glycosylation-associated ‘omics data sets to
174 provide further insights into the glycosylation process [55].
175 Although developed to provide a kinetic representation, the model by Krambeck et al. [23] was
176 used, in part, for glycoprofiling data analysis. The model builds on previous work [34] and extends
177 the number of possible reactions from ~10,000 to >50,000. Based on mass spectra data, the model
178 identifies which glycoenzymes are present, prunes the reaction network accordingly and then
179 estimates the amount of each glycoenzyme required to quantitatively match the experimental
180 data. In the context of pharmaceutical bioprocess development, this strategy, along with others
181 that link glycogene transcriptomics data [52], has the potential of aiding cell line selection by
182 identifying transfectants that are more likely to contain the glycosylation machinery required to
183 achieve desired TGP glycoforms. It may further prove useful for diagnosing and establishing
184 personalised treatments for glycosylation-associated illnesses.
185 Perspectives on unified modelling approaches for TGP glycosylation
186 Mathematical models for glycosylation have been developed to identify bottlenecks/troubleshoot
187 [41], devise improved operating strategies [44] and, in the future, be used for advanced control
188 and optimisation. Integrated models [41,46-48] which link bioprocess variables that are
189 routinely monitored (i.e. cell density, extracellular nutrient and metabolite concentrations,
190 culture temperature and pH) with outputs of interest (e.g. TGP concentration, glycoform
191 distribution) can thus serve as a mechanistic basis for pharmaceutical bioprocess control and
192 optimisation. However, implementation of such all-encompassing kinetic models will require
10 of 20 193 substantial development, particularly with respect to computational expense and experimental
194 validation.
195 We propose that the different modelling strategies reviewed herein can, in conjunction,
196 contribute to address cellular and process design challenges. Statistical models can be leveraged
197 to identify additional correlations between bioprocess conditions, culture dynamics and TGP
198 glycosylation. Through experimentation, these correlations can be translated into mechanistic
199 descriptions which could ultimately be built into the kinetic models to enhance their predictive
200 capabilities and scope. Flux-based models can be used to rationally constrain the glycosylation
201 reaction network [23] and, thus, reduce the implicit computational burden of kinetic models. In
202 addition, metamodel development, based on e.g. Machine Learning or Random Sampling-High
203 Dimensional Model Representation [56] methods could reduce computational timeframes to be
204 adequate for real-time bioprocess optimisation. Crucially, the above efforts must be underpinned
205 by the development of advanced experimental approaches, such as real-time glycan analysis [57]
206 or inducible glycogene expression [58], that will enable model validation and real-world
207 deployment of control strategies to produce TGPs with optimal and consistent glycosylation.
11 of 20 208 Figures
209
210 Figure 1. Common therapeutic glycoproteins and glycans produced by CHO cells [3]
211 Three common therapeutic glycoproteins, including N- and O-linked glycosylation sites, are
212 schematically represented. (A) is an IgG-based mAb, (B) is an Fc fusion protein (Etanercept®) and
213 (C) is erythropoietin (rHuEPO). The monosaccharide composition and glycosidic bond linkages
214 of oligomannose (D), complex biantennary (E) and complex tetra-antennary (F) N-linked glycans,
215 as well as O-linked glycans (G and H, respectively) are shown. The symbolic representation of
216 each monosaccharide present in the glycans is outlined at the bottom.
12 of 20 217
218 Figure 2. CHO cell N-linked glycosylation [22]
219 In the endoplasmic reticulum (light purple space), oligosaccharyltransferase (OST) transfers the
220 dolichol-bound precursor oligosaccharide, co-translationally, to the nascent polypeptide chain.
221 The three glucose residues ( ) serve as quality control markers for appropriate glycoprotein
222 folding via the calnexin and calreticulin chaperones and are sequentially removed by α-
223 glucosidases I and II (αGlcI and αGlcII, respectively). Absence of all three glucose residues indicate
224 correct 3D structure and enable the protein to be transferred to the Golgi apparatus (light green
225 space). There, mannose residues are removed by α-mannosidases I and II (αManI and αManII),
226 and GlcNAc is added by α-1,3 N-acetylglucosaminyltransferase (GnTI) to produce the A1G0
227 glycan. From this point on, multiple parallel reactions may occur, where α-6 fucosyltransferase
228 (α6FucT) adds core fucose, N-acetylglucosaminyltransferases II, IV and V (GnTII, GnTIV and
229 GnTV) add additional GlcNAc residues and increase glycan branching (antennarity). Beta-1,4
230 galactosyltransferase (β4GalT) and alpha-2,3 sialyltransferase (α3SiaT) may extend the antennae
231 by adding Gal and Neu5Ac residues, respectively. Shorthand notation for commonly-observed
232 glycans is presented below each structure.
13 of 20 233 References and recommended reading
234 Papers of particular interest, published within the period of review, have been highlighted as:
235 • of special interest
236 •• of outstanding interest
237 [1] EvaluatePharma: World Preview 2018, Outlook to 2023. Edited by Malik A, Urquhart L.
238 London: Evaluate Ltd.; 2018:46.
239 [2] Houel S, Hilliard M, Yu YQ, McLoughlin N, Martin SM, Rudd PM, Williams JP, Chen W: N- and
240 O-glycosylation analysis of etanercept using liquid chromatography and quadrupole
241 time-of-flight mass spectrometry equipped with electron-transfer dissociation
242 functionality. Anal Chem 2014, 86:576-584.
243 [3] Zhang P, Woen S, Wang T, Liau B, Zhao S, Chen C, Yang Y, Song Z, Wormald MR, Yu C, et al.:
244 Challenges of glycosylation analysis and control: an integrated approach to producing
245 optimal and consistent therapeutic drugs. Drug Discov Today 2016, 21:740-765.
246 [4] Razaghi A, Owens L, Heimann K: Review of the recombinant human interferon gamma as
247 an immunotherapeutic: Impacts of production platforms and glycosylation. J Biotechnol
248 2016, 240:48-60.
249 [5] Glaspy JA, Jadeja JS, Justice G, Kessler J, Richards D, Schwartzberg L, Tchekmedyian NS,
250 Armstrong S, O'Byrne J, Rossi G, et al.: Darbepoetin alfa given every 1 or 2 weeks alleviates
251 anaemia associated with cancer chemotherapy. Br J Cancer 2002, 87:268-276.
252 [6] Liu L: Antibody glycosylation and its impact on the pharmacokinetics and
253 pharmacodynamics of monoclonal antibodies and Fc-fusion proteins. J Pharm Sci 2015,
254 104:1866-1884.
255 [7] ICH: Q8(R2) Pharmaceutical Development. 2009.
256 [8] Steinke JW, Platts-Mills TA, Commins SP: The alpha-gal story: lessons learned from
257 connecting the dots. J Allergy Clin Immunol 2015, 135:589-596.
258 [9] Ghaderi D, Taylor RE, Padler-Karavani V, Diaz S, Varki A: Implications of the presence of N-
259 glycolylneuraminic acid in recombinant therapeutic glycoproteins. Nat Biotechnol 2010,
14 of 20 260 28:863-867.
261 [10] Goetze AM, Liu YD, Zhang Z, Shah B, Lee E, Bondarenko PV, Flynn GC: High-mannose glycans
262 on the Fc region of therapeutic IgG antibodies increase serum clearance in humans.
263 Glycobiology 2011, 21:949-959.
264 [11] Huang L, Biolsi S, Bales KR, Kuchibhotla U: Impact of variable domain glycosylation on
265 antibody clearance: an LC/MS characterization. Anal Biochem 2006, 349:197-207.
266 [12] Yin B, Gao Y, Chung CY, Yang S, Blake E, Stuczynski MC, Tang J, Kildegaard HF, Andersen MR,
267 Zhang H, et al.: Glycoengineering of Chinese hamster ovary cells for enhanced
268 erythropoietin N-glycan branching and sialylation. Biotechnol Bioeng 2015, 112:2343-
269 2351.
270 [13] Sondermann P, Pincetic A, Maamary J, Lammens K, Ravetch JV: General mechanism for
271 modulating immunoglobulin effector function. Proc Natl Acad Sci U S A 2013, 110:9868-
272 9872.
273 [14] Yu M, Brown D, Reed C, Chung S, Lutman J, Stefanich E, Wong A, Stephan JP, Bayer R:
274 Production, characterization, and pharmacokinetic properties of antibodies with N-
275 linked mannose-5 glycans. MAbs 2012, 4:475-487.
276 [15] Malhotra R, Wormald MR, Rudd PM, Fischer PB, Dwek RA, Sim RB: Glycosylation changes
277 of IgG associated with rheumatoid arthritis can activate complement via the mannose-
278 binding protein. Nat Med 1995, 1:237-243.
279 [16] Shinkawa T, Nakamura K, Yamane N, Shoji-Hosaka E, Kanda Y, Sakurada M, Uchida K,
280 Anazawa H, Satoh M, Yamasaki M, et al.: The absence of fucose but not the presence of
281 galactose or bisecting N-acetylglucosamine of human IgG1 complex-type
282 oligosaccharides shows the critical role of enhancing antibody-dependent cellular
283 cytotoxicity. J Biol Chem 2003, 278:3466-3473.
284 [17] Thomann M, Reckermann K, Reusch D, Prasser J, Tejada ML: Fc-galactosylation modulates
285 antibody-dependent cellular cytotoxicity of therapeutic antibodies. Mol Immunol 2016,
286 73:69-75.
15 of 20 287 [18] Peschke B, Keller CW, Weber P, Quast I, Lunemann JD: Fc-Galactosylation of Human
288 Immunoglobulin Gamma Isotypes Improves C1q Binding and Enhances Complement-
289 Dependent Cytotoxicity. Front Immunol 2017, 8:646.
290 [19] Fournier J: A Review of Glycan Analysis Requirements. Biopharm International 2015,
291 28:32-37.
292 [20] Planinc A, Dejaegher B, Vander Heyden Y, Viaene J, Van Praet S, Rappez F, Van Antwerpen P,
293 Delporte C: Batch-to-batch N-glycosylation study of infliximab, trastuzumab and
294 bevacizumab, and stability study of bevacizumab. Eur J Hosp Pharm Sci Pract 2017,
295 24:286-292.
296 [21] Springer SA, Gagneux P: Glycan Evolution in Response to Collaboration, Conflict, and
297 Constraint. J Biol Chem 2013, 288:6904-6911.
298 [22] Stanley P, Taniguchi N, Aebi M: N-Glycans. In Essentials of Glycobiology. Edited by rd, Varki
299 A, Cummings RD, Esko JD, Stanley P, Hart GW, Aebi M, Darvill AG, Kinoshita T, Packer NH, et
300 al.; 2015:99-111.
301 ••[23] Krambeck FJ, Bennun SV, Andersen MR, Betenbaugh MJ: Model-based analysis of N-
302 glycosylation in Chinese hamster ovary cells. PLoS One 2017, 12:e0175376.
303 A mechanistic model for mammalian N-linked glycosylation that considers over 50,000 possible
304 glycan structures. Based on mass spectra data inputs, it predicts which glycoenzymes are present
305 in the cell, automatically prunes the reaction network and estimates the amount/activity of each
306 glycoenzyme to quantitatively match experimentally-determined glycosylation profiles.
307 [24] Fan Y, del Val IJ, Muller C, Lund AM, Sen JW, Rasmussen SK, Kontoravdi C, Baycin-Hizal D,
308 Betenbaugh MJ, Weilguny D, et al.: A multi-pronged investigation into the effect of glucose
309 starvation and culture duration on fed-batch CHO cell culture. Biotechnol Bioeng 2015,
310 112:2172-2184.
311 [25] Fan Y, del Val IJ, Muller C, Wagtberg Sen J, Rasmussen SK, Kontoravdi C, Weilguny D,
312 Andersen MR: Amino acid and glucose metabolism in fed-batch CHO cell culture affects
313 antibody production and glycosylation. Biotechnol Bioeng 2015, 112:521-535.
16 of 20 314 [26] Villacres C, Tayi VS, Lattova E, Perreault H, Butler M: Low glucose depletes glycan
315 precursors, reduces site occupancy and galactosylation of a monoclonal antibody in
316 CHO cell culture. Biotechnol J 2015, 10:1051-1066.
317 [27] Sha S, Agarabi C, Brorson K, Lee DY, Yoon S: N-Glycosylation Design and Control of
318 Therapeutic Monoclonal Antibodies. Trends Biotechnol 2016, 34:835-846.
319 [28] Grainger RK, James DC: CHO cell line specific prediction and control of recombinant
320 monoclonal antibody N-glycosylation. Biotechnol Bioeng 2013, 110:2970-2983.
321 [29] Shelikoff M, Sinskey AJ, Stephanopoulos G: A modeling framework for the study of protein
322 glycosylation. Biotechnol Bioeng 1996, 50:73-90.
323 [30] Umana P, Bailey JE: A mathematical model of N-linked glycoform biosynthesis.
324 Biotechnol Bioeng 1997, 55:890-908.
325 [31] Krambeck FJ, Betenbaugh MJ: A mathematical model of N-linked glycosylation. Biotechnol
326 Bioeng 2005, 92:711-728.
327 [32] del Val IJ, Nagy JM, Kontoravdi C: A dynamic mathematical model for monoclonal
328 antibody N-linked glycosylation and nucleotide sugar donor transport within a
329 maturing Golgi apparatus. Biotechnol Prog 2011, 27:1730-1743.
330 [33] Hossler P, Mulukutla BC, Hu WS: Systems analysis of N-glycan processing in mammalian
331 cells. PLoS One 2007, 2:e713.
332 [34] Krambeck FJ, Bennun SV, Narang S, Choi S, Yarema KJ, Betenbaugh MJ: A mathematical
333 model to derive N-glycan structures and cellular enzyme activities from mass
334 spectrometric data. Glycobiology 2009, 19:1163-1175.
335 [35] Ohadi K, Aghamohseni H, Gädke J, Moo-Young M, Legge RL, Scharer J, Budman HM: Novel
336 Dynamic Model to Predict the Glycosylation Pattern of Monoclonal Antibodies from
337 Extracellular Cell Culture Conditions. In 12th IFAC Symposium on Computer Applications in
338 Biotechnology January 2014; Mumbai, India,Edited by Venkatesh KV, Modak J, Bullinger E,
339 Lakshminarayanan S: International Federation of Automatic Control (IFAC): 2013:30-35.
340 ••[36] Spahn PN, Hansen AH, Hansen HG, Arnsdorf J, Kildegaard HF, Lewis NE: A Markov chain
17 of 20 341 model for N-linked protein glycosylation--towards a low-parameter tool for model-
342 driven glycoengineering. Metab Eng 2016, 33:52-66.
343 A Markov chain Monte Carlo flux model for glycosylation that predicts shifts in TGP glycosylation
344 that occur upon cellular glycoengineering interventions. As a low-parameter tool, it requires
345 minimum experimental data input to make quantitative predictions.
346 •[37] Spahn PN, Hansen AH, Kol S, Voldborg BG, Lewis NE: Predictive glycoengineering of
347 biosimilars using a Markov chain glycosylation model. Biotechnol J 2017, 12.
348 The same Markov chain model as above has been used for cell line selection and to propose cell
349 glycoengineering strategies that facilitate TGP biosimilarity.
350 ••[38] Kremkow BG, Lee KH: Glyco-Mapper: A Chinese hamster ovary (CHO) genome-specific
351 glycosylation prediction tool. Metab Eng 2018, 47:134-142.
352 A genome-scale reconstruction of all pathways involved in protein glycosylation. It integrates all
353 known glycogenes with cellular metabolism to accurately predict the glycoforms that will be
354 produced upon single or multiple glycoengineering events.
355 •[39] Hutter S, Villiger TK, Bruhlmann D, Stettler M, Broly H, Soos M, Gunawan R: Glycosylation
356 flux analysis reveals dynamic changes of intracellular glycosylation flux distribution in
357 Chinese hamster ovary fed-batch cultures. Metab Eng 2017, 43:9-20.
358 A flux-based model that has been developed to represent dynamic variations in the intracellular
359 production and consumption rates of mAb glycoforms. It was used to evaluate glycoenzyme
360 availability/specific productivity relationships and their influence on mAb glycosylation.
361 [40] Sokolov M, Ritscher J, MacKinnon N, Bielser JM, Bruhlmann D, Rothenhausler D, Thanei G,
362 Soos M, Stettler M, Souquet J, et al.: Robust factor selection in early cell culture process
363 development for the production of a biosimilar monoclonal antibody. Biotechnol Prog
364 2017, 33:181-191.
365 [41] Sokolov M, Ritscher J, MacKinnon N, Souquet J, Broly H, Morbidelli M, Butte A: Enhanced
366 process understanding and multivariate prediction of the relationship between cell
367 culture process and monoclonal antibody quality. Biotechnol Prog 2017, 33:1368-1380.
18 of 20 368 •[42] Sokolov M, Morbidelli M, Butte A, Souquet J, Broly H: Sequential Multivariate Cell Culture
369 Modeling at Multiple Scales Supports Systematic Shaping of a Monoclonal Antibody
370 Toward a Quality Target. Biotechnol J 2018, 13:e1700461.
371 A novel multivariate analysis and statistical modelling workflow that quantitatively links
372 bioprocess conditions with TGP quality. It includes analysis of dynamic variations in TGP quality
373 attributes.
374 [43] Lacki KM: High throughput process development in biomanufacturing. Curr Opin Chem
375 Eng 2014, 6:25-32.
376 [44] Karst DJ, Scibona E, Serra E, Bielser JM, Souquet J, Stettler M, Broly H, Soos M, Morbidelli M,
377 Villiger TK: Modulation and modeling of monoclonal antibody N-linked glycosylation in
378 mammalian cell perfusion reactors. Biotechnol Bioeng 2017, 114:1978-1990.
379 ••[45] Jedrzejewski PM, del Val IJ, Constantinou A, Dell A, Haslam SM, Polizzi KM, Kontoravdi C:
380 Towards controlling the glycoform: a model framework linking extracellular
381 metabolites to antibody glycosylation. Int J Mol Sci 2014, 15:4492-4522.
382 A detailed kinetic model that integrates cell culture dynamics, nucleotide sugar metabolism and
383 the Golgi glycosylation process. It successfully predicts the dynamic profiles of cell growth,
384 extracellular nutrient availability, intracellular NSD concentrations and mAb glycoforms.
385 •[46] del Val IJ, Fan Y, Weilguny D: Dynamics of immature mAb glycoform secretion during
386 CHO cell culture: An integrated modelling framework. Biotechnol J 2016, 11:610-623.
387 A model that integrates cell culture dynamics with the Golgi glycoyslation process. It is the first
388 model to mechanistically describe the influence of secretory pathway dynamics on glycoyslation.
389 ••[47] Villiger TK, Scibona E, Stettler M, Broly H, Morbidelli M, Soos M: Controlling the time
390 evolution of mAb N-linked glycosylation - Part II: Model-based predictions. Biotechnol
391 Prog 2016, 32:1135-1148.
392 An integrated kinetic model that is the only one to date that mechnaistically includesthe effect of
393 pH on glycosylation. It quantitatively predicts the effects of pH shifts on mAb Fc glycosylation.
394 [48] Aghamohseni H, Spearman M, Ohadi K, Braasch K, Moo-Young M, Butler M, Budman HM: A
19 of 20 395 semi-empirical glycosylation model of a camelid monoclonal antibody under
396 hypothermia cell culture conditions. J Ind Microbiol Biotechnol 2017, 44:1005-1020.
397 [49] Hang I, Lin CW, Grant OC, Fleurkens S, Villiger TK, Soos M, Morbidelli M, Woods RJ, Gauss R,
398 Aebi M: Analysis of site-specific N-glycan remodeling in the endoplasmic reticulum and
399 the Golgi. Glycobiology 2015, 25:1335-1349.
400 [50] del Val IJ, Polizzi KM, Kontoravdi C: A theoretical estimate for nucleotide sugar demand
401 towards Chinese Hamster Ovary cellular glycosylation. Sci Rep 2016, 6:28547.
402 [51] McDonald AG, Tipton KF, Davey GP: A Knowledge-Based System for Display and
403 Prediction of O-Glycosylation Network Behaviour in Response to Enzyme Knockouts.
404 PLoS Comput Biol 2016, 12:e1004844.
405 [52] Bennun SV, Yarema KJ, Betenbaugh MJ, Krambeck FJ: Integration of the transcriptome and
406 glycome for identification of glycan cell signatures. PLoS Comput Biol 2013, 9:e1002813.
407 [53] Rivinoja A, Hassinen A, Kokkonen N, Kauppila A, Kellokumpu S: Elevated Golgi pH impairs
408 terminal N-glycosylation by inducing mislocalization of Golgi glycosyltransferases. J Cell
409 Physiol 2009, 220:144-154.
410 [54] Sou SN, Jedrzejewski PM, Lee K, Sellick C, Polizzi KM, Kontoravdi C: Model-based
411 investigation of intracellular processes determining antibody Fc-glycosylation under
412 mild hypothermia. Biotechnol Bioeng 2017, 114:1570-1582.
413 [55] Bennun SV, Hizal DB, Heffner K, Can O, Zhang H, Betenbaugh MJ: Systems Glycobiology:
414 Integrating Glycogenomics, Glycoproteomics, Glycomics, and Other 'Omics Data Sets to
415 Characterize Cellular Glycosylation Processes. J Mol Biol 2016, 428:3337-3352.
416 [56] Zuniga MM, Kucherenko S, Shah N: Metamodelling with independent and dependent
417 inputs. Comput Phys Commun 2013, 184:1570-1580.
418 [57] Tharmalingam T, Wu CH, Callahan S, C TG: A framework for real-time glycosylation
419 monitoring (RT-GM) in mammalian cell culture. Biotechnol Bioeng 2015, 112:1146-1154.
420 [58] Auslander S, Fussenegger M: From gene switches to mammalian designer cells: present
421 and future prospects. Trends Biotechnol 2013, 31:155-168.
20 of 20