Access to this full-text is provided by Wiley.
Content available from Biotechnology and Bioengineering
This content is subject to copyright. Terms and conditions apply.
Biotechnology and Bioengineering
ARTICLE
Flux Sampling Suggests Metabolic Signatures of High
Antibody‐Producing CHO Cells
Kate E. Meeson
1,2
| Joanne Watson
1
| Susan Rosser
3
| Ellie Hawke
2
| Andrew Pitt
2
| Tessa Moses
3
|
Leon Pybus
4
| Magnus Rattray
1
| Alan J. Dickson
2
| Jean‐Marc Schwartz
1
1
Faculty of Biology, Medicine and Health, University of Manchester, Manchester, UK |
2
Manchester Institute of Biotechnology, University of Manchester,
Manchester, UK |
3
EdinOmics, RRID:SCR_021838, Centre for Engineering Biology, School of Biological Sciences, University of Edinburgh, Edinburgh,
UK |
4
FUJIFILM Diosynth Biotechnologies, Billingham, UK
Correspondence: Kate E. Meeson (kate.meeson@manchester.ac.uk) | Jean‐Marc Schwartz (jean-marc.schwartz@manchester.ac.uk)
Received: 17 January 2025 | Revised: 11 March 2025 | Accepted: 23 March 2025
Funding: The authors acknowledge financial support from a Prosperity Partnership grant (EP/V038095/1) funded by EPSRC, BBSRC, and FUJIFILM Diosynth
Biotechnologies. This grant involved the University of Manchester, University of Edinburgh, University of York, and staff at FUJIFILM Diosynth
Biotechnologies.
ABSTRACT
Chinese hamster ovary (CHO) cells remain the industry standard for producing numerous therapeutic proteins, particularly
monoclonal antibodies (mAbs). However, achieving higher recombinant protein titers remains an ongoing challenge and a
fundamental understanding of the cellular mechanism driving improved bioprocess performance remains elusive. To directly
address these challenges and achieve substantial improvements, a more in‐depth understanding of cellular function within a
bioprocess environment may be required. Over the past decade, significant advancements have been made in the building of
genome‐scale metabolic models (GEMs) for CHO cells, bridging the gap between high information content 'omics data and the
ability to perform in silico phenotypic predictions. Here, time‐course transcriptomics has been employed to constrain culture
phase‐specific GEMs, representing the early exponential, late exponential, and stationary/death phases of CHO cell fed‐batch
bioreactor culture. Temporal bioprocess data, including metabolite uptake and secretion rates, as well as growth and pro-
ductivity, has been used to validate flux sampling results. Additionally, high mAb‐producing solutions have been identified and
the metabolic signatures associated with improved mAb production have been hypothesized. Finally, constraint‐based modeling
has been utilized to infer specific amino acids, cysteine, histidine, leucine, isoleucine, asparagine, and serine, which could drive
increased mAb production and guide optimal media and feed formulations.
1 | Introduction
Unlike other rodent species, such as the house mouse (Mus
musculus) or the Norway rat (Rattus norvegicus), the Chinese
Hamster (Cricetulus griseus) is not primarily utilised for in vivo
testing. Instead, its cells, CHO cells, are widely utilized as a
secretory “cell factory”for the production of biopharmaceu-
ticals. CHO cells are responsible for producing the majority of
therapeutic mAbs (Xu et al. 2023), used to treat a range of
illnesses, for example, cancer (Zahavi and Weiner 2020), auto-
immune disorders such as Crohn's disease and rheumatoid
arthritis (Hafeez et al. 2018), and potentially neurogenerative
diseases such as Parkinson's disease (Castonguay et al. 2021;
Pagano et al. 2024). The widespread use of CHO in the bio-
pharmaceutical industry is attributable to several key factors.
Their ability to grow in chemically defined suspension culture
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly
cited.
© 2025 The Author(s). Biotechnology and Bioengineering published by Wiley Periodicals LLC.
Abbreviations: FBA, flux balance analysis; FDBK, FUJIFILM Diosynth Biotechnologies, UK; FDBKA cell line, mAb expressing Apollo X CHO‐DG44 cell line; FDR, false discovery rate; FVA, flux
variability analysis; GEM, genome‐scale model; GPR, gene–protein‐reaction rule; mAb, monoclonal antibody; Qp, specific productivity.
1898 of 1942Biotechnology and Bioengineering, 2025; 122:1898–1913
https://doi.org/10.1002/bit.28982
minimises the risk of contamination, reduces batch‐to‐batch
variability, and facilitates large‐scale production. Additionally,
CHO cells have a strong regulatory track record, are resistant to
infection by human viruses and produce proteins with human‐
like glycosylation patterns. Finally, CHO cells secrete proteins
into the culture medium and optimised bioprocesses enable
high yields of recombinant protein.
Traditionally, bioreactors for clinical mAb production have
capacities of up to 250,000 L. Production typically occurs in fed‐
batch mode over 14 days, with periodic nutrient additions. Some
advanced fed‐batch processes have achieved titers exceeding
10 g/L (Xu et al. 2023). However, integrated and continuous
bioprocesses (ICB) are gaining traction. This approach employs a
continuous perfusion bioreactor coupled to a continuous down-
stream process, offering well‐documented advantages for both
clinical and, in some cases, commercial manufacturing (Pybus
et al. 2024).
Irrespective of the processing format, achieving high titers
remains a persistent challenge in the quest for more cost‐
effective biopharmaceuticals. Metabolic engineering through
media and feed optimization has been a significant strategy to
increase titers since the 1980s (Wurm 2004). However, the
impact of these optimisations can vary widely. Factors such as
the specific cell line (and even clones within a heterogenous cell
population), the stage of bioreactor culture and the nature of
the recombinant protein can influence the cellular response to a
given media formulation.
While cell metabolism has been extensively manipulated to en-
hance CHO bioreactor productivity, the equilibrium of cellular
metabolic demands has primarily been targeted through media
blending and spent media analysis, coupled with one‐factor‐at‐a‐
time (OFAT) and design of experiments (DoE)‐based approaches.
These empirical methods demand extensive resources and lack
insight into the complex and dynamic control of intracellular
metabolism with the bioreactor environment. As metabolism
supplies the “building blocks”for the protein product and the
cellular factory required for its production, a deeper compre-
hension of CHO cell metabolism is crucial to identify metabolic
bottlenecks and rationally guide more sophisticated cell and/or
media engineering strategies (Gopalakrishnan et al. 2024; Kavoni
et al. 2025;Parketal.2024).
To address the evolving demands of biopharmaceutical manu-
facturing, the industry is embracing digital innovation to es-
tablish more streamlined, reliable, and cost‐effective bioprocess
(Park et al. 2024). A deep mechanistic understanding of cellular
behavior and dynamics is critical to creating realistic digital
representations of bioprocesses. The sequencing of the CHO cell
genome was therefore a pivotal moment, unlocking the genetic
parts list which could be linked to cellular phenotypes (Xu
et al. 2011). Additional 'omics analyses (e.g., transcriptomics,
proteomics, and metabolomics) have allowed greater insight
into cellular processes and phenotypes. These data enable the
construction of GEMs, which mathematically bridge the gap
between high‐information content 'omic data sets and pheno-
types. Although maximising the variety and depth of 'omics
datasets to form model constraints would result in a more
personalised model, often transcriptomics datasets form the
initial constraints, due to their coverage and availability. The
size of transcriptomics datasets far exceeds other data types,
offering tens of thousands of unique gene IDs compared to a
few thousand proteins or up to a few hundred targeted
metabolites, depending on the analytical platform. Thus, tran-
scriptomics facilitates the tailoring of many more metabolic
enzymes in the genome‐scale model. However, it is imperative
that any model constrained with transcriptomics datasets be
compared to experimental datasets informing cell behavior, to
ensure the enzyme abundances have translated to phenotypic
behavior in a reliable manner. CHO GEMs serve as digital
twins, providing insight into cellular behavior and metabolic
state within biomanufacturing processes (Park et al. 2021). This
potentially enables systematic optimisation of bioprocesses and
the identification of engineering targets in silico.
GEMs are computational models that can be constrained by var-
ious biological data, such as metabolite concentrations, flux mea-
surements, and gene/protein expression levels, to narrow down
the infinite solution space. Genome annotation and comprehen-
sive literature review are crucial for constructing the stoichio-
metric matrix, which defines the reactants and products involved
in metabolic reactions. Furthermore, gene–protein‐reaction (GPR)
rules which describe the association among genes, proteins, and
reactions are used to further constrain the model. Popular algo-
rithms such as GIMME and iMAT can be employed to incorporate
'omic data sets (Becker and Palsson 2008;Zuretal.2010).
Over the past decade, multiple GEMs have been developed for
CHO cells, allowing constraint‐based workflows and metabolic
predictions to guide bioprocess optimisation. In 2016, the
foundational CHO GEM, iCHO1766, was introduced, accu-
rately predicting CHO growth rates and CHO‐specific amino
acid auxotrophies (Hefzi et al. 2016). This model was con-
structed by reconciling incomplete CHO GEMs and human
metabolic models (Recon1 and Recon2 [Duarte et al. 2007;
Thiele et al. 2013]). Building upon this foundation, three
subsequent GEMs (iCHO2291, iCHO2048, and iCHO2101)
were published (Fouladiha et al. 2021; Gutierrez et al. 2020;
Yeo et al. 2020). The most recent and comprehensive CHO
GEM, iCHO2441, incorporates updated elements and has been
systematically evaluated against all the previous models
(Strain et al. 2023). These GEMs serve as platforms for
constraint‐based modeling, allowing researchers to tailor
models to specific experimental conditions and generate per-
sonalised metabolic predictions.
GEMs have many applications in bioprocess optimization,
including improved media design through in silico simulation
of nutrient additions and supplementation and the identifi-
cation of targets for cell engineering based on predicted phe-
notypes. For example, constraint‐based modeling has been
used to propose optimal media conditions that extend the cell
growth phase (Ramos et al. 2022), to calculate the energetic
costs of protein secretion (Gutierrez et al. 2020) and to predict
amino acid concentrations in culture using machine learning
(Schinn et al. 2021). However, despite their great potential for
the rational design of CHO cell lines and associated cell cul-
ture processes, GEMs continue to face challenges and limita-
tions in terms of model reliability, method development, and
practical application.
1899 of 1942
Flux balance analysis (FBA) is a common method to predict
metabolic fluxes within a GEM, identifying a single optimal steady‐
state solution based on an “objective function”(Orth et al. 2010).
However,toexplorearangeofpossiblesolutions,fluxvariability
analysis (FVA) can be employed, determining the maximal and
minimum flux values for reactions while satisfying given constraints
(Mahadevan and Schilling 2003). The most common objective
function for FBA and FVA is cell growth, especially in cancer
constraint‐based modeling (Yizhak et al. 2015). However, in the
case of CHO cell modeling, the choice of an appropriate objective
function becomes more complicated. Given the distinct growth and
productivity phases of CHO cell culture, it is ambiguous at any
given stage whether the cell is focussed on either solely growth or
protein production, or a combination of both processes. Therefore,
to avoid the oversimplification and errors that optimising an
inappropriate objective function could represent, we adopt an
alternative flux analysis method that does not require an objective
function. In the flux sampling framework, the model is solved in an
unbiased manner with fluxes sampled from an ensemble of steady‐
state solutions. The marginal flux distribution can be summarised
for each reaction as a histogram and can also be investigated for flux
correlations (Herrmann et al. 2019). The coordinate hit‐and‐run
with rounding algorithm has been determined to be the most effi-
cient flux sampling algorithm (Haraldsdóttir et al. 2017;Herrmann
et al. 2019), and previously, flux sampling has been used to evaluate
the accuracy of the CHO GEM, iCHO2441 (Strain et al. 2023)andto
predict uptake and secretion behavior across distinct culture phases
(Gopalakrishnan et al. 2024). However, due to the computational
intensiveness of flux sampling, it is used much less often than FBA
and FVA, highlighting the need for new workflows employing the
methodology. Furthermore, due to its unbiased nature, flux sam-
pling represents the ideal analysis method for studying the meta-
bolic plasticity of CHO cells.
Utilising the iCHO2441 model as a framework (Strain et al. 2023),
we aimed to constrain this CHO GEM with transcriptomics data
to generate three separate culture phase‐specific models. By
comparing these models, we sought to identify metabolic features
associated with improved mAb production and identify potential
targets for directed media and feed optimisation. Our hypothesis
was that flux sampling of constraint‐based models could highlight
metabolic features associated with high mAb production, and that
thesefluxpredictionswouldberepresentativeofexperimental
bioprocess data. The practical objective here was to develop a flux
sampling workflow which could identify high mAb producing
solutions and to use statistical methods to uncover significant
metabolic signatures of high mAb producing cells and characterise
distinct phases of CHO cell bioreactor culture.
2 | Materials and Methods
2.1 | Experimental Materials and Methods
2.1.1 | Cell Line Details and Growth Conditions
The cell line used for experimental work was a mAb expressing
Apollo X CHO‐DG44 cell line (FUJIFILM Diosynth Bio-
technologies), which will be referred to as “FDBKA”. Fed‐batch
cultures of FDBKA were performed in a 2 L BIOSTAT single‐
use bioreactor (Sartorius) with an initial working volume of
1.2 L JM05B media (FUJIFILM Irvine Scientific, catalogue
number 981486) at a starting density of 0.5 × 10
6
cells/mL. All
the cell cultures were performed in triplicate. The pH of culture
was maintained at 7.0 ± 0.05 using carbon‐dioxide gas injection
and sodium carbonate base addition. Temperature was set at
37°C and controlled using a bioreactor jacket. Dissolved oxygen
was maintained at 40% of air saturation. The feeding regime
consisted of daily additions of L‐glutamine (Merck, catalogue
number 1002861000), HyClone Cell Boost 7a Supplement
(Cytiva, catalogue number SH31026.08), and HyClone Cell
Boost 7b Supplement (Cytiva, catalogue number SH31027.09).
Glucose concentrations were maintained above a threshold
value using a 50% (w/v) concentrated solution (SAFC, catalogue
number CR40138‐20B). L‐glutamine was also maintained above
a threshold value using a 200 mM solution.
2.1.2 | Bioprocess Measurements
For metabolite analysis of the cell culture, daily measurements
were taken informing lactate and ammonia production
(Bioprofile Flex 2; Nova Biomedical). Growth rate was inter-
preted from viable cell count (ViCell XR; Beckman Coulter),
and product concentration was measured using a Protein A
UPLC‐based method.
2.1.3 | Transcriptomics Generation and Analysis
Samples were taken for the transcriptomics data set between days 4
and 14 of cell culture (days 4, 5, 6, 7, 8, 11, 12, and 14); there were
two technical repeats per sample across three separate bioreactors,
(representing biological repeats). FDBKA cells were grown in
JM05B media conditions, provided with high glutamine concen-
trations. The samples went through automated preparation using
TruSeq stranded mRNA‐seq library and sequenced using NovaSeq.
100PE. Reads were trimmed, and aligned to the C. griseus genome
(222‐107), which was assembled at the Edinburgh Genomics
Facility, using STAR version 2.7.3a (Dobin et al. 2013)specifying
paired‐end reads and the option—outSAMtype BAM Unsorted
(all the other parameters were left at default). The quasi‐alignment‐
based tool Salmon was used for transcriptome quantification (Patro
et al. 2017), and the raw counts table was filtered to remove genes
consisting of near‐zero counts, filtering on counts per million. The
transcript level abundances were summed into gene level abun-
dances using the bioconductor package tximport (Soneson
et al. 2015). Reads were normalised using the weighted trimmed
mean of M‐values method (Robinson and Oshlack 2010).
The distribution of the normalized transcriptomics data set was
visualised using principal components analysis (PCA). DESeq2
differential gene expression analysis was performed (in R)
between transcriptomics samples from different points in the
time series (Love et al. 2014), including between concurrent
time points, for example, day 4 to day 6, or between distinct
culture phases, for example, day 4 to day 12 (Table 1). Differ-
ential expression was defined as an adjusted p‐value (false
discovery rate; FDR) of less than 0.05 and a minimum fold
change of 2. Heatmap visualisation (using Z‐scores) was applied
to subset of differentially expressed genes, which were selected
1900 of 1942 Biotechnology and Bioengineering, 2025
according to differential expression between concurrent time
points, regardless of direction, that is, day 4 to day 6, as these
were two neighbouring measurements, but this subset did not
include those genes differentially expressed between day 4
and day 12. Genes included in the heatmap were carried for-
ward to functional enrichment analysis via the biomaRt R
package, specifying the “KEGG_2021_Mouse”reference
genome and the “GO_Biological_Process_2023”ontology
(Durinck et al. 2005,2009).
2.2 | Computational Materials and Methods
Code for constraint‐based analysis methods detailed in this
project is available at the corresponding GitHub: https://github.
com/katemeeson/CHO_cell_modelling_2024/tree/main. A com-
bination of R and Python were used across this project.
2.2.1 | Adapting the iCHO2441 Genome‐Scale Model
The iCHO2441 GEM (Strain et al. 2023) was downloaded from
BioModels and the reactions relating to mAb production were
updated to represent the product profile of the FDBKA cell line.
Peptide sequences for heavy and light chains were converted
into the reaction definition and used to estimate ATP and GTP
hydrolysis. In the adapted version of iCHO2441 used here, the
reaction entitled “ICproduct_TRANSLATION_protein”relates
to product formation. The updated iCHO2441 was checked for
consistency and errors using the “validate_sbml_model”func-
tion in COBRA (Ebrahim et al. 2013), and in silico media con-
ditions were left at their default values.
During the curation of iCHO2441 (Strain et al. 2023), metadata
describing metabolic subsystems and compartments was taken
from iCHOv1 (Hefzi et al. 2016) and mapped to the new GEM;
this has facilitated a subsystem enrichment analysis, which has
been demonstrated in Section Results.
2.2.2 | Integrating Transcriptomics Into GEM
Before integration as model constraints, the transcriptomics
gene IDs had to be mapped to the model IDs. The iCHO2441
GEM uses NCBI gene IDs, and these were mapped to Ensembl
Gene IDs using a combination of the BioMart web‐based tool
(Durinck et al. 2005,2009) and flat files from NCBI, Ensembl,
and Uniprot. The iCHO2441 GEM also uses slightly altered
BIGG IDs for metabolites, therefore these have been adapted
using mappings against KEGG, CHEBI, and other flat files from
the BIGG website (Hastings et al. 2016; Kanehisa et al. 2022;
Kanehisa 2000)(http://bigg.ucsd.edu/).
Before transcriptomics was integrated into the GEM, visualisation
indicated that day 5 could be a batch effect, therefore it was re-
moved from further analysis (Figure S1). Gene expression values
were averaged across the days to be grouped to form the three
distinct time phase models (early exponential: day 4; late
exponential: day 6, 7, and 8; stationary/death: day 11, 12, and 14)
and these were translated to reaction constraints using a previously
published algorithm (Timouma et al. 2024). This 'omics integration
algorithm enforces gene expression values as constraints for the
lower and upper bounds of reactions within a GEM, and integration
is directed by reaction reversibility and the GPRs in the GEM
(Timouma et al. 2024). Reactions with genes not included in the
transcriptomics were left unconstrained—an approach employed by
other, existing algorithms, such as PROM (Chandrasekaran and
Price 2010)—because the removal of all these reactions would make
the model infeasible to solve.
2.2.3 | Flux Sampling and Isolating the High mAb‐
Producing Solutions
For all the methods, the Gurobi optimiser was used with an
academic license (Gurobi Optimization LLC. 2023). Using the
cobra sample function (Ebrahim et al. 2013), each of the time
phase models were sampled 50,000,000 times and solutions
were stored every 10,000 iterations, resulting in 5000 data points
per reaction, per model. This was the same approach taken by
Strain et al. (2023), and helps prevent oversampling of the same
area of solution space (Herrmann et al. 2019). The coordinate
hit‐and‐run with rounding algorithm was used for flux sam-
pling (Haraldsdóttir et al. 2017).
To identify reactions associated with high mAb‐production,
solutions were selected based on a 95th percentile threshold of
mAb production; this meant the metabolic profile of 250 high
mAb‐producing solutions were compared against a pool of 5000
solutions, per culture phase‐specific model. The Mann–Whitney
Utest was implemented using the mannwhitneyu function from
the scipy.stats library (Virtanen et al. 2020). The null hypothesis
was that the flux distribution would not change between the
high mAb‐producing solutions and all other solutions. p‐values
were adjusted using the fdrcorrection function from the scipy.-
stats library (Virtanen et al. 2020). Reactions with a false dis-
cover rate (FDR) of less than 0.05 were considered to have
skewed flux in the high mAb‐producing solutions.
3 | Results
For this project, the iCHO2441 model (Strain et al. 2023) was
adapted to represent the specific mAb produced by an
TABLE 1 | Contrasts used for DESeq. 2 analysis. Each differential
expression analysis is between two individual days, as specified by
Contrast A and Contrast B.
Variable Contrast A Contrast B
Day Day 4 Day 6
Day Day 4 Day 8
Day Day 4 Day 12
Day Day 6 Day 7
Day Day 7 Day 8
Day Day 8 Day 11
Day Day 11 Day 12
Day Day 12 Day 14
1901 of 1942
industrially relevant FDBKA CHO cell line, providing a scaffold
for transcriptomic data integration (Figure 1). The mAb
translation reaction and product composition (“ICproduct_-
TRANSLATION_protein”) required all the 20 amino acids,
ATP, GTP, and H
2
O to produce the desired protein product, and
produced, ADP, AMP, GDP, protons, inorganic phosphate, and
pyrophosphate. Separately, the reaction corresponding to pro-
duction of the mature mAb product (ID “ICproduct_Final_de-
mand”) was primarily used to study protein production.
Following model adaptation, flux sampling was performed on
three transcriptomics‐constrained models, each representing a
distinct culture phase of CHO cells (Figure 1). These models
were validated using bioprocess data and used to predict the
metabolic features of high mAb‐producing CHO cells
(Figure 1).
3.1 | Transcriptomics Analysis Predicts Features
of Distinct CHO Cell Culture Phases
Before its integration into the adapted iCHO2441 model, the
quality and depth of the transcriptomics data was analysed. After
mapping to the genome and trimming, there were 20.6–34.4 M
read pairs which remained, representing 53.3–61.4% of the initial
FIGURE 1 | Workflow for obtaining biological conclusions from flux sampling results. Time series transcriptomics has been integrated into an
updated version of iCHO2441 (Strain et al. 2023), using a recently designed universal integration algorithm (Timouma et al. 2024). Flux sampling has
been applied to three culture phase‐specific constraint‐based models and metabolic predictions have been validated against bioprocess data from
multiple bioreactors. High mAb‐producing solutions have been extracted from n= 5000 flux sampling solutions, and reactions exhibiting a change of
flux which was statistically significantly associated with high production have been used to predict metabolic signatures. Qp: cell‐specific
productivity.
1902 of 1942 Biotechnology and Bioengineering, 2025
total read pairs—with a total of 22,220 unique genes detected
after filtering.
The time series transcriptomics samples demonstrated a linear
trend over time, with the most variation having been observed
between samples taken at day 4 and stationary/death phase
time points (days 11, 12, and 14) (Figure 2A). Owing to the
divergence in gene expression observed over time, the most
differentially expressed genes were identified between days 4
and 12, where there were 2417 genes upregulated and 1306
downregulated at day 12 compared to day 4, respectively
(Figure 2B)—representing approximately 16.8% of all unique
genes detected. Concerning the difference in gene expression
between early exponential and late exponential, there were 612
genes upregulated and 129 downregulated at day 8 compared
to day 4 (Figure 2B). Furthermore, as the CHO cells moved into
the stationary/death phase, there were 623 genes upregulated
and 457 downregulated at day 11 compared to day 8
(Figure 2B). In summary, differential gene analysis indicated
that the greatest change in gene expression occurred at the
transition from the early exponential to the late exponential
phase, but there were differentially expressed genes at all time
points.
The next task was to interpret the differential gene expression
between CHO cell culture phases, to show that changes in gene
expression translated to a meaningful culture phenotype,
therefore justifying the integration of transcriptomics into the
iCHO2441 GEM for metabolic analysis. Genes showing a peak
expression at day 4 were enriched for cell cycle, ribosomal, and
nucleic acid processes (Figure 2C), which could be interpreted
as cell growth. Similarly, genes showing a peak expression
at days 6, 7, and 8 were enriched with DNA replication and
mismatch repairs processes (Figure 2C). In contrast, genes with
FIGURE 2 | Gene expression represents distinct culture phases. A. PCA of time series transcriptomics. Individual data points for the same day
represent two technical replicates across three biological replicates (separate bioreactors) B. Heatmap for the expression of genes found to be differentially
expressed at concurrent time points, regardless of direction, according to DESeq. 2 analysis (FDR< 0.05 and a minimum fold‐change of 2). Days have
been grouped into early exponential (red), late exponential (green) and stationary/death phase (purple), according to gene expression. Z‐score shown on
heatmap. Genes ordered according to time phase in which gene expression peaked. C. Functional enrichment of differentially expressed genes from B,
with GO Biological Process annotated on x‐axis. Size of bubble corresponds to number of differentially expressed genes from this GO term. Notebooks:
DeSeq. 2 analysis, and enrichment analysis on DeSeq. 2_analysis.R; PCA performed, heatmap generated, and enrichment results visualized on tran-
scriptomics_figs.ipynb.
1903 of 1942
peak expression at the stationary/death phase were enriched for
cell death and cellular stress pathways, including autophagy,
mitophagy, and lysosome pathways (Figure 2C).
3.2 | Comparing Culture Phase‐Specific Model
Predictions With Bioprocess Data
After integration into the adapted iCHO2441 model, it was neces-
sary to experimentally validate the key metabolic predictions of the
three culture phase‐specific models (Figure 3). Bioprocess mea-
surements included the following: growth, specific productivity
(Qp), and lactate and ammonia production. Bioprocess measure-
ments were compared to model predictions, following flux sampling
(Figure 3).
It was important for modeling to accurately predict CHO cell
growth rate, which experimentally, was shown to decrease over
time from the early and late exponential phases to negative growth
rates as cells move into the stationary/death phase (Figure 3A).
Constraint‐based modeling was able to recapitulate this behavior,
where the median growth rate decreased by approximately 40%
between early and late exponential, and by 50% between the late
exponential and stationary/death phases (Figure 3B).
Models should accurately represent the mAb production behavior
of CHO cells. Compared to growth rate, which experimentally
showed a decrease over time, the specific productivity was found to
be constant, as evidenced by overlapping standard deviation
(Figure 3C). Typically, CHO cells show an increased productivity
with decreased growth rate as they shift through the culture phases,
and cell cycle arrest at the later culture phases is one approach to
enhancing specific productivity of CHO cells (Kumar et al. 2007).
However, an increase in specific productivity at the late exponential
and stationary/death phases was not observed in the FDBKA cell
line, indicating a cell line‐specific phenotype (Figure 3C). Modeling
predicted a consistent productivity over time, where although there
was a decrease in the upper ranges of flux sampling solutions across
the culture phase models (Figure 3D), the standard deviation of
mAb production overlapped between all culture phase models.
CHO DG44 cells switch from lactate production to consump-
tion over time regardless of media composition (Reinhart
et al. 2015), and this has been represented by the bioprocess
data, where a negative reaction flux indicated consumption
from day 7 onwards (Figure 3e). When constrained with
transcriptomics data, the culture phase‐specific GEMs were
able to capture lactate production at the early exponential
phase, and although not able to predict consumption at the
later culture phases, they did predict a shift towards negative
values between the early and late exponential culture phases,
represented by the quartiles of the flux sampling solutions
(Figure 3f). Furthermore, the models also captured the
increase to more positive values, which was observed experi-
mentally between the late exponential and stationary/death
culture phases (Figure 3f). The fact that models were not able
to capture the lactate consumption phenotype is suspected to
be because the late exponential phase (days 6, 7, and 8) rep-
resents a transition state for the cells, where at day 6 cells are
producing lactate and day 8 consuming lactate (Figure 3e),
which is hard to capture in silico, where gene expression data
used to constrain the model was averaged across the 3 days
(Figure 3f).
Experimentally, the pattern of ammonia production over time
follows a similar trend to lactate production, where there is a
decrease in production over time (Figure 3g). Ammonia pro-
duction in CHO DG44 is mainly due to the deamination of
glutamine consumed from the feed source, and glutaminolysis
occurs at an increased rate at the early exponential phase (Ahn
and Antoniewicz 2011; Coulet et al. 2022). Although the models
have not been sensitive to this decrease in ammonia production
over time, they have been able to capture the increased varia-
bility in ammonia production behavior at the late exponential
time phase, which is evidenced by a large standard deviation
at day 6 (Figure 3g) and range of flux sampling solutions which
was the greatest for the late exponential model (0.2 mmol/
gDW/h compared to 0.11) (Figure 3h).
3.3 | Differential Regulation of Metabolic
Subsystems at Each Culture Phase
To begin with, the relationship between growth (reaction ID:
“biomass_cho_prod”) and productivity (reaction ID: “ICpro-
duct_Final_demand”) was explored across 5000 flux sampling
solutions. In agreement with the bioprocess data there was no
observable change in mAb production rate solutions as the
corresponding growth rate solution increased, per culture phase
(Figure S2).
The next step was to identify and isolate the highest mAb‐
producing solutions, representing the 95th percentile. Metabolic
flux predictions were compared from the 95th percentile to the
general pool of 5000 solutions, using the Mann–Whitney Utest
(Figure 4A) to identify those reactions that change (with either
increased or decreased flux) with increased mAb production.
When the metabolic profiles (6336 reactions across 250 high
mAb‐producing solutions) were visualised, it could be observed
that the landscape of reaction fluxes was specific to the culture
phase metabolic model, with distinct separation between the
early exponential, late exponential and stationary/death phases
(Figure 4B). Early and late exponential high mAb‐producing
solutions show separation from the stationary/death phase
along PC1 (37.8% of explained variance), while all the culture
phases can be separated using PC2 (20.5% of explained vari-
ance) (Figure 4B).
Using results from the Mann–Whitney Utest, it was possible
to determine the number of reactions with a change in flux
significantly associated with high mAb‐production, across the
different culture phases and flux sampling solutions
(Figure 4C). Results indicated that there were many reactions
with a change in flux which was associated with high mAb
production, some of which were unique to a specific culture
phase, and others which were commonly associated across all
three culture phases. There were the highest number of
reactions uniquely associated with high mAb‐production at
the stationary/death culture phase (841 reactions), suggesting
either the upstream regulation or the downstream effect of
high mAb production has the most unique profile later on in
CHO cell culture (Figure 4C). In comparison, there were a
1904 of 1942 Biotechnology and Bioengineering, 2025
FIGURE 3 | Validating metabolic predictions using bioprocess data from industrial bioreactors. It is possible to convert between pg/cell/h
(bioprocess units) and mmol/gDW/h (constraint‐based model units) using reported dry cell weights, which vary per CHO cell line and condition (Széliová
et al. 2020), and the molecular weight of the metabolite. The following formula should be used for this conversion: ∗
∗
(
)
γ(1000/
)
α
β
10−12 ,whereαis rate
of bioreactor process (pg/cell/h), βis dry cell weight (g) and γis M
r
of metabolite (g/mol). A. Growth rate measured from bioreactors at FDBK (hour
−1
).
Error bars correspond to one standard deviation from the mean of six biological replicates (individual bioreactors). Culture phase as determined by gene
expression (either early exponential, late exponential or stationary/death) has beenoverlayed onto the graphs for ease of comparison to model predictions.
Negative values indicate cell number is decreasing over time, i.e. through cell death. B. Constraint‐based modeling predictions of cellular growth rate
(hour
−1
). C. Specific productivity (Qp) measurements from bioreactors (pg/cell/day). Six biological replicates. D. Constraint‐based modeling predictions of
the mAb production rate (mmol/gDW/h). E. Rate of lactate production (pg/cell/h). Positive values correspond to metabolite production and negative to
consumption. F. Constraint‐based modeling predictions of the rate of lactate production (mmol/gDW/h), taken from model.summary() in COBRApy.
Median, 25th and 75th percentile have been indicated. G. Rate of ammonia production as measure from bioreactors (pg/cell/h). Six biological replicates.
H. Constraint‐based modeling predictions of the rate of ammonia production (mmol/gDW/h).
1905 of 1942
similar number of reactions uniquely associated with high
mAb production at the early and late exponential culture
phases: 643 and 678 reactions, respectively (Figure 4C). There
were 132 reactions associated with high mAb production that
were common to all culture phases (Figure 4C). The high
frequency of reactions predicted by models as being associated
with high mAb production indicated that no single metabolic
reaction or enzyme was driving improved production, and
supported the high‐throughput, 'omics‐led approach of meta-
bolic modeling.
Of those 132 reactions which were associated with high mAb
production across all culture phases, there were 68 reactions
with metabolic subsystem annotations in iCHO2441. Of these
68 reactions, there were 11 unique metabolic subsystems rep-
resented, with the most prominent being “Transport reactions,”
“Exchange/demand/sink reactions,”“Fatty acid metabolism,”
and “Nucleotide metabolism”(Figure 5A). Of those metabolite
transport reactions which were associated with high mAb
production, one‐third involved the transport of amino acids,
therefore highlighting amino acid transport as a principal sub-
system associated with mAb production. Other metabolites
which were described by these “Transport reactions”; included
fatty acids, molecules involved in redox signalling (namely
thioredoxin and glutathione), nucleotides and hormones, such
as L‐thyroxine.
Once the main metabolic subsystems had been identified, the
direction of regulation of flux (i.e., upregulation or down-
regulation) associated with high mAb production was explored.
Once the count of reactions per subsystem associated with high
mAb production was normalised as a portion of the total
metabolic subsystem, “Transport reactions,”“Bile, eicosanoid
and steroid metabolism,”“Fatty acid metabolism,”and “Ex-
change/demand/sink reaction”were the top subsystems asso-
ciated with high mAb production (Figure 5B). Overall, models
predicted that the early and late exponential culture phases
have contrasting metabolic flux profiles of high mAb produc-
tion. Most subsystems had the highest proportion of upregu-
lated reactions at the early exponential culture phase, while the
late exponential culture phase models predicted that cells more
often downregulate reactions to favor high mAb production
(Figure 5B). Considering the regulation of specific subsystems,
there was almost twice as high a proportion of reactions from
the “Transport reaction”subsystem which were downregulated
in the high mAb solutions for the late exponential and sta-
tionary/death culture phases as there were upregulated
(Figure 5B). This indicated that the downregulation of some
FIGURE 4 | Isolating and studying high antibody‐producing flux sampling solutions. A. Workflow for isolating the “high mAb‐producing”
solutions from flux sampling and determining whether the flux of individual reactions is skewed in these solutions, and therefore associated with
high mAb production. B. PCA of the flux through total model reactions of the high‐producing solutions (top 95th percentile, as illustrated in A), for
each time phase model. C. Upset plot to show the number of reactions associated with high mAb production for each culture phase model and where
these overlap between separate culture phases. Associated reactions have been determined using a Mann–Whitney Utest (FDR < 0.05).
1906 of 1942 Biotechnology and Bioengineering, 2025
metabolite transport was associated with high mAb production
at the later culture phases, where we would expect antibody to
be being synthesised. The downregulation of specific metabolic
reactions could explain an increase in mAb production by
suggesting there is a re‐focus in the pathways through which
energy is being transferred. Aside from transport reactions,
models predicted the greatest proportion of reactions down-
regulated in high mAb‐producing solutions for the “Bile, eico-
sanoid, and steroid metabolism”subsystem, where roughly 30%
of the entire subsystem showed downregulation with high mAb
production at the late exponential culture phase (Figure 5B).
Furthermore, between 15% and 20% of all fatty acid metabolism
reactions were downregulated with high mAb production at the
late exponential and stationary/death culture phases, while less
than 10% of these reactions were downregulated at the early
exponential culture phase, suggesting downregulation of fatty
acid metabolism is associated with high mAb production more
so at the later time points (Figure 5B). Other observations
included the fact that “Vitamin and cofactor metabolism”
showed variable levels of downregulation with high mAb pro-
duction for each culture phase, and there was a roughly three
times as great proportion of “Glycerophospholipid, sphingoli-
pid, and inositol”reactions upregulated at the early exponential
compared to both other culture phases (Figure 5B). Overall, the
regulation of reactions associated with high mAb production
showed differences across each culture phase, with specific
culture phase‐specific signatures emerging.
3.4 | Amino Acid Transport Changes Associated
With High mAb Production Have Culture Phase‐
Specific Signatures
The next step was to use model predictions to decipher a culture
phase‐dependent amino acid transport signature with the aim of
influencing medium optimisation strategies. There were over 600
reactions at the early and late exponential and over 800 reactions at
the stationary/death culture phases which were altered in high
FIGURE 5 | Metabolic subsystems associated with increased mAb production rates. A. The metabolic subsystem annotations of those reactions
found to be associated with high mAb production, and which are shared across all culture phases (n= 132 reactions from Figure 4C) B. The
proportion of total reactions in a subsystem which have had a change of flux associated with increased mAb production flux (decreased: left‐hand
side, increased: right‐hand side). Data from flux sampling (n= 5000) solutions. Significance was determined using Mann–Whitney Utest
(FDR < 0.05). Colour of point refers to time phase model, as indicated in key. Notebook: Subsystem dot plot visualized on FluxSampling_highIgG_
reactions_subsystemsVis.R.
1907 of 1942
mAb solutions (Figure 4C). The most dysregulated subsystem in the
high mAb solutions was “Transport reactions”and as mentioned
earlier, the type of molecule occupying most of these reactions
related to amino acid transport (i.e., one‐third of transport reactions
associated with high mAb production involved amino acids).
Therefore, those reactions involving amino acid transport, and
which were highly correlated with the principal component features
(from Figure 4B), were used to propose culture phase‐specific
amino acid transport signatures of high mAb production (Figure 6).
In total, the top 50 loadings for both PC1 and PC2 were com-
piled for this analysis. Of these 100 total loadings, there were 65
unique reactions which were significantly associated with high
mAb production. There were 40 amino acid transport reactions
contained within this subset, and these described 15 different
amino acids across the three culture phase models. These 40
amino acid transport reactions were studied to develop amino
acid transport signatures associated with high mAb production,
and which showed a unique pattern of regulation across the
culture phases (Figure 6). For the purpose of directing amino
acid supplementation for CHO cell culture media, those amino
acids which models predicted overall increased import in
association with high mAb production will be focussed on here.
Using the principal component loadings and results from the
Mann–Whitney Utest, of the original 40‐reaction subset
described previously, there were 22 amino acid transport
reactions which were associated with high mAb production at
the early exponential time phase (Figure 6A). There were 10
different amino acids described by this subset of transport
reactions, with the highest frequency of reactions—four
reactions—describing cysteine transport (Figure 6A). When
the import and export predictions were considered together,
models suggested that overall, the increased import of cysteine
is favored at the early exponential phase by high mAb‐
producing cells. This conclusion was reached because there was
increased import of cysteine via the APC transporter (reaction
ID: AAPAOC69) coupled with decreased export via the L‐
cysteine/L‐glutamine reversible exchanger (reaction ID: CY-
SGLUexR) and cotransport with sodium and counter transport
of a proton (reaction ID: CYSSNAT5tc) in the high mAb‐
producing solutions (Figure 6A). In addition, the increased
import of histidine was associated with high mAb production at
the early exponential phase, involving sodium symport
(reaction ID: HISt4) (Figure 6A). Another observation at the
early exponential phase was that there could be overall less
isoleucine transported into the cell, owing to decreased import
(reaction ID: ILEt4) and increased export (reaction ID: AA-
PAOC11), coupled to the increased import of its structural
isomer, leucine (reaction ID: AAPAOC6) (Figure 6A).
Moving onto the late exponential culture phase, in parallel to
the early exponential model, results predicted that there was
increased import of cysteine in high mAb‐producing solutions,
through the CYSATB0tc reaction (Figure 6B). Another simi-
larity between the early and late exponential culture phases was
that there was decreased import of isoleucine into the cell
(reaction ID: AAPAOC11) coupled with increased import of
leucine (reaction ID: AAPAOC6) (Figure 6B). The late ex-
ponential phase model also suggested that the export of aspar-
agine from the cell could be decreased in association with
improved mAb production, exemplified by decreased flux in the
high mAb‐producing solutions across two exporting reactions
(reaction IDs: ASN_Lte and AAPAOC8) (Figure 6B). Decreased
export of asparagine from the cell favors increased intracellular
concentrations of this amino acid in high mAb‐producing cells.
Finally, the stationary/death phase was studied for amino acid
signatures associated with high mAb production (Figure 6C).
On top of the decreased export of asparagine at the late ex-
ponential phase, this behavior was observed at the stationary/
death phase, where there was increased import of asparagine
associated with high mAb production (reaction ID: AA-
PAOC42) (Figure 6C). Another amino acid which was associ-
ated with high mAb production at the stationary/death culture
phase, but which was unique to this culture phase, was serine,
where there was increased import via two separate reactions
(reaction IDs: AAPAOC39 and AAPAOC42) (Figure 6C).
In summary, “Transport”reactions was defined as a key subsystem
associated with high mAb production following enrichment anal-
ysis and modeling has suggested multiple reactions describing six
different amino acids: cysteine, histidine, leucine, isoleucine,
asparagine, and serine which may show a culture phase‐dependent
signature and could be explored by media optimization experiments
(Figure 6D). Of these six amino acids, four are essential to CHO
DG44 cells and need to be included in the feed (cysteine, leucine,
isoleucine, and histidine) and asparagine and serine are non‐
essential amino acids, which CHO DG44 cells can produce them-
selves (Figure 6D).
4 | Discussion
This study adapted the iCHO2441 (Strain et al. 2023)GEMto
FDBKA, an industrially relevant mAb producing CHO cell line
cultured in a fed‐batch bioreactor. The goal was to predict
metabolic signatures and amino acid uptake patterns, associated
with distinct culture phases. Constraints were applied to the
iCHO2441 GEM based on transcriptomic data, and flux sampling
was used to compare different culture phases and identify high
mAb‐producing subpopulations. Our analysis revealed multiple
metabolic subsystems linked to improved mAb production, with
a particular emphasis on metabolite transport and amino acid
metabolism. Six specific amino acids (His, Cys, Leu, Ile, Asn, and
Ser) were identified as key targets for future optimization.
To validate these predictions and further optimize mAb pro-
duction, we propose targeted experiments that perturb media
conditions during specific culture phases. By tailoring nutrient
availability to the metabolic needs of the cell cultutre at dif-
ferent stages of the bioreactor process, we aim to further en-
hance mAb productivity.
4.1 | Transcriptomics Analysis and Bioprocess
Validation Support Reliable Constraint‐Based
Models
Initially, transcriptomic analysis identified three distinct
culture phases based on gene expression patterns (Figure 2),
aligning with the recognised early exponential, late
1908 of 1942 Biotechnology and Bioengineering, 2025
exponential, and stationary/death phases of CHO cell cul-
ture. Previous studies have linked transcriptomic signatures
to these culture phases, with late exponential phase
characterised by differential regulation of DNA synthesis,
protein transport, and the cell cycle leading to an
increase in cell proliferation and cell size increase (Pan
et al. 2017,2019). Our results corroborate these findings,
showing differential regulation of DNA replication, RNA
transport, and the cell cycle pathways between the early and
late exponential time points (Figure 2). This congruence
between transcriptomic data and GEM constraints indicates
the identification of relevant and expected pathways
through functional enrichment.
Regarding the validation of flux sampling predictions against
bioprocess data (Figure 3), a strong agreement was observed for
the majority of sampling predictions and experimental bio-
reactor measurements. This provides confidence in the suit-
ability of these models for intracellular flux analysis. Notably,
the models accurately captured the decrease in growth rate over
time and the consistent mAb production during the bioreactor
process. However, the model failed to predict lactate con-
sumption during the late exponential phase, potentially due to
the transitional nature of this phase. This suggests that aver-
aging expression data across multiple days for GEM model
constraint may limit the predictive capability for metabolic
reactions that exhibit significant metabolic shifts. This
FIGURE 6 | Proposed amino acid transport signatures for distinct CHO cell culture phases. A. Model‐predicted amino acid signature for the early
exponential culture phase metabolic model. Amino acids have been grouped on the x‐axis; import has been shown in the right‐hand side of the graph as
positive reaction flux (mmol/gDW/h); export has been shown in the left‐hand side of the graph as negative reaction flux (mmol/gDW/h; darker bars
correspond to the mean reaction flux for the high‐producing 95th percentile of flux sampling solutions, while the paler bar corresponds to the mean reaction
flux for the general pool of flux sampling solutions.Reactionformulahavebeenshownintheleft‐hand x‐axis, with amino acid three‐letter codes; “[c]”: cytosol,
“[e]”: extracellular. For the amino acid of interest, the forward reaction in the left‐hand x‐axis has been written with forward flow from extracellular to cytosol.
B. Model‐predicted amino acid signature for the late exponential culture phase metabolic model. C. Model‐predicted amino acid signature for the stationary/
death culture phase metabolic model. D. Summary illustration of the model‐predicted amino acid uptake signatures which have been associated with
increased mAb production using flux sampling. Amino acids have been overlaid onto a schematic for CHO cell growth in culture. Yellow‐lined boxes
correspond to CHO or mammalian auxotrophies and blue‐lined boxes correspond to CHO proxotrophies.
1909 of 1942
limitation is not unique to our study, as similar challenges have
been observed in other bioprocess modeling studies, particu-
larly in capturing the transition from lactate production to
consumption during late exponential phase. For example, a
systematic evaluation of 'omics integration algorithms showed
that models tended to predict a higher (more positive) rate of
lactate production than was measured experimentally
(Machado and Herrgård 2014).
4.2 | The Stationary/Death Phase of CHO Fed‐
Batch Culture Has a Unique Metabolic Signature
The high mAb‐producing flux sampling solutions demonstrated a
culture phase dependent metabolic phenotype (Figure 4). Notably,
the stationary/death phase was represented by more than 800 un-
ique reactions, associated with high mAb production, significantly
more than the early and late exponential phases (~600). This is
expected, given the bioprocess evidence showing that the uptake
and secretion behavior of the FDBKA cells changes the most dra-
matically between the early exponential and stationary/death
phases.
Enrichment analysis of flux sampling solutions identified
metabolite transport (where 30% of these reactions described
amino acid transport), fatty acid metabolism, nucleotide
metabolism, and lipid metabolism as the subsystems most
strongly associated with increased mAb production
(Figure 5). While the changing dynamics of glucose and
amino acid consumption have been extensively studied in
CHOproteinproduction(Couletetal.2022), nucleotide
metabolism has received less attention. However, the accu-
mulation of glycerol‐3‐phosphate and glycerol (contained in
the glycerophospholipid, sphingolipid, and inositol sub-
system in iCHO2441) in the stationary/death has been pre-
viously reported (Coulet et al. 2022). Additionally, when
normalised for subsystem size, our results highlighted central
carbon and energy metabolism as a top‐ranked subsystem
associated with increased mAb production. This aligns with
previous studies demonstrating cell engineering strategies
targeting genes involved in the tricarboxylic acid cycle can
enhance mAb productivity (Zhang et al. 2021). Furthermore,
overexpression of the mitochondrial pyruvate carrier and
pyruvate oxaloacetate has been shown to increase mAb pro-
duction (Zhang et al. 2021).
4.3 | Amino Acid Consumption Behavior
Regulates mAb Production and Varies With CHO
Fed‐Batch Culture Phase
Media optimisation is a well‐established approach to max-
imise protein production in CHO cell culture. A previous
study applied FVA on the iCHO1766 CHO GEM and pre-
dicted that threonine and arachidonate would be effective
media supplementations. This prediction was validated
in vitro resulting in a a two‐fold increase in mAb production
(Fouladiha et al. 2020). Our work further supports this
approach by predicting a preference for leucine uptake over
isoleucine in high mAb‐producing solutions (Figure 6). This
aligns with the observation that balancing leucine and
arginine is crucial for controlling mAb productivity due to
their presence in the protein product (Fan et al. 2015).
Consequently, the six amino acids highlighted by our
modeling analysis, along with 14 others, are included in the
stoichiometric equation for the mAb product expressed in
this study. Due to this direct relation between amino acid
demand of a CHO cell line and its specific product profile, it
is important to note that the findings made in this study are
specific to the FDBKA CHO cell line and product formula-
tion and should be investigated using a parallel workflow
before being translated to other cell line scenarios.
Research has indicated that amino acid consumption, which
impacts mAb productivity, can be influenced by the ex-
pression of amino acid transporters. For instance, the up-
regulation of genes encoding transporters of alanine,
cysteine, and glutamate during the stationary phase, has
been linked to increased glutathione synthesis, a process
associated with high mAb production (Kyriakopoulos
et al. 2013). This upregulation is thought to be a response to
decreasing intracellular concentrations of amino acids
(Kyriakopoulos et al. 2013).
Further evidence for a tailored feed strategy comes from studies
demonstrating changes in amino acid consumption behavior
over time. Non‐essential amino acids like asparagine, gluta-
mine, and cysteine exhibit high consumption during the ex-
ponential growth phases (Pan et al. 2017). Additionally, the
consumption of essential amino acids is more closely tied to cell
number than cell volume (Pan et al. 2017). These findings
support the notion of a phase‐specific feed strategy, as certain
amino acids, such as arginine, can accumulate towards the later
culture phases, indicating an oversupply in the fed‐batch
medium (Coulet et al. 2022).
However, it is essential to consider the solubility limits and
potential toxic or inhibitory effects of amino acid additions.
Excessive supplementation at any given time point should
be avoided, as exemplified by the association between
asparagine supplementation and ammonia accumulation
(Pereira et al. 2018). Previous studies also advise that levels
of leucine and serine above 0.5–1 mM could result in growth
inhibition (Pereira et al. 2018) and high cysteine concen-
trations can lead to oxidative stress and reduced bioprocess
performance (Komuczki et al. 2022). However, negative
impacts of excessive isoleucine and histidine have not been
reported.
5 | Conclusions
This study has provided valuable insight into the dynamic
metabolic changes that occur during CHO cell fed‐batch bio-
reactor culture, however, several questions still remain. A
critical future direction is to elucidate the causal relationships
between these metabolic shifts and improved mAb production.
Specifically, it is essential to determine whether alterations in
amino acid consumption are a direct driver of increased mAb
synthesis or a secondary consequence of cellular adaptation in a
bioprocess environment.
1910 of 1942 Biotechnology and Bioengineering, 2025
To address this, carefully designed experiments involving tai-
lored media formulations and genetic engineering approaches
are necessary. By manipulating specific amino acid uptake and
metabolic pathways, we can directly assess their impact in mAb
productivity. Constraint‐based modeling will provide a power-
ful tool to predict and simulate the effects of these genetic and
metabolic perturbations, and although these modeling predic-
tions can only be theoretical before experimental mechanisms
and evidence have been provided, this study highlights the
hypothesis‐generating potential of constraint‐based models
while prioritising the experimental validation of initial predic-
tions to bioprocess data.
It is important to acknowledge that cellular responses to media
composition can be complex and dynamic, involving feedback
mechanisms and adaptation strategies that are not fully captured
by current models. Therefore, a synergistic approach combining
computational modeling and experimental validation will be
essential to fully unravel the underlying metabolic mechanisms.
In conclusion, the findings presented in this study highlight the
potential of metabolic engineering and systems biology to op-
timise mAb production processes. By leveraging advanced
modeling techniques and experimental validation, we can
develop more precise and effective strategies to enhance bio-
pharmaceutical manufacturing.
Author Contributions
K.E.M. and J.W. developed the methodology, analysed data, and wrote
the manuscript. S.R., E.H., A.P., T.M., and L.P. generated data and
assisted in the interpretation of results. M.R., A.J.D., and J.M.S. edited
the manuscript and supervised the project. All the authors have read
and approved the final manuscript.
Acknowledgments
The authors acknowledge financial support from a Prosperity Partner-
ship grant (EP/V038095/1) funded by EPSRC, BBSRC, and FUJIFILM
Diosynth Biotechnologies. This grant involved the University of Man-
chester, University of Edinburgh, University of York, and staff at FU-
JIFILM Diosynth Biotechnologies and we acknowledge the Edinburgh
Genomics department for transcriptomics data and analyses. We thank
Ellie Hawke (University of Manchester), John Raven (FUJIFILM), and
Robyn Hoare (FUJIFILM) for bioprocess data generation and many
researchers from all collaborating universities for scientific discussion.
References
Ahn, W. S., and M. R. Antoniewicz. 2011. “Metabolic Flux Analysis of
CHO Cells at Growth and Non‐Growth Phases Using Isotopic Tracers
and Mass Spectrometry.”Metabolic Engineering 13, no. 5: 598–609.
https://doi.org/10.1016/j.ymben.2011.07.002.
Becker, S. A., and B. O. Palsson. 2008. “Context‐Specific Metabolic
Networks Are Consistent With Experiments.”PLoS Computational
Biology 4, no. 5: e1000082. https://doi.org/10.1371/journal.pcbi.1000082.
Castonguay, A.‐M., C. Gravel, and M. Lévesque. 2021. “Treating Par-
kinson's Disease With Antibodies: Previous Studies and Future Direc-
tions.”Journal of Parkinson's Disease 11, no. 1: 71–92. https://doi.org/
10.3233/JPD-202221.
Chandrasekaran, S., and N. D. Price. 2010. “Probabilistic Integrative
Modeling of Genome‐Scale Metabolic and Regulatory Networks in Es-
cherichia Coli and Mycobacterium Tuberculosis.”Proceedings of the
National Academy of Sciences 107, no. 41: 17845–17850. https://doi.org/
10.1073/pnas.1005139107.
Coulet, M., O. Kepp, G. Kroemer, and S. Basmaciogullari. 2022.
“Metabolic Profiling of CHO Cells During the Production of Biother-
apeutics.”Cells 11, no. 12: 1929. https://doi.org/10.3390/cells11121929.
Dobin, A., C. A. Davis, and F. Schlesinger, et al. 2013. “STAR: Ultrafast
Universal RNA‐Seq Aligner.”Bioinformatics 29, no. 1: 15–21. https://
doi.org/10.1093/bioinformatics/bts635.
Duarte, N. C., S. A. Becker, N. Jamshidi, et al. 2007. “Global
Reconstruction of the Human Metabolic Network Based on Genomic
and Bibliomic Data.”Proceedings of the National Academy of Sciences
104, no. 6: 1777–1782. https://doi.org/10.1073/pnas.0610772104.
Durinck, S., Y. Moreau, and A. Kasprzyk, et al. 2005. “Biomart and
Bioconductor: A Powerful Link Between Biological Databases and Mi-
croarray Data Analysis.”Bioinformatics 21, no. 16: 3439–3440. https://
doi.org/10.1093/bioinformatics/bti525.
Durinck, S., P. T. Spellman, E. Birney, and W. Huber. 2009. “Mapping
Identifiers for the Integration of Genomic Datasets With the R/Bio-
conductor Package Biomart.”Nature Protocols 4, no. 8: 1184–1191.
https://doi.org/10.1038/nprot.2009.97.
Ebrahim, A., J. A. Lerman, B. O. Palsson, and D. R. Hyduke. 2013.
“COBRApy: COnstraints‐Based Reconstruction and Analysis for
Python.”BMC Systems Biology 7, no. 1: 74. https://doi.org/10.1186/1752-
0509-7-74.
Fan, Y., I. Jimenez Del Val, and C. Müller, et al. 2015. “Amino Acid and
Glucose Metabolism in Fed‐Batch CHO Cell Culture Affects Antibody
Production and Glycosylation.”Biotechnology and Bioengineering 112,
no. 3: 521–535. https://doi.org/10.1002/bit.25450.
Fouladiha, H., S.‐A. Marashi, and S. Li, et al. 2021. “Systematically Gap‐
Filling the Genome‐Scale Metabolic Model of CHO Cells.”Biotechnology
Letters 43, no. 1: 73–87. https://doi.org/10.1007/s10529-020-03021-w.
Fouladiha, H., S.‐A. Marashi, F. Torkashvand, F. Mahboudi, N. E. Lewis,
and B. Vaziri. 2020. “A Metabolic Network‐Based Approach for Devel-
oping Feeding Strategies for CHO Cells to Increase Monoclonal Antibody
Production.”Bioprocess and Biosystems Engineering 43, no. 8: 1381–1389.
https://doi.org/10.1007/s00449-020-02332-6.
Gopalakrishnan, S., W. Johnson, and M. A. Valderrama‐Gomez, et al.
2024. “Multi‐Omic Characterization of Antibody‐Producing CHO Cell
Lines Elucidates Metabolic Reprogramming and Nutrient Uptake Bot-
tlenecks.”Metabolic Engineering 85: 94–104. https://doi.org/10.1016/j.
ymben.2024.07.009.
Gurobi Optimization LLC. (2023). Gurobi Optimizer Reference Manual
(Version 11.0) [Computer Software]. https://www.gurobi.com.
Gutierrez, J. M., A. Feizi, S. Li, et al. 2020. “Genome‐Scale Re-
constructions of the Mammalian Secretory Pathway Predict Metabolic
Costs and Limitations of Protein Secretion.”Nature Communications
11: 68. https://doi.org/10.1038/s41467-019-13867-y.
Hafeez, U., H. K. Gan, and A. M. Scott. 2018. “Monoclonal Antibodies
as Immunomodulatory Therapy Against Cancer and Autoimmune
Diseases.”Current Opinion in Pharmacology 41: 114–121. https://doi.
org/10.1016/j.coph.2018.05.010.
Haraldsdóttir, H. S., B. Cousins, I. Thiele, R. M. T. Fleming, and
S. Vempala. 2017. “CHRR: Coordinate Hit‐And‐Run With Rounding for
Uniform Sampling of Constraint‐Based Models.”Bioinformatics 33,
no. 11: 1741–1743. https://doi.org/10.1093/bioinformatics/btx052.
Hastings, J., G. Owen, A. Dekker, et al. 2016. “ChEBI in 2016: Improved
Services and an Expanding Collection of Metabolites.”Nucleic Acids
Research 44, no. D1: D1214–D1219. https://doi.org/10.1093/nar/gkv1031.
Hefzi, H., K. S. Ang, M. Hanscho, et al. 2016. “A Consensus Genome‐
Scale Reconstruction of Chinese Hamster Ovary Cell Metabolism.”Cell
Systems 3, no. 5: 434–443.e8. https://doi.org/10.1016/j.cels.2016.10.020.
1911 of 1942
Herrmann, H. A., B. C. Dyson, L. Vass, G. N. Johnson, and J.‐M.
Schwartz. 2019. “Flux Sampling Is a Powerful Tool to Study Metabolism
Under Changing Environmental Conditions.”NPJ Systems Biology and
Applications 5, no. 1: 32. https://doi.org/10.1038/s41540-019-0109-0.
Kanehisa, M. 2000. “KEGG: Kyoto Encyclopedia of Genes and
Genomes.”Nucleic Acids Research 28, no. 1: 27–30. https://doi.org/10.
1093/nar/28.1.27.
Kanehisa, M., M. Furumichi, Y. Sato, M. Kawashima, and M. Ishiguro‐
Watanabe. 2022. “KEGG for Taxonomy‐Based Analysis of Pathways and
Genomes.”Nucleic Acids Research 51: D587–D592. https://doi.org/10.
1093/nar/gkac963.
Kavoni, H., I. S. P. Savizi, N. E. Lewis, and S. A. Shojaosadati. 2025.
“Recent Advances in Culture Medium Design for Enhanced Production
of Monoclonal Antibodies in CHO Cells: A Comparative Study of
Machine Learning and Systems Biology Approaches.”Biotechnology
Advances 78: 108480. https://doi.org/10.1016/j.biotechadv.2024.108480.
Komuczki, D., A. Stadermann, and M. Bentele, et al. 2022. “High
Cysteine Concentrations in Cell Culture Media Lead to Oxidative Stress
and Reduced Bioprocess Performance of Recombinant CHO Cells.”
Biotechnology Journal 17, no. 11: 2200029. https://doi.org/10.1002/biot.
202200029.
Kumar, N., P. Gammell, and M. Clynes. 2007. “Proliferation Control Strat-
egies to Improve Productivity and Survival During CHO Based Production
Culture: A Summary of Recent Methods Employed and the Effects of Pro-
liferation Control in Product Secreting CHO Cell Lines.”Cytotechnology 53,
no. 1–3: 33–46. https://doi.org/10.1007/s10616-007-9047-6.
Kyriakopoulos, S., K. M. Polizzi, and C. Kontoravdi. 2013. “Comparative
Analysis of Amino Acid Metabolism and Transport in CHO Variants
With Different Levels of Productivity.”Journal of Biotechnology 168,
no. 4: 543–551. https://doi.org/10.1016/j.jbiotec.2013.09.007.
Love, M. I., W. Huber, and S. Anders. 2014. “Moderated Estimation of
Fold Change and Dispersion for RNA‐Seq Data With DESeq. 2.”
Genome Biology 15, no. 12: 550. https://doi.org/10.1186/s13059-
014-0550-8.
Machado, D., and M. Herrgård. 2014. “Systematic Evaluation of
Methods for Integration of Transcriptomic Data Into Constraint‐Based
Models of Metabolism.”PLoS Computational Biology 10, no. 4:
e1003580. https://doi.org/10.1371/journal.pcbi.1003580.
Mahadevan, R., and C. H. Schilling. 2003. “The Effects of Alternate
Optimal Solutions in Constraint‐Based Genome‐Scale Metabolic Mod-
els.”Metabolic Engineering 5, no. 4: 264–276. https://doi.org/10.1016/j.
ymben.2003.09.002.
Orth, J. D., I. Thiele, and B. Ø. Palsson. 2010. “What Is Flux Balance
Analysis?”Nature Biotechnology 28, no. 3: 245–248. https://doi.org/10.
1038/nbt.1614.
Pagano, G., K. I. Taylor, and J. Anzures Cabrera, et al. 2024. “Prasine-
zumab Slows Motor Progression in Rapidly Progressing Early‐Stage
Parkinson's Disease.”Nature Medicine 30, no. 4: 1096–1103. https://doi.
org/10.1038/s41591-024-02886-y.
Pan, X., A. A. Alsayyari, C. Dalm, J. A. Hageman, R. H. Wijffels, and
D. E. Martens. 2019. “Transcriptome Analysis of CHO Cell Size Increase
During a Fed‐Batch Process.”Biotechnology Journal 14, no. 3: 1800156.
https://doi.org/10.1002/biot.201800156.
Pan, X., C. Dalm, R. H. Wijffels, and D. E. Martens. 2017. “Metabolic
Characterization of a CHO Cell Size Increase Phase in Fed‐Batch Cul-
tures.”Applied Microbiology and Biotechnology 101, no. 22: 8101–8113.
https://doi.org/10.1007/s00253-017-8531-y.
Park, S.‐Y., D.‐H. Choi, and J. Song, et al. 2024. “Driving Towards Digital
Biomanufacturing by CHO Genome‐Scale Models.”Trends in Biotechnology
42: 1192–1203. https://doi.org/10.1016/j.tibtech.2024.03.001.
Park, S.‐Y., C.‐H. Park, D.‐H. Choi, J. K. Hong, and D.‐Y. Lee. 2021.
“Bioprocess Digital Twins of Mammalian Cell Culture for Advanced
Biomanufacturing.”Current Opinion in Chemical Engineering 33:
100702. https://doi.org/10.1016/j.coche.2021.100702.
Patro, R., G. Duggal, M. I. Love, R. A. Irizarry, and C. Kingsford. 2017.
“Salmon Provides Fast and Bias‐Aware Quantification of Transcript
Expression.”Nature Methods 14, no. 4: 417–419. https://doi.org/10.
1038/nmeth.4197.
Pereira, S., H. F. Kildegaard, and M. R. Andersen. 2018. “Impact of
CHO Metabolism on Cell Growth and Protein Production: An
Overview of Toxic and Inhibiting Metabolites and Nutrients.”
Biotechnology Journal 13, no. 3: e1700499. https://doi.org/10.1002/
biot.201700499.
Pybus, L. P., C. Heise, T. Nagy, et al. 2024. “A Modular and Multi‐
Functional Purification Strategy That Enables a Common Frame-
work for Manufacturing Scale Integrated and Continuous Bioma-
nufacturing.”Biotechnology Progress 40, no. 4: e3456. https://doi.
org/10.1002/btpr.3456.
Ramos, J. R. C., G. P. Oliveira, P. Dumas, and R. Oliveira. 2022. “Genome‐
Scale Modeling of Chinese Hamster Ovary Cells by Hybrid Semi‐Parametric
Flux Balance Analysis.”Bioprocess and Biosystems Engineering 45, no. 11:
1889–1904. https://doi.org/10.1007/s00449-022-02795-9.
Reinhart, D., L. Damjanovic, C. Kaisermayer, and R. Kunert. 2015.
“Benchmarking of Commercially Available CHO Cell Culture Media for
Antibody Production.”Applied Microbiology and Biotechnology 99,
no. 11: 4645–4657. https://doi.org/10.1007/s00253-015-6514-4.
Robinson, M. D., and A. Oshlack. 2010. “A Scaling Normalization
Method for Differential Expression Analysis of RNA‐Seq Data.”Genome
Biology 11, no. 3: R25. https://doi.org/10.1186/gb-2010-11-3-r25.
Schinn, S.‐M., C. Morrison, W. Wei, L. Zhang, and N. E. Lewis. 2021. “A
Genome‐Scale Metabolic Network Model and Machine Learning Predict
Amino Acid Concentrations in Chinese Hamster Ovary Cell Cultures.”
Biotechnology and Bioengineering 118, no. 5: 2118–2123. https://doi.org/
10.1002/bit.27714.
Soneson, C., M. I. Love, and M. D. Robinson. 2015. “Differential
Analyses for Rna‐Seq: Transcript‐Level Estimates Improve Gene‐Level
Inferences.”F1000Research 4: 1521. https://doi.org/10.12688/
f1000research.7563.2.
Strain, B., J. Morrissey, A. Antonakoudis, and C. Kontoravdi. 2023.
“How Reliable Are Chinese Hamster Ovary (CHO) Cell Genome‐Scale
Metabolic Models?”Biotechnology and Bioengineering 120, no. 9:
2460–2478. https://doi.org/10.1002/bit.28366.
Széliová, D., D. E. Ruckerbauer, and S. N. Galleguillos, et al. 2020.
“What CHO Is Made of: Variations in the Biomass Composition of
Chinese Hamster Ovary Cell Lines.”Metabolic Engineering 61: 288–300.
https://doi.org/10.1016/j.ymben.2020.06.002.
Thiele, I., N. Swainston, R. M. T. Fleming, et al. 2013. “A Community‐
Driven Global Reconstruction of Human Metabolism.”Nature
Biotechnology 31, no. 5: 419–425. https://doi.org/10.1038/nbt.2488.
Timouma, S., L. N. Balarezo‐Cisneros, J.‐M. Schwartz, and D. Delneri.
2024. “Development of a Genome‐Scale Metabolic Model for the Lager
Hybrid Yeast S. pastorianus to Understand the Evolution of Metabolic
Pathways in Industrial Settings.”mSystems 9, no. 6: e0042924. https://
doi.org/10.1128/msystems.00429-24.
Virtanen, P., R. Gommers, and T. E. Oliphant, et al. 2020. “SciPy 1.0:
Fundamental Algorithms for Scientific Computing in Python.”Nature
Methods 17, no. 3: 261–272. https://doi.org/10.1038/s41592-019-0686-2.
Wurm, F. M. 2004. “Production of Recombinant Protein Therapeutics in
Cultivated Mammalian Cells.”Nature Biotechnology 22, no. 11: 1393–
1398. https://doi.org/10.1038/nbt1026.
Xu, W.‐J., Y. Lin, C.‐L. Mi, J.‐Y. Pang, and T.‐Y. Wang. 2023. “Progress
in Fed‐Batch Culture for Recombinant Protein Production in CHO
Cells.”Applied Microbiology and Biotechnology 107, no. 4: 1063–1075.
https://doi.org/10.1007/s00253-022-12342-x.
1912 of 1942 Biotechnology and Bioengineering, 2025
Xu, X., H. Nagarajan, N. E. Lewis, et al. 2011. “The Genomic Sequence
of the Chinese Hamster Ovary (CHO)‐K1 Cell Line.”Nature
Biotechnology 29, no. 8: 735–741. https://doi.org/10.1038/nbt.1932.
Yeo, H. C., J. Hong, M. Lakshmanan, and D.‐Y. Lee. 2020. “Enzyme
Capacity‐Based Genome Scale Modelling of CHO Cells.”Metabolic
Engineering 60: 138–147. https://doi.org/10.1016/j.ymben.2020.04.005.
Yizhak, K., B. Chaneton, E. Gottlieb, and E. Ruppin. 2015. “Modeling
Cancer Metabolism on a Genome Scale.”Molecular Systems Biology 11,
no. 6: 817. https://doi.org/10.15252/msb.20145307.
Zahavi, D., and L. Weiner. 2020. “Monoclonal Antibodies in Cancer
Therapy.”Antibodies 9, no. 3: 34. https://doi.org/10.3390/antib9030034.
Zhang, H.‐Y., Z.‐L. Fan, and T.‐Y. Wang. 2021. “Advances of Glyco-
metabolism Engineering in Chinese Hamster Ovary Cells.”Frontiers in
Bioengineering and Biotechnology 9: 774175. https://doi.org/10.3389/
fbioe.2021.774175.
Zur, H., E. Ruppin, and T. Shlomi. 2010. “iMAT: An Integrative
Metabolic Analysis Tool.”Bioinformatics 26, no. 24: 3140–3142. https://
doi.org/10.1093/bioinformatics/btq602.
Supporting Information
Additional supporting information can be found online in the
Supporting Information section.
1913 of 1942
