Novel Data‐Driven Mechanistic Modeling of Untargeted Metabolome Data Reveals Feed Component Effects in CHO Cell Bioprocess Using Column Generation‐Based EFMs

ABSTRACT

This study presents a novel approach for applying mechanistic metabolic modeling to untargeted metabolomics data. The approach was applied to the production process of a difficult‐to‐express enzyme by CHO cells, to identify key feed medium component candidates responsible for improved productivity through feed modification. The exploitation of untargeted metabolomics implies no prior decision of the metabolites or pathways and thus allows screening of metabolic phenomena and bringing an objective perspective. However, such exploitation is challenging due to the high‐dimensionality, complexity, relative quantitative information, and high analysis cost of the data, leading to data scarcity. A combination of untargeted metabolomics data exploration and mechanistic modeling was developed to leverage metabolomics data. The study analyzed LC/MS/MS metabolomics data (563 cellular and 386 supernatant metabolites) to determine the key metabolites involved in the productivity increase associated with a feeding modification. The metabolome data was utilized to expand the original stoichiometric reaction network of 127 reactions to 370 reactions. Mechanistic modeling using elementary flux modes‐based column generation identified and simulated the underlying metabolic pathways. Twenty‐one key metabolites significant for productivity improvement were revealed. This included several unexpected metabolites, such as citraconate and 5‐aminovaleric acid, in addition to well‐known components, as well as their underlying metabolic pathways. This study offers a novel approach for investigating nutrient supplementation in terms of metabolic fluxes and process performance, paving the way for rational process optimization supported by mechanistic understanding.

Graphical Abstract and Lay Summary

This study presents a novel approach for applying mechanistic metabolic modeling to untargeted metabolomics data from the production process of a difficult‐to‐express enzyme by CHO cells, to identify key feed media component candidates responsible for improved productivity through feed modification. The study analyzed LC/MS/MS metabolomics data (563 cellular and 386 supernatant metabolites) to determine the key metabolites involved in the productivity increase. The metabolome data was utilized to expand the original stoichiometric reaction network of 127 reactions to 370 reactions. Mechanistic modeling using elementary flux modes (EFM)‐based column generation identified and simulated the underlying 300 metabolic pathways. Twenty‐five production associated pathways were selected and used to limit the feed candidates even further. The list included several unexpected metabolites, such as citraconate and 5‐aminovaleric acid, in addition to well‐known components. This study offers a novel approach for investigating nutrient supplementation in terms of metabolic fluxes and process performance, paving the way for rational process optimization supported by mechanistic understanding.

Abbreviations

5AVA

5‐Aminovaleric acid

AA

amino acids

C02

carbon dioxide

CG

column generation

CHO cells

Chinese hamster ovary cells

EFM

elementary flux mode

FDR

false discovery rate

HPLC

high pressure liquid chromatography

LC

liquid chromatography

LC/MS/MS

liquid chromatography/Mass spectrometry/Mass spectrometry

MVC

million viable cells

P1

principal component 1

P2

principal component 2

pC02

partial pressure of carbon dioxide

PCA

principal component analysis

PEA

pathway enrichment analysis

PoI

product of Interest

PTM

posttranslational modification

qGlc

cell‐specific glucose consumption

qLac

cell‐specific lactate consumption

RMSE

root mean square error

SD

standard deviation

UPLC

ultra high‐pressure liquid chromatography

1. Introduction

Chinese hamster ovary (CHO) cells have been used in bioprocesses for over 30 years to produce complex therapeutic recombinant proteins. Fed‐batch is the conventional production method for biopharmaceuticals using mammalian cell lines. The extensive use of fed‐batch mode in industry [1] is due to the robustness and proven track record. Fed‐batch bioprocesses use concentrated feeds to deliver nutrients that support cell growth, viability, and production of the product of interest (PoI). Balancing the amino acids and vitamins in the feed is critical for bioprocess productivity and optimization [2].

Although the final product of a fed‐batch process can be very consistent between batches, many metabolites and even product attributes can vary over time during the cultivation. A common metabolite shift during a fed‐batch process is the change from lactate production to consumption, which can correlate with higher productivity [3, 4]. Another common shift involves three phases during the fed‐batch culture: exponential growth, stationary, and death phases [1]. However, mechanisms controlling these switches are not completely understood.

Omics data offer insights into the metabolic status of the cells, providing insights into biochemical pathways and the effect of environmental factors. Furthermore, an insight into the switches in cellular metabolism enables knowledge‐based process improvements [5, 6]. Various omics analyses, such as transcriptomics, proteomics, metabolomics, lipidomics, and their combinations, are increasingly used in bioprocess studies [7, 8, 9, 10]. The differences between utilizing different feeds can be assessed by both intracellular and extracellular metabolite profiling [11]. Thus, metabolomics data are promising for identifying key metabolites and metabolic pathways associated with high yields of target molecules.

Metabolomic analysis can be carried out through targeted or untargeted methods. Targeted methods study specific metabolites or pathways, whereas untargeted methods aim to study the entire cell metabolome. Untargeted metabolomic studies have no predetermined metabolites or pathways and require the identification of key metabolites and pathways. However, the analysis is challenging due to high‐dimensional, complex, scarce, and relative data produced by untargeted methods.

CHO cell metabolism is complex, and mathematical modeling can provide a better understanding [12]. Mechanistic metabolic modeling, an alternative to black‐box modeling, is a powerful tool for predicting metabolic fluxes. In mechanistic modeling, the cell is considered a catalyst, and a common approach is to simplify the cell metabolism into main pathways or macro‐reactions. In pathway analysis, macro‐reactions can be obtained using the elementary flux modes (EFMs), which are stoichiometrically balanced linear combinations of metabolic reactions, connecting the components, for example, extracellular substrates to products [12, 13, 14, 15]. Solving the EFM‐based metabolic flux analysis problem provides these linear combinations in a data‐driven manner [16].

Column generation (CG) is an EFM data driven‐based approach introduced in ref. [17], iteratively identifying the pathways without limitation by the reaction network size. CG has been shown to successfully construct mechanistic metabolic models for bioprocesses with fluxes for amino acids, glucose, lactate, ammonia, biomass and PoI as inputs [17, 18].

In this study, we aimed to apply mechanistic metabolic modeling to untargeted metabolomics data generated in processes for the bioproduction of a difficult‐to‐express enzyme. A framework for identifying potential key factors responsible for process improvements was created. We combined metabolomics data exploration with mechanistic modeling. First, we identified key metabolites involved in the increase in productivity associated with a process modification from the metabolomics data. Next, we simulated metabolic pathways involved in CHO cell metabolism with mechanistic modeling. A workflow was demonstrated for investigating nutrient supplementation effects on metabolic fluxes and productivity in the cells. Metabolic pathways and medium components associated with high bioproduction yields were identified.

2. Materials and Methods

2.1. Cell Line and Expansion

CHO‐K1 cells producing a difficult‐to‐express enzyme as PoI were expanded in shake flasks (Corning) at 37°C in a Kühner shaker incubator (5% CO₂) using a proprietary chemically‐defined medium, with passaging every 3/4 days. For production scale, a similar shake flask seed train was used to inoculate the first of two additional seed bioreactors.

2.2. Bioreactor Cultures and Process Parameters

Two fed‐batch production processes, named “Process A” and “Process B,” differing only by the inoculation cell density of 0.4 and 0.5 × 10⁶ cells/mL (MVC/mL) and the feed, were studied. Dissolved oxygen was controlled at 40% by sparging O₂ and pH by sparging CO₂. Both processes were cultured in batch mode until Day 2, when the temperature was shifted from 37°C to 31°C. Three runs (triplicates) of Process A and B were carried out. Duplicate runs for both processes were performed in 3L‐glass stirred tank bioreactors (Applikon). Metabolomics data were collected from Process B, carried out in a 15 L‐glass stirred tank bioreactor (Applikon), and Process A, carried out in a 1000 L single‐use bioreactor (Sartorius). The processes were fully scalable between small and commercial scale bioreactors (data not shown).

2.3. Feed and Glucose Addition

The same proprietary chemically defined base and feed media were used in Processes A and B, but Process B was also supplemented with galactose. A feed bolus of 1.5% [or 3%] v/v based on the initial working volume was added everyday between Day 2 and 16 in Process A [and Process B]. Glucose was supplemented separately with a target concentration of 12 mM, based on mass balance calculation using the glucose concentration and cell specific glucose consumption.

2.4. Process Monitoring

Samples were collected for offline process parameter measurements. pH, pCO₂, lactate, and glucose concentrations were immediately analyzed with a blood gas analyzer (Radiometer Medical ApS, Brønshøj, Denmark). Viable cell density (VCD) and viability were analyzed with a ViCELL XR Cell Viability Analyzer (Beckman Coulter, Brea, CA). Samples were stored at ‐70°C before product analyses with in‐house assays for the product concentration and enzyme activity, analyzed with a fluorometric activity assay.

Amino acid analysis was performed based on a Waters UPLC Amino Acid Analysis applications kit using the vendor's protocol [19]; briefly: Cell culture supernatants were filtered through 10 kDa filters, and filtrates were diluted using HPLC‐grade water to appropriate concentrations. Amino Acid Standard (Waters) was diluted and used as a standard curve. Diluted samples and standards were combined with internal standard Norvaline (Sigma‐Aldrich) and AccQ Tag Derivatization Kit reagents (Waters). Samples were run on an Acquity UPLC system (Waters) using an AccQ‐Tag Ultra RP Column 130 °A 1.7 µm, 2.1 mm, 100 mm (Waters). Data were analyzed using EMPOWER software (Waters).

2.5. Metabolomics Data Analysis

Samples for metabolomic analyses were taken from the 1000 L bioreactor for Process A and from the 15 L bioreactor for Process B. Samples were collected from five time points (Days 2/3, 6, 10, 14, and 17), to capture typical dynamic behavior throughout the process. Ten million cells were collected and centrifuged (755 × g, 8 min). Supernatant samples were collected after the first centrifugation. Cells were resuspended into the remaining supernatant (0.75 mL) and recentrifuged. The supernatant was then completely removed, and the dry cell pellet was stored at −70°C.

Global untargeted metabolomic analyses of cell pellet and supernatant samples were performed using Metabolon's global metabolomics platform. Samples were extracted and analyzed on the LC/MS/MS and Polar LC platforms. Metabolites were identified by matching ions to an in‐house library of standards, identifying 563 cellular and 386 supernatant metabolites. Metabolites were quantified by peak area integration. Integrated peak areas were normalized by the total amount of protein in each sample to ensure the sample comparability, with the assumption that the total amount of protein was equal in all the cells [20] and where the total amount of protein was quantified by Bradford assay. Finally, both intracellular and supernatant metabolites were rescaled to have a median of 1. “Intracellular” refers to metabolites contained in the cell pellets, while “extracellular” refers to metabolites contained in the cell culture supernatant.

2.6. Analysis of Metabolomics Data

The subsequent bioinformatics analyses were performed in R (version 4.2.1, June 23, 2022) and RStudio (version 2022.12.0 + 353) or MATLAB (2022a, The MathWorks Inc., Natick, MA, USA). In R the following libraries were used: readxl, broom, genefilter, dplyr, corr, and ggplot2.

Processes A and B exhibited biphasic growth behavior, with a growth phase until Day 7, followed by a stationary phase until the end of the culture (see section Results). The data from each process were divided into two groups, generating a total of four groups: (1) Process A growth phase (Days 3 and 6); (2) Process A stationary phase (Days 10, 14, and 17); (3) Process B growth phase (Days 2 and 6); 4) Process B stationary phase (Days 10, 14, and 17). To identify significantly changing metabolites between groups, Student t‐test was performed separately for the intracellular and supernatant metabolites. p values were adjusted using the FDR method to reduce false positive results [21] and filtered with a p value cutoff of 0.05 for statistical significance.

Metabolites in the supernatant can originate from excretion and/or leakage from dead cells. Therefore, correlation between the metabolites and dead cell concentration was calculated with the MATLAB “corr” function, where metabolites with a correlation coefficient ≥ 0.9 [22] were considered as leaking from dead cells.

The feed medium volume was doubled in Process B compared to Process A, which also impacted the metabolite concentrations. To assess this, the ratio of metabolite concentration in Process B to Process A was calculated (Equation 1):

$$\frac{(\frac{1}{n}{\sum }_{k=1}^n{b}_k)-c_k}{(\frac{1}{n}{\sum }_{k=1}^n{a}_k)-c_k}$$ (1)

The average concentration of each metabolite (k) across Process A (a_k ) and Process B (b_k ) was calculated. The concentration of metabolite in the base medium (c_k ) was subtracted from both averages, and the value for Process B was divided by the value for Process A. Metabolites with values close to 2 were considered primarily affected by feeding.

2.7. Modeling

Metabolomics data are often scarce due to the high costs associated with the method. Furthermore, the metabolite concentrations from untargeted metabolomics are relative, posing challenges for modeling. To deal with these challenges, we combined daily process data (metadata) with metabolomic information, identified important metabolites to understand the differences between Processes A and B, and aligned the metabolomic information to a central metabolism reaction network by using a systematic modeling approach. This modeling approach identified plausible pathways linking the metabolites by iteratively identifying EFMs based on the experimental data.

2.7.1. Data Pretreatment

Due to external constraints, the measurements of ammonia and glutamine in Process B run, sampled for the metabolomic analysis, were incomplete. These values were obtained by averaging two previous Process B runs together. Additionally, values for missing days for cell count, cell viability, glutamine, glucose, ammonia, lactate, amino acids, and metabolomics data were linearly approximated (Microsoft Excel, 2019). Subsequent data handling was performed in MATLAB. All concentrations were smoothed with the MATLAB “smooth” function (moving average, window of 3). The cell density and PoI concentrations were converted to mM by multiplying with factors of 12.115 and 0.044, respectively (Equations 2 and 3), similar to ref. [18]:

$$X_{ext}=X_{MVC/ml}\ast \frac{315\frac{pg}{cell}\ast 0.93}{24.18\frac{g}{mol}}\approx X_{\frac{MVC}{ml}}\ast 12.115$$ (2)

where X represents the biomass and MVC/mL stands for million (10⁶) viable cells per mL. X_ext is assumed to have the mass of 24.18 g/mol [23] and the dry weight of X_ext is assumed to be 315 pg/cell [24]. The macro‐molecules making up biomass sum up to 93%:

$$P_{ext}=\frac{P_{mg/L}}{22.5\frac{g}{mol}}=P_{\frac{mg}{L}}\ast 0.044$$ (3)

where POI (P_ext ) is assumed to have a weight of 22.5 g/mol and $P_{\frac{mg}{L}}$ is the measured POI concentration.

Metabolic fluxes were calculated (Equation 4) from smoothed concentrations and normalized by dividing each flux by the average of the fluxes. Normalized data were used for model generation:

$$q_{ext}=\frac{c_{k,t}{V}_t-c_{k,t-1}{V}_{t-1}}{\frac{1}{2}\left(X_{t-1}+X_t\right)\left(T_t-T_{t-1}\right)}$$ (4)

where X_{t − 1} is the VCD before feeding and X_t is the VCD before the following feeding, c_{k,t − 1} and c_k,t are the metabolite k concentrations, V _{t − 1} and V_t are the volumes and T_{t − 1} and T_t are the time points of sample collection.

2.7.2. Generation of EFMs by CG

The purpose of the CG method is to identify metabolic pathways through the cellular metabolic reaction network. It identifies a subset of EFMs without systematic enumeration of all the EFMs, which is computationally impossible for reaction networks larger than some 40 reactions. The CG uses an algorithm of a data‐fitting problem posed as a dual optimization problem by iteratively solving a master problem (Equation 6) and a subproblem (Equations 5a–5d) to find a subset of EFMs and their associated pathway rates ( w ) that optimally fits the experimental data; see a more extensive description in previous work [17, 18]. The algorithm treats the experimental data sequentially to achieve this EFM subset and terminates when the solution is optimal. The solution is globally optimal but not necessarily unique.

In the present report, the CG (Equations 6 and 5a to 5d) was modified by adding the new constraints 5e to 5j, generating a modified CG, modCG (Equations 6 and 5a to  5j). In the CG, Equations (5a) to (5d) identify an EFM e by selecting a set of reactions that best fit the metabolite consumption and production rates (5a), subject to constraints 5b–5d. The steady state assumption is expressed by constraint 5b, where   A _int,nm is the stoichiometry matrix of produced intracellular metabolites. (  A _int,nm  e ) must be equal to 0, requiring that the steady state metabolite consumption and production have the same rate. However, we observed changes in several internal metabolite levels and added new constraints for the non‐steady‐state metabolites, partially abolishing the steady state assumption. In the modCG, the produced intracellular metabolites (matrix A _int,prod) were required to increase in an individual EFM, e (constraint 5e), while the consumed intracellular metabolites (matrix A _int,cons) were required to decrease (constraint 5f). Furthermore, for the internal metabolites, observed to diminish or accumulate, a consumption or production reaction was required to be included in the EFM, e , which led to constraints 5 g and 5 h. The introduction of these latter EFMs generated short EFMs containing only internal reactions or transport reactions; therefore, constraints 5i and 5j were added to impose that EFM e included both transport and internal reactions:

$$\underset{\mathbf{e}}{\min }\sum _{k=1}^D({({\mathbf{A}}_{meas}\mathbf{E}\mathbf{w}-{\mathrm{q}}_k^T)}^T{\mathbf{A}}_{meas})\mathbf{e}$$ (5a)

subject to

$${\mathbf{A}}_{\mathrm{int},\mathrm{nm}}\mathbf{e}=0$$ (5b)

(5c)

$$\mathbf{e}\ge 0$$ (5d)

$$-{\mathbf{A}}_{\mathrm{int},\mathrm{prod}}\mathbf{e}\le 0$$ (5e)

$${\mathbf{A}}_{\mathrm{int},\mathrm{cons}}\mathbf{e}\le 0$$ (5f)

(5g)

(5h)

$$1^T{\mathcal{J}}_{ext}\mathbf{e}>0$$ (5i)

$$1^T{\mathcal{J}}_{int}\mathbf{e}>0$$ (5j)

where A _meas is the stoichiometry matrix of measured metabolites, both intracellular and extracellular; E  is the set of EFMs generated; w is the flux through the EFMs, q_k is the measured flux of the kth metabolite; A _int,nm is the stoichiometry matrix of unmeasured steady‐state intracellular metabolites; A _int,prod is the stoichiometry matrix of produced intracellular metabolites; A _int,cons is the stoichiometry matrix of consumed intracellular metabolites; ${\mathcal{I}}_{prod},$ ${\mathcal{I}}_{cons}$, ${\mathcal{I}}_{ext}$, and ${\mathcal{I}}_{int}$ are identity vectors for reactions consuming/producing a consumed/produced intracellular metabolite and for reactions with extracellular or intracellular metabolites as reactants, respectively.

The master problem, given by Equation (6), was the same as reported in ref. [17], with the exception of the introduction of a multiplier m for the internal metabolites q_int , to accommodate for normalized intracellular concentrations. The multiplier m , representing an unknown normalization factor, was solved at the same time as the rates of the pathways. Constraint 6c imposed that the normalization factor was positive:

$$\underset{\mathbf{w},\mathbf{m}}{\min }\frac{1}{2}\sum _{k=1}^D{\parallel \left[\begin{array}{c}{\mathrm{q}}_{ext,k}\\ {\mathrm{q}}_{int,k}\mathbf{m}\end{array}\right]-{\mathbf{A}}_{meas}\mathbf{E}\mathbf{w}\parallel }_2^2$$ (6a)

subject to

$$1000\ge \mathbf{w}\ge 0$$ (6b)

$$\mathbf{m}>0$$ (6c)

The data from Processes A and B were processed together by CG or modCG to provide one set of EFMs for the growth and another one for the stationary phase. The union of these two sets of EFMs was then taken, and duplicate EFMs were removed, as previously reported [18]. Finally, the union of the obtained EFMs was used as input for the master problem (Equation 6) for fitting pathway rates or fluxes ( w ) for the combined set of EFMs.

2.7.3. Network Modifications

In comparison with our previous work [18], the stoichiometric network of 126 reactions was expanded to include the large number of metabolites provided in the metabolomics data as detailed below. For internal metabolites measured in the supernatant samples and for which transport was described in the literature or in databases, transport reactions were added to the reaction network. A network with reactions connected to differing‐A‐vs‐B metabolites was generated by adding reactions from the BiGG model iCHOv1 [25], Brenda [26], [brenda‐enzymes.org], Metabolic Atlas [27], [metabolicatlas.org], ChEBI [28], www.ebi.ac.uk/chebi/init.do], KEGG [29], [genome.jp/kegg/pathway.html], MGI [30], [informatics.jax.org], and Reactome [31], [reactome.org]. Additional reactions were taken from literature [32, 33]. For four of the differing‐A‐vs‐B metabolites (2‐hydroxy‐4‐(methylthio) butanoic acid, N‐acetylproline, 2R,3R‐dihydroxybutyrate, and 2‐hydroxyhippurate (salicylurate)), no descriptions of metabolic reactions occurring in mammals were found; therefore, no reactions were added for these metabolites.

3. Results

The objective of the present study was to create a framework for identifying key elements for process improvement from metabolomics data. Data were collected from multiple runs of two bioproduction processes, Process A and Process B. Process B was a bioproduction process characterized by higher VCD and productivity than Process A (Figures 1A,B,G), with slightly higher inoculation cell density and a larger volume of the feed medium. Both processes had typical fed‐batch phases. Following an 8‐day growth phase, the cells entered a stationary phase with slowly declining VCD. Process B achieved a slightly higher VCD (max 8.8 MVC/mL), compared to Process A (max 8.3 MVC/mL). Viability exceeded 95% throughout cultivation in both processes. Process B led to a 35% higher PoI concentration, associated with the increased feed rate, compared to Process A. Process B also had an increased cell‐specific productivity, 25% higher than Process A, during the stationary phase (Figure 1H). The PoI from Process B had slightly reduced specific activity and posttranslational modifications (PTM) (Figures 1I,J).

FIGURE 1.

Cultivation, productivity, and product quality parameters and metabolomics data collected for Processes A and B. Data (excluding metabolomics) were collected from triplicates (n = 3) reflecting the general performances differentiating Process A and B viable cell density (VCD, open circles) and viability (black circles) for triplicate (n = 3) runs of Process A (A) and Process B (B). (C), (D) Glucose (open circles), lactate (black squares), and ammonium (grey circles) concentrations. (E), (F) Cell‐specific lactate production (circles) and glucose (asterisks) consumption. (G) Titer at Days 14 and 17. (H) PoI cell‐specific productivity at Days 14–17. (I) Specific activity of the PoI at Days 14 and 17. (J) Posttranslational modifications of the PoI (PTM) at Days 14 and 17. (K) Heat plot of 563 intracellular metabolites. (L) Heat plot of 386 supernatant and medium metabolites. Heat plot values were normalized, and the samples were labeled for each sampling day (D) of Process A (A) and Process B (B).

Processes A and B differed in their cell metabolism. Process B had an increased lactate accumulation (15.6 mM) during the growth phase compared to Process A (12.2 mM). However, both processes had similar cell‐specific lactate production rates, q_lac . The cell specific glucose consumption, q_glc, increased after Day 10 in both processes, being higher for Process B (0.69 pmol/cell/day) than Process A (0.52 pmol/cell/day). The pCO₂ concentration profiles were comparable for both processes (data not shown).

Metabolomics data were collected for 386 supernatant and 563 intracellular metabolites from single runs of both Process A and Process B as well as the feed and base media (Figures 1K–L). An untargeted global metabolomics platform identifies the metabolites based on a library of metabolite standards, while some metabolites not included in the library remain unidentified. These were labeled ’unidentified’ in this study.

To understand the process differences leading to an increased production, an analysis framework was created, where the modeling investigation focused on the main metabolism of the cells and the PTM's were excluded. An overview of the overall approach is given in Figure 2A. First, the key metabolites involved in the increase in productivity in Process B versus A were identified from the metabolomics data—Section 3.1, “Identification of Key Metabolites,” presents the different methods used for this identification. Then a mechanistic model of the cell metabolism based on the metabolome was obtained by the identification of the main metabolic pathways using the method of CG for a network of reactions including information relevant for the key metabolites—in Section 3.2 “Generation of Mechanistic Metabolic Models Using Metabolomics Data,” different models involving different types of data and complexities of the reaction network are presented and compared. Finally, one selected model was exploited to identify the metabolic mechanisms and components of the feed medium related to the improvement of Process B versus A as well as potential improvements in the future—Section 3.3 “Interpretation of the Model for PoI Production.”

FIGURE 2.

Identification of the key metabolites from the metabolomics data. (A) Illustration of the whole workflow. (B) Illustration of the key metabolite identification workflow and data used. (C) PCA of the metabolomics data (1050 metabolites) from supernatant and intracellular samples—the labels Dxx indicate the days of culture. (D) PCA of the diminished metabolites. (E) PCA of the highly varying SD metabolites. (F) PCA of the metabolites differing between Processes A and B. (G) Overlapping and uniquely identified metabolites. (H) Significantly differing metabolites issued from feed additions or generated by leakage from dying cells.

3.1. Identification of Key Metabolites

Principal component analysis (PCA) was used to determine if metabolomics data could distinguish the processes (Figure 2B). Principal components P1 and P2 explained 62% and 17% of sample variation across both processes. No correlation was observed between the principal components and the processes. The process day, however, correlated with P1, suggesting that metabolites with a high coefficient for P1 related to the culture age in both processes. A separate PCA of the supernatant and the intracellular metabolites showed similar results (not shown). Next, we investigated different rationales to identify a subset of metabolites related to process differences.

3.1.1. Metabolites With Decreasing Concentration With Time—“Diminished”

Drastic decrease in metabolites can cause process differences [2]. We identified the metabolites decreasing with time during the process by visualizing their distribution in Process A supernatant (see Figure S1). There was a clear segregation between the metabolites present at levels ≤ or > 20% of their original concentration in the base medium, and the metabolites ≤ 20% were then considered as “diminished.” According to this criterion, 21 metabolites (including 5 unidentified) were diminished during Process A of which 7 were not diminished during Process B (Table S1). The PCA of the diminished metabolites, represented in Figure 2C, showed a clear separation between Process A and B samples, supporting that the doubled feed in Process B had an important impact on the process. Eleven metabolites (including 3 unidentified) were depleted below the detection limit, but this was not sufficient to explain the difference between the processes.

3.1.2. Metabolites With High Level Variation During Process—“Highly SD Variable”

Cell metabolism is primarily responsible for the variations in cell culture component concentrations. To quantify this, we calculated the standard deviation (SD) of each metabolite in the extracellular metabolomics data, and then performed PCA on metabolite subsets of the 5, 10, 20, 30, 40, or 50 highest SDs, to investigate if these metabolites provided clustering of Processes A and B (see an example in Table S2). The subsets with high SD variability (“highly SD variable”) were, however, unable to provide such clustering on the two first principal components (see Figure 2D), suggesting that using the SD was not a useful approach here.

3.1.3. Metabolites Significantly Differing Between Processes A and B—“Differing‐A‐vs‐B”

A Student's t‐test was performed for each metabolite (see Section 2.6) to determine the significantly differing metabolites between Processes A and B (referred to as “differing‐A‐vs‐B”). The test indicated that the largest number of significantly differing metabolites were found in the stationary phase, with 26 metabolites in the supernatant and 3 in the intracellular samples (Table S3 and Figure S3). One metabolite, the intracellular acetylcarnitine, was found to be differing‐A‐vs‐B during the growth phase between the processes. PCA of the differing‐A‐vs‐B metabolites (Figure 2E, Table S3) separated Process A and Process B with respect to the second principal component, P2, (8.8% variation), thus capturing the variation between the processes. The highest loading values of P2 were for guanine (0.708), 5AVA (0.289), and harmane (0.249), suggesting differences in guanine, 5AVA, and harmane metabolisms between processes A and B. Along P1, taurine had the highest loading value (0.962), indicating a strong contribution to the separation between phases, with the next highest contributors being guanine (−0.125) and homocystine (−0.088).

3.1.4. Evaluating the Effects of Feeding—“Feed”

The metabolite concentrations were also affected by the feed additions (referred to as “feed”) and by metabolite leakage from dying cells (Figure 2G). The feed volume in Process B was double Process A's, so it was assessed if the observed metabolite level changes were likely resulting from this double feeding (Table S4; method see Section 2.6). HEPES, a metabolically inactive pH buffer, was nearly doubled (2.15) in Process B compared to Process A, suggesting that the identification of the feeding effect was successful. Several differing‐A‐vs‐B supernatant metabolites were found to be nearly doubled: homocystine (1.97), ornithine (1.90), betaine (2.20), and 2‐hydroxyhippurate (2.22).

3.1.5. Evaluating the Effect of Cell Death

The correlation between the dead cell count and the supernatant metabolites was calculated to separate the metabolites leaking from dying cells and the ones related to the cell metabolism (see Section 2.6). One hundred fifty‐two metabolites had a high correlation (≥ 0.9) to the dead cell count, including one differing‐A‐vs‐B metabolite, homocystine (0.904, p = 0.00033). However, the most differing‐A‐vs‐B metabolites lacked correlation to the dead cell count, suggesting these were not attributable to cell death leakage.

3.2. Generation of Mechanistic Metabolic Models Using Metabolome Data

Next, mechanistic metabolic modeling was carried out to provide insight into the differences between Processes A and B (Figure 3A). The metabolic reaction pathways and pathway rates were identified using the CG (see Section 2.7.2). Prior to model generation, the data were pretreated (see Material and Methods) to obtain normalized metabolite fluxes.

FIGURE 3.

Mechanistic metabolic modeling with column generation. (A) Illustration of the steps taken during modeling. m1 was generated using a network of 127 reactions with 26 input metabolites. m2 was generated using the expanded reaction network and metabolomics data. m3 was generated using the expanded reaction network and metabolomics data with column generation. (B) Number of reactions in the reaction network for each model. (C) Number of input metabolites. (D) Number of pathways (EFMs). (E): Root mean square error of the model fitting for the inputs (RMSE). (F) Relative RMSE for the PoI. (G) Relative RMSE for the biomass formation.

3.2.1. Using the Metabolome Data as Input in Modeling

The metabolome data provided information from a much larger metabolite set compared to our earlier work [17, 18], which involved solely the metadata of amino acids, glucose, lactate, ammonia, biomass, and PoI. In line with ref. [18], first, we generated a reference model (m_CG_RN‐127) using CG with the reaction network RN‐127, composed of 126 reactions as in ref. [18] and an additional 127th reaction describing the cell death. The amino acid, glucose, lactate, ammonia, cell, and PoI concentrations were the inputs. The model m_CG_RN‐127 had 92 EFMs and a root mean square error of the model fitting (RMSE_abs) of 0.00034 for the inputs, relative RMSE of 0.3% for the PoI (RMSE_{rel_POI}), and 1.5% for the biomass (RMSE_{rel_X}); see Figures 3B–G.

Since the RN‐127 lacked reactions for several differing‐A‐vs‐B metabolites, it was expanded by adding the reactions connecting the differing‐A‐vs‐B metabolites, and adding missing transport reactions for the supernatant metabolites reported to have such transport from literature, generating a network of 370 reactions, RN‐370. A model, m_CG_RN‐370_OMX, was generated by CG using RN‐370 and the metabolomics data, both intracellular and extracellular (Figure 3A). It had increased complexity (502 EFMs) and a higher number of input metabolites, suggesting that additional metabolites led to a higher complexity. The RMSE_abs of m_CG_RN‐370_OMX was 0.00593, thus larger than for m_CG_RN‐127 (0.00034), and the model fit was also compromised for PoI (0.7%) and biomass (2.1%); however, the relative RMSE remained less than 5% for both.

3.2.2. Using the Intracellular Metabolome Data to Replace the Steady State Assumption

In line with the observation that some intracellular metabolites varied with time, the steady state assumption was discarded and several constraints were added, leading to the modCG, a modified CG (see Section 2.7.2). A model m_modCG_RN‐370_OMX was generated with modCG using the network RN‐370 and the metabolomics data (Figures 3B–G). The removal of the steady state assumption led to a less complex model (300 EFMs) with a better fit (RMSE_abs 0.00205). The RMSE_{rel_POI} and RMSE_{rel_X} were 6.5% and 10%, representing an increase compared to m_CG_RN‐127; however, these values were both still within the measurement error. Notice that the same results were obtained when using either the data from Process A or Process B only (not shown). Other variant models were explored to catch the impact of the reaction network and the input parameters and are detailed in Table S3.

3.3. Interpretation of the Model for PoI Production

Lastly, we used the mechanistic model to obtain suggestions for feed improvements, by identifying pathways related to the PoI and supernatant metabolites.

3.3.1. Identifying Pathways Relevant to PoI Production

We identified the pathways related to the PoI production by computing the correlation between the model pathway rates and the PoI formation rate, using the m_modCG_RN‐370_OMX model (Figure 4A). Seventy pathways had a significant correlation (p adj. < 0.05) with the PoI formation rate, including 21 pathways with negative correlation and 49 with positive correlation. The pathway fluxes ( w ) were plotted against this correlation (Figure 4B), revealing that several pathways were markedly high. Focusing on these, the pathways (25) with flux larger than 20 were selected for further analysis. In general, the rates of positively correlated pathways were higher for Process B, compared to Process A (Figure S5).

FIGURE 4.

Interpretation of model m_modCG_RN‐370_OMX. (A) Steps taken to identify the pathways related to the PoI production. (B) Pathway (EFM) activity (flux, y‐axis) and correlation to PoI formation (x‐axis) for each pathway significantly (p adj. < 0.05) negatively (blue) or positively (red) correlated to PoI formation. (C) Metabolites in positively correlated pathways (consumed+, produced+) and negatively correlated pathways (consumed−, produced−). “Feed” refers to metabolites fed more during Process B than A; “SD” to highly SD variable metabolites; “Diminished” to diminished metabolites.

3.3.2. Metabolites in Pathways Related to PoI Production

The extracellular metabolites involved in the identified pathways were listed (Table S5), and it was determined which metabolites were included in the subsets of feed, differing‐A‐vs‐B, diminished, and/or highly SD variable (see table of Figure 4C). From this, several main conclusions were drawn for these identified pathways. There were 7 differing‐A‐vs‐B, 2 diminished, and 2 highly SD variable metabolites, as well as 7 in the feed. Among the metabolites in the feed, consumed with positive correlation with PoI formation, putrescine was diminished and highly SD variable during the process but not differing‐A‐vs‐B, while serine and ornithine were differing‐A‐vs‐B. This indicated that feeding these metabolites improved the process performance, and that increasing the feed of putrescine could be favorable since it was close to depleted. Present as well in the feed, aspartate, 5AVA, and citraconate were differing‐A‐vs‐B, and consumed in both positively and negatively correlated pathways to the PoI, indicating that feeding aspartate, 5AVA, and citraconate should be further investigated due to this dual correlation. The differing‐A‐vs‐B guanine and highly SD variable uridine were both accumulating during the process. Interestingly, they were produced by positively correlated pathways to the PoI and consumed by negatively correlated pathways to the PoI, suggesting that reducing these metabolites in the base medium (not present in feed) or their precursors in the feed could be beneficial.

3.4. Reactions of Pathways Related to the PoI Production

To elucidate mechanistic details, we investigated the reaction rates in the identified pathways (Figures S6–S14). These reactions were involved in the central metabolism and in the extended reactions, added to network RN‐127, and generated RN‐370. Reactions from the central metabolism involved parts of glycolysis, the TCA‐cycle, anaplerotic reactions, nonessential amino acid metabolism, and the urea cycle. The extended reactions involved 5AVA, putrescine/spermidine, citraconate, and purine and pyrimidine metabolism. The reactions of the central metabolism provided mechanistic connection to aspartate and ornithine, while the extended reactions connected ornithine, N‐acetyl‐putrescine, N1N8‐spermidine, citraconate, guanine, deoxycytidine, putrescine, and spermine, highlighting the role of this extension to achieve mechanistic explanations. Reactions in the significant pathways were interconnecting several parts of the metabolism, for example, pyruvate was interconnecting glycolysis, the TCA cycle, pyrimidine, citraconate, and nonessential amino acid metabolism, while acetylCoA was interconnecting the TCA cycle, 5AVA, spermidine, citraconate metabolism, and ornithine was interconnecting nonessential amino acid metabolism, the urea cycle, and spermidine metabolism. This suggests the presence of coordinated metabolic changes as a response to these feed medium components.

Positively related reactions were participating in acetylCoA to CoA cycling by utilizing carnitine, putrescine, spermidine, and citraconate (Figure 5A), suggesting that acetyl transfers could support increased productivity. These reactions were connected to several differing‐A‐vs‐B and diminished metabolites: acetyl carnitine, N‐acetyl‐putrescine, putrescine, N1N8‐spermidine and citraconate. AcetylCoA levels remained more stable during Process B, and levels of acetylated metabolites were different between processes (Figure 5B).

FIGURE 5.

Investigating individual model reactions. A: Reactions in pathways (EFMs) positively correlated to PoI formation. Bar intensity indicates the reaction rate. Metabolites differing between Processes A versus B, (magenta), diminished (grey), or extracellular (blue) are highlighted. B: Levels of acetylated metabolites: acetylCoA (AcCoA), acetyl carnitine (Acrn), N‐acetyl‐putrescine (NaPutr), and N1N8‐acetyl‐spermidine (N1N8Sperm). Extracellular metabolites are marked with “ext.”

4. Discussion

Adding metabolome data to metadata can potentially support identifying key metabolites and metabolic pathways associated with high bioproduction yields. However, untargeted metabolome datasets are large, do not provide absolute concentrations, and are scarce due to their high cost. In this study, we designed an analysis framework to assess nutrient supplementation effects, Process B versus A, on increased PoI productivity observed in an industrial fed‐batch process. We investigated methods to identify key components involved in productivity change and utilized mechanistic modeling to simulate the involved metabolic pathways. Additionally, metabolic pathways and individual reactions related to bioproduction yield were identified.

4.1. Identification of Key Metabolites

No clear distinction between the two processes emerged from the full metabolome dataset, highlighting the need to identify key metabolites. However, a clear distinction between process days was observed in the full dataset, which is in line with others [34, 35, 36]. Thus, identifying the critical metabolites among all the wide dataset metabolites was crucial.

Detecting diminished metabolites from one process and differing‐A‐vs‐B metabolites from the two‐process dataset generated the identification of metabolites related to variations between processes. Statistical testing alone [35, 37], or paired with fold change criteria (differential analysis [38]), has been proven successful. Utilizing statistical tests alone avoids arbitrary cutoff definitions. Furthermore, approaches for selecting relevant subsets provide an actionable number of target medium components for feed engineering. However, the metabolite variability was not relevant to segregate the processes here, which is why we do not recommend this approach.

Only a few intracellular metabolites were differing‐A‐vs‐B, contrasting with 10 times more differing‐A‐vs‐B extracellular metabolites. Others have also reported infeasibility of exploiting intracellular metabolites [35, 36]. This is likely due to the effort of the cells to maintain internal homeostasis [39]. Therefore, we recommend prioritizing the extracellular metabolome for process optimization. Intracellular metabolite data were, however, were important for the modeling, as discussed below.

Diminished and differing‐A‐vs‐B metabolites included unidentified metabolites, consistently appearing as key metabolites, for which their identification could be important for process improvement overall, benefiting academia and industry. Some metabolites, such as o‐Tyrosine, accumulated with time, and reducing their feeding is recommended.

We identified N‐acetylated amino acids, N‐acetylserine and N‐acetylproline, related to process differences. Previously, N‐acetylated amino acids have been related to CHO cell culture phases [40]. Acetylated amino acids can arise from acetylated terminal amino acids during protein degradation. Interestingly, the acetylation of nonproteogenic amino acid taurine N‐acetyltaurine by AcetylCoA and acetate [41] indicates that proteinogenic amino acids could act similarly. Acetate arises from ketogenesis, suggesting N‐acetylated amino acids as ketogenic state indicators and potential cell markers for process monitoring. As excess AcetylCoA suppresses the flux from glycolysis to the TCA cycle [42], increased taurine feeding could be beneficial for maintaining the AcetylCoA level.

4.2. Generation of Mechanistic Metabolic Models Using Metabolomics Data

For the first time, we created EFM‐based modeling with metabolome data. Considering the metabolome key metabolites as input and extending accordingly the network to 370 reactions, RN‐370, generated model m_modCG_RN‐370_OMX of moderate complexity and fitting error ≤ than the measurement error. However, considering these key metabolites and the metadata data with the 127‐reaction network, RN‐127, generated a higher model complexity and error, highlighting the benefit of the m_modCG_RN‐370_OMX modeling approach. Previous studies [40], have expanded the reaction network with metabolomics data; however, studies utilizing large sets of untargeted metabolomics data as modeling input are lacking. Furthermore, while the intracellular metabolites alone give unexploitable information, in combination with the present modeling approach, they provide valuable support and are recommended to be included for modeling purposes.

Similar to ref. [35], several identified key metabolites lacked knowledge about their metabolic connections, and thus important metabolic reactions might be missed during modeling. Genome‐scale networks contain extensive reaction sets; however, with some incompleteness, appealing for further case‐by‐case investigation to assess the metabolite relevance to a process. However, the very detailed genome‐scale networks and increasingly more extensive small‐scale networks complement and mutually benefit each other. Fully mapping the reactions of all the key metabolites associated with process changes is crucial for bioproduction process understanding. A good example is taurine, which was quickly depleted here, yet lacks consumption reactions in literature.

4.3. Interpretation of the Model for PoI Production

We achieved understanding of the metabolism and identified potential feed improvements by performing correlation‐based analysis of model modCG_RN‐370_OMX. Pathways positively (or negatively) correlated with PoI production exhibited higher (or lower) rates during Process B than A, suggesting that they captured production differences distinguishing Process B from A. Thus, these correlated pathways mechanistically bridged diminished and differing‐A‐vs‐B metabolites to productivity differences. This approach highlighted aspartate, putrescine, 5AVA, citraconate, ornithine, and serine for feed improvement.

Here we identified a limited set of metabolic reactions that were related to PoI productivity; the metabolites key for the PoI were linked to the central cell metabolism via nonessential amino acid metabolism, the urea cycle, an extended network related to 5AVA, spermidine, citraconate, purine and pyrimidine metabolisms, and included glycolysis, the TCA cycle, and anaplerotic reactions, where pyruvate, acetylCoA, and ornithine were interconnecting intermediates. It remains to be studied whether targeting these metabolic systems is a medium development strategy that could be used for other bioproduction processes as well. As reported by others, the TCA cycle is commonly associated with increased productivity [43, 44, 45]. The addition of nonessential proteinogenic amino acids, such as asparagine, aspartate, and serine, has been shown to impact bioproduction performance [46]. Furthermore, polyamines, such as putrescine and spermidine, have been shown to improve bioproduction process performance [47], therefore suggesting that the addition of metabolites part of spermidine metabolism could be a generally feasible strategy. To our knowledge, the addition of citraconate/clutaconate as well as 5AVA has not been studied and there is scarce information regarding utilizing purine and pyrimidine metabolism intermediates as bioprocess optimization targets [48].

PoI positively correlated reactions were associated with acetyl transfers and connected to several differing‐A‐vs‐B and diminished metabolites. Furthermore, there were differences in the acetylated metabolites between the processes. Acetylated metabolites and AcCoA are also identified in many other CHO‐cell productivity studies, yet their exact role and mechanism remain unresolved [43, 45].

Partial least squares‐based methods can connect metabolites to productivity [43, 44]. They do not require reaction network knowledge and rely purely on correlation, which lacks mechanistic understanding. Another approach, pathway enrichment analysis (PEA), can highlight metabolic changes [35, 36, 37, 43, 44, 49]. but is unable to illustrate detailed mechanistic linking to cell culture medium components. On the other hand, PEA does not require a reaction network, unlike mechanistic metabolic models. Both PLS‐based and PEA can be used in parallel [43, 44]. The approach described in this paper enables inspecting reactions connecting key metabolites via the reaction network, in contrast to the identification of individual pathways in PEA.

Mechanistic modeling provides a valuable tool to simulate metabolic pathways but relies on the accuracy and completeness of the underlying metabolic reaction network. The incomplete actual knowledge of reactions and regulations involved in CHO (or mammal) cell metabolism led here to a partial lack of identified pathways in case reactions or metabolites were missing from the network, leaving an incomplete understanding of the involved metabolic mechanisms. Due to economical and practical constraints, it was not possible to experimentally validate the present findings of feed improvement and to repeat the metabolomics. Worth noting, several of the identified metabolites were already known to be important for the process performance, which supports the present approach. The runs of processes A and B, which were sampled for the metabolome data, had different scales; however, the present study focused on aspects of the metabolic behavior that were known to be similar between different scales for both processes and thus were not affected by the process scale. Therefore, it was assumed that the effect of the scale in the analysis framework was negligible in the present study. Despite these limitations, the demonstrated analysis framework successfully identified possible beneficial process changes from large metabolome data.

5. Conclusions

In conclusion, we successfully demonstrated an analysis framework integrating metabolite identification, mechanistic modeling, and pathway identification for assessing the effects of nutrient supplementation on CHO cell metabolism and productivity in a fed‐batch process.

Key metabolites can be identified from single‐ or two‐process metabolome data by investigating diminished levels or differing‐A‐vs‐B concentrations between processes. Interestingly, they are more consistently identified from supernatant compared to the intracellular metabolome, and mechanistic modeling relies mainly on the extracellular metabolome, while intracellular metabolome data bring support. We demonstrated a model interpretation method based on identifying pathways correlated to PoI productivity, providing mechanistic connection between medium components and cell metabolism. Thus, here the framework reduced the data dimensionality, revealing key metabolites and associated metabolic pathways linked to PoI production. Our findings highlight the potential of integrating omics data and modeling for bioproduction improvement. Future studies could address the limitations of this study by incorporating additional omics data and improving the completeness of the metabolic reaction network.

Author Contributions

Meeri E.‐L. Mäkinen: Conceptualization (lead); methodology (lead); formal analysis (lead); visualization (equal); writing–original draft (lead); writing–original draft (lead); Writing–review and editing (equal). Markella Zacharouli: Formal analysis (supporting); visualization (equal); writing–original draft (supporting), Sigrid Särnlund: Investigation (supporting); data curation (supporting); visualization (equal); writing–original draft (supporting), Yun Jian: Investigation (lead); data curation (lead); Conceptualization (supporting); funding acquisition (equal). Veronique Chotteau: Conceptualization (supporting); funding acquisition (equal); visualization (equal); project Administration (lead); writing–original draft (supporting); Writing–review and editing (equal).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting information

Data Availability Statement

Data sharing is not applicable to this article, as no new data were created or analyzed in this study.