All model validations were performed with a physiological pH of 7.0. The growth behavior was tested in several media with access to varying carbon sources under aerobic and anaerobic conditions to validate the predictive power of the curated model iCGB21FR.

As C. glutamicum is widely used in biotechnology, the model's capabilities can be used to yield hints for metabolic engineering. All subsequent analyses were performed using the CGXII minimal medium with D-glucose as the sole carbon source under aerobic conditions.

The new GEM of Corynebacterium glutamicum constructed in this work is named iCGB21FR. This name follows the latest recommended naming conventions, which are part of the community standardization of metabolic models (Carey et al., 2020). The lower-case “i” in italics means in silico, followed by the species indicator “CG” for C. glutamicum. “B” represents the city where the particular strain ATCC 13032 was sequenced (Bielefeld, Germany, see also Kalinowski et al., 2003). The three-letter code “CGB” also serves the corresponding strain identifier in the KEGG pathway database (Kanehisa et al., 2019). It follows an iteration identifier, in this case, the year 21 of this century. The last two characters, “FR,” refer to the last names of the primary model curators.

The model iCGB21FR is available in the SBML Level 3 Version 1 format (Hucka et al., 2019) with the fbc plugin (Olivier and Bergmann, 2018) and the groups plugin (Hucka and Smith, 2016) enabled. It contains 1042 metabolites, 1539 reactions, and 805 genes. Thus, further 42 reactions, 13 metabolites, and 25 genes were added to the model following the initial draft reconstructed with CarveMe (Machado et al., 2018). All metabolites and reactions have a human-readable, descriptive name and a chemical formula. The model comprises the cytosolic, periplasmic, and extracellular compartments.

Its overall Memote score amounts to 87 %. The Memote score of the initial draft model created with CarveMe was 33 %. With intensive manual curation, the number of mass and charge imbalanced reactions could be diminished from an initial 170 to 19 imbalanced reactions. These represent 1.2 % of the total number of reactions. The model has a stoichiometric consistency of 99.7 % and does not contain any energy-generating cycles, dead-end metabolites, nor orphan metabolites.

Seventeen different databases are cross-referenced in the model's instances, yielding a Memote annotation score of 84 % for reactions, 84 % for metabolites, and 49 % for genes. Genes include cross-references to the three databases KEGG (Kanehisa et al., 2019), NCBI genes (Maglott et al., 2005), and NCBI proteins (Pruitt et al., 2005). Metabolites and reaction annotations contain cross-references to 13 and seven different databases, respectively. The databases BiGG (Norsigian et al., 2019), BioCyc (Karp et al., 2019), KEGG (Kanehisa et al., 2019), MetaNetX (Moretti et al., 2021), Reactome (Croft et al., 2010), and ModelSEED (Henry et al., 2010) are referenced for metabolites and reactions. Reactions also have cross-references to the RHEA database (Lombardot et al., 2019) and EC numbers. Metabolites have additional cross-references to the ChEBI database (Hastings et al., 2016), the Human Metabolome Database (HMDB) (Wishart et al., 2007), BioPath (Brandenburg et al., 2004), InChIKey (Heller et al., 2015), UniPathway (Morgat et al., 2011), lipid maps structure database (Sud et al., 2007), and the University of Minnesota Biocatalysis/Biodegradation Database (UM-BBD) (Ellis et al., 2003).

All model instances were further annotated using SBO terms (see Figure 2). While genes and metabolites received general SBO terms for genes and simple chemicals, the model's reactions were annotated with 23 different SBO terms. The most prominent ontology group is “biochemical reactions”: 313 reactions in the model hold the general SBO term for reactions. The number of biochemical reactions is followed by the group of exchange reactions with 181 reactions. The transport reactions are described more precisely by the SBO terms for active, passive, co-, symporter-mediated, antiporter-mediated, or general transport. For all other reactions, we identified SBO terms that describe the occurring biochemical reaction more precisely. In terms of ontology, these SBO terms are child nodes of the SBO term for biochemical reactions. These terms include, for example, redox reactions, the transfer of a chemical group, hydrolysis, or phosphorylation. The SBO terms for ATP maintenance and biomass production occur only once in the model. Figure 2 gives an overview of all 23 added reaction SBO terms and their occurrence in the model.

The plugins fbc (Olivier and Bergmann, 2018) and groups (Hucka and Smith, 2016) are enabled in iCGB21FR, thus allowing information such as metabolic charges, chemical formulas, or gene products to be stored. All identified KEGG pathways (Kanehisa et al., 2019) were added as a group to the model, with all reactions participating in the pathway as group members. In total, 102 groups were added to the model. The group with the most members is the “metabolic pathways” group with 563 members, followed by the “biosynthesis of secondary metabolites” group with 297 members. Other groups with more than 100 members are the “biosynthesis of amino acids” with 104 associated reactions, the “biosynthesis of cofactors” with 151 members, and the group of reactions associated with “microbial metabolism in diverse environments” with 171 members. With the help of these groups, reactions of a particular pathway can easily be extracted and analyzed.

The biomass objective function (BOF) created by CarveMe was refined in several steps to obtain realistic growth rates for the tested media. With the help of BOFdat and the nucleotide sequence of C. glutamicum ATCC 13032, the stoichiometric coefficients of the DNA nucleotides were adapted. The following seven metabolites were added as coenzymes and inorganic ions to the biomass objective function: NADH, NADPH, adenosine monophosphate (AMP), pyruvate, ammonium, sodium, and nickel. The stoichiometric coefficients of 16 further coenzymes and inorganic ions were adapted using BOFdat. The inorganic ion cobalt was removed from the BOF based on the elemental composition of C. glutamicum ATCC 13032 cells, as described by Eggeling and Bott (2005, p.16, 18). After including these changes, the simulated biomass production is in the range of a reasonable growth rate with no blocked biomass precursors in both the default and the complete medium.

We simulated the growth of iCGB21FR in different media under aerobic and anaerobic conditions, and with access to different carbon sources (see Figure 3). Growth was tested on the two minimal media, M9 and CGXII, and the complete LB medium. The heat map in Figure 3A gives an overview of the growth behavior of C. glutamicum in the three different media under aerobic and anaerobic conditions. With 1.0266 mmol gDW^-1 h^-1 the biomass consistency of iCGB21FR is close to 1 mmol gDW^-1 h^-1. Consequently, it is approximately possible to directly compare the in-silico growth rate to an experimentally obtained growth rate given in 1 h^-1. The simulated aerobic model growth on the minimal medium M9 with glucose as a single carbon source resulted in a maximal growth rate of 0.57 mmol gDW^-1 h^-1. A maximal realistic aerobic growth rate of 0.57 mmol gDW^-1 h^-1 on CGXII was obtained using the simulation tools COBRApy (Ebrahim et al., 2013) and confirmed with the Systems Biology Simulation Core Library (SBSCL) (Panchiwala et al., 2021). The simulated value is only slightly lower than the growth rate of 0.61 h^-1 that Unthan et al. (2014) could experimentally obtain. The model simulates growth on the complex LB medium with a growth rate of 1.0214 mmol gDW^-1 h^-1 under aerobic conditions without further adjustments or refinements. As expected, the growth rate in the complex medium (LB) is approximately twice as high as in the two minimal media (M9 and CGXII). The aerobic growth conditions in all three media show a higher simulated growth rate compared to the anaerobic conditions, as anticipated. All growth rates are within a realistic range (Unthan et al., 2014).

The growth of C. glutamicum in CGXII minimal medium under aerobic and anaerobic conditions with varying carbon sources was tested. Under aerobic conditions, the model simulated biomass production on all carbon sources. Aerobic growth was possible on all carbon sources. The growth rates varied between 0.8437 mmol gDW^-1 h^-1 on sucrose and 0.2401 mmol gDW^-1 h^-1 on acetate. In our anaerobic in silico experiments, biomass production was also possible on three of the experimentally validated carbon sources, but not on ribose. Additionally, gluconate can be used as carbon sources. The biomass production on gluconate yielded a rate of 0.0945 mmol gDW^-1 h^-1.

The new model of C. glutamicum can simulate the production of all 20 canonical amino acids while growing on the CGXII medium with D-glucose as the carbon source under aerobic conditions. In Figure 4, each square represents the result of a single FBA with the objective to maximize the corresponding amino acid production. The color indicates the amino acid production rate with respect to the glucose uptake rate. The positioning represents the ATP requirements and CO₂ production in relation to the amino acid production rate. The two amino acids, l-aspartate (asp) and l-alanine (ala), have the highest absolute amino acid production rates with 11.77 mmol gDW^-1 h^-1 and 13.73 mmol gDW^-1 h^-1, respectively (see also Zelle et al., 2015). In contrast, the amino acids l-histidine, l-arginine, and l-tryptophan have the lowest amino acid production rate and the highest ATP requirements. A relationship exists between the yield of amino acid production, energy expenditure, and CO₂ production: The more ATP is required, the more CO₂ produced and the lower is the amino acid production rate. L-glutamate is of particular interest for metabolic engineering in C. glutamicum. Its total production rate under the selected conditions yields 8.7 mmol gDW^-1 h^-1.

C. glutamicum is a well-known L-glutamate producer. However, L-glutamate is also required for the growth or maintenance function. L-glutamate accounts for the growth function with a stoichiometric coefficient of 0.0149. Thus, a trade-off between growth requirement and the production of L-glutamate is expected. This trade-off is depicted in Figure 5. For growth rates between 0 mmol gDW^-1 h^-1 and 0.4 mmol gDW^-1 h^-1, the L-glutamate production rate remains comparably high. It only decreases by 5 mmol gDW^-1 h^-1. With increasing growth rates greater than 0.4 mmol gDW^-1 h^-1, the L-glutamate production rate decreases rapidly.

The PC plays a pivotal role in L-glutamate production (Peters-Wendisch et al., 2001). The effect of a knock-out of the PC on the flux distribution is depicted in Figure 6. Knocking out the PC decreases the L-glutamate production only to a small extent (from 7.31 mmol gDW^-1 h^-1 to 7.26 mmol gDW^-1 h^-1). The limiting factor in L-glutamate production is the availability of a carbon source (in this example, D-glucose) and, as shown above, the growth rate. The PC in silico knock-out experiment indicates that C. glutamicum can compensate for the knocked-out reaction.

Performing FVA helps to identify potential the ranges of each flux. Reactions relevant for optimizing the objective function can be identified by filtering for reactions with almost identical minimal and maximal flux values. With loopless FVA, we identified six highly relevant reactions for L-glutamate production in C. glutamicum. Among these six reactions were two pseudo-reactions: the exchange reaction of D-glucose and the sink reaction for L-glutamate. Glucose is the sole carbon source in the in silico experiment. Therefore, its strong influence on L-glutamate production is apparent. The same holds for the sink reaction that was used as the objective function in the FVA. The other four relevant reactions include the aconitate hydratase (ACONT), which converts citrate to isocitrate, the citrate synthase (CS), which converts acetyl-CoA and oxaloacetate to citrate and coenzyme A, the glucose transport via phosphoenolpyruvate, and the isocitrate dehydrogenase (ICDHyr), which converts isocitrate to 2-oxoglutarate (see also Table 1). The reactions ACONT, CS, and ICDHyr represent the fragile connection between glycolysis and L-glutamate biosynthesis. This connection can additionally be seen in Figure 6, where the three mentioned reactions are also illustrated.

The comprehensive annotations of all model components, including metabolites, reactions, and genes, contribute to the high quality of the reconstruction of iCGB21FR. Each instance is uniquely referenced to at least one database, thus providing a permanent link to clearly and uniquely identify this instance with its attributes (Juty et al., 2012). Almost all model instances are annotated with references to a minimum of one database, allowing more precise cross-referencing and interoperability between different databases. This high level of annotations is advantageous, as findable, accessible, interoperable, reusable (FAIR) data principles allow fellow scientists to conduct research on and with this model continuously (Dräger and Palsson, 2014; Wilkinson et al., 2016). Erroneous or missing information, incompatible data formats, or missing annotations significantly hamper the reuse of GEMs (Ravikrishnan and Raman, 2015). Missing annotations can lead to identification problems of compounds and reactions. GPRs are added for different reactions, and all instances are equipped with SBO terms, facilitating FAIR data principles.

Generally, the high degree of annotations in iCGB21FR is confirmed by the high Memote scores within the different categories. In terms of the presence of annotations, almost all Memote scores of the annotation categories rank close to 100 %. This high score implies that almost all suggested standards concerning annotations are met for this GEM (see also Carey et al., 2020). The current version of Memote does not include all C. glutamicum-specific databases, while other organism-specific databases with less relevance for our model are incorporated. One example can be found in the annotations section of the genes, where cross-references to different Escherichia coli databases are checked. In iCGB21FR, the Memote score for the presence of SBO terms for biochemical reactions sticks out due to its comparatively low value. The current version of Memote checks every reaction for the annotation with the most general SBO term (SBO:0000176), “biochemical reaction.” This check implies that Memote can not yet capture the fine-grained description of biochemical reactions in this model. Thus, the score of the metabolic reactions of 33.4 % diminishes the overall Memote SBO term annotation score.

Two typical ways exist to calculate the biomass objective function (BOF) of an organism. These are the macromolecular-based and the sequence-based approach. A typical biomass objective function (BOF) comprises the cell's primary macromolecules, essential coenzymes, inorganic ions, and species-specific metabolites, including the cell wall components. Additionally, the energy requirements for growth and non-growth associated maintenance costs are included (Lachance et al., 2019). Using an experimentally derived biomass composition implies that its cellular composition depends on the experimental conditions under which it was obtained. For example, the availability of nutrients and the resulting growth rate influence the ratio between DNA, RNA, and proteins (Scott et al., 2010). It thus represents a biased approach to compute the biomass. When no species-specific experimental data for the sequence-based approach is available to calculate a species-specific biomass function, a universal bacterial biomass function is included. Adapted biomass composition of a highly developed and curated GEM of E. coli is often used (Orth et al., 2011; Xavier et al., 2017). The new model possesses a biomass function adapted specifically to C. glutamicum. This conceptual approach differs from previous works by Kjeldsen and Nielsen (2009) and Zelle et al. (2015). For the BOF of Kjeldsen and Nielsen (2009), a biomass equation and the corresponding energy consumption associated with each reaction were formulated for each macromolecule. No C. glutamicum-specific data for the energy requirement of the polymerization of macromolecules was available; thus, E. coli data was used instead. The BOF of iEZ482 is based on the BOF of Kjeldsen and Nielsen (2009). Using species-specific data forms the basis for models with high predictive value. BOFdat enables the curation and refinement of a species-specific BOF by incorporating various -omics data into its calculation. In this study, genomics data were available and applied to refine the species-specific BOF.

The model iCGB21FR was validated by simulating growth on three different media under aerobic and anaerobic conditions. C. glutamicum is also known to grow in the brain heart infusion (BHI) medium. Modeling requires chemically defined media for growth simulations. We could not test the growth of iCGB21FR in BHI as no exact composition of the chemical definition of this medium exists. Simulating aerobic growth on LB complete medium was possible without any additional refinement of the model. Aerobic growth on the two minimal media was only possible by adapting the biomass function for growth on CGXII and adding missing reactions to the model for growth on M9. Missing reactions were identified by literature research. Anaerobic growth was enabled after adding six reactions. By this, the model was refined step by step to simulate and confirm already known growth conditions.

In silico growth rates were higher on the complete medium compared to the two minimal media. The aerobic growth rates were higher than the anaerobic growth rates, with all growth rates within a realistic range (Unthan et al., 2014; Michel et al., 2015). Both findings are expected, as complete media provide more nutrients and biomass precursors than minimal media. As their name suggest, minimal media only provide minimal required nutrients for the organism to grow. C. glutamicum uses oxygen and the more efficient aerobic respiration. It is even often regarded as aerobe (Takeno et al., 2007). However, as C. glutamicum is facultatively anaerobe, it can also switch to fermentation and anaerobic respiration if oxygen is absent. Anaerobic growth by nitrate respiration is limited, as nitrate accumulates and inhibits growth. Additionally, glucose is converted to l-lactate and succinate without the growth of the organism (Inui et al., 2004b; Koch-Koerfges et al., 2013).

Two observations stick out from these growth results. First, the current tools for the reconstruction of GEMs still demand subsequent manual refinement. Even though automated tools, such as CarveMe (Machado et al., 2018), reduce the amount of time spent on the reconstruction dramatically, manual refinement remains a pivotal part of the reconstruction process. The necessity of manual curation becomes particularly apparent when comparing the growth predictions of our model for the different media. The draft model created by CarveMe enabled growth on the complete medium without further ado. However, manual refinement was essential for the simulation of growth on the two minimal media. The second interesting observation is the organism-specific gap-filling, which appears to be more fruitful when applied to media that specifically compensate for certain physiological or metabolic oddities of the organism. In our case, knowledge gap-filling was most fruitful on the CGXII medium which, for example, compensates for the limited ability of C. glutamicum to synthesize and excrete siderophores (Budzikiewicz et al., 1997). This makes sense, as the minimal medium provides the microbe's bare necessities to grow. Potentially lacking compounds could be compensated by the composition of the complete medium.

We validated our model by testing the growth rate on the CGXII minimal medium under aerobic and anaerobic conditions using different experimentally validated carbon sources. Aerobic growth was possible on all experimentally validated carbon sources. In addition to the experimentally confirmed anaerobic growth on glucose, fructose, sucrose as carbon sources, our in silico model also grew on gluconate. We verified that all genes required to utilize gluconate as carbon sources exist in C. glutamicum. The rate of NADPH reoxidation could be a potential explanation for the in silico growth on the additional anaerobic carbon source. The enzyme 6-phosphogluconate dehydrogenase oxidizes 6-phosphogluconate to ribulose 5-phosphate. This enzyme is inhibited by NADPH, which is essential for the cellular control of the NADPH synthesis (Moritz et al., 2000). The rate of NADPH re-oxidation represents a critical element of this process. Gluconate is phosphorylated after uptake and then catabolized in the pentose phosphate pathway. If NADPH re-oxidation was too low under anaerobic conditions, NADPH could accumulate and result in complete inhibition of 6-phosphogluconate dehydrogenase activity. This accumulation would lead to a stop in growth on gluconate, as was shown by experimental data (Michel et al., 2015). If, however, NADPH re-oxidation is sufficiently fast and no NADPH accumulates, the activity of the 6-phosphogluconate dehydrogenase could remain active and allow anaerobic growth on gluconate by simulation studies with this in silico model.

As a final validation step of the metabolic model, its ability to produce amino acids was examined. In complex bacterial systems, amino acid production co-occurs with growth (Marx et al., 1996). The biosynthesis of amino acids requires a lot of the carbon source's total budget, usually used for bacterial growth (Neidhardt et al., 1990). The growth rate was fixed to 0.4 mmol gDW^-1 h^-1 to ensure the microbe's viability while producing amino acids. With increasing CO₂ production rate and ATP requirements, the amino acid production yield decreases. Especially smaller amino acids with only a few carbon atoms, like l-alanine with only three carbon atoms, or glycine with two carbon atoms, have a low CO₂ production and ATP requirement rate.

In contrast, amino acids with more carbon atoms, such as l-tryptophan, have a much higher ATP requirement and CO₂ production rate, while the amino acid yield is relatively low. This leads to the conclusion that building more extensive and more complex amino acids needs more energy. An increasing proportion of the consumed sugar will be used to produce more energy, is then lost as CO₂, and cannot be used for amino acid formation (Gourdon et al., 2003).

One might expect a linear correlation between the L-glutamate production and the growth rate, where the slope is related to the amino acid's stoichiometric coefficient in the biomass function. The trade-off between the production of L-glutamate and the growth rate was investigated by fixing the growth rates and maximizing the L-glutamate production (see Figure 5). The system moves between these two boundaries: the maximal possible growth rate and the maximal possible production of L-glutamate. The closer the values get to either maximum, the greater is the influence on the respective other value. When ceasing growth, the theoretical production of L-glutamate would reach a maximum since the available metabolic capacity is invested in the L-glutamate production. The reverse situation occurs with ceasing L-glutamate production and maximizing growth where all energetic demand is invested. L-glutamate production reaches its maximum when no growth occurs, and the available glucose is completely used to produce L-glutamate. Thus, the growth rate is the limiting factor for our in silico model, independent of the L-glutamate production.

The PC was first investigated to study relevant reactions for L-glutamate production (see Figure 6). The PC has been described as the bottleneck in the production of L-glutamate (Peters-Wendisch et al., 2001). Knocking out the PC in laboratory experiments leads to ambivalent results: Both drastic decrease (Peters-Wendisch et al., 2001) and increase (Sato et al., 2008) in L-glutamate production were reported after a disruption of the PC. Pyruvate is part of the complex network responsible for carboxylation and decarboxylation reactions, which connect the glycolysis and TCA cycle (Becker and Wittmann, 2020). In C. glutamicum, the PC represents one of the carboxylation enzymes, the other being the phosphoenolpyruvate carboxylase (Eikmanns, 2005). The carboxylation and decarboxylation enzymatic complex in C. glutamicum is a highly flexible network that enables several pathways to respond to varying metabolic circumstances (Möllney et al., 2000; Becker et al., 2008; Becker and Wittmann, 2020)—knocking out the PC in our model still allowed L-glutamate production. We also found that the amount of produced L-glutamate does not vary significantly with the PC being knocked. According to our simulation, the limiting factors in L-glutamate production are access to carbon sources and the growth rate.

Four reactions were identified that play a pivotal role in the production of L-glutamate in our model. The first is a glucose transporter, which uses the phosphoenolpyruvate (PEP)-dependent sugar phosphotransferase system. The L-glutamate yield decreased with sugar consumption rates in laboratory experiments (Gourdon et al., 2003). This seems reasonable, as glucose is the sole carbon source and starting point for glutamate production. Increasing glucose availability is only expedient if the glucose transporters' capacity is given to take up the enhanced supply of glucose. The three remaining reactions, aconitase, citrate synthase, and isocitrate dehydrogenase, are all part of the tricarboxylic acid (TCA) cycle. The TCA cycle is a complex regulated amphibolic pathway with L-glutamate and l-lysine as derived intermediate products (Bott, 2007).

The aconitase gene is regulated by four transcriptional regulators, indicating a tight control of this enzyme. A C. glutamicum mutant lacking the aconitase gene was glutamate auxotrophic in the CGXII minimal medium with glucose as the carbon source (Baumgart et al., 2011). The model iCGB21FR could replicate the finding that aconitase is essential for L-glutamate production. It remains to be further experimentally validated how the interplay between the aconitase within the TCA cycle in terms of L-glutamate production can be optimized.

The citrate synthase catalyzes the initial reaction of the TCA cycle. Overexpression of the citrate synthase can redirect more carbon flux into the cycle and result in higher l-arginine production (Man et al., 2016). l-arginine is synthesized from the precursor l-glutamic acid (Utagawa, 2004). Thus, higher production of L-glutamate might also be dependent upon the activity of the citrate synthase. The role of the citrate synthase in L-glutamate production might be an interesting topic to investigate since it might represent a target for metabolic engineering of C. glutamicum's TCA cycle. Since the citrate synthase is the initial reaction of the TCA cycle with L-glutamate and l-lysine as intermediates, its activity might prove particularly fruitful.

Isocitrate dehydrogenase catalyzes the oxidative decarboxylation of isocitrate. Becker et al. (2009) found in their investigation of the effects of the isocitrate dehydrogenase on l-lysine production that decreased activity of the isocitrate dehydrogenase improves the l-lysine production. This decrease induced a flux shift from the TCA cycle to anaplerotic carboxylation (van Ooyen et al., 2012). However, the PC functions as an anaplerotic enzyme in L-glutamate production (Peters-Wendisch et al., 1997). In other words, the isocitrate dehydrogenase has different functions in L-glutamate production than in l-lysine production. This differing function becomes more apparent when looking at the effects of isocitrate dehydrogenase inactivation in C. glutamicum: Inactivation of the NADP-dependent isocitrate dehydrogenase in C. glutamicum leads to L-glutamate auxotrophy (Eikmanns et al., 1995). This connection between the PC and the isocitrate dehydrogenase in L-glutamate production might be an interesting target for metabolic engineering.