Background

Corynebacterium glutamicum, a Gram-positive soil bacterium, is an important platform organism for industrial biotechnology and environmental remediation [1, 2]. This microorganism is generally recognized as safe and is used in the fermentative production of l-glutamate and l-lysine on a million-ton scale and l-threonine, l-leucine and l-valine on a thousand-ton scale in the food, feed and pharmaceutical industries [1, 3]. Due to the metabolic flexibility, broad substrate spectrum, and fermentative robustness of C. glutamicum cells, this bacterium has attracted attention as a potential cell factory to manufacture organic acids, bioplastic monomers (e.g., putrescine and cadaverine), and biofuels including ethanol, isobutanol, 1,2-propanediol, and 1-propanol [1, 4, 5]. Based on the knowledge of biochemical pathways and microbial fermentation processes, new technologies such as systems metabolic engineering and high-throughput analyses at the genome, transcriptome, proteome, and fluxome levels have been applied to strain development to improve the yield and productivity of desired products [6,7,8,9,10]. In silico genome-wide predictive simulation of metabolic profiles is expected to enable the design of artificial metabolic networks for desired product synthesis [11].

A genome-scale metabolic model (GEM) is constructed based on experimental data and information on gene annotations and functions, metabolites, reactions, enzymes, and their interactions [11, 12]. GEM has become an essential tool for understanding cellular metabolism, characterizing cell phenotypes, designing mutant strains with desired properties, and assessing the effects of genetic intervention and environmental perturbation on cellular metabolism [13]. The quality of a GEM determines the accuracy of outcomes for describing the genotype–phenotype relationship of a given strain [13, 14]. The first GEM of Escherichia coli (iJE660) consisted of 660 genes, 627 reactions, and 438 metabolites. iJE660 was constantly updated to iJO1366 by increasing the numbers of new genes and reactions, expanding the gene-protein-reaction associations (GPRs), balancing the reactions and compartmentalizing metabolites, among other updates [15,16,17,18]. The updated GEM accurately characterized E. coli growth metabolic profiles and was used to perform computational and quantitative analyses to resolve problems [14]. When coupled with metabolic flux constraints and a cellular objective, GEM analysis using computational algorithms [e.g., flux balance analysis (FBA), OptKnock, or OptForce] can predict the modification of targets in a metabolic network and successfully improve the production of succinic acid, ethanol, and malonyl-CoA in E. coli [19,20,21].

The initial GEM of C. glutamicum ATCC13032 (ModelCg 1) consisted of 247 genes, 446 reactions, and 411 metabolites and was updated to ModelCg 2 (277 genes, 502 reactions, and 423 metabolites) by revising inadequate reaction loops [22, 23]. Although these two models are useful, they have several limitations due to previously inadequate information. Annotated genes known to participate in different metabolic pathways, such as pyruvate kinase isoenzyme, glycerol-3-phosphatase, and (S,S)-butanediol dehydrogenase, were not included in ModelCg 1 or in ModelCg 2 [24,25,26]. The deficiency of corresponding metabolic pathways in the two models prevented the prediction of cell growth on certain carbon sources (e.g., acetate, lactate, and xylose). The imbalance of the mass and charge in a few reactions (e.g., synthesis of RNA, protein, and biomass) and the directionality of certain reactions without thermodynamic constraints had a negative effect on genome-wide prediction using the two models. Moreover, the two models have incomplete l-glutamate metabolic systems, which prevented their application in studies of the metabolic state of l-glutamate production in the stationary phase [22, 27]. In 2016, Mei et al. constructed iJM658 based on the genome information of C. glutamicum S9114, which was generated by mutagenesis and is widely used for industrial l-glutamate production [27]. iJM658 contains l-glutamate secretion and uptake systems, which could be more suitable for studying l-glutamate production [27]. iJM658 is not applicable for the prediction of metabolic profile of C. glutamicum ATCC13032 due to the presence of numerous differences in the genome sequences between C. glutamicum S9114 and ATCC13032. Therefore, it is necessary to expand the scope of C. glutamicum ATCC13032’ GEM to metabolic engineering for the desired products.

In this study, we present an updated model of the C. glutamicum ATCC13032 metabolic network reconstruction. This new model, named iCW773, accurately predicted the growth ability and flux distributions of cells cultivated under different growth conditions. iCW773 simulations of genetic interventions for the overproduction of l-lysine, 1,2-propanediol, and isobutanol were highly consistent with experimental data. Along with wet-lab practices, genetic modifications were performed to generate a de novo l-proline-overproducing strain that exhibited the highest reported titer and productivity using a minimal medium in fed-batch fermentation.

Results and discussion

A new genome-scale metabolic model of C. glutamicum ATCC13032

We constructed a new GEM for ATCC13032, iCW773, which consisted of chemical reactions and interconverted metabolites. This network reconstruction contained 773 genes, 951 metabolites, and 1207 reactions, representing considerable increases in the numbers of genes, reactions and metabolites by 2.79-, 2.41-, and 2.25-fold, respectively, over ModelCg 2 (Additional file 1: Table S1). In the internal reactions of iCW773, 709 ORFs (89.18%) were extracted from the KEGG and UniProt databases [28]. The total ORF coverage was 25.75% in iCW773, which was 16.52% higher than that in ModelCg 2 (Additional file 1: Table S1). Five lumped reactions for fatty acids biosynthesis in ModelCg 2 were refined in iCW773 by replacing each lumped reaction with several sequential reactions in iCW773 (Additional file 2). Additionally, the charge and mass of reactions in iCW773 were balanced, and the mass of reactions related to biomass was tested to be 1.00 g/g dry cell weight (DCW) (Additional file 3). The directionality of reactions was constrained by the Gibbs free energy Δ r G′, which was calculated using eQuilibrator2.0 [29]. A complete list of all the reactions and metabolites for iCW773 is provided in Additional file 3, and the central carbon metabolic pathway is presented in Additional file 4: Figure S1.

Prediction of growth rates under different culture conditions

To evaluate the quality of iCW773, the growth rates of C. glutamicum ATCC13032 cultivated on different carbon sources were predicted through FBA simulations using biomass maximization as the objective function. First, we calculated the growth rates on glucose using GEMs by setting the glucose and oxygen consumption rates as the experimental values [22, 30,31,32]. As shown in Fig. 1a, the Pearson’s correlation coefficient (PCC) between the experimental and iCW773-simulated data was consistent with that between the experimental and ModelCg 2-simulated data. When fructose, fructose mixed with glucose, sucrose, acetate, and lactate were set as the carbon sources, the cell growth rates on acetate or lactate could be calculated only by iCW773, rather than ModelCg 2 (Fig. 1b) [33,34,35,36,37]. It has been reported that C. glutamicum was able to grow on non-native glycerol or xylose as the sole carbon source after expressing heterologous-related enzymes [38,39,40]. When the three reported reactions catalyzed by glycerol kinase, glycerol 3-phosphate dehydrogenase, and aquaglyceroporin from E. coli were added, both GEMs could calculate the biomass from glycerol [38, 39]. In contrast, only iCW773 could simulate the biomass accumulation from xylose (Fig. 1b) when two reported reactions catalyzed by xylose isomerase and xylulokinase from E. coli were added [40]. These results provided evidence that supplemental reactions in iCW773 coupled the metabolism of the corresponding substrates (e.g., acetate, lactate, and xylose) with the central carbon metabolism (Additional file 5), resulting in the realization of growth simulations on these substrates.

Fig. 1
figure 1

Comparison of growth rates between experimental data and in silico simulation results at various glucose and oxygen supply levels (a) and different carbon sources (b) using iCW773 (left) and ModelCg 2 (right). PCC is the Pearson’s correlation coefficient between the predicted product yield values and experimental data. The line corresponding to x = y is also shown. a The glucose uptake rate is shown in regular font in parentheses, and the oxygen uptake rate is shown in italics in parentheses. b The uptake rate of the corresponding carbon source is shown in parentheses

Full size image

Metabolic profile prediction using iCW773

Next, we compared model-predicted metabolic flux distributions to 13C-tracer metabolic flux data from ATCC13032 cultivated in minimal medium using glucose as the sole carbon source [8, 31, 41]. The biomass production was used as the objective function, and all flux data were normalized to the glucose uptake rate (100%). Notably, the PCC between the experimental and iCW773-simulated fluxes was 0.99, suggesting that iCW773 provided an accurate intracellular flux distribution of the WT strain growing on glucose (Fig. 2a).

Fig. 2
figure 2

Comparison of the metabolic flux profiles of the central metabolism between the 13C-tracer experimental WT (a) and WTΔpgi (b) data for the exponential growth phase and in silico simulation results using iCW773 (left) and ModelCg 2 (right). Fluxes were normalized to the glucose uptake rates, which were set to 100. PCC is the Pearson’s correlation coefficient between the predicted product yield values and experimental data. The line corresponding to x = y is also shown in the inserted figure in the top right corner. EMP: Embden–Meyerhof pathway (glycolytic pathway); PPP: pentose phosphate pathway; TCA: tricarboxylic acid cycle. a The experimental data were derived from three studies, and the error bars represent the standard deviations [8, 31, 41]

Full size image

It has been reported that the inactivation of pgi-encoding glucose-6-phosphate isomerase can lead to a significant metabolic flux disturbance, blocking the EMP pathway and redirecting the carbon flux from glucose-6-phosphate toward the PPP pathway in C. glutamicum [31]. To analyze the metabolic disturbance by gene knockout, the pgi gene was deleted in iCW773 and ModelCg 2. When FBA was used to calculate the intracellular metabolic flux, the iCW773-predicted flux toward the PPP pathway increased from 38.49 to 98.59%, which was consistent with the experimental values (from 44.10 to 97.80%) (Fig. 2b) [31]. Moreover, the simulated amount of NADPH generated from the PPP pathway in WT∆pgi was 1.97 mol/mol glucose, which was 2.30-fold higher than that generated from the PPP pathway in WT (0.77 mol/mol glucose). Due to the requirement for a certain amount of NADPH to generate biomass in the simulation (e.g., 4 mol NADPH per mol l-lysine and 2 mol NADPH per mol l-valine), the NADPH generated from isocitrate dehydrogenase in the TCA cycle decreased from 0.41 mol/mol glucose to 0.22 mol/mol glucose to maintain a stoichiometric balance [42, 43]. This difference in NADPH generation resulted in a smaller deviation from the iCW773-simulated TCA flux to the experimental value, as shown in Fig. 2b (left). Notably, the flux of pyruvate carboxylase (PYRC) in iCW773 was efficiently generated, while no PYRC flux from pyruvate to oxaloacetate was calculated in ModelCg 2. These results illustrate the suitability of iCW773 for predicting intracellular metabolic flux distribution.

Validation of iCW773 simulation for amino acid overproduction

To investigate whether iCW773 is an efficient tool to design strategies for the metabolic engineering of C. glutamicum, the ranges of flux variability in the network for the WT and l-lysine overproducing strains were calculated using OptForceMUST (Additional file 6). As shown in Fig. 3, the gap between the flux ranges of the WT and overproducing network quantified the degree of required reaction flux modification. In the simulations of iCW773, the upregulation of aspartokinase (encoded by lysC), dihydrodipicolinate synthase (encoded by dapA), dihydrodipicolinate reductase (encoded by dapB), diaminopimelate dehydrogenase (encoded by ddh), and diaminopimelate decarboxylase (encoded by lysA) in the l-lysine synthesis pathway were consistent with previous reports [8, 44, 45]. Downregulation of homoserine dehydrogenase (encoded by hom) decreased the flux toward the competitive pathway [8, 44, 46]. Moreover, deletion of phosphoenolpyruvate carboxykinase (encoded by pck) and upregulation of pyruvate carboxylase (encoded by pyc) directed the flux toward oxaloacetate formation at the pyruvate node for l-lysine synthesis [8, 47]. Downregulation of aconitase (encoded by acn) and isocitrate dehydrogenase (encoded by icd) further minimized carbon loss from pyruvate [48]. In addition, upregulation of the entire transketolase operon and fructose-1,6-bisphosphatase (encoded by fbp) enhances the NADPH supply via PPP to increase l-lysine production [8, 49]. In contrast, ModelCg 2 failed to predict the deletion of pck, the downregulation of hom and the upregulation of lysC, ddh, dapB, and fbp. Therefore, the predictions by iCW773 were highly consistent with experimental results. For l-valine and l-serine overproduction, the modification targets predicted by iCW773 were identical to the reported experimental results (Additional file 1: Table S2).

Fig. 3
figure 3

Application for overproduction of l-lysine using OptForceMUST. Comparison of the flux ranges for reactions in iCW773 and ModelCg 2. Genes encoding relevant enzymes are depicted with italics next to the arrows: acn: aconitase; dapA: dihydrodipicolinate synthase; dapB: dihydrodipicolinate reductase; ddh: diaminopimelate dehydrogenase; fbp: fructose-1, 6-bisphosphatase; hom: homoserine dehydrogenase; icd: isocitrate dehydrogenase; lysA: diaminopimelate decarboxylase; lysC: aspartokinase; pck: phosphoenolpyruvate carboxykinase; pyc: pyruvate carboxylase; tal: transaldolase; tkt: transketolase; zwf: glucose 6-phosphate dehydrogenase. 26dap-m: meso-2,6-diaminoheptanedioate; 23dhdp: (2S,4S)-4-hydroxy-2,3,4,5-tetrahydrodipicolinate; accoa: acetyl-CoA; akg: α-oxoglutarate; aspsa: l-aspartate 4-semialdehyde; cit: citrate; e4p: erythrose-4-phosphate; f6p: fructose-6-phosphate; fdp: fructose 1,6-bisphosphate; g3p: glyceraldehyde phosphate; g6p: glucose-6-phosphate; Glc: glucose; icit: isocitrate; l-asp: l-aspartate; l-hom: l-homoserine; l-lys: l-lysine; oaa: oxaloacetate; pep: phosphoenolpyruvate; pyr: pyruvate; ru5p-D: ribulose-5-phosphate; s7p: sedoheptulose-7-phosphate; succ: succinate; thdp: 2,3,4,5-tetrahydrodipicolinate; xu5p-D: xylulose-5-phosphate

Full size image

Validation of iCW773 simulation for non-native 1,2-propanediol and isobutanol synthesis

Corynebacterium glutamicum has previously been engineered for non-native 1,2-propanediol and isobutanol production [50, 51]. Therefore, we performed a comparison between iCW773-simulated and experimentally performed genetic modification strategies with respect to 1,2-propanediol and isobutanol overproduction [50, 51]. As previously reported, a three-reaction synthesis route by the heterologous genes mgsA, gldA, and yqhD, which encode methylglyoxal synthase, glycerol dehydrogenase, and alcohol dehydrogenase from E. coli, was added to iCW773 to synthesize 1,2-propanediol from dihydroxyacetone phosphate (dhap) [50]. Then, the ranges of flux variability in the network for the WT and 1,2-propanediol-overproducing strains were calculated by OptForceMUST (Additional file 7). As shown in Fig. 4a, upregulation of mgsA, gldA, and yqhD directed the flux toward 1,2-propanediol synthesis, which was consistent with the finding that the heterologous expression of these three genes in the WT strain led to a product yield of 0.13 mol/mol (1,2-propanediol/glucose) [50]. Moreover, model simulation showed that the deletion of hdpA encoding dihydroxyacetone phosphate phosphatase could increase the precursor (dhap) supply by preventing glycerol formation from dhap, which was supported by the experimental results that 1,2-propanediol production increased by 90.08% in strain WT∆hdpA (pEKEx3-mgsA-gldA-yqhD) [50]. The OptForceMUST analysis indicated that knockout of ldh encoding lactate dehydrogenase could further decrease the formation of byproducts and increase 1,2-propanediol production. This prediction was validated by the experimental finding that ldh deletion increased 1,2-propanediol by 37.75% [50]. In addition, downregulation of tpiA, gap, pgk, and gpmA would be novel targets for 1,2-propanediol overproduction.

Fig. 4
figure 4

Application for overproduction of the non-natural products 1,2-propanediol (a) and isobutanol (b) using OptForceMUST. Genes encoding relevant enzymes are depicted with italics next to the arrows: aceE: pyruvate dehydrogenase complex; adhA: alcohol dehydrogenase; alaS: acetolactate synthase; gap: glyceraldehyde-3-phosphate dehydrogenase; gldA: glycerol dehydrogenase; gpmA: phosphoglycerate mutase; hdpA: dihydroxyacetone phosphate phosphatase; ilvC: acetohydroxyacid isomeroreductase; ilvD: dihydroxyacid dehydratase; ilvE: valine transaminase; kivD: keto acid decarboxylase; ldh: lactate dehydrogenase; mgsA: methylglyoxal synthase; pgi: glucose-6-phosphate isomerase; pgk: phosphoglycerate kinase; tpiA: triose-phosphate isomerase; yqhD: aldehyde reductase. For other genes, see Fig. 3. 12ppd: 1,2-propanediol; 13dpg: 3-phospho-d-glyceroyl phosphate; 23dhmp: 2,3-dihydroxy-3-methylbutanoate; 2pg: d-glycerate 2-phosphate; 3mob: 3-methyl-2-oxobutanoate; 3pg: 3-phospho-d-glycerate; 6pgl: 6-phospho-d-glucono-1,5-lactone; acetol: acetol; alas-S: 2-acetolactate; dha: dihydroxyacetone; dhap: dihydroxyacetone phosphate; ibth: isobutyraldehyde; isob: isobutanol; l-lac: l-lactate; mthgxl: methylglyoxal; l-val: l-valine. For other abbreviations, see Fig. 3

Full size image

As reported previously, the isobutanol pathway in iCW773 was constructed by kivd-encoding keto acid decarboxylase from Lactococcus lactis and native adhA-encoding alcohol dehydrogenase catalyzing 3-methyl-2-oxobutanoate to isobutanol [52]. The predicted knockout, upregulation, and downregulation of target genes by OptForceMUST are shown in Fig. 4b (Additional file 7). The simulation was consistent with the experimental results that the upregulation of four genes (alsS, ilvCD, kivd, and adhA) could endow C. glutamicum with isobutanol synthesis capability [52]. The OptForceMUST analysis indicated that the decrease in pyruvate consumption caused by inactivating ldh-encoding lactate dehydrogenase and downregulating the aceE-encoding pyruvate dehydrogenase subunit would increase isobutanol synthesis. In vivo ldh deletion increased the isobutanol yield by 23%, while deletion of aceE led to pyruvate overflow to lactate formation rather than increased isobutanol synthesis [52]. It was reported that aceE deletion caused C. glutamicum to be unable to grow with glucose unless supplemented with acetate, indicating that knockout of aceE has a negative effect on cell growth and that knockdown of aceE, as suggested by the iCW773 simulation, might be a better alternative [53]. In addition, the iCW773-simulated strategy to improve isobutanol synthesis included an increase in the intracellular NADPH pool by downregulation of pgi in combination with upregulation of zwf. This combined regulation mode for pgi and zwf had been demonstrated to increase intracellular NADPH levels by redirecting carbon flux toward PPP [9]. In contrast, deletion of pgi together with aceE severely inhibited growth and eliminated isobutanol production [52]. Taken together, these data indicate that downregulation of aceE and pgi together with upregulation of zwf would contribute to isobutanol synthesis and simultaneously maintain growth. In summary, the iCW773-designed strategy could provide insights into future work for the optimization of 1,2-propanediol and isobutanol production in C. glutamicum.

In silico design and wet-lab construction of an l-proline-overproducing strain

Despite the strong flux toward l-glutamate formation in C. glutamicum, l-proline converted from l-glutamate under three sequential catalytic reactions by γ-glutamyl kinase (proB), glutamate-5-semialdehyde dehydrogenase (proA), and pyrroline-5-carboxylate reductase (proC) has not been efficiently synthesized by the engineered strain derived from C. glutamicum ATCC13032 [54,55,56]. The heterologous expression of ornithine cyclodeaminase merely made an l-ornithine-producing strain overproduce l-proline by six enzymatic reactions from l-glutamate, which was not an economically feasible means of carbon utilization in l-proline synthesis [56, 57]. Thus, iCW773 was applied to identify metabolic interventions that led to the overproduction of l-proline via OptForceMUST (Additional file 8). As shown in Fig. 5a, the upregulation of five reactions in the glycolytic pathway and citrate synthase (CS, encoded by gltA) and downregulation of α-oxoglutarate dehydrogenase (AKGDH, encoded by kgd) predicted by iCW773 were consistent with those predicted by ModelCg 2. The differences between the iCW773 and ModelCg 2 simulations were the downregulation of reactions from pyruvate to l-valine and l-alanine and upregulation of aconitase [ACONTa(b), encoded by acn] and isocitrate dehydrogenase (ICDHyr, encoded by icd), which converts citrate to α-oxoglutarate. The largest difference between the two models was the knockout of the putA gene to block the conversion of Δ1-pyrroline-5-carboxylate to glutamate. PutA is a bifunctional enzyme with both proline dehydrogenase and Δ1-pyrroline-5-carboxylate dehydrogenase activities, which catalyzes the two-step oxidation of l-proline to glutamate in E. coli [58]. In iCW773, the reaction from l-glutamate to Δ1-pyrroline-5-carboxylate formed a cycle in the presence of PutA, which made it inefficient to convert Δ1-pyrroline-5-carboxylate to l-proline. Blocking this conversion and upregulation of Δ1-pyrroline-5-carboxylate reductase (encoded by proC), the flux could be efficiently directed toward l-proline synthesis. Then, the PutA reactions were deleted in both models to predict the metabolic profile of overproducing l-proline (Additional file 9). As shown in Fig. 5b, one difference between the iCW773 and ModelCg 2 simulations was the upregulation of ProB to direct the metabolic flux from l-glutamate toward l-proline synthesis. However, the reaction from glutamate-5-semialdehyde to Δ1-pyrroline-5-carboxylate, which differed in the ModelCg 2 simulation, was impossible to control by genetic modification because this reaction occurred spontaneously without enzymatic catalysis.

Fig. 5
figure 5

Metabolic interventions predicted using OptForceMUST for l-proline overproduction in wild-type C. glutamicum ATCC13032 (a) predicted by iCW773 (left) and ModelCg 2 (right) and in the mutant strain C. glutamicum ATCC13032 ΔputA (b) predicted by iCW773 (left) and ModelCg 2 (right). Genes encoding relevant enzymes are depicted in italics next to the arrows: fda: fructose-bisphosphate aldolase; gltA: citrate synthase; kgd: α-oxoglutarate dehydrogenase; pfkA: phosphofructokinase; proA: glutamate-5-semialdehyde dehydrogenase; proB: γ-glutamyl kinase; proC: pyrroline-5-carboxylate reductase; putA: proline dehydrogenase and Δ1-pyrroline-5-carboxylate dehydrogenase. For other genes, see Figs. 3 and 4. 1pyr5c: Δ1-pyrroline-5-carboxylate; glu5p: glutamate-5-phosphate; glu5sa: glutamate-5-semialdehyde; l-ala: l-alanine; l-glu: l-glutamate; l-pro: l-proline. For other abbreviations, see Figs. 3 and 4

Full size image

To validate the accuracy of the iCW773 prediction for l-proline overproduction, different targets between the iCW773 and ModelCg 2 simulations were chosen for genetic modification. Because the enzymatic activity of ProB is strictly restricted to feedback inhibition by l-proline, no accumulation of l-proline will occur in WT [59]. Therefore, the feedback inhibition of ProB by l-proline was first released through the introduction of the nucleotide substitution G446A into the chromosomal proB gene [60]. The Pro-1 strain accumulated a small amount of l-proline compared with the WT (Table 1). Notably, the l-proline titer dropped after 36 h (Fig. 6a), indicating that l-proline may be converted to l-glutamate under PutA catalysis as simulated by iCW773. When putA was knocked out in Pro-1, the l-proline production of Pro-2 was 1.92-fold higher than that of Pro-1 (Table 1), which was consistent with a previous report that deletion of putA significantly increased the formation of trans-4-l-hydroxyproline derived from l-proline in recombinant E. coli BL21 [61]. In addition, the product yield of l-proline from glucose was improved 1.94-fold upon deleting putA (Fig. 6a).

Table 1 Comparison of fermentation parameters of different strains in shake flask cultivation
Full size table
Fig. 6
figure 6

Analysis of l-proline-overproducing strains. a Flask cultivation of different C. glutamicum strains. b Simulation of the impact of increased ACONTa(b) flux on cellular metabolism for robustness analysis was investigated with increasing relative ACONTa(b) fluxes (between 1.0 and 5.0). X-axis relative ACONTa(b) flux (the ratio of the test ACONTa(b) flux to the native ACONTa(b) flux). Relative ACONTa(b) flux = 1.0 served as the control (or native GPD flux). Y-axis relative flux of the pathways (the ratio of the test flux to the native flux). Yellow represents no change, red represents upregulation, and blue represents downregulation of the flux. Descriptions of abbreviated reaction names are provided in Additional file 3. c Relative transcript expression of related genes of the analyzed C. glutamicum strains in the exponential growth phase in glucose. Significant differences in the data were determined using Student’s t test (*P < 0.05, **P < 0.01). The data were derived from experiments performed at least three times, and the error bars represent the standard deviations

Full size image

Upregulation of gltA, acn, and icd expression was simulated by iCW773 to drive more carbon flux to α-oxoglutarate formation for l-proline synthesis. It has been demonstrated that CS flux in vivo is not strictly controlled by the amount of gltA expression [62], while the transcription of the acn gene is subjected to complex regulations mediated by the AcnR, RipA, and GlxR repressors [63, 64]. Thus, aconitase-catalyzed reactions should be a major control point in the TCA cycle during growth on glucose [64, 65]. The effects of increased ACONTa(b) flux on cell metabolism and l-proline synthesis were simulated using FBA. The fluxes toward CS and ICDHyr were enhanced with the increase in the relative ACONTa(b) flux, and maximum extracellular l-proline production was achieved when the relative ACONTa(b) flux was improved by 2.5-fold compared with WT (Fig. 6b). Conversely, the fluxes toward l-valine and l-alanine synthesis decreased. After the native promoter of the acn gene was replaced with three promoters of varying strength (P glyA , P pck , and P eftu ) and the start codon was changed from TTG to ATG in Pro-2, the mRNA levels of the acn gene increased by 1.47-, 1.82-, and 2.83-fold in the Pro-3, Pro-4 and Pro-5 strains, respectively (Fig. 6c). With upregulation of the acn gene, the mRNA levels of gltA in Pro-4 and Pro-5 increased by 50.82 and 88.52%, respectively, consistent with a previous observation that the expression of gltA and acn were simultaneously increased and might be controlled by a similar mechanism [65]. However, no significant changes were observed in the mRNA level of icd or kgd (Fig. 6c). When the mRNA levels of acn were increased, Pro-5 produced 2.26 ± 0.03 g/L l-proline with a 33.73% increase over Pro-2. Correspondingly, the byproducts of l-alanine and l-valine decreased by 16.95 and 8.09%. The extracellular metabolites exhibited a 73.34% decrease in citrate and a 1.37-fold increase in α-oxoglutarate.

Finally, to validate the positive correlation between GLU5K flux and l-proline yields (Additional file 4: Figure S2), the mutant proB (G446A) under the control of P tac on a plasmid was induced to overexpress the feedback-resistant γ-glutamyl kinase. When the mRNA level of proB* was significantly increased, the mRNA levels of the proA, proC, and gdh genes were simultaneously increased (Fig. 6c). The resultant Pro-6 strain showed an 8.27-fold increase in l-proline production (18.71 ± 0.21 g/L), with 8.18- and 9.03-fold increases in yield (0.36 g/g) and productivity (0.32 g/L/h), respectively, compared with Pro-5 (Fig. 6a). In addition, there were 85.55 and 45.42% decreases in extracellular α-oxoglutarate and succinate, respectively.

The performance of Pro-6 in fed-batch fermentation

The production performance of the final Pro-6 strain was investigated in fed-batch fermentation. As shown in Fig. 7, the strain grew continuously from 0 to 32 h and reached a cell concentration of OD600 = 109.02 at 32 h. The assimilated glucose was efficiently channeled to the l-proline biosynthetic pathway; l-proline production began at an early stage and continuously increased throughout the fermentation period. A major and constant increase in l-proline production was achieved during the feeding phase, which was initiated after the initial sugar (40 g/L) in the batch medium was consumed. The maximal specific growth rate of strain Pro-6 was 0.40 h−1, and the maximal specific glucose consumption rate was 5.46 mmol/gDCW/h. The maximal l-proline titer reached 66.43 g/L at 60 h with a yield of 0.26 g/g glucose (0.41 mol/mol glucose) and a productivity of 1.11 g/L/h. Consequently, the genetically engineered Pro-6 strain in this study was demonstrated to be an efficient l-proline-producing strain without any nutrition-auxotroph and with the highest titer and productivity reported on minimal medium with glucose as the sole carbon source (Table 2).

Fig. 7
figure 7

Fed-batch culture of Pro-6 in a 7.5-L bioreactor

Full size image
Table 2 Overview of l-proline-producing strains
Full size table

Conclusions

In this study, we reconstructed a genome-scale metabolic model of C. glutamicum ATCC13032 with an expanded size and scope compared with the previous ModelCg 2. The iCW773-simulated results of growth rates, metabolic flux profiles, and modification targets for overproducing natural and non-natural products were much more consistent with experimental data than those of ModelCg 2. Moreover, in silico analysis of l-proline overproduction and modifications of only three genes (putA, acn, and proB) cooperatively forced carbon flux toward l-proline, with the highest titer and productivity reported on minimal medium. This work revealed that the genome-scale metabolic model successfully guided the wet-lab experiments. This development will further boost C. glutamicum, one of the most promising and valuable workhorses, into a new era of industrial microbial biotechnology beyond the classical field.

Methods

Genome-scale metabolic model reconstruction

The reconstruction of the C. glutamicum ATCC13032 metabolic network was conducted based on the model of Shinfuku et al. [22]. A 96-step procedure for metabolic network reconstruction was recently published [66], and the appropriate steps were followed after new genes, reactions, and metabolites were added to generate iCW773. The KEGG (http://www.genome.jp/kegg/) and UniProt (http://www.uniprot.org/) databases were referenced to obtain reaction information and reaction-gene associations during the draft reconstruction stage, and iJO1366 was referenced to obtain the abbreviations of the same metabolites between C. glutamicum and E. coli [17]. The charge and mass of reactions in iCW773 were automatically balanced using the CheckMassChargeBalance command in Cobra Toolbox, and reversibility was assigned based on the Gibbs free energy by the eQuilibrator database (http://equilibrator2.milolab.webfactional.com/). Cobra Toolbox 2.05 with MatLab 2010b (MathWorks Inc.) was used for additional model testing [67], and the glpk solvers were used for the optimization procedures [68].

Flux balance analysis

The growth rates and metabolic flux distribution of C. glutamicum ATCC13032 were calculated using FBA with commercially available glpk and Matlab software [16, 18]. To simulate growth rates and metabolic flux distributions, biomass production was set as the objective function. For the simulation of growth rates under different glucose and oxygen levels, glucose and oxygen uptake rates were set as the experimental values, as shown in Fig. 1a. For the simulation of growth rates on different carbon sources, the uptake rate of related carbon source was set as the experimental value, as shown in Fig. 1b. Small molecular external metabolites such as CO2, H2O, SO3, NH3, and PO4 were allowed to be freely transported across the cell membrane.

OptForceMUST

The flux variability analysis was conducted to calculate each reaction’s flux span for both the wild-type and overproducing strains. A 13C flux and a relatively high secretion rate analyses were employed to constraint the wild-type and overproducing-type simulations, respectively [69]. The flux ranges for each reaction were compared, and some reactions of the overproducing strain need to be genetically modified through up/downregulation or knockout if no overlap between the flux range of the wild-type strain and that of the overproducing strain was found. A pool of modified reactions to construct the overproducing-type strain was produced.

To analyze the overproductions of the native and non-native metabolites, 13C flux data of the wild-type strain on glucose under aerobic condition were used as the constraints of the maximal range of flux variability [8, 31, 41]. For the non-native product, simulations of 1,2-propanediol and isobutanol production were performed after the heterogeneous reactions were added to iCW773, as described in Additional file 10. We imposed a minimum production yield of 98% of the theoretical maximum for l-lysine, l-valine, l-serine, l-proline, 1,2-propanediol, and isobutanol; the biomass flux was constrained to at least 1% of its theoretical maximum; and the l-lysine, l-valine, l-serine, l-proline, 1,2-propanediol, and isobutanol transport reactions were set as the target reactions, respectively. The upper and lower bounds of the 13C-tracer experimental fluxes are listed in Additional file 10 [8, 31, 41]. OptForceMUST was used to identify the reactions/genes that required to be up/downregulated or knocked out to maximize the production of the targeted metabolites.

Strains and plasmids

Wild-type C. glutamicum ATCC13032 (American Type and Culture Collection, Manassas, VA, USA) was used as the parental strain for strain engineering. The E. coli strain EC135 was used as the cloning host, and the pXMJ19 and pK18mobsacB plasmids were applied for gene overexpression with induction of isopropyl-β-d-thiogalactopyranoside (IPTG) and gene deletion, respectively [70]. All of the strains and plasmids used in this study are listed in Additional file 1: Table S3.

Constructions of plasmids and strains

Total genomic DNA was extracted from C. glutamicum according to a previously described procedure [71]. To construct a C. glutamicum strain carrying a point mutation resulting in an amino acid substitution from glycine to aspartate at position 149 of the ProB polypeptide, the proB gene was first deleted in the WT strain. The upstream and downstream homologous fragments of proB were amplified using PCR with primers P1/P2 and P3/P4. The amplified DNA fragments were spliced using overlap extension PCR and were ligated into the suicide vector pK18mobsacB [72]. The plasmids were verified by DNA sequencing and then transformed into the WT strain through electroporation to generate the WTΔproB strain, which was further verified by sequencing. In a second step, the mutation was introduced into the proB gene by PCR amplification with the primers P1/P7 and P8/P4. The amplified DNA fragments were spliced using overlap extension PCR and ligated into the suicide vector pK18mobsacB. The obtained plasmid was transformed into the WTΔproB strain to generate the Pro-1 (WTproB*) strain [60]. For deletion of the putA gene, previously isolated WT genomic DNA was amplified using PCR with primers P9/P10 and P11/P12. The amplified DNA fragments were spliced using overlap extension PCR and ligated into the suicide vector pK18mobsacB. The plasmids were verified by DNA sequencing and then transformed into the Pro-1 (WTproB*) strain through electroporation to generate the Pro-2 (WTproB*∆putA) strain, which was further verified by sequencing.

To replace the native promoter of the acn gene with the promoter of pck, the RBS sequence with ‘AAAGGAGGA,’ and the start codon with ATG, three fragments corresponding to the upstream region of acn, the introduced modulation element including the pck promoter and the RBS as well as the acn gene were amplified with primers P15/P16, P17/P18, and P19/P20, respectively. These fragments were spliced using overlap extension PCR and ligated into the suicide vector pK18mobsacB. The plasmids were verified by DNA sequencing and then transformed into the Pro-2 (WTproB*∆putA) strain through electroporation to generate the Pro-3 (WTproB*P pck ::P acn acn*∆putA) strain, which was further verified by sequencing. Similarly, the promotor of the acn gene was altered in Pro-2 (WTproB G446A∆putA), which generated Pro-4 (WTproB*P glyA ::P acn acn*∆putA) and Pro-5 (WTproB*P eftu ::P acn acn*∆putA).

The C. glutamicum and E. coli shuttle vector pXMJ19 was used for the overexpression of the proB* gene [73]. For amplification of the proB* gene, genomic DNA isolated from Pro-1 (WTproB*) was amplified using PCR with primers P31/P32. Then, the amplified DNA fragment was ligated into the shuttle vector pXMJ19. The plasmids were verified by DNA sequencing and then transformed into the Pro-5 (WTproB*P eftu ::P acn acn*∆putA) strain through electroporation to generate the Pro-6 (WTproB*P eftu ::P acn acn*∆putA/pXMJ19-proB*) strain. When needed, 0.1 mM IPTG was added to the culture medium to induce target gene overexpression. The primers used in this study are provided in Additional file 1: Table S4.

Media and cultivation

The E. coli strains were aerobically grown at 37 °C in Luria–Bertani medium [74]. The C. glutamicum strains were cultivated in brain heart infusion (BHI) medium (37 g/L brain heart infusion with 91 g/L sorbitol) at 30 °C for genetic disruption and complementation [75]. As a minimal medium, CGXII medium was used with 40 g/L glucose [76]. When necessary, antibiotics were added at the following concentrations: 50 μg/mL kanamycin or 20 μg/mL chloramphenicol for E. coli and 25 μg/mL kanamycin or 10 μg/mL chloramphenicol for C. glutamicum.

In the shake-flask growth experiment, C. glutamicum strains were precultured in the CGIII seed medium at 30 °C and 200 rpm until the OD600 reached 12. One milliliter of seed culture was inoculated in a 500-mL baffled shake flask with 30 mL of CGXII minimal medium. The cells were cultured in triplicate at 30 °C and shaken at 220 rpm. The pH was maintained at 7.0–7.2 via ammonia supplementation.

The fed-batch fermentation was performed in a 7.5-L bioreactor (BioFlo®/CelliGen®115, New Brunswick, USA) with a working volume of 2 L of CGX medium containing 40 g/L glucose. After 3 h of growth, 0.1 mmol/L IPTG was added for the induction of P tac . The concentration of glucose over all of the fed-batch cultures was maintained within the range of 5 ± 5 g/L by supplying 800 g/L of glucose reservoir. The glucose reservoir was fed into the fermenter at the rate of 0.2–0.5 mL/min according to the residual glucose concentration. The temperature was maintained at 32 °C using cold water circulation. The pH was maintained at 6.9 via the automated addition of ammonia and 100 mmol/L H3PO4. Dissolved oxygen was determined using a pO2 electrode and maintained above 30% saturation via variation of the stir speed.

Analytical methods

The glucose concentration was measured using an SBA-40D biosensor analyzer (Institute of Biology of Shandong Province Academy of Sciences, Shandong, China). The cell concentration was determined by measuring the absorbance at 600 nm (OD600) using a spectrophotometer (V-1100D; Mapada Instruments, Shanghai, China). The dry cell weight (DCW) per liter was calculated using an experimentally determined formula: DCW (g/L) = 0.27 × OD600 [77]. The amino acids in the culture supernatant were determined using high-performance liquid chromatography with a Zorbax Eclipse XDB-C18 column (4.6 mm × 150 mm, 5 μm; Agilent) at 40 °C and 360 nm after derivatization with 2,4-dinitrofluorobenzene. Mobile phase A was 55% (v/v) acetonitrile, and mobile phase B consisted of 40 mmol/L KH2PO4 at pH 7.0–7.2. The flow rate of the mobile phase was 1 mL/min. The organic acids in the culture supernatant were determined using high-performance liquid chromatography with a SB-Aq column (4.6 × 250 mm; 5 μm; Agilent) at 40 °C and 210 nm. Mobile phase A was acetonitrile, and mobile phase B consisted of 20 mmol/L KH2PO4 at pH 2.3. The flow rate of the mobile phase was 1 mL/min.

RNA preparation and quantitative RT-PCR

Corynebacterium glutamicum strains were grown to the exponential phase in CGXII minimal medium as described above. The cells were harvested and the total RNA was isolated using an RNAprep Pure Cell/Bacteria Kit (Tiangen, China). Reverse transcription of approximately 400 ng of RNA was performed with the primers listed in Additional file 1: Table S4 using a FastQuant RT Kit (Tiangen, China). Quantitative PCR was performed using GoTaq qPCR master mix (Promega, USA) in a 20-μL mixture with a LightCycler® 96 Real-Time PCR System (Roche, Switzerland). The C. glutamicum rpoB gene was used as the reference gene to normalize the gltA, acn, icd, kgd, gdh, proB, proA, and proC mRNA levels [70]. Negative controls were used in each PCR run to exclude DNA and other contaminants. The qPCR products were verified via a melting curve analysis. Data collection and analyses were conducted using LightCycler® 96 software (Roche, Switzerland) with the 2−∆∆CT method [78].