Systems Analyses Reveal the Resilience of Escherichia coli Physiology during Accumulation and Export of the Nonnative Organic Acid Citramalate

Citramalate is an attractive biotechnology target because it is a precursor of methylmethacrylate, which is used to manufacture Perspex and other high-value products. Engineered E. coli strains are able to produce high titers of citramalate, despite having to express a foreign enzyme and tolerate the presence of a nonnative biochemical. A systems analysis of the citramalate fermentation was undertaken to uncover the reasons underpinning its productivity. This showed that E. coli readily adjusts to the redirection of metabolic resources toward recombinant protein and citramalate production and suggests that E. coli is an excellent chassis for manufacturing similar small, polar, foreign molecules.

recombinant protein/product accumulation in high-cell density fermentations at a system level.
To investigate cellular adaptations to production of citramalate, samples of the citramalate production fermentation were taken at representative time points before and after induction of CimA3.7 or CimA3.7 dead . Samples were used for comprehensive multiomic analyses of mRNA, protein, lipid and metabolite levels.
Citramalate production results in an increased intracellular pyruvate pool. Before induction of CimA, the control and test fermentation endo-metabolomes were similar, clustering together in principal-component analysis (PCA) plots (Fig. 2). After induction, the endo-metabolomes of the control and test fermentations were clearly differentiated by citramalate (bin mass of 148.08 Da) ( Fig. 2A). Reanalysis of these data after removal of the citramalate and phosphate mass bins revealed further differences between the control and test fermentations after induction (Fig. 2B). These included bin masses of 88.05 Da (identified as pyruvate by tandem mass spectrometry [MS-MS]), 130.08 Da (citraconic acid), 149.08 Da (possibly methionine), and 162.08 Da (2ethylmalate). Citraconic acid is formed from citramalate by the E. coli 3-isopropylmalate dehydratase LeuCD (Fig. 2C) (3). However, previous attempts to remove this activity via individual leuC and leuD mutations increased fermentation cost and complexity through the requirement to feed leucine and did not enhance citramalate production in shake flask cultures (7). In an industrial process, citraconate is not problematic since it can be converted to methacrylic acid using the same process as that developed for citramalate (4,5). However, the formation of 2-ethylmalate was unexpected, but resolved by finding novel activity of CimA3.7 when 2-oxobutyrate was supplied in place of the cognate substrate, pyruvate, in vitro (see Fig. S3 in the supplemental material). Hence, 2-ethylmalate was formed by CimA3.7-catalyzed condensation of acetyl-CoA with 2-oxobutyrate (Fig. 2C). 2-Oxobutyrate can be formed via the citramalate pathway catalyzed by CimA and LeuBCD or in reactions catalyzed by threonine deaminase and O-succinylhomoserine lyase (Fig. 2C). Although relatively low, the intracellular concentration of 2-ethylmalate was ϳ1.7-fold greater than the con-centration of pyruvate. Production of 2-ethylmalate represents an additional drain on acetyl-CoA pools, reducing the availability of acetyl-CoA for citramalate production. Possible interventions to limit production of the 2-ethylmalate by-product would be deletion of 3-isopropylmalate dehydratase, threonine deaminase, and O-succinylhomoserine lyase activities or improve the substrate specificity of CimA3.7. However, it should be noted that 2-ethylmalate is not detectable in the culture supernatant and would not, therefore, affect the purity of the citramalate product. Furthermore, elimination of these 2-ethylmalate-forming reactions would require expensive nutritional supplementation to compensate for the ensuing auxotrophy (5). Therefore, further metabolic engineering to eliminate 2-ethylmalate would not provide any benefits.
The clear differences between the endo-metabolome of the test fermentations postinduction compared to the controls suggested that there are changes that extend beyond simply redirecting pyruvate and acetyl-CoA away from aerobic respiratory metabolism and biomass production toward citramalate synthesis, including the intracellular production of 2-ethylmalate. Therefore, changes in the chassis transcriptome, proteome, and lipidome during control and test fermentations were analyzed for responses to the metabolic burdens imposed by heterologous protein synthesis and citramalate production.
Expression of recombinant citramalate synthase affects gene expression more than the production of citramalate. The transcriptional response to CimA3.7 and citramalate production was analyzed. In the control fermentation, 303 genes exhibited altered transcription (Ն2-fold, adjusted P value of Յ0.05), 4 h after induction of CimA3.7 dead synthesis and 36 genes after 21.5 h, representing a total of 316 genes (see Table S1A and Fig. S4 in the supplemental material). For the control fermentations, only genes linked to flagellar biosynthesis were identified as being significantly enriched (GO:0071978; 7.79, P ϭ 0.024, Fisher's exact test with false-discovery rate [FDR] multiple-test correction) by the AmiGO gene ontology software (9). For the test fermentations, only 15 genes met the significance criteria 4 h after induction and 42 genes 21.5 h postinduction, representing a total of 53 different genes, of which 39 were also differentially regulated in the control (Table S1B). As observed for the control fermentations, flagellar genes were significantly enriched (GO:0071978; 29.02, P ϭ 0.0007) in the test fermentations, suggesting that downregulation of flagellar biosynthesis is an adaptation to high-cell density recombinant protein production. Downregulation of flagellar synthesis has been observed during recombinant protein synthesis in other E. coli strains, and it has been suggested that chassis performance might be enhanced by shutting down the flagellar machinery (10)(11)(12)(13).
Genes that were only upregulated in the test fermentations were deemed to be potentially specific responses to citramalate production (Table 1). These included two efflux pumps (alaE and mdtE) and components of the E. coli acid response (gadC and hdeB). This suggested that citramalate production was associated with acid stress and that AlaE and/or MdtE might contribute to efflux of citramalate from the cytoplasm. The global gene expression data described above suggested that the transcriptional response to citramalate synthesis was limited. To complement that analysis, the TFInfer software was used to infer any changes in the activities of 201 regulators during the control and test fermentations (14). The statistical inference of changes in the activities of transcription factors has the potential to extract hidden information in the transcriptomic data by revealing the regulators (and by extension their cognate signals) responsible for the observed changes in transcript abundance. This probabilistic approach requires a connectivity matrix linking transcription factors to transcription units, and this was constructed using the list of E. coli transcription factors and the operons that they regulate in RegulonDB (15).
The transcription factors exhibiting the greatest changes in activity (signal/noise ratio of Ն4) in the control fermentations were AraC, BirA, ExuR, FlhDC, and PurR ( Fig. 3A and C; see Table S2 in the supplemental material). For the test fermentations, five transcription factors, including BirA and FlhDC (BirA, FlhDC, GadW, MalT, PaaX, and PuuR), responded with a signal-to-noise ratio of Ն4 ( Fig. 3B and C; Table S2). The TFInfer software also reveals the target genes that are most heavily influenced by each transcription factor (Table S2). These analyses further confirm that inactivation of FlhDC and the consequent inhibition of flagellar gene expression are a core adaptation to recombinant protein production and that citramalate production initiates a GadWmediated acid response and modulation of MalT-regulated sugar uptake. The latter is notable because enhanced succinate production by E. coli strain HX024, an evolved derivative of strain Suc-T110, was associated with increased expression of the malK-lamB-malM operon (16). Thus, changes in MalT-mediated regulation might be a common response to the production of organic acids using an E. coli chassis. Nevertheless, none of the major stress-responsive transcription factors (e.g., cyclic AMP receptor protein [CRP], glucose starvation; fumarate and nitrate reductase [FNR], oxygen starvation; OxyR/SoxRS, oxidative stress; CpxAR, envelope stress; CspA, cold shock; EvgSA/ PhoPQ/GadE/GadX/EnvZ-OmpR, low pH; NhaR, sodium stress; LexA, DNA damage; CusR/CueR/Fur/Zur/ZntR, metal ion homeostasis) were significantly activated in response to CimA3.7 dead /CimA3.7 and/or citramalate production.
An alternative analytical approach is to calculate the differences in transcript abundance between the samples CS2 and TS2 and CS3 and TS3 after normalization to the respective preinduction samples CS1 and TS1 (Fig. 1). The data were then filtered to include transcripts exhibiting differences in abundance of Ն2-fold (adjusted P value of Յ0.05) in the test samples compared to the control (Table S1C). These criteria were met by only seven genes: two associated with sulfur metabolism (cysD and cysW), two with transporter functions (exuT and malK-lamB-malM), one coding for a hypothetical protein (ygbK), and genes associated with flagellar (fliA) and tryptophan (trpD) biosynthesis ( Table 2). Despite the diversion of pyruvate and acetyl-CoA away from central metabolism toward citramalate production and the potential for citramalate-induced osmotic stress in the postinduction production cultures, extensive changes in gene expression were not observed (Table S1C).
Roles of sRNAs during citramalate production. As well as permitting changes in the expression of mRNAs to be detected, the microarray design included 132 custom probes to report on the abundances of 77 noncoding RNAs. The abundances of 13 small regulatory RNAs (sRNAs) changed significantly upon induction of CimA3.7 (Table 3). Three of these sRNAs (ChiX, RyhB, and RyjA) were regulated similarly in the control and citramalate production fermentations. Increased abundance of ChiX results in inhibition of chitosugar uptake and utilization, enhanced levels of RyhB are associated with iron starvation and lowered synthesis of iron-containing proteins, and RyjA (also known as SarL) inhibits the synthesis of trigger factor (a ribosome associated protein-folding chaperone) in response to lower protein synthesis in stationary phase (17). Nine sRNAs were significantly regulated in the control fermentations, but not in the test fermentations, postinduction, but these sRNAs were not significantly differentially regulated when their abundances in the test fermentations were compared to those in the controls ( Table 3). The 10th sRNA, CsrB, exhibited opposite regulation in the control and production fermentations, being upregulated in the former and downregulated in the latter postinduction. The enhanced levels of CsrB should decrease the activity of CsrA, which acts as a repressor of gene expression in stationary phase, consistent with the higher growth rate/yield of the control fermentations (18) (Fig. 4). More specifically, CsrA maintains carbon flux toward pyruvate, rather than to storage as glycogen. Therefore, the decreased abundance of CsrB may increase glycolytic flux beyond the capacity of the pyruvate dehydrogenase complex. This increase in glycolytic flux would contribute to the observed increase in intracellular pyruvate. An increase in glycolytic flux might be triggered by a decreased capacity to generate ATP by oxidative phosphorylation created by the diversion of pyruvate and acetyl-CoA away from the citric acid cycle toward citramalate production. This diversion likely constrains the ability to generate reducing equivalents via the tricarboxylic acid (TCA) cycle and subsequently ATP by oxidative phosphorylation. Under such circumstances, lower ATP availability might invoke increased glycolytic flux, as observed when ATP concentrations are manipulated by expression of an uncoupled ATPase (19). Two sRNAs (GadY and RprA) were differentially upregulated in the production strain compared to the control ( Fig. 4 and Table 3). Both GadY and RprA sRNAs are associated with acid tolerance in E. coli (20,21). GadY increases the stability of the gadX mRNA, and GadX, along with GadE and RcsB, activates expression of several components of the E. coli acid response. The E. coli acid fitness island consists of a 12-gene locus that contains a Differences in the abundances of the indicated sRNA species relative to the preinduction samples (CS1 and TS1). Only those sRNA species that exhibited Ն2-fold change (FDR adjusted P value of Յ0.05) for one or more samples are shown. b Differences in the abundances of the indicated sRNA species in the production strain (TS2 and TS3) relative to those in the control strain (CS2 and CS3). Only those sRNA species that exhibited Ն2-fold change (unadjusted P value of Յ0.05) for one or more comparisons are shown. c NS, not significant.

FIG 4
Responses of the E. coli chassis to production of citramalate. Inference of changes in transcription factor activities suggested that the activity of GadW decreased upon synthesis of citramalate (Fig. 3B). Accordingly, the abundance of the GadW-repressed sRNA GadY increased (Table 3). GadY acts to stabilize the mRNA encoding GadX, which along with GadE activates expression of the gadBC and hdeAB operons (coding for the glutamate decarboxylase-glutamate:4-aminobutyrate antiporter acid tolerance system and the periplasmic chaperones [ Table 1]). Citramalate synthesis was also associated with enhanced abundance of the RprA sRNA, which enhances S -dependent acid tolerance by increasing translation of RpoS (Table 3). The redirection of carbon from biomass to citramalate production resulted in decreased abundance of the sRNA CsrB, which is an antagonist of the translational regulator CsrA (Table 3). Higher CsrA activity in the production strain maintains carbon flux toward pyruvate, rather than to storage as glycogen, thereby enhancing citramalate productivity. Transcription factors are boxed, sRNAs are in bold, operons are italic, processes are circled, increased abundance is red, and lower abundance is blue.
the gadE-mdtEF operon that was upregulated in response to citramalate production (see above). Most of the genes in this island were also upregulated in response to citramalate, although most did not meet the criteria to be regarded as significant (see Fig. S5 in the supplemental material). Increased osmolarity enhanced RprA-dependent translation of the rpoS mRNA, and RprA is known to be upregulated when batch cultures are exposed to acetic, succinic, or itaconic acid (21,22), suggesting that the enhanced abundance of RprA observed in the production fermentations may be a response to citramalate accumulation.
The chassis proteome responds to induction of recombinant protein production but is unperturbed by citramalate. To identify changes in the E. coli proteome in response to heterologous protein production, a label-free quantification approach was used to compare preinduction (CS1) and postinduction (CS2 and CS3) samples from the control fermentations expressing recombinant CimA3.7 dead . Combining CS2 and CS3 revealed that 109 of the 1,413 detected proteins exhibited a change (64 increased, 45 decreased) in abundance postinduction (Ն2-fold change in abundance relative to CS1; P Ͻ 0.05) ( Fig. 5; Table S1D and Fig. S4). Of the 1,304 proteins that exhibited no significant change, 1,193 of the corresponding transcripts also did not significantly change. Fourteen differentially regulated genes and proteins were correlated (11 upregulated, araG, betB, exuT, gabT, nanE, rspB, uxaB, uxaC, yafV, ycjX, and ygeA; 3 downregulated, cysD, purM, and pyrI). The enhanced abundance of the arabinose transporter (AraG, 91.4-fold; AraF, 65.4-fold; and AraC, 12.4-fold) was expected in response to arabinose-mediated induction of cimA3.7 dead expression, resulting in an ϳ95-fold increase in CimA3.7 dead protein. A major feature of the transcript profiling was the downregulation of flagellar and chemotaxis genes, but this was not significant in the proteome. In both control and test fermentations, flagellar gene expression was downregulated after induction, and this is interpreted as a "fermentation" response to the production of recombinant protein and was not specific to citramalate production. Examination of the proteomic data for control and test fermentations showed that the only motility-related protein that was significantly downregulated after induction was Tar (8.3-fold and 6.2-fold, respectively). Nevertheless, all the other flagellum-related proteins that were detected exhibited lower abundance after induction (shown as protein, control fold decrease, test fold decrease: FliC, 2.7, 2.1; FliY, 2.9, 2.0; FlgK, 2.8, 2.7; FlgH, 2.0, 1.8; Tsr, 5.2, 4.9; CheA, 1.5, none; CheW, 1.2, none) but did not meet the criteria for statistical significance. In addition, not all flagellar proteins were detected in the proteomic analysis, and so changes in abundances of these could not be established. Thus, qualitatively the transcriptomics and proteomics of motility-related proteins are in agreement.
The rspB transcript and the RspA and RspB proteins were upregulated in both control and production strains upon induction of CimA. Co-overexpression of rspAB improved recombinant production of a model protein, ␤-galactosidase, by lowering accumulation of lactone by-products of amino acid metabolism and dysregulation of homoserine lactone-associated stationary-phase adaptations (23,24). Thus, similar to the Ibp response (see below), upregulation of rspAB is likely associated with recombinant protein synthesis and not citramalate production.
The test fermentations, which produce both heterologous protein (active CimA3.7) and a chemical product (citramalate), also showed increased abundance of CimA3.7 (up to 76-fold), AraF (arabinose uptake), and UxaC and UxuAB (hexauronate catabolism) among a total of 152 proteins (74 more abundant and 80 less abundant from a total of 1,103 proteins detected; Ͼ2-fold; P Ͻ 0.05) ( Fig. 5; Table S1E). Of the 951 proteins that exhibited no significant change, 944 also showed no changes at the level of transcription. The regulation of only four genes/proteins was positively correlated (araG, ibpB, nanE, and rspB).
In the context of recombinant protein production, the ϳ100-fold increased abundances of the heat shock proteins IbpA and IbpB in both the control and test fermentations is of interest (Table S1D and E) (25). IbpA and IbpB have been also previously found as components of inclusion bodies formed during recombinant protein production and can act to stabilize loosely aggregated proteins and facilitate effective disaggregation by the DnaK and ClpB chaperone systems.
Differences (Ͼ2-fold, adjusted P value of Ͻ0.05) in protein abundance between samples TS2 and CS2 and TS3 and CS3 after normalization to the respective preinduction sample CS1 were observed for only five proteins: Slp (32-fold upregulated), YbjX (7.5-fold downregulated), OppD (5.4-fold upregulated), DeoA (5-fold upregulated), and Upp (4.8-fold upregulated) (Table S1F). Interestingly, the slp transcript, which encodes an outer-membrane-stabilizing protein that accumulates as growth rate decreases (26), was ϳ3-fold upregulated when TS2 was compared to CS2, with an unadjusted P value of 0.016. No clear phenotypes have been associated with a slp mutant, which was unaltered in its resistance to hydrogen peroxide, ethanol, and SDS (26). There is evidence that Slp is enhanced in response to N limitation and temperature downshift and that it might be associated with resistance to low pH (27,28). The relative upregulation of Slp at both the transcript and protein levels might suggest that this is a response specifically associated with citramalate production.
Overall, our data suggest that the chassis proteome is largely "blind" to citramalate accumulation as most reproducible changes appear related to protein production stress and occur under both test and control conditions without citramalate accumulation.
Citramalate production did not alter the chassis lipidome, but strong growth cycle effects were observed. The lipids present in the chassis (E. coli) membranes during high-cell density batch fermentation and citramalate production were determined by liquid chromatography (LC)-MS-MS with positive-and negative-mode heated electrospray ionization (HESI). In addition, gas chromatography (GC)-MS of fatty acid methyl esters (FAMES) was used to identify fatty acids present and any changes in the proportions of cis/trans isomers. LC-MS-MS analysis resulted in 190 peaks by positive ESI (see Table S3A in the supplemental material) and 41 by negative HESI (Table S3B). Chassis lipid composition was largely unaffected by production of citramalate but did change as both fermentations proceeded (Fig. 6A and B). Principal-component analysis (PCA) showed a separation between the control and test fermentations for the positivemode ESI data, but this change was not evident in the negative-mode data set. The separation observed in the larger positive-mode data set arose from small changes in numerous lipids (Fig. 6C and D). Nevertheless, a minor increase in cyclopropane fatty acid content was observed for both fermentations in CS3 and TS3 (Fig. 6E). Although distinct changes in the E. coli lipidome were observed during the fermentation, these changes were not responses to citramalate production.
Citramalate excretion is likely to be nonspecific and mediated by several efflux systems. The test fermentations yielded citramalate in the culture medium at a concentration of 25 g liter Ϫ1 . As a charged molecule, citramalate secretion was anticipated to require the action of one or more efflux pumps. The efflux pump genes alaE and mdtE were upregulated after induction of CimA3.7, the nmpC gene coding for an outer membrane protein was downregulated, Slp an outer membrane-stabilizing protein, was significantly upregulated in the test fermentation proteome compared to the controls, and acrA appeared in both test and control fermentations as the gene most heavily influenced by MprA in the TFInfer analysis. Therefore, the corresponding mutant E. coli strains and the parent were transformed with pBAD24-cimA3.7, and small-scale citramalate production assays were performed using harvested cells. Analysis of the culture supernatants showed that the extracellular citramalate concentrations were similar for all strains tested (see Table S4 in the supplemental material). This suggested that individually acrA, alaE, mdtE, nmpC, and slp were not essential for citramalate excretion.
It is possible that intracellular concentrations of citramalate result in efflux as a result of low-affinity interactions with multiple transporters. Data for pyruvate excretion by E. coli were used to assess the likelihood of this suggestion (29). If the steady-state intracellular-to-extracellular pyruvate ratio of ϳ50:1 (0.16 mM outside, 7.5 mM inside; dilution rate of 0.1 h Ϫ1 ; biomass, 2.5 g liter Ϫ1 ) results from a passive and unsaturated process (i.e., E. coli has no specific pyruvate efflux system), the rate of pyruvate export (6.4 ϫ 10 Ϫ6 mol h Ϫ1 g Ϫ1 ) is equal to k[7.5 ϫ 10 Ϫ3 ], yielding an export rate constant k equal to 8.5 ϫ 10 Ϫ4 h Ϫ1 . The maximum rate of citramalate appearance during the fermentations was ϳ0.3 ϫ 10 Ϫ3 mol h Ϫ1 g Ϫ1 . Using the value for k calculated for pyruvate, the maximum intracellular concentration of citramalate is estimated to be ϳ300 mM. The total E. coli metabolome concentration has been estimated at 300 mM (30). Assuming that E. coli is essentially "blind" to citramalate, as implied by the omic responses, then the percentage of total ion counts for citramalate (ϳ40%) suggests that the intracellular concentration reaches ϳ200 mM, similar to the estimate calculated above and the measured extracellular citramalate concentration. These analyses suggested that citramalate concentrations inside and outside the cell are similar, which might be consistent with the operation of mechanosensitive channels opening in response to increased turgor pressure as the citramalate accumulates in the cytoplasm. However, citramalate production assays using harvested cells with single and double mscL and mscS mutations did not lower the extracellular concentration of citramalate, suggesting that citramalate export probably occurs through promiscuous major facil-itator superfamily proteins when the intracellular citramalate concentration reaches ϳ200 mM (Table S4).
Conclusion. Characterization of the stresses experienced by bacterial cell factories during synthesis of recombinant proteins and chemical products is essential for improving process efficiency and designing new bacterial chassis. Many studies have investigated the potential use of engineered E. coli strains to produce native organic acids (e.g., lactate, malate, and succinate [31]). While high-titer processes for production of these organic acids have been developed, multiomic analyses of high-cell-density fed-batch cultures to determine the cellular responses during organic acid production are relatively few. Such information could be valuable in guiding further process and strain developments. Here, a comprehensive omic analysis of E. coli fed-batch fermentations for recombinant protein synthesis (CimA3.7 dead ) and for the production of a nonnative organic acid, citramalate, revealed minimal effects on gene expression, metabolism, or lipid composition, thus providing a simple explanation for the exceptional product titers and yields. Although it was anticipated that production of citramalate would require major rewiring of cellular metabolism rooted in reprogramming of transcription and consequent changes in the proteome and lipidome, multiomic analyses revealed that few adaptations were required to permit high production of citramalate (Fig. 7).
The multiomic analysis of E. coli high-cell-density batch cultures expressing inactive recombinant CimA3.7 dead (ϳ20 mg g Ϫ1 cdw) permitted identification of core stress responses associated with recombinant protein expression (Fig. S1B). Transcriptomic and proteomic analyses indicated that recombinant protein expression was associated with downregulation of flagellar genes and induction of the ibpAB-encoded chaperones. It has been shown that induced expression of the ibpAB operon can be a useful intervention to enhance recombinant protein yields, especially when combined with other chaperones, such as ClpB (25,32,33).
As expected, the postinduction exo-and endo-metabolite profiles of the production strain were dominated by citramalate. Two minor by-products, citraconate and 2-ethylmalate, were detected at concentrations that would not warrant further intervention to eliminate their presence. During citramalate production, the transcript, protein, and lipid profiling indicated that the level of stress imposed on the E. coli chassis was low, but some responses were detected that could be attributed to the citramalate synthesis, including roles for sRNA molecules in maintaining carbon flux through glycolysis (CsrB) and in acid tolerance (GadY and RprA). Optimizing these sRNA responses could represent a possible intervention to further enhance citramalate yields. The analyses of citramalate production reported here suggest that E. coli is a robust chassis organism and is likely to be an excellent choice for production of other organic acids.    (35). A control plasmid (pBAD-cimA3.7 dead ) expressing an inactivated CimA3.7 by virtue of a His192Ala substitution was constructed using Quickchange II XL site-directed mutagenesis kit (Agilent Technologies). The manufacturer's protocol was followed with primers (5=-ACCTGCCGGTTAGC GTGGCCTGCCATAACGATTTCGGC-3= and 5=-GCCGAAATCGTTATGGCAGGCCACGCTAACCGGCAGGT-3=) and pBAD-cimA3.7 (5) as the template. Electro-competent E. coli BW25113 ldhA was transformed to create strains expressing active or inactive CimA; transformants were selected on LB agar supplemented with carbenicillin (50 mg liter Ϫ1 ). Escherichia coli BW25113 ldhA was used in the fermentation processes to minimize lactate formation based on unpublished results (J. Webb, personal communication). Other E. coli BW25113 mutants were also obtained from the Keio collection, except the mscL mscS double mutant, which was constructed using pCP20 (36) to cure the kanamycin cassette from E. coli BW25113 mscL::Kan r , followed by deletion of mscS from the resulting strain (E. coli BW25113 mscL::FRT) by P1 transduction of the mscL mutation from E. coli BW25113 mscL::Kan r .

MATERIALS AND METHODS
Citramalate production using harvested cells. Single colonies of E. coli BW25113 ldhA expressing either active or inactive CimA3.7 were grown overnight (37°C, 250 rpm) in SM glycerol medium. Cultures were diluted to a starting OD 600 of 0.1 in SM glycerol medium, and the culture was incubated (250 rpm, 37°C) until the OD 600 reached 0.6. CimA3.7 expression was induced by the addition of L-arabinose (0.2 g liter Ϫ1 ). Cells were harvested by centrifugation (4,000 ϫ g, 20 min, 4°C) 4 h postinduction (OD 600 of 3 to 4) and concentrated to a dry cell weight of 15 g liter Ϫ1 in SM medium without NH 4 Cl or yeast extract but with glucose (20 g liter Ϫ1 ). The cell suspensions (approximately 10 ml) were incubated in baffled flasks (250 ml, 24 h, 250 rpm, 37°C) before analysis of culture supernatants by high-pressure liquid chromatography (HPLC).
Fermentations. Inocula for fed-batch fermentations were produced by inoculating E. coli BW25113 ldhA pBAD24-cimA3.7 or E. coli BW25113 ldhA pBAD24-cimA3.7 dead into ML medium (50 ml) supplemented with glucose (10 g liter Ϫ1 ) and carbenicillin (50 g ml Ϫ1 ) and incubating overnight (200 rpm, 37°C). The cultures were diluted to OD 600 of 0.1 in sterile water (50 ml) and used to inoculate ML medium (1 liter) supplemented with glucose (11.9 g liter Ϫ1 ) and carbenicillin in a 3-liter BioFlo/CelliGen 115 bioreactor (New Brunswick Scientific). Cultures were grown at 37°C, and the pH was maintained at 7.0 Ϯ 0.1 by the addition of NH 4 OH (28 to 30%) and 2 M H 2 SO 4 (2 M). The airflow rate was initially set at 1 liter min Ϫ1 , and the dissolved oxygen (dO 2 ) was maintained at 30% of saturation by automatic control of stirrer speed between 400 and 1,200 rpm and an airflow cascade between 1 and 7 liter min Ϫ1 . The airflow cascade was only implemented when maximum agitation was reached. Drops of polypropylene glycol (100%) were manually added to the medium when required to avoid foaming. When the glucose in the batch medium had been consumed, indicated by a sharp increase in dO 2 and confirmed using glucose test strips (117866; Merck Millipore), a feed of glucose was started containing glucose (650 g liter Ϫ1 ), yeast extract (5.9 g liter Ϫ1 ), MgSO 4 (7.2 g liter Ϫ1 ), and trace elements (11.8 ml). The flow rate was adjusted manually as required to maintain a pseudoexponential growth rate of approximately 0.25 h Ϫ1 and avoid the accumulation of excess glucose in the culture, using the test strips as confirmation as described above. Protein expression was induced by the addition of L-arabinose (0.02% wt/vol) when the culture OD 600 was 50.
Analytical methods. Growth was monitored by measuring OD 600 . Samples were diluted in deionized water when the OD 600 was Ͼ0.8. Dry cell weight was measured by centrifuging 1-ml samples in preweighed polypropylene tubes, removing the supernatant, and drying the pellets to a constant weight. D-Glucose, (R)-citramalate, and other organic acids were quantified by HPLC using an Agilent 1200 series HPLC system equipped with both UV (215 nm) and refractive index detectors. Samples were resolved using a Rezex ROA organic acid Hϩ column (Phenomenex) at 55°C with 0.01 N H 2 SO 4 (0.5 ml min Ϫ1 ) as the mobile phase. Samples were prepared for HPLC analysis by centrifuging (12,000 ϫ g, 5 min) and filtering the supernatants (0.2-m-pore filter). Data analysis was performed with ChemStation software using calibration curves prepared using authentic standards of each compound (0.1 to 200 mM).
Transcriptomics. Fermentation samples (3 ϫ 0.1 ml) were mixed with RNAprotect (Qiagen [0.2 ml]), vortexed, and incubated at room temperature (5 min). The mixture was then centrifuged (16,000 ϫ g, 2.5 min, 4°C) and the supernatant discarded, and pellets were stored at Ϫ80°C. Cell pellets were thawed at room temperature and normalized in TE buffer (10 mM Tris-HCl at pH 8.0, 0.1 mM EDTA), such that 1 ml of TE buffer contained 2 OD 600 units. Aliquots (1 ml) were harvested by centrifugation (12, 000 ϫ g, 1 min), and the pellets were resuspended in 100 l TE buffer containing lysozyme (15 mg ml Ϫ1 ). These were incubated at room temperature (10 min) with vortexing every 2 min. Total RNA was prepared using the RNeasy RNA purification kit (Qiagen) according to the manufacturer's protocol (including the on-column DNase treatment step). Labeled cDNA was produced using SuperScriptIII reverse transcriptase (Invitro-gen) with the Cy3-dCTP included in the dNTP mixture. Labeled E. coli genomic DNA was produced using the BioPrime DNA labeling kit (Invitrogen) with Cy5-dCTP included in the dNTP mixture. Labeled genomic DNA and cDNA were combined and hybridized overnight to an oligonucleotide microarray (Agilent Technologies). Quantification of cDNA, hybridization of cDNA to microarrays, microarray processing, and microarray scanning were performed as described in the Fairplay III labeling kit (Agilent Technologies, 252009, version 1.1). Microarrays were scanned with a high-resolution microarray scanner (Agilent Technologies). Features with background intensities exceeding 10 times the array median, or with a signal/background ratio below 3 were excluded from further analysis. Background correction (37), within-array Loess normalization (38), and between-array quantile normalization were applied to the remaining features using the R statistical package LIMMA from Bioconductor (39). Moderated t statistics were calculated using gene-wise linear models with an empirical Bayes approach (40,41). P values were adjusted for multiple testing using the Benjamini-Hochberg method (42). Transcripts exhibiting Ն2-fold change in abundance with an adjusted P value of Ͻ0.05 were deemed to be differentially regulated. The data are available in ArrayExpress. The relative activities of transcription factors were inferred using the TFInfer software package as previously described (14).
Metabolomics. Quenching of fermentation samples (0.5 ml) was performed by mixing ethanol solution (40% vol/vol) prepared in NaCl (0.8% wt/vol) at Ϫ20°C (43). The mixture was centrifuged (16,000 ϫ g, 2.5 min, 4°C). The supernatant was discarded, and the pellets were stored (Ϫ80°C) until required. Cell pellets were resuspended to an OD 600 of 1 in chloroform-methanol (1:1, Ϫ20°C), and 1 ml was transferred to a fresh polypropylene tube. An ice-cold ball bearing was added to each tube before incubation (Ϫ80°C, 1 h). The samples were then vortexed twice (30 s) and returned to the Ϫ80°C freezer for 1 h. Milli-Q water (400 l) was added to each sample, followed by vortex mixing (30 s). Samples were then centrifuged (4,000 ϫ g, 1 min, 4°C). The upper layer was transferred to a fresh tube. This step was repeated, and the two organic phases were combined. Electrospray ionization time-of-flight (ESI-TOF) MS was performed on a Hybrid quadrupole time-of-flight (TOF) LC-MS-MS spectrometer (Waters, Ltd., Manchester, United Kingdom) based on the methods described by Davey et al. (44) and Walker (45). Data acquisition and processing were performed on MassLynx (version 4) to create centroid peak lists (m/z accurate to 4 decimal places versus ion counts), which were then transferred to Microsoft Excel (Microsoft Corp., USA) as text files. The mass spectrometer was operated in negative-ion mode at a rate of 1 scan per s for 6 min. Samples were loaded using a syringe pump (Razel, CT, USA) at a flow rate of 20 l min Ϫ1 . A Lockspray interface was used to give an external standard and allow automated correction of mass measurements (5 ng liter Ϫ1 sulfadimethoxine, giving a lock mass of 309.0653). Samples were analyzed in a randomized order to minimize effects of day-to-day machine variation. Data processing and downstream analysis were performed in R, using Bioconductor package XCMS (46). Peaks were aligned across analytical replicates and grouped into 0.2-m/z-width bins (47). Peaks were rejected if all three replicates were not present or the mass variance fell outside an acceptable range defined as a function of the m/z (formula modified from Overy et al. [48]; H. Walker, personal communication).
Proteomics. (i) Protein extraction and peptide sample preparation. Fermentation samples (500 l) were harvested by centrifugation (16,000 ϫ g, 2.5 min, 4°C). The supernatant was discarded and the pellet stored (Ϫ80°C). Cells were embedded in a 20% polyacrylamide gel matrix and digested and washed following the gel-aided sample preparation (GASP) protocol with three minor changes (49). First, samples were reduced using 10 mM Tris(2-carboxyethyl)phosphine (TCEP); second, they were desalted after gel extraction on Sola C 18 cartridges following the manufacturer's instructions; third, after the samples were dried to near dryness, they were resuspended in 0.1% formic acid for injection into the mass spectrometry-liquid chromatography system. Injections were normalized by a quantitative colorimetric assay (Thermo) that uses a modified bicinchoninic acid (BCA) chemistry for sensitive and accurate peptide sample quantification to keep samples comparable and to optimize signal intensity.
The buffers used for the GASP method were as follows. Lysis was performed on ice (30 min) in a buffer containing urea (6 M), thiourea (1.5 M), SDS (4%), TCEP (10 mM), 20 mM Tris (20 mM, pH 8.0). Alkylation was achieved on ice (30 min) using an equal volume of monomeric acrylamide mix solution (40%, 37.5:1 acrylamide-bisacrylamide solution). Polymerization was initiated at room temperature by addition of N,N,N=,N=-tetramethylethylenediamine (TEMED [5 l]) and ammonium persulfate (APS [10%]) followed by rapid mixing by vortexing. After polymerization was complete (ϳ30 min), gel plugs were cut by centrifugation through a membrane-less SpinX centrifugation filter support device (Costar) to achieve identical sizes of gel pieces. Gel pieces were fixed by addition of fixation buffer (1 ml) containing methanol-acetic acid-water (50:40:10) and overhead rotation (10 min). After brief pulse centrifugation, the supernatant was discarded and gel pieces were rehydrated with urea (6 M, 0.5 ml, 10 min). Acetonitrile (1 ml) was used to dehydrate gel pieces. Following removal of the supernatant using a gel loading pipette tip (200-l-volume pipette tip with a long narrow end to avoid loss of gel pieces), two more rounds of urea and acetonitrile washes were performed to minimize carryover of SDS and other contaminants. After these washes, the gel pieces were pH adjusted using triethylammonium bicarbonate (TEAB [500 l, 50 mM, pH 8.0]) and rotation (10 min). Gel pieces were dehydrated once more using acetonitrile (1 ml), supernatant was removed, and acetonitrile (0.5 ml) was used for more thorough dehydration of the gel pieces. Dried-out gel pieces were soaked with trypsin solution (1/50 enzyme/ substrate ratio based on expected protein content from known cell mass in Tris buffer [20 mM, pH 8.0]). Digestion was performed overnight with shaking (1,000 rpm, 37°C). Elutions were performed using one gel volume: elution 1 with acetonitrile, 2 with formic acid (5%), and 3 with acetonitrile. Elutions 2 and 3 were combined.
(ii) Mass spectrometry analysis. Peptides (1 g) from each label-free sample were injected and separated via 1D-ultrahigh-pressure liquid chromatography (UHPLC) using a Dionex Ultimate 3000 RSLC nitrogen gas flows for sheath, auxiliary, and sweep gases were set to 42, 14, and 1 arbitrary unit, respectively. The ion transfer tube was set to 300°C and the HESI vaporizer to 280°C. Data were acquired in a set 1-s cycle time, with high resolution (240,000 full width at half-maximum [FWHM] at m/z 200) MS1 profile spectra acquired over the m/z 200 to 1,600 scan range, with an AGC target of 200,000 and a maximum ion injection time of 100 ms. MS1 data were internally calibrated during acquisition to Ͻ1-ppm mass error using the installed Thermo Easy IC option. Data-dependent MS2 scans were collected in parallel using a 1-m/z quadrupole isolation window and scanned at approximate unit mass resolution in centroid mode from the ion trap, alternating between higher-energy collisional dissociation (HCD) and collision-induced dissociation (CID) fragmentation modes. Only MS1 masses of Ͼ400 m/z were selected for fragmentation, with an AGC target of 10,000 and max ion time of 50 ms. HCD spectra were collected in stepped collision energy mode centered at 40% and CID spectra at a fixed 40%. Dynamic exclusion was enabled, with exclusion after n ϭ 1 times, for 6 s.
For analysis of LC-MS-MS data, Xcalibur .raw files were converted to centroided. mzML format using the msconvert function of ProteoWizard (release 3.0.18114). Peak processing workflows were conducted in R 3.5.0. MS1 peaks were extracted using the "centWaveWithPredictedIsotopeROIs" method from the XCMS package (version 3.2.0 [46,51]). Peaks were grouped across samples, missing peaks imputed by reintegration within group boundaries, and group median m/z values further processed using the CAMERA package (version 1.23.3 [52]), to identify isotopes and adducts. Candidate formulas were generated with adapted code from the rcdk package (version 3.4.9 [53]), with the following limits: positive mode, C10-300, H20-500, O0-20, N0-3, P0-2, Na02, RDBE Ϫ0.5-18, 2-ppm error; negative mode, C10-300, H20-500, O0-20, N0-3, P0-2, S0-2, RDBE Ϫ0.5-18, 20-ppm error. Consensus peak groups were then filtered with custom R scripts to (i) exclude any peak groups where any peak in the group was present with an area less than the mean ϩ 3 standard deviations of the value from blank extracts, (ii) only retain the most intense monoisotopic peak identified by CAMERA, and (iii) only retain peaks with valid molecular formulas. MS1 peaks were searched against downloaded local copies of the E. coli metabolome database (ECMDB [54]) and the Lipid Maps Structure Database (LMSD [http://www.lipidmaps.org]). Consensus HCD and CID MS2 spectra were extracted with custom scripts and searched against the in silico LipidBlast (55) and LipidMatch (56) databases. Peaks were annotated following manual examination of MS2 spectra in consensus with returned database hits. All retained peaks were normalized to the SPLASH deuterated PC [15:0_18:1(d7)] internal standard and then to sample dry weight. Statistical analyses were carried out on glog normalized data (lambda value in glog transform taken to be 1/10 of nonzero minimum), using time-series ANOVA2 models from the online MetaboAnalyst resource (https:// www.metaboanalyst.ca).
Data availability. Transcriptomics data are available in ArrayExpress under accession no. E-MTAB-7257. The mass spectrometry proteomics data have been deposited in the ProteomeXchange Consortium database via the PRIDE partner repository under accession no. PXD013088 (50).