In Vivo Thermodynamic Analysis of Glycolysis in Clostridium thermocellum and Thermoanaerobacterium saccharolyticum Using 13C and 2H Tracers

Thermodynamics constitutes a key determinant of flux and enzyme efficiency in metabolic networks. Here, we provide new insights into the divergent thermodynamics of the glycolytic pathways of C. thermocellum and T. saccharolyticum, two industrially relevant thermophilic bacteria whose metabolism still is not well understood. We report that while the glycolytic pathway in T. saccharolyticum is as thermodynamically favorable as that found in model organisms, such as E. coli or Saccharomyces cerevisiae, the glycolytic pathway of C. thermocellum operates near equilibrium. The use of a near-equilibrium glycolytic pathway, with potentially increased ATP yield, by this cellulolytic microbe may represent an evolutionary adaptation to growth on cellulose, but it has the drawback of being highly susceptible to product feedback inhibition. The results of this study will facilitate future engineering of high-performance strains capable of transforming cellulosic biomass to biofuels at high yields and titers.

T hermoanaerobacterium saccharolyticum and Clostridium thermocellum are thermophilic, anaerobic bacteria with complementary metabolic capabilities being developed for industrial-scale production of biofuels, such as ethanol, from lignocellulosic biomass (1)(2)(3). C. thermocellum readily solubilizes lignocellulosic biomass and ferments cellulose-derived sugars, including oligomers, to ethanol, acetate, lactate, formate, and H 2 . However, C. thermocellum is incapable of utilizing the hemicellulose fraction of biomass (containing xylose, arabinose, mannose, and galactose) (4,5). Metabolic engineering efforts have increased the native ethanol yields and titers of C. thermocellum, but the best strains, with a titer of ϳ25 g/liter and 75% theoretical yield, do not yet match the productivity of noncellulolytic ethanologens such as Saccharomyces cerevisiae and Zymomonas mobilis (6)(7)(8)(9). In contrast, T. saccharolyticum cannot solubilize lignocellulosic biomass or cellulose but readily ferments hemicellulose-derived sugars, including oligomers, to ethanol, acetate, lactate, and H 2 (4,10). T. saccharolyticum has been engineered to produce ethanol at greater than 90% theoretical yields and at titers of up to 70 g/liter (3,10). These two thermophilic bacteria have been used in cocultures that seek to combine the cellulolytic capability of C. thermocellum with the higher ethanol productivity and hemicellulose-consuming capability of T. saccharolyticum (11). Several previous studies have directly compared the fermentation capabilities of C. thermocellum and T. saccharolyticum, and T. saccharolyticum has been used frequently as a source of genes to engineer enhanced ethanol productivity in C. thermocellum (12)(13)(14)(15)(16)(17)(18)(19).
Recently developed experimental approaches for estimating in vivo Gibbs free energies (⌬G) of metabolic reactions have been applied to investigate the thermodynamics of glycolytic pathways in a few model organisms and biofuel producers (27)(28)(29). The estimation of ⌬G by these approaches relies on the fundamental relation of ⌬GϭϪRT ln(J ϩ /J Ϫ ) (30) and the determination of forward (J ϩ ) to backward (J Ϫ ) flux ratios (J ϩ /J Ϫ ) from isotope tracer experiments (R is the gas constant and T is the temperature in kelvin) (27,29). These studies, together with theoretical and computational advances, have provided new insights on the connection between pathway thermodynamics, flux, and enzyme efficiency (31)(32)(33)(34).
Here, we integrated quantitative metabolomics with 2 H and 13 C metabolic flux analysis to investigate the in vivo reversibility and thermodynamics of the glycolytic pathways and central metabolic networks of T. saccharolyticum and C. thermocellum. We show that glycolysis in C. thermocellum is highly reversible at every step of the pathway and operates close to thermodynamic equilibrium. In comparison, the glycolytic pathway in T. saccharolyticum is highly thermodynamically favorable, comparable  to that of anaerobically grown Escherichia coli. We also found that the ethanol fermentation pathway of C. thermocellum was substantially more reversible than that of T. saccharolyticum. Besides revealing differences in pathway reversibility, our analyses also represent the first experimental isotope-based reconstruction of the T. saccharolyticum central metabolic network. We found that this bacterium metabolizes glucose exclusively via the EMP pathway, while the Entner-Doudoroff (ED) pathway and the oxidative pentose phosphate pathway (oxPPP) are inactive. Our analyses also revealed a bifurcated tricarboxylic acid (TCA) cycle and a sedoheptulose bisphosphate bypass within the pentose phosphate pathway (PPP). Finally, we present evidence supporting the activity of several currently unannotated amino acid biosynthetic pathways in T. saccharolyticum.

RESULTS
Reconstruction of the T. saccharolyticum central metabolic network. The application of 2 H and 13 C metabolic flux analysis to measure in vivo reversibility and thermodynamics of metabolic reactions requires prior knowledge of the underlying metabolic network topology. Recent studies have used isotope tracers, metabolic flux analysis (MFA), proteomics, or biochemical assays to elucidate the central metabolic network of C. thermocellum (14,18,20,21,24,26,(35)(36)(37)(38). However, although the genome sequence of T. saccharolyticum is available and systems-level approaches were used to characterize its physiology (22,23,39,40), an isotope-based reconstruction of its metabolic network is currently lacking. To address this knowledge gap, we performed parallel steady-state labeling experiments by growing cells in [1-13 C 1 ]glucose, [6-13 C 1 ]glucose, [1,2-13 C 2 ]glucose, or a 50:50 mixture of [U-13 C 6 ]glucose and unlabeled glucose. Using these 13 C labeling data, we first carried out a manual evaluation and reconstruction of the T. saccharolyticum central metabolic network. In a subsequent section (i.e., " 2 H and 13 C metabolic flux analysis," below), we present a quantitative MFA that corroborates the metabolic network reconstruction presented in this section.
(i) Glucose catabolism. The Embden-Meyerhof-Parnas (EMP) and the Entner-Doudoroff (ED) pathways are the two most common glycolytic pathways found in microbial species, although several other variants have been described (20,31,(41)(42)(43). The EMP and ED pathways share a common set of reactions that metabolize glyceraldehyde-3-phosphate (GAP) into pyruvate (Pyr) (i.e., lower glycolysis), but they differ in their initial steps ( Fig. 2A and B). In the ED pathway, the intermediate 2-keto-3-deoxy-6-phosphogluconate (KDPG) is cleaved into Pyr (first three carbons) and GAP (last three carbons). Therefore, unlike in the EMP pathway (in which two GAP molecules are produced from glucose), the first three carbons of glucose bypass lower glycolysis in the ED pathway. Additionally, carbons 1, 2, and 3 of glucose catabolized via the ED pathway become the carboxyl, carbonyl, and methyl carbons, respectively, of Pyr, which is the reverse order of that obtained via the EMP pathway (Fig. 2C). T. saccharolyticum has genes encoding all of the enzymes in the EMP pathway, but its genome annotation does not support a complete ED pathway. When T. saccharolyticum was grown on [1-13 C 1 ]-or [6-13 C 1 ]glucose, the glycolytic intermediates 3-phosphoglycerate (3PG) and phosphoenolpyruvate (PEP) both were ϳ50% M ϩ 0 (i.e., unlabeled; all carbons are 12 C, where M represents the unlabeled parent mass) and ϳ50% M ϩ 1 (i.e., containing one 13 C carbon), which was consistent with their production via EMP glycolysis ( Fig. 2A (27). Finally, as shown in Table S1 in the supplemental material, high intracellular levels of the EMP intermediates FBP and dihydroxyactone phosphate (DHAP), which were comparable to those in anaerobic E. coli, a known EMP-utilizing organism, together with nondetectable levels of the ED pathway intermediates 6PG and KDPG, which are present in the ED pathway, utilizing organisms such as Z. mobilis, was consistent with the existence of an active EMP pathway and an inactive ED pathway in T. saccharolyticum (27,44).
Interestingly, we found the production of a large fraction (30 to 40%) of M ϩ 0 S7P in both [1,2-13 C 2 ]-and [1-13 C 1 ]glucose that could not be explained by canonical nonoxidative PPP reactions ( Fig. 3A and B). For example, taking into account the observed labeling patterns of PPP intermediates during growth in [1,2-13 C 2 ]-or [1-13 C 1 ]glucose, no more than 10% M ϩ 0 S7P should be present under either condition if Tal and Tkt1 were the only reactions producing S7P. Therefore, the large unexpected fraction of M ϩ 0 S7P suggested that another route for S7P production should be present. Specifically, these data suggested that T. saccharolyticum possesses a sedoheptulose bisphosphate (SBP) bypass comprising an SBP aldolase (Sba; DHAP ϩ E4P¡SBP) and an SBP phosphatase (SBPase; SBP ϩ P i ¡S7P ϩ pyrophosphate [PP i ]) (24,26). While there are no genes associated with SBPase activity in T. saccharolyticum, several genes (i.e., Tsac_0260, Tsac_0328, and Tsac_2313) are annotated as both Fbas and Sbas (22,23), and one or more of these could be responsible for the observed Sba activity (48).
(iii) TCA cycle. Several TCA cycle enzymes remain unannotated in the T. saccharolyticum genome, including citrate synthase and malate dehydrogenase. Our 13 C tracer data indicated that the TCA cycle in T. saccharolyticum functions as a branched pathway to produce essential biosynthetic precursors: ␣-ketoglutarate (AKG) was produced oxidatively despite the lack of an annotated citrate synthase, while oxaloacetate (OAA), malate (Mal), and succinate all were produced reductively. During growth in [6-13 C 1 ]glucose, ␣-ketoglutarate (AKG) labeling patterns were consistent with its production from oxaloacetate and acetyl-coenzyme A (AcCoA) via citrate and isocitrate (Fig. 4). However, although AKG was ϳ23% M ϩ 2, succinate was Ͻ0.6% M ϩ 2, indicating negligible oxidative succinate production from AKG via an AKG dehydrogenase. Although several genes in T. saccharolyticum are annotated as a 2-oxoacid/AKG:ferredoxin dehydrogenase, there is no gene specifically annotated as an AKG dehydrogenase (22,23,45). Instead, labeling patterns in succinate, Mal, and aspartate (here being
(v) Amino acid synthesis and one-carbon metabolism. Our 13 C-tracer data confirmed that aspartate, asparagine, and threonine were produced via canonical routes starting from oxaloacetate (Fig. S2A). Similarly, our data supported production of glutamate, glutamine, ornithine, citrulline, and proline via canonical pathways starting with AKG (Fig. S2B). The amino acids valine and leucine were also produced via canonical branched-chain amino acid synthesis pathways from the intermediate ketoisovalerate (Fig. S2C). Finally, 13 S2). In addition to its production via Fumh, fumarate can be produced by deamination of aspartate, which occurs in biosynthetic pathways such as purine biosynthesis. This route of fumarate production would result in the same labeling patterns as its production from oxaloacetate via Mdh and Fumh. The increased proportion of M ϩ 0 pyruvate, succinate, ␣-ketoglutarate, and malate was due to extracellular pools of these metabolites that do not become labeled. CO 2 labeling was inferred from the mass isotopomer distributions (MIDs) of citrulline and ornithine, as described in Materials and Methods. The diagram on the left shows predicted metabolite labeling with [6-13 C 1 ]glucose. Pyruvate formate lyase (Pfl) was omitted from the figure but serves the same role as Pfor. Solid arrows represent a single reaction; dashed arrows represent multiple reaction steps. 13  Thermodynamics of Glycolysis in Thermophilic Bacteria phenylalanine and tyrosine were consistent with their production via the shikimate pathway (Fig. S2D).
Although the genes encoding the enzymes responsible for producing serine from 3PG are not annotated in T. saccharolyticum, our data indicated that serine was produced primarily from 3PG ( Fig. 6A and Fig. S3A and B). However, serine can also be synthesized from glycine via serine hydroxymethyltransferase (Shmt; 5,10methylenetetrahydrofolate [MeTHF] ϩ glycine¡serine ϩ tetrahydrofolate; Tsac_1185 [22,23]), and our data suggested that a measurable portion of the serine pool (i.e., ϳ22%) was derived (either net production or exchange) from glycine via Shmt (Fig. 6A and Fig. S3A and B).
T. saccharolyticum has annotated genes encoding most of the enzymes required for isoleucine biosynthesis from threonine, but it does not have an annotated threonine deaminase, which is the first step in this pathway. Similarly, most of the enzymes required for isoleucine synthesis via citramalate are annotated, but citramalate synthase is missing (22,23). Of these routes, the citramalate pathway (49) was necessary to explain isoleucine labeling patterns (Fig. 6B). During [1,2-13 C 2 ]glucose labeling, the production of M ϩ 6 isoleucine can only be explained by citramalate pathway activity, and the close to equal proportions of M ϩ 0 to M ϩ 6 and M ϩ 2 to M ϩ 4 labeled forms also support isoleucine production via this pathway. Labeling using an equimolar mixture of [U-13 C 6 ]glucose and unlabeled glucose supported primary production of isoleucine via the citramalate pathway and indicated that less than 10% of isoleucine was synthesized from threonine (Fig. S3D).
Reversibility of glycolytic reactions. Having reconstructed the central metabolic network of T. saccharolyticum, we sought to compare the reversibility of glycolytic reactions in T. saccharolyticum to those in C. thermocellum to identify potential factors influencing the greater ethanol productivity in T. saccharolyticum. As previously detailed in the introduction, the glycolytic pathway in C. thermocellum has several features that set it apart from the canonical EMP glycolytic pathway, including distinct cofactor utilization at specific steps in the pathway and the use of alternative routes for the conversion of PEP to pyruvate (Fig. 1). We hypothesized that the nonstandard glycolytic pathway of C. thermocellum will display substantial differences in reaction reversibility and thermodynamics compared to those of the canonical EMP glycolytic pathway, such as that possessed by T. saccharolyticum. We used 13 C and 2 H steady-state labeling experiments to directly compare reaction reversibility between the glycolytic pathways of C. thermocellum, T. saccharolyticum, and anaerobically grown Escherichia coli as an additional point of reference.

FIG 6
One carbon metabolism and aromatic amino acid production. (A) [6-13 C 1 ]glucose labeling indicated that in T. saccharolyticum serine was produced from both 3PG and glycine. In addition, the production of a small fraction of M ϩ 2 methionine indicated the presence of 13 C-labeled C1 units in 5-methyltetrahydrofolate (MTHF), which is produced from 5,10-methylenetetrahydrofolate (MeTHF). This implied partial production of C1 units from serine via SMHT. These conclusions were supported by [1-13 C 1 ] and [1,2-13 C 2 ]glucose labeling data (Fig. S3). (B) Isoleucine may be produced from threonine or via the citramalate pathway. [1,2-13 C 2 ]glucose labeling data were consistent with production of isoleucine via the citramalate pathway in T. saccharolyticum. Specifically, production of M ϩ 6 isoleucine can only be explained by citramalate pathway activity, and the close-to-equal proportions of M ϩ 0 to M ϩ 6 and M ϩ 2 to M ϩ 4 labeled forms also support isoleucine production via this pathway. Labeling experiments using an equimolar mixture of [U-13 C 6 ]glucose and [U-12 C 6 ]glucose supported this conclusion (Fig. S3D). CO 2 labeling was inferred from the MIDs of citrulline and ornithine. The diagrams on the left show predicted metabolite labeling with the specified isotope tracer. Solid arrows represent a single reaction; dashed arrows represent multiple reaction steps. 13  (Continued on next page) Thermodynamics of Glycolysis in Thermophilic Bacteria saccharolyticum, and E. coli were ϳ38%, ϳ37%, and ϳ24%, respectively, indicating that Fba is highly reversible in all three organisms but significantly less so in E. coli (P Ͻ 0.05) (Fig. 7A). Once produced, M ϩ 3 FBP may propagate back to F6P and G6P. Therefore, the reversibility of the upstream reactions with Pfk and phosphoglucose isomerase (Pgi; G6P¡F6P) is informed by the production of M ϩ 3 F6P and M ϩ 3 G6P. The substantially larger M ϩ 3 F6P fraction in C. thermocellum (ϳ23%) versus T. saccharolyticum (ϳ3%) and E. coli (ϳ5%) indicates that phosphofructokinase is substantially more reversible in this bacterium (P Ͻ 0.005). This observation is consistent with PP i -Pfk in C. thermocellum being less energetically favorable than the ATP-Pfk used by T. saccharolyticum and E. coli (20,25,50,51). As shown in Fig. 7B and C, differences in the reversibility of the Pfk and Fba reactions across the three bacteria were also evident   (Fig. 7D). For example, during [6-13 C 1 ]glucose labeling, forward Fba activity first produces M ϩ 0 DHAP and M ϩ 1 GAP. Forward Tpi activity then produces M ϩ 0 GAP from M ϩ 0 DHAP. However, reverse Tpi flux results in the production of M ϩ 1 DHAP from M ϩ 1 GAP. Thus, the closer the M ϩ 1 DHAP fraction is to 50%, the closer Tpi is to equilibrium (i.e., equal forward and reverse flux). The fractions of M ϩ 1 DHAP in C. thermocellum, T. saccharolyticum, and E. coli were 49.8%, 46.4%, and 43.5%, respectively, indicating that Tpi is highly reversible in all three organisms. The data indicated that Tpi was significantly more reversible in C. thermocellum than E. coli (P Ͻ 0.005), although the difference between C. thermocellum and T. saccharolyticum did not reach statistical significance (P ϭ 0.07).
(ii) Reversibility of lower glycolysis (GAP to PEP). 2 H tracers can provide information on the reversibility of reactions where a C-H bond is broken or formed via dehydration, isomerization, or dehydrogenation (27)(28)(29). Hence, we used [4-2 H 1 ]glucose and [5-2 H 1 ]glucose for investigating reversibility of lower glycolytic reactions (i.e., GAP to PEP) in which carbon rearrangements do not occur.
As shown in Fig. 8A, during [4-2 H 1 ]glucose labeling, [1-2 H 1 ]GAP (i.e., GAP with a deuterium atom bound to its first carbon) is produced via Fba. Conversion of [1-2 H 1 ]GAP to 1,3-bisphosphoglycerate (BPG), catalyzed by glyceraldehyde-3-phosphate dehydrogenase (GAPDH), removes the deuterium from [1-2 H 1 ]GAP and places it on NAD ϩ , generating 2 H-labeled NADH. Reverse GAPDH flux then can produce unlabeled GAP if unlabeled NADH is used. The M ϩ 0 GAP produced can propagate to DHAP (via reverse Tpi) and FBP (via reverse Fba), as well as to other upstream glycolytic intermediates. We observed that during growth on [4-2 H 1 ]glucose, C. thermocellum displayed a larger loss of 2 H labeling across all glycolytic intermediates than T. saccharolyticum and E. coli, which indicated increased reversibility of glycolytic reactions in this bacterium. For example, although we could not reliably measure GAP, higher fractions of M ϩ 0 DHAP and M ϩ 0 FBP in C. thermocellum, i.e., 92% and 72%, respectively, versus 82% and 62% in T. saccharolyticum or 89% and 43% M ϩ 0 in E. coli, suggested greater reversibility of its GAPDH/Tpi and GAPDH/Fba pairs of reactions (P Ͻ 0.05 for all pairwise comparisons).
During growth on [5-2 H 1 ]glucose, the intermediates GAP, BPG, 3PG, and 2-phosphoglycerate (2PG) become deuterated at the second carbon (Fig. 8B). The loss of deuterium can occur from the conversion of [2-2 H 1 ]GAP to M ϩ 0 DHAP via reverse Tpi or by the forward phosphopyruvate hydratase reaction (Eno) converting [2-2 H 1 ]2PG to M ϩ 0 PEP. Once produced, propagation of M ϩ 0 DHAP or M ϩ 0 PEP to upstream metabolites provides a measure of the reversibility of Tpi, Eno, and other glycolytic reactions. As with [4-2 H 1 ]glucose, the greater loss of 2 H labeling across glycolytic intermediates in C. thermocellum during [5-2 H 1 ]glucose labeling indicated increased reversibility of glycolysis in this bacterium compared to that of T. saccharolyticum or E. coli. For example, taking Tpi reversibility into account (Fig. 7D), the larger fractions of M ϩ 0 3PG in C. thermocellum (98.4%) and T. saccharolyticum (96.6%) suggested increased reversibility of the Pgm/Eno reaction pair in these thermophilic organisms compared to that in E. coli (88.4%; P Ͻ 0.05 for both pairwise comparisons).
When cells are grown in an equimolar mixture of [U-13 C 6 ]glucose and unlabeled glucose, the Pyr initially produced via glycolysis is either M ϩ 0 or M ϩ 3, and its decarboxylation produces acetyl-CoA containing an M ϩ 0 or M ϩ 2 acetyl group. Complete or partial (i.e., decarboxylation step only) reversibility of Pfor or Pfl resulted in the interchange of the carboxylic group in Pyr with unlabeled CO 2 or formate to
Once formed, M ϩ 2 Pyr may propagate to upstream metabolites to produce M ϩ 2 PEP, indicating reversibility of PEP-to-Pyr conversion by whichever route the organism has available. Despite a potentially longer path, we observed that PEP-to-Pyr conversion was significantly more reversible (P Ͻ 0.01) in C. thermocellum (10.2% M ϩ 2 PEP) than in E. coli (2.0% M ϩ 2 PEP). Data suggested that PEP-to-Pyr conversion was also more reversible in C. thermocellum than T. saccharolyticum (3.9% M ϩ 2 PEP), but this difference was not significant at P Ͻ 0.05. Considering the reported absence of pyruvate carboxylase or oxaloacetate decarboxylase activity in C. thermocellum (18,26), the production of M ϩ 2 Mal (14.0%) and M ϩ 2 aspartate (12.0%; surrogate for OAA) suggested considerable reversibility of the malate shunt. Another potential alternative for the production of M ϩ 2 aspartate and M ϩ 2 Mal in C. thermocellum is a reversible citrate synthase, which has been shown to be reversible in other thermophilic bacteria, such as Desulfurella acetivorans and Thermosulfidibacter takaii (61,62).

H and 13 C metabolic flux analysis.
To determine in vivo flux ratios (J ϩ /J -) and compare the thermodynamics of the distinct glycolytic pathways of C. thermocellum and T. saccharolyticum, we integrated data from five 13 Table S2). MFA was performed using the INCA software suite, which utilizes the elementary metabolite unit (EMU) framework to model isotopic distributions (63,64). Reaction directionalities were not predetermined for most reactions in central carbon metabolism (glycolysis, TCA cycle, and PPP) to allow combined 2 H and 13 C MFA to determine forward and reverse fluxes. Reaction directionality was preassigned primarily for transport reactions, dilution fluxes, and combined amino acid biosynthesis reactions.
For T. saccharolyticum, 2 H and 13 C MFA best-fit solutions validated the qualitative assessment of its metabolic network, presented earlier. We constructed alternative models containing all possible glycolytic routes (e.g., EMP, ED, and oxidative pentose phosphate pathways) as well as models lacking reactions that we identified as necessary to explain labeling patterns of specific metabolites (e.g., citramalate pathway for isoleucine production and SBP bypass for S7P production). The results of this analysis agreed with our previous conclusions and predicted an inactive ED pathway (i.e., Ͻ0.25% glycolytic flux), an inactive oxidative PPP (i.e., 0% flux), and a bifurcated TCA cycle with oxidative production of AKG and reductive production of succinate. Although Pepck typically is considered a gluconeogenic reaction (i.e., transforming OAA to PEP), our MFA results supported the activity of a PEP carboxylating enzyme producing OAA from PEP. We also found that the SBP bypass was necessary to explain labeling patterns in PPP intermediates and aromatic amino acids. This analysis also supported qualitative conclusions regarding amino acid biosynthetic routes, such as isoleucine synthesis via the citramalate pathway instead of from threonine. Additionally, the production of serine from 3PG and its consumption by Shmt to form labeled THF intermediates was necessary to achieve a statistically acceptable model fit.
For C. thermocellum, our MFA results were also consistent with the previously reported central metabolic network (26). Specifically, our analysis predicted glycolysis via the EMP pathway with no ED pathway activity, a solely nonoxidative PPP containing a SBP bypass, and a bifurcated TCA cycle. Our results were also consistent with reported amino acid synthesis pathways in C. thermocellum.
Comparative thermodynamic pathway analysis. Using the relation ⌬GϭϪRT ln(J ϩ /J Ϫ ), we estimated ⌬G of glycolytic reactions (i.e., from G6P to PEP) from MFAderived flux ratios (J ϩ /J Ϫ ) in C. thermocellum and T. saccharolyticum (Fig. 10B and  into a single, statistically acceptable flux map. Arrow thickness is proportionate to net flux. Numbers next to each arrow represent net flux, with reverse fluxes in parentheses. Fluxes are represented as a proportion of glucose uptake, which was normalized to 100. Malic enzyme was omitted from the C. thermocellum model due to the analysis being unable to resolve flux through malic enzyme versus Ppdk. Omission of malic enzyme had a negligible effect on flux ratios in glycolysis (see Materials and Methods). *, see Table S2 for complete MFA results, including 95% confidence intervals. (B) MFA-derived ΔG values for glycolytic reactions (Table 1) revealed that the glycolytic pathway in C. thermocellum was remarkably close to thermodynamic equilibrium, with an overall drop in Gibbs free energy of Ϫ4.55 kJ/mol between G6P and PEP. In contrast, the overall drop in free energy of glycolysis in T. saccharolyticum was substantially greater, ΔG ϭ Ϫ28.67 kJ/mol. The limited thermodynamic driving force of the glycolytic pathway in C. thermocellum could be attributed in large part to the small driving force of the Pfk reaction, which had a ΔG of only Ϫ1.45 kJ/mol compared to Ϫ22.57 kJ/mol in T. saccharolyticum. Intracellular metabolite concentrations provided crossvalidation of ΔG estimates for the Pfk reaction in T. saccharolyticum (see Materials and Methods). Error bars shown represent the 95% confidence intervals of the free energy estimates. Error bars smaller than the icon used to represent the average are not visible.
Thermodynamics of Glycolysis in Thermophilic Bacteria Table 1). For a small number of reactions, we used intracellular metabolite concentrations to improve ⌬G estimates obtained from MFA-derived flux ratios (Table 1). Estimated ⌬G values aligned with the qualitative assessment of reaction reversibility presented above and showed that the glycolytic pathway in C. thermocellum, with an overall ⌬G of Ϫ4.55 kJ/mol, operates remarkably close to thermodynamic equilibrium. In comparison, the glycolytic pathway in T. saccharolyticum was substantially more thermodynamically favorable (overall ⌬G ϭ Ϫ28.67 kJ/mol) and was comparable to the glycolytic pathway in anaerobically grown E. coli (overall ⌬G ϭ Ϫ23.47 kJ/mol) (29) (Fig. S4). The limited thermodynamic driving force of the glycolytic pathway in C. thermocellum could be attributed in large part to the small driving force of the Pfk reaction, which had a ⌬G of only Ϫ1.45 kJ/mol compared to Ϫ22.57 kJ/mol in T. saccharolyticum and Ϫ18.84 kJ/mol in E. coli. The small ⌬G of Pfk in C. thermocellum was likely due to the use of PP i instead of ATP as the high-energy phosphate donor. Indeed, the standard free energy for the ATP-dependent Pfk (⌬G ϭ Ϫ15.0 kJ/mol) is ϳ10.6 kJ/mol more favorable than the PP i -dependent Pfk (⌬G ϭ Ϫ4.4 kJ/mol) due to the greater free energy of ATP hydrolysis than PP i hydrolysis (50).
Consistent with our qualitative observations regarding its reversibility, the Fba reaction operated close to thermodynamic equilibrium in both T. saccharolyticum (Ϫ0.70 kJ/mol) and C. thermocellum (Ϫ0.47 kJ/mol) but was more thermodynamically favorable in anaerobic E. coli (Ϫ3.64 kJ/mol) (29). Also in agreement with qualitative observations, the Pgi reaction was slightly more favorable in T. saccharolyticum (Ϫ1.93 kJ/mol) than in C. thermocellum (Ϫ0.74 kJ/mol). The ⌬G of lower glycolytic reactions (GAPDH to Eno) were close to equilibrium in both C. thermocellum and T. saccharolyticum but were generally more favorable in T. saccharolyticum. For example, the ⌬G of the Pgm/Eno pair of reactions was Ϫ0.49 kJ/mol in C. thermocellum compared to Ϫ0.77 kJ/mol in T. saccharolyticum.

DISCUSSION
Metabolite concentrations and thermodynamics of glycolysis. Intracellular concentrations of glycolytic intermediates across the bacteria in this study reflected the thermodynamic profile of their glycolytic pathways (see Table S1). For example, the limited driving force of the PP i -Pfk reaction in C. thermocellum was associated with a low FBP concentration, which was about one-third of that in T. saccharolyticum or E. coli. In contrast, G6P and F6P levels in C. thermocellum were higher than those in T. saccharolyticum (6.8-fold and 1.2-fold, respectively) and also higher than those in E. coli (4.1-fold and 5.0-fold, respectively). Higher levels of G6P and F6P in C. thermocellum Interestingly, ATP/ADP and GTP/GDP ratios were considerably lower in C. thermocellum (25.7 and 28.7, respectively) than E. coli (66.0 and 42.4) and T. saccharolyticum (59.9 and 61.4). The higher ATP/ADP ratios in E. coli and T. saccharolyticum likely contribute to the favorability of the ATP-Pfk reaction and other reactions in upper glycolysis in these organisms. In contrast, the low ATP/ADP and GTP/GDP ratios in C. thermocellum may help make ATP/GTP-generating reactions in lower glycolysis, such as GAPDH and Ppdk, more favorable but may also slow biosynthetic reactions that rely on ATP/GTP hydrolysis, such as amino acid, protein, fatty acid, nucleotide, and mRNA synthesis.
Enzyme efficiency, metabolic flux, and product titers. Thermodynamics constitutes a key determinant of enzyme efficiency and flux within metabolic pathways. As determined by ⌬GϭϪRT ln(J ϩ /J Ϫ ), enzyme efficiency (i.e., the fraction of enzyme used in the forward versus reverse direction) is directly proportional to the drop in Gibbs free energy (⌬G) of a biochemical reaction (31,32). Thus, the fraction of enzyme that counterproductively catalyzes the reverse reaction (J Ϫ ) increases exponentially as ⌬G approaches equilibrium, decreasing the net reaction rate (J ϩ -J Ϫ ) (33,65). This implies that a metabolic pathway that is close to thermodynamic equilibrium will require substantially greater amounts of catalytically active enzyme to maintain a given flux than a more thermodynamically favorable pathway. Therefore, considering cellular limits on total enzyme abundance (66), the small thermodynamic driving force of C. thermocellum glycolysis could represent a limiting factor for engineering ethanologen strains with high glycolytic rates.
Besides limiting flux, a small thermodynamic driving force also can make a pathway more susceptible to end product inhibition, since only a comparatively small amount of product would need to accumulate before some reaction within the pathway approaches thermodynamic equilibrium (⌬G¡0) and effectively stops further substrate utilization and product formation. Therefore, it is likely that the limited thermodynamic driving force of the C. thermocellum glycolytic pathway represents a key contributor to its low ethanol titer (25 g/liter) (6) compared to that of organisms with more thermodynamically favorable glycolysis, such as T. saccharolyticum (70 g/liter) (3), S. cerevisiae (75 g/liter) (67), and Z. mobilis (Ͼ85 g/liter) (27,68,69). Indeed, a recent computational evaluation that used maximum-minimum driving force (MDF)-based analysis to model the thermodynamics of the glycolytic and ethanol fermentation pathways in C. thermocellum predicted that the thermodynamic driving force of these pathways is rapidly depleted as ethanol accumulates extracellularly (70). This limitation may help explain the difficulties encountered by groups attempting to engineer C. thermocellum to increase product titers compared to the successes at engineering T. saccharolyticum (6,10,13,16,70).
Our results suggest that a promising strategy for increasing glycolytic flux and ethanol titers in C. thermocellum is to replace PP i -Pfk with an ATP-dependent Pfk. In addition, replacing the malate shunt and Ppdk routes for conversion of PEP to Pyr with Pyk also may increase flux and ethanol titers by providing a single more thermodynamically favorable route. Indeed, C. thermocellum strains expressing an exogenous Pyk achieve higher ethanol titers (13), and elimination of malate shunt activity improved ethanol production in Clostridium cellulolyticum, another cellulolytic ethanol producer (71). Finally, adding a mechanism for ATP hydrolysis could increase the thermodynamic favorability of ATP-harvesting reactions in glycolysis (Pgk and Pyk) (31). ATP demand has been implicated in glycolytic flux control in a number of organisms, including Z. mobilis, S. cerevisiae, and E. coli. (72)(73)(74). This strategy has a potential additional benefit in that it may limit cellular resources that are spent on replication by limiting access of biosynthetic reactions to ATP (72,74).
The previously reported MDF-based analysis of C. thermocellum also produced a computationally derived thermodynamic profile of glycolysis (60). Our in vivo measurement of pathway free energies agreed with the general thermodynamic profile produced by MDF, with most of the free energy released by the Pfk and Ppdk reactions and the rest of the reactions closer to equilibrium. Although the overall profiles were similar, their analysis predicted more favorable PP i -Pfk (Ϫ6.67 kJ/mol) and Ppdk (Ϫ6.92 kJ/mol) reactions than the values we report in this study (Ϫ1.45 and Ϫ3.35 kJ/mol, respectively). This discrepancy may be due to in vivo constraints not accounted for in the MDF analysis, such as limits on enzyme abundance and the need to maintain adequate metabolite pools to drive biosynthetic pathways (31,33,34). In addition, the cellular objective in C. thermocellum may not be completely aligned with the objective function implicit in the MDF algorithm (i.e., to maximize the lowest free energy given a set of reaction steps).
A trade-off between thermodynamic driving force and energy yield. Interestingly, the limited driving force of the glycolytic pathway in C. thermocellum mirrors that of C. cellulolyticum, another cellulose-degrading organism that also utilizes PP i -Pfk, Ppdk, and the malate shunt (29,71,75). A potential advantage of glycolytic pathways with low thermodynamic driving force is increased energy efficiency, i.e., increased ATP or GTP yield per glucose. The use of alternative enzymes, such as PP i -Pfk (which uses PP i in place of ATP) and Ppdk (which produces ATP from AMP and PP i ), by cellulolytic microbes represents a potential mechanism for increased energy yield. However, recent research suggests that the PP i needed for glycolysis is in excess of what can be generated as a byproduct of biosynthetic pathways and is instead generated by ATP-consuming glycogen cycling reactions (20). This casts some doubt on the use of PP i -dependent enzymes as a mechanism of ATP conservation. However, PP i also could be generated by a membrane-bound proton-pumping pyrophosphatase (PP i ase, Cthe_1425), which is transcribed and expressed in C. thermocellum (24,37). The proton gradient required for this process may originate from ferredoxin oxidation by an ion-tranlocating reduced ferredoxin:NAD ϩ oxidoreductase (Rnf) (76,77), with reduced ferredoxin provided by Pfor in the ethanol fermentation pathway (12,18). Thus, Pfor, Rnf, and PP i ase may work together to harvest additional energy from glycolysis by coupling pyruvate decarboxylation to proton translocation and subsequent PP i production. Therefore, PP i generation from a proton gradient represents a potential mechanism for energy conservation and increased ATP yield by a pathway that utilizes PP i -dependent enzymes.
The use of near-equilibrium glycolytic pathways with increased ATP yield by cellulolytic bacteria may represent an evolutionary adaptation to growth on cellulosic substrates. Specifically, microbes utilizing soluble substrates can maximize either the specific substrate consumption rate (grams of substrate ϫ gram cells Ϫ1 hour Ϫ1 ) or the cell yield (grams of cells ϫ gram substrate Ϫ1 ), the product of which is the specific growth rate (per hour). For microbes growing on cellulosic biomass, however, the specific substrate consumption rate is highly constrained, leading to strong selective pressure for cellulolytic bacteria to maximize the cell yield by increasing glycolytic ATP yield. Conversely, the use of highly thermodynamically favorable pathways with lower ATP yield, such as the ED pathway in Z. mobilis, likely represents an adaptation to glucose-rich environments (27,31,78).
In conclusion, our analysis revealed that the glycolytic pathway in C. thermocellum is remarkably close to thermodynamic equilibrium compared to the glycolytic pathways of T. saccharolyticum and E. coli. The primary contributor to the large reversibility of glycolysis in C. thermocellum is the Pfk reaction, but AdhE in the ethanol fermentation pathway is likely also an important contributor. Our findings help explain the low ethanol titer in C. thermocellum and suggest engineering strategies that can be used to increase its ethanol productivity and glycolytic rate. calculated concentrations in E. coli, were used to measure metabolite concentrations in C. thermocellum and T. saccharolyticum. Intracellular metabolite quantitation was performed in biological triplicate.
Aniline derivatization. To enhance chromatographic separation and detection of some metabolites, we performed an aniline derivatization protocol for some samples used for metabolite quantitation (46). After the extraction of metabolites as described above, 100 l extract was dried under N 2 and then reconstituted in water. Ten microliters N-(3-dimethylaminopropyl)-N=-ethylcarbodiimide hydrochloride (EDC) solution and 10 l 6 M aniline solution was added. Samples were vortexed for 2 h, and then 5 l triethylamine was added to stop the reaction. Samples were centrifuged to remove debris and then subjected to LC-MS analysis as described above.
Positional labeling in pyruvate and acetyl group labeling. Positional labeling in pyruvate was determined by comparing the labeling patterns in ornithine (Orn) with those in acetylornithine (AcOrn). These compounds differ by a single acetyl group, received from acetyl-CoA, which is itself produced by decarboxylation of pyruvate. Thus, the acetyl group always contains only the 2nd and 3rd carbons of pyruvate, and the gain of labeled carbons by Orn during acetylation to AcOrn can be used as a readout of the position of labeled carbons in pyruvate. Acetyl group labeling was also measured by comparing the labeling patterns of glucosamine-6-phosphate to those in N-acetyl-glucosamine-6-phosphate. CO 2 labeling. During 13 C labeling, unlabeled CO 2 may come from two sources: (i) glucose dissimilation and (ii) carryover from inoculation. For glucose dissimulation, in cultures fed an equimolar mixture of [U 13 C 6 ]glucose and unlabeled glucose, ϳ50% of CO 2 released during glucose fermentation to ethanol or acetate is expected to be unlabeled; the decarboxylation of pyruvate to form acetyl-CoA likely releases most of the cell-produced CO 2 . In cultures fed [1-13 C 1 ]-or [6-13 C 1 ]glucose, almost all CO 2 released by the culture would be unlabeled. The 13 C carbon under these conditions becomes the methyl carbon of pyruvate (Fig. 2), leaving the carboxyl carbon released by decarboxylation unlabeled. For carryover from inoculation, cultures were inoculated from a preculture containing unlabeled carbon sources, allowing unlabeled CO 2 to be produced in the preculture and transferred to the working culture by syringe during inoculation.
To determine the labeling of intracellular CO 2 , we compared the labeling patterns of Orn with those of citrulline (Citr), which differs from Orn by incorporation of a single CO 2 molecule. Thus, the difference between the labeling patterns of the two metabolites is due to the labeling of CO 2 . We used CO 2 labeling derived this way solely to assist in our qualitative analysis of metabolism; CO 2 was not assigned an MID during MFA.
THF intermediate and mock purine labeling. THF intermediate labeling was measured by comparing the labeling patterns of threonine (Thr) to the labeling pattern of methionine (Met). Met is formed from threonine by the addition of a single carbon from 5-methyltetrahydrofolate (MTHF), derived from the same pool as FTHF. Thus, the difference between Thr and Met labeling allowed us to determine FTHF labeling. Labeling of "mock purine" was produced by combining the labeling patterns of all precursors to purine formation, ribose-5-phosphate (R5P), glycine (gly), 2 THF intermediates, and CO 2 , into a single hypothetical MID for purines. This simulated MID showed agreement with measured labeling of purines ATP and GTP (Fig. S3C).
Metabolic network model construction. The T. saccharolyticum metabolic model was constructed from the KEGG genome-scale metabolic model, manually curated (10,12,22,23) (Table S5A). The model contains the major central metabolic reactions from the EMP, nonoxidative pentose phosphate, and TCA pathways, along with fermentative, amino acid biosynthetic, and biomass-forming reactions. We constructed models containing alternative versions of pathways (i.e., an Entner-Doudoroff glycolytic pathway in addition to or instead of EMP glycolysis, a complete oxidative TCA cycle, or an oxidative PPP) to confirm that these pathways lacked activity in T. saccharolyticum. Our largest additions to the final genome-scale model were the inclusion of a complete set of reactions producing serine from 3-phosphoglycerate, gap filling of the citramalate pathway for isoleucine biosynthesis, and addition of an SBP bypass for the PPP, all of which were supported by qualitative analysis of steady-state labeling experiments and better model fits.
Metabolic flux analysis. MFA was performed using the INCA software suite (63). INCA is implemented in Matlab and simulates isotopic distributions according to the elementary metabolite unit (EMU) framework (64). We estimated intracellular fluxes by solving a nonlinear least-squares regression problem that minimizes the variance-weighted sum of squared residuals (SSR) between simulated and measured isotopic distributions of intracellular and extracellular metabolites. We combined all tracer data sets, together with uptake, excretion, and growth rates, for each organism into a single, statistically acceptable flux map using the COMPLETE-MFA technique (85). Labeling data from 13 C and 2 H tracer experiments were entered into INCA without correction for naturally abundant heavy isotopes (Table S6). We restarted flux estimation 25 times using random initial parameters to ensure a global SSR minimum had been reached. Reversible reactions were modeled as a forward and backward reaction. Net fluxes (J net ) equals forward flux (J ϩ ) minus backward flux (J Ϫ ), and exchange flux (J exch ) equals min(J ϩ , J Ϫ ). Using the optimal solution, we calculated 95% confidence intervals for all estimated fluxes by performing a parameter continuation, which varies each flux to determine how sensitive the optimal SSR is to that flux. Upper and lower bounds are assigned to each flux by finding how far they can be varied before the SSR is perturbed past a critical point, corresponding to a chi-square distribution with a single degree of freedom.
Some metabolites, such as pyruvate, malate, and fumarate, were present in both the cytosol and fermentation broth but did not appear to be taken up and used by the cell, just excreted over time. To account for the extracellular fractions of these metabolites, we modeled them in such a way as to allow dilution with their naturally labeled equivalents without allowing other reactions to access the naturally labeled fractions of these metabolites. Dilution is also present due to the incorporation of unlabeled atmospheric CO 2 , which was modeled by allowing intracellular CO 2 to exchange with a pool of unlabeled CO 2. We used a previously published biomass equation (86) and ethanol excretion rate (26), together with our measured uptake and excretion fluxes for lactate, acetate, glucose, and formate in C. thermocellum. The remaining glucose consumption was assigned to pyruvate, valine, asparagine, and alanine proportional to previous reported amounts (87).
Measured uptake and excretion fluxes for biomass, lactate, ethanol, acetate, and glucose in T. saccharolyticum were assigned based on previously reported end product titers (10,39). Our analysis was unable to fully resolve the contributions of Ppdk versus the malate shunt regarding PEP-to-pyruvate flux. We therefore constructed two versions of the C. thermocellum model, whose only difference was the inclusion or omission of malic enzyme, and compared glycolytic flux ratios (J ϩ /J Ϫ ) and the derived free energies across them. We found a Ͻ1.0% cumulative difference in free energy of glycolytic reactions between the two models, with the single most affected reaction having only a 5.5% difference in free energy. Thus, the omission of malic enzyme had a negligible effect on ⌬G measurements, and this model was used for all flux and free energy calculations presented in the text and figures.
Goodness-of-fit analysis. A 2 test was used to determine whether the estimated fluxes adequately describe the measured data. The optimized SSR of a correct model and data set is a variable with a 2 distribution with degrees of freedom equal to the number of fitted measurements (n) minus the number of estimated independent parameters (p). Fitted measurements are fitted external fluxes, namely, uptake, excretion, growth rates, and all non-zero MIDs. Estimated parameters are all free fluxes, including net, exchange, and dilution fluxes. We required that our models pass the 2 test with a critical threshold of 0.05 (95% confidence), meaning the optimized SSR fell between ␣⁄2 2 ͑n Ϫ p͒ and 1Ϫ␣⁄2 2 ͑n Ϫ p͒. For C. thermocellum, the acceptable SSR range was 422.1 to 543.7, and the SSR after convergence was 499.2. For T. saccharolyticum, the acceptable SSR range was 434.3 to 557.5, and the SSR after convergence was 507.3.
Estimation of Gibbs free energies. Reaction free energies were calculated from MFA-derived flux measurements using the relation ⌬GϭϪRT ln(J ϩ /J Ϫ ). For two reactions (Pfk and Pyk) in the glycolytic pathway of T. saccharolyticum, the 95% confidence interval lower bound for the reverse flux was 0, implying the reaction free energy is unbounded in the negative direction (i.e., the free energy is determined to be less than a discrete value with no lower bound). For these reactions, we supplemented our flux-based ⌬G values with those calculated from metabolite concentrations using the relation ⌬G ϭ ⌬G°= ϩ RT lnQ, where Q is the ratio of the concentrations of products to reactants, R is the gas constant, and T is temperature in kelvin. Reaction standard free energies (⌬G°=) were retrieved from the online tool eQuilibrator, using a pH of 7 and ionic strength of 0.1 as settings (50,88,89). Metabolite concentrations provided a tighter bound on reaction free energies. The published reaction free energies for anaerobically grown E. coli (29) combine Pgi and Pfk as a single lumped reaction (Pgi-Pfk), so we supplemented these values with the free energy of Pgi and Pfk calculated from our measured intracellular metabolite concentrations for display (Fig. S4).

SUPPLEMENTAL MATERIAL
Supplemental material is available online only.