Optimizing Strategies for Bio-Based Ethanol Production Using Genome-Scale Metabolic Modeling of the Hyperthermophilic Archaeon, Pyrococcus furiosus

ABSTRACT

A genome-scale metabolic model, encompassing a total of 623 genes, 727 reactions, and 865 metabolites, was developed for Pyrococcus furiosus, an archaeon that grows optimally at 100°C by carbohydrate and peptide fermentation. The model uses subsystem-based genome annotation, along with extensive manual curation of 237 gene-reaction associations including those involved in central carbon metabolism, amino acid metabolism, and energy metabolism. The redox and energy balance of P. furiosus was investigated through random sampling of flux distributions in the model during growth on disaccharides. The core energy balance of the model was shown to depend on high acetate production and the coupling of a sodium-dependent ATP synthase and membrane-bound hydrogenase, which generates a sodium gradient in a ferredoxin-dependent manner, aligning with existing understanding of P. furiosus metabolism. The model was utilized to inform genetic engineering designs that favor the production of ethanol over acetate by implementing an NADPH and CO-dependent energy economy. The P. furiosus model is a powerful tool for understanding the relationship between generation of end products and redox/energy balance at a systems-level that will aid in the design of optimal engineering strategies for production of bio-based chemicals and fuels.

IMPORTANCE The bio-based production of organic chemicals provides a sustainable alternative to fossil-based production in the face of today’s climate challenges. In this work, we present a genome-scale metabolic reconstruction of Pyrococcus furiosus, a well-established platform organism that has been engineered to produce a variety of chemicals and fuels. The metabolic model was used to design optimal engineering strategies to produce ethanol. The redox and energy balance of P. furiosus was examined in detail, which provided useful insights that will guide future engineering designs.

INTRODUCTION

Climate change and global socio-political instability have resulted in increased cost of fuels, incentivizing the development of alternative, renewable sources of energy. One solution lies in the ability of microorganisms to convert renewable plant biomass into valuable organic chemicals. Compared to conventional methods of chemical manufacturing from fossil fuels, biological routes offer several advantages such as milder reaction conditions, fewer unit operations, and higher selectivity (1, 2). Nevertheless, new strategies are required to achieve economically viable yields of key products. The use of extreme thermophilic microorganisms (≥ 70°C) as metabolic engineering platforms is promising due to several theoretical benefits, including reduced risk of contamination, reduced cooling costs, continuous distillation of volatile products, and improved solubility of carbohydrates (3–5).

Pyrococcus furiosus is a hyperthermophilic marine archaeon (T_opt = 100°C) that is being developed as a metabolic engineering platform. It is an obligate heterotroph which grows anaerobically by fermenting starch-derived sugars into acetate, CO₂, and H₂ (6). One key development in the use of P. furiosus as an engineering platform was the emergence of a naturally competent variant, COM1, which enables easy and rapid recombination and construction of mutant strains (7, 8). Since then, P. furiosus has been engineered with a variety of synthetic pathways allowing the production of ethanol (9), lactate (10), butanol (11), and 3-hydroxypropionate (12, 13). In addition, it was engineered to use CO as an energy source by heterologous expression of a carbon monoxide dehydrogenase (CODH) from Thermococcus onnurineus (14). The ease of manipulating P. furiosus along with its rapid doubling time (37 min: [6]) and ability to grow at extreme temperatures makes P. furiosus an excellent platform for metabolic engineering.

P. furiosus possesses several unique features in its central carbon and redox metabolism that dictate its ability to produce pyruvate-derived products. Glycolysis in P. furiosus follows a modified Embden-Meyerhof (EM) pathway with only a single ATP-generating step, pyruvate kinase (PYK), which converts phosphoenolpyruvate (PEP) and ADP to pyruvate and ATP. In the canonical EM pathway, ATP is also generated during the conversion of glyceraldehyde-3-phosphate (GAP) to 3-phosphoglycerate (3PG), which is catalyzed in 2 steps by an NAD⁺-dependent GAP dehydrogenase (GAPDH) and an ATP-generating phosphoglycerate kinase (PGK). However, this pathway is active only in the gluconeogenic direction in P. furiosus (15, 16) and instead the conversion of GAP to 3PG is catalyzed in a single step by GAP ferredoxin oxidoreductase (GAPOR) using ferredoxin as an electron carrier (17). Alternatively, GAP can be converted to 3PG by a non-phosphorylating GAP dehydrogenase (GAPN), which uses NADP⁺ as the electron acceptor (18). However, GAPN is nonfunctional in the genetically-tractable P. furiosus COM1 strain (19) due to a 300 bp deletion in the gene that encodes it in the wild-type strain, indicating that GAPOR alone is capable of sustaining glycolysis. Another key feature of glycolysis in P. furiosus is that the phosphorylation reactions of glucose and fructose-6-phosphate are catalyzed using ADP-dependent enzymes (20, 21). Overall, the metabolism of glucose to pyruvate in P. furiosus consumes 4 ADP and generates 2 AMP and 2 ATP, and the ADP can be replenished by adenylate kinase, which converts 1 ATP and 1 AMP into 2 ADP, resulting in a stoichiometrically balanced cycle. Instead of generating ATP, the central carbon metabolism of P. furiosus results in the net production of reduced ferredoxin by GAPOR and pyruvate ferredoxin oxidoreductase (POR), which oxidizes pyruvate to acetyl-CoA (22).

Ferredoxin plays a key role in driving the synthesis of products and the overall energy economy of P. furiosus. Aldehyde ferredoxin oxidoreductase (AOR) reversibly catalyzes the oxidation of aldehydes to corresponding carboxylic acids, using ferredoxin as an electron carrier (23). With the heterologous expression of a bacterial alcohol dehydrogenase (Adh), AOR was shown to significantly contribute to the production of ethanol at 70°C (the optimum temperature for the AdhA) in P. furiosus by converting acetate to acetaldehyde, the precursor to ethanol (9). A similar result was found by upregulating the expression of a gene encoding the native NADPH-dependent Adh enzyme (AdhF) at 75°C, but not at 95°C ([24]; companion paper). At the higher temperature, acetaldehyde was instead supplied by the decarboxylative side reaction of POR, which converts acetyl-CoA directly to acetaldehyde without ferredoxin (25). This temperature-driven shift in pathway usage has implications for the overall energy balance of the cell, since ATP is generated when acetaldehyde is produced from acetate via AOR, but not when it is produced from acetyl-CoA via the POR side reaction (Fig. 1).

FIG 1.

Overview of the central carbon and energy metabolism of P. furiosus. Enzymes in yellow boxes are nonfunctional in P. furiosus COM1, and dotted arrows are reactions that are missing in the COM1 strain due to loss of enzymes through gene disruptions. The CODH complex (black, dotted) is not present in P. furiosus but is a potential engineering design. Electron carriers are color coded to highlight the interconversions between oxidized and reduced forms: red–oxidized/reduced ferredoxin (Fd_ox/Fd_red), blue–NADP⁺/NADPH, and green–NAD⁺/NADH. Created with BioRender.com.

Ferredoxin is also linked to the production of H₂ via a membrane-bound hydrogenase (MBH), which contributes to the generation of a sodium gradient across the membrane (26). MBH is coupled to the production of ATP via ATP synthase (27). Given that P. furiosus uses a sodium-dependent ATP synthase (28), it is likely that the H⁺/Na⁺ antiporter (Mrp) module of the MBH complex allows it to exchange the proton gradient for a sodium gradient which can be used for ATP production (29, 30). In a closed system, the H₂ produced by MBH can be used to replenish NADPH and NADH for biosynthesis using 2 soluble hydrogenases, SHI and SHII, respectively (31, 32). The NADH-dependent ferredoxin NADP⁺ oxidoreductase I (NfnI) couples NADP⁺ reduction to the oxidation of NADH and reduced ferredoxin simultaneously using an electron bifurcation mechanism (33, 34). Taken together, an understanding of the complex redox machinery of P. furiosus is required to design optimal metabolic engineering strategies to produce target industrial chemicals.

Genome-scale metabolic models (GEMs) are frequently used to investigate metabolism from a systems level, aiding in the design of optimal strains and growth conditions for bio-based chemical productions. Prior to this study, there were only 3 existing genome-scale models of hyperthermophilic archaea: Thermococcus paralvinellae ES1 (35) (T_opt = 82°C [36]), Sulfolobus solfataricus P2 (37) (T_opt = 80°C [38]), and Methanocaldococcus jannaschii DSM 2661 (39) (T_opt = 85°C [40]). Here, we developed a genome-scale metabolic model of P. furiousus, GEM-iPfu, the first model of an organism that can grow at 100°C. The model was used to characterize the overall redox and energy balance of P. furiosus growing on disaccharides. This knowledge was applied to design optimal engineering strategies that promote high ethanol yields while optimizing growth.

RESULTS

Overview of the genome-scale metabolic model GEM-iPfu.

The metabolic model of P. furiosus contains 623 genes, 865 metabolites, and 727 total metabolic reactions (Data set S1). The model achieved a high consistency score of 99.32% based on the Memote consistency check, which considers the stoichiometric consistency, mass and charge balance, metabolite connectivity, and presence of reactions with unbounded fluxes (41). Compared to other models of hyperthermophilic archaea, GEM-iPfu contains the highest percentage of reactions with gene associations (88.6%) and has the highest percentage of genes from its genome represented in the model (31.5%) (Table 1).

TABLE 1.

Comparison of the P. furiosus model with other thermophiles

Organism	Pyrococcus furiosus DSM 3638	Thermococcus paralvinellae ES1	Sulfolobus solfataricus P2	Methanocaldococcus jannaschii DSM 2661
Type	Hyperthermophile	Hyperthermophile	Hyperthermophile	Hyperthermophile
T opt (°C)	100	82	80	85
Genome
Size (Mb)	1.89	1.96	2.99	1.66
CDS No.	1,983	2,074	2,834	1,711
Model ID	iGEM_Pfu	MODEL1707280000	iTU515	iTS436
yr	This study	2018	2012	2004
Metabolites	865	825	705	510
Total reactions (exclude exchange)	730	725	718	609
Genes in Model
No.	624	341	515	436
Percent of total genes	31.47%	16.44%	18.17%	25.48%
Gene-associated Reactions
No.	647	475	606	297
Percentage of total reactions	88.63%	65.52%	84.40%	48.77%

Metabolic reconstruction of the P. furiosus model involved the use of an automated annotation pipeline, along with extensive manual curation combining evidence from putative gene orthology and biochemical/physiological experiments in other members of the Thermococcales (e.g., Thermococcus kodakarensis), other Euryarchaeota (e.g., Methanocaldococcus, Methanosarcina), and thermophilic bacteria (Thermatoga maritima, Thermus thermophilus). The manual curation specifically addressed 3 limitations of the automatic pipeline: (i) functions that are present in the metabolism but not in public databases like KEGG, e.g., Nfn1 and MBH, (ii) reactions where the automatic annotation could not determine the exact cofactors requiring further experimental validation, e.g., GAPDH, and (iii) confirmation of annotations supported by biochemical experiments e.g., ACS, AOR and POR. The full details of the curated enzymes and the associated genes and reactions are provided in Data set S2. The manual curation led to the incorporation of 237 genes (38% of the model) associated with 261 metabolic reactions, including 87 new reactions that are not represented in the current KEGG database. The curated pathways included glycolysis/gluconeogenesis, the pentose biphosphate pathway, energy metabolism, purine metabolism, pyruvate metabolism, amino acid metabolism, polyamine biosynthesis, nicotinate and nicotinamide metabolism, CoA biosynthesis, organic sulfur metabolism, membrane lipid biosynthesis, chitin catabolic pathways, folate biosynthesis, compatible solutes production, and the citric acid cycle.

GEM-iPfu incorporates the representation of 2 P. furiosus strains, wild type (DSM 3638) and COM1 (Table 2). COM1 is the parent strain for genetic engineering owing to its amenability to transformation and genetic manipulations. It includes the deletion of pyrF, resulting in uracil auxotrophy (7). Comparison of the COM1 and DSM 3638 genomes found 55 genes with major disruptions in their coding sequences (< 90% identity to the original protein) that are likely to be nonfunctional in COM1 (19). Of these genes, 11 had associations with reactions in GEM-iPfu. Deletion of these 11 genes resulted in the removal of 15 reactions including the non-phosphorylating GAPDH (GAPN) and a beta-glucosidase (CelB) (Table S1). These changes to the model did not have significant effects on the overall growth on cellobiose, maltose, or pyruvate (Table S2), showing a general consistency of the model with experimentally measured physiology (7). When mutant strains are developed, pyrF is reintroduced, allowing for prototrophic selection. In this study, we name the strain of COM1 complemented with pyrF as COM1c and use it as the parent strain for the simulation of engineering designs (Table 2).

TABLE 2.

Summary of strains modeled or referenced in this study

Index	Genotype	Parent	Strain ID	Reference	Simulations
WT	Wild type		DSM 3638	6	Model validation, growth simulations
COM1	ΔpyrF		MW002	7	-
COM1c	ΔpyrF::pyrF	COM1	MW003	62	Flux sampling, end product yield ratios
OE-AdhF	ΔpyrF PslpadhF::Pgdh-pyrF	COM1	MW631	24; Companion paper	Flux sampling, end product yield ratios, testing of engineering designs
OE-AdhF-CO	ΔpyrF Pmbh-TonCODH Pslp-adhF::PgdhpyrF	OE-AdhF	MW650	24; Companion paper	Flux sampling, end product yield ratios
OE-AdhF-N	GAPN+ ΔpyrF PslpadhF::Pgdh-pyrF	OE-AdhF	This study	This study	Flux sampling, testing of engineering designs
OE-AdhF-N/CO	GAPN+ CODH+ ΔpyrF PslpadhF::Pgdh-pyrF	OE-AdhF	This study	This study	Flux sampling, testing of engineering designs

Growth simulations.

The P. furiosus model was applied to simulate growth yields of the wild-type strain on different carbon substrates, and the simulation results were validated using experimental growth data of P. furiosus DSM 3638 with 3 different carbon sources, maltose, cellobiose, and pyruvate, at different concentrations (42). Metabolic simulations were configured in the wildtype (WT) representation of the model by optimizing the biomass yield using flux balance analysis (FBA) (see Materials and Methods). Exchange constraints were defined based on substrate concentration of the experimental growth media, and the exchange constraints of major products (acetate, CO₂, ethanol, and H₂) were constrained within a range based on the measurements from each substrate concentration. Additionally, growth associated ATP maintenance (GAM) and non-growth associated ATP maintenance (NGAM) costs were estimated by examining a linear correlation between growth and ATP hydrolysis under different carbon and product constraints. This resulted in a GAM value of 74.73 mmol ATP/g CDW, and an NGAM value of 2.34 mmol ATP (Fig. S1). The model accurately simulated the biomass production of P. furiosus under all the examined substrates, with the linear fitting between modeled biomass yield and experimental growth measurement achieving a coefficient of near 1 (R² = 0.95) (Fig. 2).

FIG 2.

Comparison of experimental and model-simulated growth of the WT strain grown on cellobiose, maltose, or pyruvate (42). Media formulation, substrate, and product concentrations were constrained in the model based on experimental values for the simulation of growth yields (Data set S3).

Additional growth simulations were performed to assess the model by providing carbohydrates or peptides as carbon sources and in the presence or absence of elemental sulfur (S⁰). If P. furiosus is grown in the presence of elemental sulfur, the MBH complex is not expressed and is instead replaced by a homologous respiratory complex known as MBS (membrane-bound, sulfane sulfur reductase), which reduces S⁰ to H₂S. Like MBH, MBS oxidizes ferredoxin and generates a sodium ion gradient; this is also used to generate ATP (43). Elemental sulfur is essential for growth on gluconeogenic substrates, such as peptides, but is not needed for growth on carbohydrates (44). One important gene that mediates metabolic responses of P. furiosus to different carbon substrates encodes the TrmBL1 transcriptional regulator. TrmBL1 acts as a global regulator of genes involved in glycolysis and gluconeogenesis, such as GAPOR, which is suppressed by TrmBL1 when gluconeogenic carbon substrates are used in P. furiosus (15). Therefore, we decided to model the suppression of GAPOR under gluconeogenic conditions by setting its reaction flux to zero when peptides were used as the sole carbon substrate. We applied the P. furiosus GEM to simulate its growth in the presence of a carbohydrate (maltose) or peptides, and with or without the presence of elemental sulfur (Fig. S2). Each simulation was performed by optimizing growth yields of P. furiosus at a designated carbon and elemental sulfur condition (Materials and Methods). We observed no growth on peptides without elemental sulfur (maximum growth yield of 0.02 g CDW/L), and growth on peptides when elemental sulfur was added (maximum growth yield of 0.34 g CDW/L), matching the experimentally observed phenotype (44).

Simulation of product yield ratios.

The model was also applied to assess the generation of end products, such as ethanol and acetate, in COM1c and genetically engineered strains of P. furiosus. Based on our experimental study, the P. furiosus genome encodes an aldehyde dehydrogenase gene (adhF, PF0608), that can be overexpressed to increase the production of ethanol ([24]; companion paper). By default, our model does not incorporate gene expression into the prediction of metabolic fluxes. To simulate the phenotypes observed in the control (COM1) and overexpressed ethanol (OE-AdhF) strains, we calibrated the model to represent adhF baseline- and high-expression states, respectively, based on experimentally measured differences in ethanol production by the 2 strains (Materials and Methods). As expected, the calibrated models demonstrated consistent prediction of ethanol:acetate ratios in the COM1 and OE-AdhF strains, respectively, when compared to experimental data (Fig. 3A).

FIG 3.

Comparison of experimental and modeled ethanol:acetate ratios. All simulations were constrained with a minimal maltose medium containing 14.6 mM (5 g/L) Maltose and 0.5 g/L yeast extract. (A) Select genotypes from Lipscomb et al. ([24]; companion paper). Overexpression of AdhF was modeled as described in Materials and Methods. (B) Modeling of the open and closed systems compared to experimental data obtained for this study.

The calibrated OE-AdhF model was applied to simulate additional engineering designs, including the OE-AdhF Δaor, where a knockout of the aor gene was introduced to the OE-AdhF background, and the OE-AdhF-CO, where a knock-in of the CODH complex was introduced to the OE-AdhF background (Table 2). The model simulations were compared with experimental results obtained from the 2 corresponding strains, OE-AdhF Δaor and CODH-OE-AdhF ([24]; companion paper), showing a good agreement between modeled and experimental data in the relative composition of ethanol and acetate end products. The OE-AdhF Δaor had lower ethanol:acetate ratios compared to the OE-AdhF parent strain, whereas the OE-AdhF-CO had higher ethanol:acetate ratios (Fig. 3A). The quantitative estimation of the OE-AdhF Δaor model, however, showed slightly lower ethanol:acetate ratios than the experimental observation. Further examination suggested a heavy reliance of the model on acetate production for energy metabolism and ATP generation, and the Δaor mutation blocks the path for converting acetate to acetaldehyde, a precursor for ethanol production. This may indicate an overestimation of the energy budget and a potential misdirection of the carbon fluxes by the OE-AdhF Δaor model. The quantitative estimations by the OE-AdhF-CO model had slightly higher ethanol:acetate ratios and higher uncertainty (indicated by large error bars) than the experimental data. However, the model predictions are generally on par with experimental results. Overall, metabolic modeling of the engineered strains, OE-AdhF Δaor and OE-AdhF-CO, demonstrated qualitative agreement with experimental data while slight variations in the quantitative precision. Our model, along with the experimental data, suggests OE-AdhF-CO as a good candidate parent strain for enhancing ethanol productions.

Simulation of open versus closed systems.

The metabolism and product generation of P. furiosus can be influenced by the partial pressure of H₂, which is produced by the MBH complex using reduced ferredoxin (Fd_red) generated through the fermentation process. P. furiosus can potentially recycle the H₂ produced from MBH using the soluble hydrogenases SHI and SHII, producing NADPH and NADH, respectively. To assess the potential effect of H₂ in mediating redox metabolism in P. furiosus, we created two distinct model settings to account for H₂ diffusion in the two extremes: (i) a closed model, where H₂ produced by MBH is not diffused to the environment and is available to SHI and/or SHII for producing reducing equivalents; and (ii) an open model, where H₂ produced by MBH escapes into the environment and is not available for producing reducing equivalents (NAD[P]H) by SHI or SHII (Table 3 and Materials and Methods). We expect the closed model would represent systems where gas exchange is inhibited, such as batch cultures in sealed bottles, and the open model would represent systems with constant mixing and gas exchange, such as flushed and stirred fermenters. We acknowledge that the two models represent extreme conditions that may under- or overestimate the physiology of P. furiosus, and the actual metabolic state of the organism is invariably in between the open and closed models. However, modeling these extreme cases will provide potential insights into the metabolic responses of P. furiosus under different cultivation environments.

TABLE 3.

Description of closed and open model configurations

Model configurations	Definition	Model implementation
Closed	H2 accumulates in headspace, e.g., batch cultures with sealed bottles	No change
Open	H2 is flushed from headspace, e.g., stirred fermenter	SHI and SHII are deactivated

Our models of the open and closed systems were validated experimentally using the P. furiosus OE-AdhF (MW631) strain. Cultures were set up in bottles with the headspace continually flushed with argon during growth, representing an open system, or in closed bottles with no flushing, representing a closed system (see Materials and Methods). A general consistency was observed between either open or closed models and their corresponding experimental measurements of the products’ ethanol:acetate ratios (Fig. 3B). As expected, the open model slightly underestimated the ethanol:acetate ratios compared to the experimental data. This is because of the stringent settings that block SHI/SHII reactions in the open model, whereas in the experiment, we expect some SHI/SHII activity. In parallel, the closed model slightly overestimated the ethanol:acetate ratios compared to the experimental data. This is because the closed model permits all H₂ produced by MBH to be accessible for SHI/SHII, whereas in the experiment, the accessibility of H₂ might be limited due to the diffusion of H₂ to the gas phase in the headspace of the bottle. It is worth noting that a near 2-fold increase in ethanol:acetate ratio was observed based on experimental measurements in the closed system compared to the open system. This general trend of higher ethanol production in closed systems was replicated by the P. furiosus model. Taken together, the experimental validation supports our design of the open and closed models and suggests that the two distinct model settings can be used to simulate product compositions by P. furiosus under different cultivation environments.

Identifying redox limitations for ethanol production.

The P. furiosus model was applied to explore the relationship between product generation and the general redox and energy balance. Here, we focused our analysis on the OE-AdhF strain as it will be used as the base strain for metabolic engineering. We first summarized the contribution of each reaction on the production and consumption of key energy and redox compounds, using 2500 random metabolic flux simulations that satisfy constraints in the OE-AdhF model (Materials and Methods). For each random sample, the proportional contribution of different reactions to the total production of ATP, Fd_red and NADPH were calculated. Proportions for each reaction were averaged across all random simulations and visualized in Fig. 4.

FIG 4.

Average contribution of key reactions to the compound balance of ATP/GTP, reduced ferredoxin (Fd_red), and NADPH based on 2500 random simulations of the P. furiosus OE-AdhF model, using a minimal maltose medium containing 14.6 mM (5 g/L) maltose and 0.5 g/L yeast extract (Materials and Methods). The proportion of reactions contributing to compound productions in (A) the closed system and (B) the open system is shown. Reactions with proportions less than 0.1 were combined into “Other”. The relevant reactions/enzymes are as follows: ATPS, ATP synthase; ACS_acetate, Acetyl-CoA synthetase (acetate-forming); PYK, pyruvate kinase; POR, pyruvate ferredoxin oxidoreductase; GAPOR, glyceraldehyde-3-phosphate ferredoxin oxidoreductase; R00248, L-Glutamate:NADP⁺ oxidoreductase; NfnI, NADH-dependent ferredoxin NADP⁺ oxidoreductase I; SHI, NADP⁺-dependent soluble hydrogenase.

We observed that P. furiosus was able to use ATP and GTP somewhat interchangeably, since it possesses both ATP and GTP-dependent acetyl-CoA synthetases (ACS) (45), along with a GTP-dependent adenylate kinase. Therefore, we analyzed the production or consumption of ATP and GTP jointly within each random simulation. This analysis identified three main sources of ATP/GTP in both open and closed systems: pyruvate kinase (PYK), acetyl-CoA synthetase (ACS), and ATP synthase (ATPS) (Fig. 4). Note that the ACS function highlighted in this figure refers to the predicted activity of the ACS isoenzymes producing acetate. The activities of ACS on other substrates in the amino acid metabolism were also accounted for in the model, but they contributed less than 10% of the total ATP production. As expected, GAPOR and POR are the two major sources of Fd_red in the P. furiosus metabolism. Other potential sources of Fd_red included the 2-keto acid oxidoreductases Fig. 4, ‘Other’, which play a role in amino acid catabolism and were active due to the consumption of amino acids from yeast extract in the medium. Finally, the production of NADPH had distinct compositions in open versus closed systems. In the closed system, SHI and NfnI are the main sources of NADPH (Fig. 4A). In the open system, NfnI is the main source of NADPH due to the inhibition of SHI, and an additional small contribution was observed to the production of NADPH by L-Glutamate:NADP⁺ oxidoreductase (R00248) of the amino acids metabolism (Fig. 4B).

In a closed system, we hypothesize that ethanol production is not limited by NADPH given that excess H₂ was released even in the highest ethanol yielding random simulations. However, this may not be true in an open system, where H₂ is expected to rapidly escape the cellular environment and not be available for NADPH production by SHI. Our simulation of the open system suggested a decrease of ethanol production to less than half of what was observed in the closed system (Fig. 3), indicating that ethanol production may be limited by NADPH availability in open systems. We also hypothesize that ethanol production is affected by the activity of the acetate-producing pathway (POR-ACS), which was shown to be a key source of ATP/GTP and Fd_red in our simulations. Although ethanol can be generated from acetate using AOR and the native AdhF that is overexpressed in the OE-AdhF strain, this strategy might be disrupted under the physiological temperature of P. furiosus due to an alternative pyruvate decarboxylase reaction catalyzed by POR ([24, 25]; companion paper).

To investigate potential redox and energy limitations and their effects on product fluxes, we introduced a set of artificial reactions to the OE-AdhF model that could freely add energy or convert reducing equivalents in the cell (Table 4). These reactions served to eliminate any redox or energy constraints that may affect the production of ethanol and acetate. In both the open and closed systems, adding free ATP or free Fd_red resulted in large, significant reductions to the production of acetate (Fig. 5 and Table S3), suggesting that acetate production is driven by the need to generate energy. In the closed system, ethanol production saw small but significant increases of 0.69 to 1.64 mM from its median value (7.8 mM) across the free ATP, Fd_red, and NADPH conditions (Table S3). In the open system, ethanol production was much lower on average in all cases and saw a large increase from 2.5 to 5.3 mM (P < 1e⁻¹⁰, d = 1.75) when free NADPH was added (Table S3). Ethanol production in the open system was not significantly different when free ATP was added, and decreased by a marginal amount when free reduced ferredoxin was added. We conclude that the ATP need explains the high acetate production by P. furiosus but adding an external source of ATP alone does not induce higher ethanol production (Fig. 5). Providing a source of NADPH seems to facilitate higher ethanol production in the open system. However, it does not seem to be sufficient for increasing ethanol production in the closed system (Fig. 5). These observations suggest that ethanol production is not solely limited by the availability of NADPH or ATP in P. furiosus, and that the metabolism must be altered to incentivize higher ethanol yields. Given this understanding, we decided to search for engineering designs that (i) incorporate ethanol as an essential product for balancing redox compounds from the central metabolism, and (ii) provide alternative methods for ATP production to disincentivize acetate production.

TABLE 4.

Artificial reactions added to evaluate redox and energy limitations in the model

Reaction namea	Equation
Free ATP	ADP[c]b + Phosphate[c] => ATP[c] + H2O[c]
Free FdRed	Oxidized ferredoxin[c] => Reduced ferredoxin[c] + H+[c]
Free NADPH	NADP+[c] => NADPH[c] + H+[c]

^aAdded reactions were freely allowed to take flux in the range (0, 1000). The redox reactions are not formula balanced as they are intended to emulate the free addition of target compounds.

^b[c] indicates compounds in the cytoplasm.

FIG 5.

Comparison of flux distributions for (A) acetate and (B) ethanol production (mM) in the P. furiosus OE-AdhF model growing on a minimal maltose medium containing 14.6 mM (5 g/L) maltose and 0.5 g/L yeast extract (Materials and Methods), with the inclusion of artificial reactions providing free redox/energy equivalents (Table 4). Simulations of the closed system are shown in blue, and those of the open system are shown in red (Table 3). Asterisks indicate conditions where the production is significantly different from the baseline unaltered model. Statistical significance was assessed using P-values from the Mann-Whitney U test and the Cohen’s d effect sizes, using the following cutoffs: small (*): P < 0.05 and d < 0.5, medium (**): P < 0.05 and d < 0.8, large (***): P < 0.05 and d > 0.8. Non-significant differences (P > 0.05) were denoted as ns.

New engineering designs to enhance ethanol production.

To obtain higher levels of NADPH production through the central carbon metabolism of P. furiosus, we added GAPN into our model of the OE-AdhF strain (Table 2). GAPN serves as an alternative pathway to GAPOR in glycolysis. With its NADPH producing capability, GAPN can be paired with the NADPH-dependent AdhF in the OE-AdhF strain to facilitate ethanol formation. We also evaluated the addition of CODH which enables the use of CO as an additional energy source (14). Ten different gene deletion strategies were analyzed in terms of their ethanol and acetate production using the random simulation approach (Materials and Methods) on each of the three different genetic backgrounds: OE-AdhF, OE-AdhF-N, which includes an insertion of GAPN to OE-AdhF, and OE-AdhF-N/CO, which includes an insertion of CODH to OE-AdhF-N (Table 2). A total of 2,500 random simulations were computed for each knockout strategy using exchange conditions set based on the minimal maltose medium ([24]; companion paper) (Data set S4). Ethanol and acetate production estimates from each random simulation were scaled per disaccharide (maltose), and then multiplied by the biomass to screen against strains with little or no growth. Strains were analyzed in both the closed and open systems (Fig. 6 and Fig. S3, respectively) and yielded comparable results. Almost all of the deletion strains tested in our screening were non-viable in the OE-AdhF background except for the OE-AdhF Δaor strain, which showed higher selectivity for acetate than the OE-AdhF strain (Fig. 6A and Fig. S3) and is consistent with our experimental data (Fig. 3A). In the OE-AdhF-N background, 4 viable strains were predicted with relatively high ethanol production and little or no acetate production. These were the OE-AdhF-N Δmbh, Δgapor, Δgapor Δmbh, and Δgapor ΔnfnI strains (Fig. 6B). Finally, in the OE-AdhF-N/CO background, all deletion strategies became viable (Fig. 6C). In the closed system, the OE-AdhF-N/CO Δgapor Δmbh strain appeared to perform the best, with twice the median growth yield of the OE-AdhF strain, and a median ethanol yield of 29.06 mM, 3.8 times higher than the yield in the original OE-AdhF strain. In the open system, the OE-AdhF-N/CO Δacs Δaor Δmbh strain performed best, with a median ethanol yield of 14.46 mM.

FIG 6.

Distributions of ethanol and acetate production fluxes for 10 different knockout strategies each in the (A) OE-AdhF, (B) OE-AdhF-N, and (C) OE-AdhF-N/CO parents (Table 2) based on random simulations in the closed system using a minimal maltose medium containing 14.6 mM (5 g/L) maltose and 0.5 g/L yeast extract (Materials and Methods). Fluxes of ethanol and acetate were scaled to mM production per mM maltose consumed, and then multiplied by the growth yield. Strains with high, non-zero ethanol fluxes (blue) and low acetate fluxes (light gray) represent potential engineering targets.

To further examine the potential of bioproduction by top-ranking engineering designs, we calculated the range of putative ethanol yields and their correlation to biomass production using the analysis of production envelopes (Fig. 7 and Fig. S4). While the random simulation approach describes a statistical sampling of the metabolic solution space, the production envelopes analysis looks at the feasible range of metabolic reactions with the goal of identifying strains where the minimal ethanol production becomes non-zero with increasing biomass production. Strains with minimum ethanol production greater than zero at their highest biomass yields are considered favorable designs because this indicates that ethanol production is required with maximum growth. Three different engineering strategies were investigated in the OE-AdhF-N and OE-AdhF-N/CO backgrounds: (i) Δmbh, (ii) Δgapor, and (iii) Δacs Δaor. The OE-AdhF-N Δmbh and OE-AdhF-N/CO Δmbh strains had the most promising phenotypes, with high minimum ethanol production under a wide range of growth yields, indicating a positive correlation between biomass and ethanol production (Fig. 7A and Fig. S4A). The OE-AdhF-N Δgapor or OE-AdhF-N/CO Δgapor strains also demonstrated high minimum ethanol production across a wide range of biomass yields. However, the minimum ethanol production was not above zero until the biomass production reached at least half of the predicted maximum (Fig. 7B and Fig. S4B). The OE-AdhF-N/CO Δgapor Δmbh strain had an identical production envelope to the OE-AdhF-N/CO Δmbh strain, indicating that there is no benefit to the combined deletion of GAPOR and MBH. Finally, the OE-AdhF-N Δacs Δaor or OE-AdhF-N/CO Δacs Δaor was used to examine the effect of eliminating acetate productions. Note that Δacs specifically refers to the deletion of ACS isoenzymes that produce acetate. Of the 10 isoenzymes comprising 5 α-subunits and 2 β-subunits, the ACS 1α-, 4α-, and 5α-subunits were deleted as this combination would remove most acetate activity while eliminating the activity on only 3 other substrates in the amino acids’ oxidation pathways (45) (Table S4 and Data set S1). ACS activity on other substrates was retained in the Δacs models as the activity of ACS 2α- and 3α- has been experimentally demonstrated on these substrates (45). The only mutant that had a non-zero minimum ethanol production in the Δacs Δaor background was the OE-AdhF-N/CO Δacs Δaor Δmbh design. The model predicted relatively high ethanol production even under low biomass yields in OE-AdhF-N/CO Δacs Δaor Δmbh (Fig. 7C and Fig. S4C). However, it seems that the maximum growth of OE-AdhF-N/CO Δacs Δaor Δmbh was lower than in the OE-AdhF-N/CO Δmbh design, indicating that the Δacs Δaor mutant offers little benefit in metabolic engineering for ethanol production. It is worth noting that with the OE-AdhF-N background, many mutants (e.g., Δmbh, Δgapor, and Δacs Δaor) were predicted to have their maximum growth yields limited compared to their corresponding mutants in the OE-AdhF-N/CO background (Fig. 7 and Fig. S4). Therefore, OE-AdhF-N/CO could be a preferred engineering design as it optimizes ethanol production without limiting the maximum growth.

FIG 7.

Production envelopes of selected high-performing engineering designs identified in Fig. 6. Each production envelope was simulated in the closed system using corresponding engineering designs with a minimal maltose medium containing 14.6 mM (5 g/L) maltose and 0.5 g/L yeast extract (Materials and Methods). The lines of an envelope represent the range of possible ethanol fluxes with varying biomass production (x axis). Designs with the lower bounds of ethanol production positively correlated with biomass yields indicate the selection for ethanol production under optimal growth.

DISCUSSION

P. furiosus is a promising metabolic engineering platform with a robust genetic system that has been developed for the synthesis of bioproducts from plant-derived polysaccharides. This study provides a systems-level overview of P. furiosus metabolism with a genome-scale metabolic model, GEM-iPfu. The GEM-iPfu applies to 2 distinct P. furiosus strains that have been sequenced to date, wild type (DSM3638) and COM1, with the latter having 11 metabolic genes disrupted resulting in the loss of 15 metabolic reactions in the model. However, these strain-level differences caused little change in the predicted growth phenotypes between WT and COM1 (Table S2), which aligns well with experimental results (7).

The model was compared to experimental data to assess its predictions of P. furiosus growth and end product production under diverse treatments, such as different carbon substrates with or without the presence of elemental sulfur (Fig. 2 and Supplemental Fig. S2), genetic perturbations (Fig. 3A), and variations in the cultivation condition (Fig. 3B). Model calibrations included the incorporation of growth associated maintenance (GAM) and non-growth associated maintenance (NGAM) costs (Fig. S1). The estimated GAM and NGAM values of the P. furiosus model are comparable to other archaeal models, such as those of the Methanosarcina barkeri model iAF692 (46) and the Methanosarcina acetivorans model iVS941 (47), which each had GAM values of 70 mmol ATP/gDCW and NGAM values of 1.75 mmol ATP/gDCW/hr (48), indicating potential conservation of energy cost and efficiency across different archaea. To establish a base strain for bio-based ethanol optimization, an OE-AdhF calibration was introduced to the model to simulate the overexpression of AdhF (Fig. 3A), a native NADPH-dependent alcohol dehydrogenase of P. furiosus ([24]; companion paper). The OE-AdhF model was applied to simulate metabolic responses of P. furiosus under open or closed cultivation conditions, which respectively corresponds to fermenters with mixing and rapid diffusion of H₂ gas, and batch cultures with sealed bottles and limited gas exchange (Fig. 3B). Overall, the P. furiosus model demonstrated good consistency with experimental data across all the evaluated conditions.

The redox and energy balance of P. furiosus was investigated using random simulations of the OE-AdhF model to identify key limiting compounds that influence the production of ethanol and acetate. Analysis of the stoichiometric balance of the redox and energy cofactors, ATP, Fd_red, and NADPH, showed that 2 enzymes in the acetate production pathway, POR and ACS, contributed greatly to the overall production of reduced ferredoxin and ATP/GTP, respectively (Fig. 4). This could explain the organism’s preference for acetate as a main product. Additionally, when ATP or Fd_red were freely provided to the model, acetate production was significantly reduced to values similar to ethanol production (Fig. 5). This supports the notion that adding additional sources of ATP or of Fd_red would potentially alleviate the costs that led to high acetate production in the model. As expected, Fd_red production was tightly associated with the production of ATP via MBH and ATP synthase, therefore both free ATP and free Fd_red served a similar purpose of removing the model’s energy demands. Surprisingly, while the median acetate production was greatly reduced when free ATP/Fd_red were added, the median ethanol production saw only small increases in comparison. In open systems, ethanol production was limited by the availability of NADPH due to restrictions of SHI in recycling H₂. This indicates that enhancing NADPH production could be a good strategy when growing P. furiosus under conditions where H₂ is constantly flushed from the headspace. However, the limits in NADPH availability could only partially explain the low ethanol yield in closed systems (Fig. 5), indicating additional limitations related to the overall redox balance of the P. furiosus metabolism.

Our observations suggested that the ferredoxin-based energy economy of P. furiosus is naturally synergistic with acetate production, and to achieve higher ratios of ethanol:acetate requires a fundamental restructuring of redox cofactors in the central metabolism. It has been hypothesized that reducing the efficiency of MBH to recycle reduced ferredoxin could lead to increased ethanol production, since the slower supply of oxidized ferredoxin would lead POR to favor the oxidative decarboxylation side reaction (producing acetaldehyde) rather than the ferredoxin-dependent oxidoreductase reaction (producing acetyl-CoA) ([24]; companion paper). This suggests a fundamental trade-off between acetate and ethanol production in the ferredoxin-dependent energy economy of P. furiosus, where ethanol production will only be favored when the growth-critical enzymes GAPOR and MBH are less active.

The GAPN strain (OE-AdhF-N) proposed here has potential to induce demand for ethanol as a growth-dependent product by introducing NADPH cycling into the central metabolism of P. furiosus. The gene encoding NADP⁺-dependent GAPN is present in wild type P. furiosus (DSM 3638) but was disrupted by an insertion sequence and had no measurable activity in the COM1 strain (19). A similar strategy was tested previously by Straub et al. (16), where GAPOR was replaced with GAPN in P. furiosus COM1 alongside heterologous expression of an NADPH-dependent alcohol dehydrogenase (AdhA) from Caldaenerobacter subterraneus. They found that the gapn⁺ Δgapor strain did not affect the ethanol:acetate ratio compared to the parent COM1 strain, conflicting with our predictions for the OE-AdhF-N Δgapor strain, which had increased selectivity toward ethanol. These differences could be a result of temperature-dependent metabolic changes, as the AdhA activity requires growing P. furiosus at a suboptimal growth temperature (80°C). Therefore, further testing is required to evaluate whether the model’s predictions would hold in the OE-AdhF genetic background, which permits ethanol production at standard P. furiosus growth temperatures (95°C). One key benefit of the GAPN addition strategy is that by offering an alternative pathway to GAPOR in glycolysis, our model predicted that GAPN would enable the deletion of MBH for enhancing ethanol production, which was previously thought to be impossible in P. furiosus COM1 (31). We found multiple knockout strategies in the GAPN⁺ strain that achieved high selectivity toward ethanol and were still capable of growth despite lacking GAPOR or MBH (Fig. 6). In the OE-AdhF-N Δmbh strain, ATP/GTP production was supported by the POR-ACS-AOR pathway, which produces acetaldehyde, a precursor of ethanol, while cycling the oxidized and reduced forms of ferredoxin between POR and AOR (Fig. 1). While this mutant did tend to have a lower growth yield on average compared to the OE-AdhF parent strain, this could be salvaged by the addition of CODH (Fig. 7), a strategy which has been successfully employed in P. furiosus ([14, 24]; companion paper).

Weighing the complexity of each engineered strain, we propose multiple strategies based on the model simulations. The OE-AdhF-N Δmbh was 1 candidate strain, with higher minimum ethanol production coupled to increasing biomass yield (Fig. 7A). Optimization of the OE-AdhF-N Δmbh design would require pyruvate to be channeled through the POR-ACS-AOR pathway, which results in the production of acetaldehyde (via AOR) and ATP while being ferredoxin balanced (Fig. 1). However, at high temperatures (95°C), acetaldehyde is produced directly from pyruvate decarboxylation by POR and AOR favors the acetaldehyde-consuming direction ([24]; companion paper), limiting ATP production in the OE-AdhF-N Δmbh design. This problem may be solved with the OE-AdhF-N/CO Δmbh design, which introduces CODH to enable the use of CO as an additional source of ATP and reductant, alleviating potential dependency on the POR-ACS-AOR pathway. Another feasible design is the OE-AdhF-N/CO Δgapor strain, which supports a higher maximum growth due to the introduction of CODH, and the GAPOR deletion helps strengthen the coupling of GAPN with ethanol production through the cycling of NADP⁺/NADPH. Strategies involving the deletion of ACS and AOR did not show equal promise, likely due to growth limitations caused by the loss of ACS (Fig. 7C). While this growth phenotype can be alleviated with the expression of CODH, Δacs Δaor alone does not support higher minimum ethanol production with increasing biomass. The OE-AdhF-N/CO Δacs Δaor Δmbh could be a useful strategy for coupling glycolysis and ethanol production in the presence of CO. However, given the promising results from the simpler OE-AdhF-N/CO Δmbh strategy in our simulations, it is unclear whether additional deletion of Δacs Δaor would provide sufficient benefit to justify the engineering investment.

Overall, the analyses presented here demonstrate the utility of genome-scale modeling to inform engineering designs. Sampling of flux distributions throughout the model solution space revealed key redox and energetic tradeoffs that can be leveraged for the optimization of ethanol production. Multiple engineering designs have been proposed based on our model to reconfigure the redox balance of P. furiosus such that ethanol became a core product. Despite the accuracy of the P. furiosus model in providing qualitative predictions, especially in predicting the ethanol:acetate ratios, we recognize that high precision quantitative predictions of product yields may be still beyond reach with the current model. Several factors that may influence the quantitative precision of our model predictions include the consideration of thermodynamics, enzyme kinetics, and regulation, both allosteric and of gene expression. These are currently not captured by the P. furiosus model but are the subject of ongoing work to enhance the model. Future studies will validate our predicted metabolic engineering strategies, and improvements to the model will allow it to capture other bottlenecks that may limit production capabilities of P. furiosus.

MATERIALS AND METHODS

Genome-scale model reconstruction.

The genome-scale metabolic model of P. furiosus was reconstructed using the complete genome sequence of strain DSM 3638 (GenBank assembly ASM730v1). Automatic curations were performed using homology searches to the following public data sets: BioCyc (49), KEGG (50), EggNog (51), and TCDB (52). Additionally, extensive manual curation of pathways from the literature was performed, with a total of 237 gene associations verified by experimental studies in P. furiosus or other related species in the order Thermococcales (Data set S2).

The model was represented in YAML format following guidelines of the PSAMM software package (53, 54). Additional curation of the model was performed using the “masscheck,” “dupcheck,” “formulacheck,” and “chargecheck” functions of PSAMM, which check the model to eliminate duplicate reactions and ensure elemental and charge balance of reactions. The stoichiometric consistency of the model was calculated using memote version 0.13.0 (41), using the online interface at memote.io.

A biomass equation was formulated, combining DNA, RNA, protein, lipids, carbohydrates, trace minerals, and polyamines with stoichiometry reflecting the gram composition of each macromolecule in 1 g cell dry weight. The mass compositions of DNA, RNA, protein, lipids, carbohydrates, and trace minerals in the overall biomass were estimated based on the iAF692 model of M. barkeri (46), and the mass composition of polyamines was determined based on experimental measurements in P. furiosus (55). Synthesis equations were defined for each biomass component, where the stoichiometry of the synthesis equation represents the millimolar composition of individual building blocks in a gram of the biomass component. The stoichiometries of each biomass component were determined as follows: DNA composition was defined as a combination of ATP, H₂O, and nucleotides in a ratio based on the overall GC content of the genome, RNA composition was defined by the nucleotide frequencies in the coding sequences of the genome, protein composition was estimated by the amino acid frequencies in the proteome, lipids were estimated based on experimental determination of the composition of core lipid composition in P. furiosus (56), carbohydrates and trace metabolites were estimated based on the composition in the iAF692 model (46), and polyamines based on experimental determination of the polyamine composition in P. furiosus (55) (Data set S1).

Simulation of growth phenotypes.

Model validation was performed using experimental product and growth measurements collected with P. furiosus DSM 3638 growing on 3 different carbon sources: cellobiose, maltose, and pyruvate, each grown at 4 different concentrations (42). For each condition, substrate concentrations were converted to millimolar units and growth yields were converted to g cell dry weight, to match the units that the biomass equation is calibrated for in the model (Data set S3). A simulated media file was created to match the experimental media formula, with either cellobiose, maltose, or pyruvate as the sole carbon source (EX_Kengen_1994) (Data set S4). To estimate the ATP demands of the cell, we calculated the maximum flux of ATP hydrolysis under each condition while constraining the products and growth yields based on experimental data. A linear model was established to describe the relationship between the maximum ATP hydrolysis and the growth yield, where the slope of this line represents the growth associated ATP maintenance (GAM) cost, and the y-intercept represents the non-growth associated maintenance (NGAM) cost (Fig. S1). These values were incorporated into the model and applied to all future simulations. Biomass simulations were performed using the “fba” function implemented in PSAMM version 1.2.1 with the IBM ILOG CPLEX Optimizer version 22.1.0 and Python version 3.9.15, with minimum and maximum flux constraints of product yields defined by the range observed in experimental measurements. Detailed simulation settings for the model validation are provided in Data set S3.

Additional growth simulations were performed on the WT model using peptides and maltose as carbon substrates in the presence and absence of sulfur. Media formulations were based on Adams et al. (44), emulating the maltose, peptides, and maltose+peptides conditions used in that study, with and without the added 0.1% (wt/vol) elemental sulfur (Data set S4). Constraints were set on the central carbon metabolism as follows to represent TrmBL1 regulation of glycolysis/gluconeogenesis: GAPOR (R07159) fixed to zero during growth on peptide but was unconstrained during growth on maltose or maltose+peptides. Growth was simulated with the “fba” function of PSAMM as described above.

OE-AdhF calibration.

Random flux sampling analysis (see below) was used to generate a set of metabolic fluxes for all reactions using an initial version of the COM1c model, where no constraint was applied to the AdhF flux. The overall distribution of ethanol production values was analyzed to develop constraints for the representation of strains with high AdhF expression (OE-AdhF) and baseline AdhF expression (COM1c). The COM1c strain (baseline adhF expression) was calibrated by setting the upper bound of ethanol production to the 0.75 quantile of the sampled overall distribution. The OE-AdhF strain (high adhF expression) was calibrated by setting the lower bound of ethanol production to the 0.25 quantile of the sampled overall distribution. All metabolic simulations in this paper used the OE-AdhF or control constraints on ethanol production. However, note that these constraints are calibrated for the exact carbon concentration of the minimal maltose media ([24]; companion paper). Recalibration based on adhF expression levels might be necessary when applied to other growth conditions.

Experimental validation of the open and closed models.

Growth experiments were performed using P. furiosus MW631 (OE-AdhF) in bottles on a minimal maltose medium at 95°C as described in Lipscomb et al. ([24]; companion paper) with the following changes. Two experiments were designed to mimic open and closed systems. In the open system, the headspace was continually flushed with argon during growth with a bubbler and ethanol and acetate were analyzed after 21 h of growth. In the closed system, H₂ was allowed to accumulate inside the bottles during growth and products and H₂ concentrations were sampled every 3 h for 21 total hours of growth. Modeling of the open and closed systems was performed with the following constraints: in the open model, SHI (R07181) and SHII (R00700) were fixed to zero; in the closed model, no constraints were applied to SHI and SHII (Table 3).

Model-based simulation of engineering strategies.

Random sampling of flux distributions was performed using cobrapy version 0.26.2 (57) with the IBM ILOG CPLEX Optimizer version 22.1.0 and Python version 3.9.15. The GEM-iPfu model was converted from YAML to SBML format using the “sbmlexport” option in PSAMM to be imported into cobrapy. A medium file was defined based on a minimal maltose medium used in ([24]; companion paper) (Data set S4, EX_YM_adhf), with 5 g/L of maltose as the carbon source and 0.5 g/L yeast extract supplied. For all random simulations, the flux of GAPDH (R01063 and R01061) was set to 0 to prevent thermodynamically infeasible looping of the GAPDH, GAPOR, and GAPN pathways (e.g., without the constraint, GAPN may reach fluxes of 300 while growing on 14.6 mM disaccharides). GAPDH has extremely low expression during growth on glycolytic substrates and is active only in gluconeogenesis in P. furiosus (15, 58) and thus should not be carrying flux during growth on maltose. All random flux simulations were performed in both the closed and open model configurations defined above. Flux distributions were randomly sampled using optGpSampler (59), an artificial centering Hit-and-Run algorithm, as implemented in the “sample” function in cobrapy. Whenever random fluxes were collected, the total sample size was 2,500, split between 10 independent sampling runs of size 250. All samples were validated to ensure that they fell within the solution space using the “validate” function of the cobra.sampling.OptGPSampler module. Statistical differences in flux samples were assessed using the Mann-Whitney U test with a P-value under 0.05 indicating significance (Table S3). Additionally, Cohen’s d was used to measure effect sizes of shifting flux distributions. Effect sizes are referred to as small (d = 0.2), medium (d = 0.5), or large (d = 0.8) based on benchmarks suggested by (60).

Random simulations of the OE-AdhF model were examined to quantify the contribution of each reaction to the production of key energy and redox carriers in open and closed systems, using model settings calibrated under 95°C (Fig. 4). For each random simulation, we used the following procedure to calculate the proportion of individual reactions to the total production of ATP, Fd_red, and NADPH. First, for each reaction that contains a compound of interest (e.g., ATP), we calculate a production/consumption index, i, by multiplying the stoichiometric coefficient of that compound by the flux of that reaction. Negative products represent consumption of the compound while positive products represent production of the compound. Second, the total production or consumption of a given compound $t=\frac{1}{2}\times {\displaystyle \sum }_{}^{}\left|i\right|$, was calculated by taking the sum of absolute values of the production/consumption indices over all reactions containing the compound divided by 2. Finally, the contribution of a reaction over the total production was calculated as the ratio between the production index of a reaction i and the total production t. Figure 4 shows the proportion of each reaction averaged across all random simulations. Since the proportions plotted in Fig. 4 represent averages across all solutions, they do not necessarily add to one for each compound.

All simulations of the engineering designs were configured based on medium and environmental settings of the OE-AdhF model and product yields were simulated using the random sampling approach described above. Engineered knockout strains were simulated by constraining the flux of reactions controlled by a gene deletion to zero. Engineered knock-in strains were simulated by manually adding reactions to the model using the add_reaction function of cobrapy. For the addition of GAPN, reaction ‘R01058’ was added to the model and for the addition of CODH, reactions ‘CODH’ and ‘TP_CO’ were added (Data set S1). The CODH reaction combines the net functionality of the respiratory Na⁺-pumping enzyme MBH and CODH subunits of the Thermococcus onnurineus CODH complex into a single stoichiometric equation. Additionally, free uptake of CO was enabled in the exchange constraints when the CODH knock-in strains were simulated (Data set S4). A summary connecting the names of genes used in this study to their corresponding reaction IDs in the model is given in Table S4.

Production envelopes (Fig. 7 and Fig. S4) were generated using the phenotypic_phase_plane method of the cameo python package version 0.13.6 (61). Simulations were performed with the EX_YM_adhf media (Data set S4) and carbon uptake forced to its maximum value.

Data availability.

The P. furiosus genome-scale model, GEM-iPfu, is publicly available on GitHub at the following address: https://github.com/zhanglab/GEM-iPfu. Both YAML and SBML formatted models along with all inputs and analysis scripts used in this study are released at https://doi.org/10.5281/zenodo.7915813.