We ran in silico experiments to determine the nutritional requirements of the O. oeni PSU-1 strain and then we compared the results with experimental data to validate iSM454. For this purpose, we used the 44 single omission experiments described by Terrade and Mira de Orduña (2009) and the 17 experiments for alternative carbon sources described by Beelman et al. (1977). Specific restrictions were set to simulate each medium (Supplementary Data Sheet 1). For each medium, we defined the restrictions required to simulate the nutrients included in the medium by allowing flux only through exchange reactions corresponding to those nutrients. Otherwise, lower and upper bounds of exchange reactions representing substrate uptake were set to zero.

For single omission experiments, each of the 44 nutrients was removed, one by one, and an optimization run was carried out each time. Nutrients that inhibited growth when removed were considered to be essential. We considered that growth was inhibited when the specific growth rate of the auxotrophic mutant was less than 20% that of the wild type. For the second set of experiments, the 17 alternative carbon sources were tested independently by performing an optimization run in each case. Carbon sources that allowed growth without the presence of other carbon sources were considered to sustain growth by themselves.

For both sets of experiments, results were classified as true positives (growth observed both in vivo and in silico), true negatives (no growth observed neither in vivo nor in silico), false positives (growth in silico but not in vivo) and false negatives (growth in vivo but not in silico).

From these predictions, we calculated critical statistical parameters that define model performance, i.e., sensitivity, specificity, precision, negative predictive value, accuracy, and the F-score, as follows:

Following standard procedures, we added an equation (Equation 7) to represent non-growth associated maintenance (NGAM).

The values of NGAM were determined from the model for each experimental condition. For this purpose, we first fixed the consumption and production rates of different metabolites and then we progressively increased the NGAM from 0 to 5 mmol gDW⁻¹ h⁻¹. At each iteration, the growth rate was maximized and the error between the experimental and predicted growth rate was calculated. The value that minimized the error between the experimental and predicted growth rate was chosen as the ATP required for maintenance of cellular processes.

The experimental rates included in the model were specific consumption rates of glucose, fructose, citrate, L-malate, L-cysteine, L-serine, L-threonine; it also included specific production rates of D-mannitol, L-lactate, D-lactate, acetate, erythritol, and ethanol. They were calculated from experimental data of two batch cultures containing 0 and 12% ethanol, respectively, run in duplicate (see below).

An O. oeni PSU-1 preculture was prepared from a frozen stock by inoculating 100-ml Erlenmeyer flasks containing 75 ml MRS (Man, Rogosa, and Sharpe) medium (De Man et al., 1960), supplemented with 0.5 g L⁻¹ of cysteine. Before inoculation, the cells were subjected to ethanol adaptation. For this purpose, we serially passaged every culture, starting from 1% ethanol (v/v) to reach 0 or 12% ethanol concentration (v/v) in each culture.

The adapted cells were inoculated in 50 mL flasks containing 35 mL of a chemically defined, wine-like, culture medium to achieve an initial optical density at 600 nm (OD600) of approximately 0.2. We employed the modified culture medium described by Terrade and Mira de Orduña (2009), at an initial pH adjusted to 4.8.

The flasks were incubated at 25°C, without stirring. OD600 was periodically measured to calculate the specific growth rate. At the same time, the content from each flask was centrifuged, the supernatant was collected and an aliquot was injected in a Lachrom L-700 HPLC system (Hitachi, Japan) equipped with a Diode Array and a Refractive Index detectors (Merck Hitachi, Japan). Organic acids, alcohols and sugars were separated using an Aminex HPX-87H ion exchange carbohydrate-organic acid column (Bio-Rad, USA) and quantified, as described previously (Varela et al., 2003).

An O. oeni PSU-1 preculture was prepared from a frozen stock as described above. The cells (not pre-adapted in ethanol) were inoculated in 50 mL flasks containing 35 mL of the same chemically defined, wine-like, culture medium, but lacking the amino acids evaluated for essentiality (glutamate, glutamine, asparagine, and threonine), one amino acid per flask, in duplicate. The flasks were incubated at 25°C during 13 days, without stirring, and OD600 was periodically measured to calculate the specific growth rate.

For each optimization using experimental data, non-zero reduced costs were extracted from the solver solution and employed for quantifying the impact of changing a capacity constraint on the objective flux. Scaled reduced costs were calculated as follows:

Where w_i represents the reduced cost, q_i the flux through exchange reaction i, and μ the specific growth rate. W_i, the scaled reduced cost of the exchange reaction i, represents the fractional change in biomass obtained by a fractional variation in compound i. Reactions that showed both, a non-zero reduced cost and a non-zero scaled reduced cost, were further analyzed.

FVA was carried out by minimizing and maximizing the flux through each reaction, under either unconstrained or constrained conditions. Span range was sorted by magnitude and plotted.

Reactions unable to carry flux were considered blocked. For each of these, the cause of the obstruction was investigated by finding dead-end metabolites; these were determined searching for those metabolites that were only consumed or produced in the stoichiometric matrix. Additionally, we determined dead-ends by adding a demand (maximizing the flux) or a sink (minimizing the flux) reaction for each metabolite. Metabolites were considered dead-ends if the model was unable to produce, nor consume them.

We conducted a random sampling analysis using optGpSampler (Megchelenbrink et al., 2014), an efficient algorithm based on the Monte Carlo procedure hit and run (Smith, 1984). For each experimental condition, we set the algorithm parameters in order to sample 100.000 points using 500 steps between each point.

We applied the algorithm to explore the solutions at an optimal specific growth rate for each experimental condition. For this purpose, the model was restricted with the calculated optimal growth rate and the corresponding experimental consumption/production rates. Then, we applied the algorithm for determining the 100,000 flux distributions that accomplished these restrictions. For every condition, we found the 50 reactions that showed the greatest flux variations among the distributions. We classified these reactions according to pathways and then we sorted pathway frequency.

We also applied this algorithm to explore the solutions near the optimal specific growth rate, by following the same procedure described above. For this purpose, the lower bound for specific growth rate was fixed at 90% of the optimal, and the upper bound, at the optimal specific growth rate.

Oenococcus oeni is a heterofermentative bacterium. It consumes hexoses through the 6-phospho-gluconate pathway and produces carbon dioxide, D-lactate, acetate, and/or ethanol. The main metabolized hexoses are glucose and fructose. The latter can also be transformed into mannitol or erythritol to fulfill the demand for NAD⁺ required in the heterolactic fermentative pathway. Even though the genes related to mannitol and erythritol biosynthetic pathways were not found in the PSU-1 genome, these pathways were included in the model to account for reported experimental data (Beelman et al., 1977). The membrane transporters of these and other carbohydrates—arabinose, ribose, melibiose, mannose, fucose, xylose, and galactose—were found using PathoLogic (Dale et al., 2010), which is provided by Pathway Tools, and included in iSM454.

Meanwhile, as O. oeni synthesizes exopolysaccharides (EPS) (Ciezack et al., 2010; Dimopoulou et al., 2012, 2014), we included 7 reactions responsible for EPS biosynthesis, associated with 22 genes annotated in the genome.

Finally, we also curated the pathways related to peptidoglycan biosynthesis. The draft reconstruction contained three alternative pathways to synthesize peptidoglycan (pathways I, III, or V). We only left pathway I in the model because it was the most complete, i.e., 9 out of the 10 reactions of this pathway were associated with genes.

The iSM454 model contains the whole biosynthetic pathways for 6 amino acids (alanine, aspartate, glutamine, lysine, proline, and glycine), as arisen from genome annotation (Mills et al., 2005). In particular, the gene dapX encoding for the enzyme diaminopimelate epimerase was absent in the genome annotation and therefore, was added to iSM454 in order to complete lysine biosynthesis pathway (Rodionov et al., 2003). The biosynthetic pathways of the remaining 14 amino acids are incomplete (arginine, asparagine, cysteine, glutamate, histidine, isoleucine, leucine, methionine, phenylalanine, serine, threonine, tryptophan, tyrosine, and valine), in accordance with genomic analysis (Mills et al., 2005).

Oenococcus oeni does not store triglycerides as an energy reserve. Instead, fatty acids are mainly utilized for the construction of the cytoplasmic membrane. The lipid fraction of O. oeni is mainly composed by saturated fatty acids (laurate, myristate, palmitate, stearate), unsaturated fatty acids (palmitoleate, oleate, cis-vaccenate) and cyclopropane fatty acids (lactobacillate and dihydrosterculate) (Tracey and Britz, 1989b; Lonvaud-Funel and Desens, 1990; Garbay et al., 1995; Guerrini et al., 2002). Biosynthesis of saturated fatty acids was automatically included into the model by Pathway Tools. Meanwhile, the biosyntheses of unsaturated and cyclopropane fatty acids were manually added to the model with their respective gene associations.

The synthetic routes for cardiolipin, 3-D-glucosyl-1,2-diacylglycerol, L-1-phosphatidyl-glycerol, and lysophosphatidylglycerol, starting from dihydroxyacetone phosphate, were also included.

In relation to β-oxidation of fatty acids, Pathway Tools assigned two genes, OEOE_1366 and OEOE_1263, to the same acyl-CoA synthetase (EC number 6.2.1.3) associated with a generic acyl fatty acid. We therefore manually included these genes and reactions for the canonical catabolism of the above-mentioned saturated fatty acids.

A modified version of the reported biomass equation of L. lactis (Oliveira et al., 2005) was included in iSM454, according to some unique features reported in the literature for O. oeni. For example, deoxyribonucleotide content was taken from the genomic analysis for O. oeni PSU-1 (Makarova et al., 2006). Fatty acid composition was determined as the average of each individual molecule (Tracey and Britz, 1989b; Lonvaud-Funel and Desens, 1990; Garbay et al., 1995; Guerrini et al., 2002). Amino acids, ribonucleotides and lipids composition correspond to L. lactis biomass composition (Oliveira et al., 2005). Similarly, macromolecular elements (proteins, lipids, DNA, and RNA) were included (Oliveira et al., 2005). Finally, the lipoteichoic acid (LTA) synthetic pathway present in the L. lactis genome was eliminated because the genes for its synthesis were absent in the O. oeni's genome and its presence has not been described in this microorganism (Ribéreau-Gayon et al., 2006).

Malate metabolism was added to the model, considering the transformation of malic acid to lactic acid by the malolactic enzyme (malate decarboxylase), codified by the OEOE_1564 gene. In this reaction, a cytosolic proton is consumed, and lactic acid diffuses outside of the cell (Figure 2) (Salema et al., 1996; Konings et al., 1997). The net result of this process is a decrease in the concentration of intracellular protons, contributing to the formation of an electrochemical gradient. Additionally, the citrate lyase complex was lumped into one reaction, directly allowing the conversion of citrate to oxaloacetate. A stoichiometric equation to account for the diffusion of citrate inside O. oeni was also added. The model also contains a functional ATP synthase system.

Model outputs of the essentiality of some amino acids differed from literature data (Garvie, 1967; Tracey and Britz, 1989a; Fourcassie et al., 1992; Mills et al., 2005; Terrade and Mira de Orduña, 2009). Therefore, we addressed these differences by experimentally evaluating their role on cell growth (Figure 3).

For example, Mills et al. (2005) reported that the genes related to the threonine biosynthetic pathway were all present in PSU-1. However, we identified a pseudogene within this pathway and experimentally demonstrated its essentiality for O. oeni PSU-1. On the contrary, even though several genes of the asparagine biosynthetic pathway were not found in the genomic sequence, our experimental results confirmed that O. oeni PSU-1 could synthesize this amino acid (Figure 3); and the whole pathway was included in the reconstructed model. In the case of glutamine, both our experimental results and model reconstruction confirmed that this amino acid is not essential, at least for this strain. Finally, O. oeni PSU-1 showed auxotrophy for glutamate, in agreement with previous results (Garvie, 1967; Tracey and Britz, 1989a; Fourcassie et al., 1992; Mills et al., 2005; Terrade and Mira de Orduña, 2009).

First, we carried out an in silico single omission experiment to compare the nutritional requirements predicted by the model with the experimental data obtained after growth of O. oeni in the culture medium of Terrade and Mira de Orduña (2009). Second, we tested alternative and independent carbon sources to evaluate growth and then we compared these results with the ones obtained by Beelman et al. (1977) (Table 4). Additionally, a confusion matrix was constructed to measure the performance of our predictions (Figure 4). This approach has been used before to assess the GEM quality of S. cerevisiae iIN800 (Nookaew et al., 2008) and iLL672 (Kuepfer et al., 2005), as well as for L. plantarum (Teusink et al., 2006) and Y. lypolitica (Loira et al., 2012).

The in silico analysis of nutritional requirements showed that at least one carbon source is required for growth. Thus, one out of the 9 following carbon sources could be employed to sustain growth: glucose, fructose, ribose, galactose, arabinose, cellobiose, trehalose, melibiose, or gluconate. Additionally, the model predicted that 14 amino acids are essential for growth (arginine, cysteine, histidine, isoleucine, methionine, phenylalanine, tryptophan, tyrosine, valine, leucine, threonine, serine, glutamate, and asparagine), and that the 6 remaining ones (alanine, aspartic acid, glutamine, glycine, lysine, and proline) were not.

Single nucleotide omission experiments of iSM454 showed that these metabolites were not essential for O. oeni. On the other hand, nicotinic acid and pantothenate were predicted to be essential nutrients. Moreover, several vitamins, such as biotin, folic acid, pyridoxine, riboflavin, and thiamin were not essential.

The model was able to identify that O. oeni was able to grow in 91% of the cases in which growth has been observed experimentally (sensitivity); whereas it identified correctly 96% of the cases where O. oeni did not grow (specificity). Furthermore, 97% of the experiments in which O. oeni was predicted to grow, O. oeni actually grew (precision). Additionally, 90% of the experiments in which O. oeni was not predicted to grow, O. oeni actually did not grow (NPV). The accuracy of the model was 93%, i.e., the proportion of correct results to total predictions. By comparison, the GEM model of L. plantarum presents an accuracy of 86% (Teusink et al., 2005); and the iIN800 model of S. cerevisiae, 89% (Nookaew et al., 2008). Finally, the F-score, a measure of the accuracy that can be interpreted as a weighted average of the sensitivity and the precision, was 94%, indicating that overall the model has a very good performance.

We obtained one false positive related to prediction of growth in the absence of L-glutamate, because of the presence of transamination reactions in the network, which artificially allowed the production of this amino acid. On the other hand, we obtained three false negatives related to growth in the absence of L-serine, and growth with esculin and salicin as sole carbon sources.

In order to predict the NGAM values for each experimental data set, we optimized the growth rate considering a range of possible NGAM values. The value of NGAM that allowed the minimal error at each specific growth rate prediction was selected, reaching an average error in the biomass formation of 0.14% in the two conditions analyzed. These NGAM values accounted for 0.07 and 2.3 mmol of ATP gDW⁻¹ h⁻¹ at 0 and 12% ethanol, respectively. Thus, when exposed to 12% ethanol, O. oeni PSU-1 spends 30 times more ATP to maintain the cellular machinery than in the absence of ethanol.

FBA of experimental data showed that a significant redistribution of intracellular fluxes occurs in the cell when O. oeni is grown in the absence of ethanol or under 12% ethanol content. To compare these two conditions, fluxes were standardized by growth rate. The glucose uptake rate is similar for 0 and 12% ethanol (Figure 5). On the contrary, significant changes occur in the consumption rates of fructose, malate, and citrate. The uptake rate of these compounds increases 102, 169, and 127%, respectively, when the bacterium is cultivated with 12%v/v ethanol. The net result is an increase in the fluxes through the heterolactic pathway. Consequently, a higher production rate of D-lactate (279%), L-lactate (144%), acetate (150%), mannitol (39%), and erythritol (7%), was achieved.

Despite the fructose uptake rate more than double in cultures with 12% ethanol, the mannitol production rate only increased by 39%. Indeed, in the absence of ethanol, Y_{mannitol/fructose} was 0.82; meanwhile, at 12% ethanol, this yield decreased to 0.56. Thus, fructose in cultures with ethanol is preferentially transformed to fructose-6-phosphate—and then to glucose-6-phosphate—compared to those without ethanol, which subsequently leads to a higher production of D-lactate and acetate. In fact, considering total carbon source as the sum of glucose, fructose, citrate, and L-malate, resulting Y_{D−lactate/total C} and Y_{acetate/total C} were 0.044 and 0.35 in the absence of ethanol; and 0.086 and 0.46 at 12% ethanol, respectively.

The flux through the malolactic reaction was also much faster in ethanol-containing cultures. The uptake rate of L-malate increased 169%, and the concomitant production rate of L-lactate, 144%. It is worthy to note that in both cases, not all the L-malate was transformed to L-lactate. A minor fraction is transformed to oxaloacetate through the malate dehydrogenase. Interestingly, Y_{L−lactate/L−malate} slightly decreased from 0.79 to 0.71, suggesting that for cells grown at 12% ethanol, a higher part of L-malate is destined to oxaloacetate.

Ethanol content significantly impacts the production rate of diacetyl, which increases from 0 to 4.23 mmol gDW⁻¹ h⁻¹. Regarding erythritol, even though its production remains almost the same in both conditions Y_{erythritol/glucose+fructose} decreased from 0.36 to 0.25, suggesting that a higher extent of fructose and glucose is transformed into other metabolites than erythritol.

As expected from the higher carbon flow through the heterolactic pathway in ethanol-containing cultures, the corresponding ATP specific production rate was 3-fold faster than in the absence of ethanol, passing from 0.74 to 1.98 mmol gDW⁻¹ h⁻¹ for 0 and 12% ethanol, respectively. These were calculated by adding the fluxes of acetate kinase (EC 2.7.2.1), pyruvate kinase (EC 2.7.1.40), and phosphoglycerate kinase (EC 2.7.2.3); and subtracting the fluxes through hexokinase (EC 2.7.1.1/E.C 2.7.1.2) and fructokinase (EC 2.7.1.4). On the other hand, by using the usual method for ATP determination through heterolactic fermentation, which consists of adding the total D-lactate and acetate produced, the resulting ATP production rates were 1.38 and 2.12 mmol gDW⁻¹ h⁻¹, for 0 and 12% ethanol-containing cultures, respectively.

Total ATP production increases as the concentration of ethanol in the medium increases (Figure 6). The model includes ATP generation by both, heterolactic fermentation and ATP synthase. At high ethanol content, the percentage of ATP produced via ATP synthase slightly decreases, from 48 to 45%; while the ATP formed through the heterolactic fermentation increases, correspondingly.

We assessed the impact of exchange reactions on the growth of O. oeni, by conducting a sensitivity analysis (Table 5) through estimation of the reduced costs, as well as of the scaled reduced costs, associated with these constrained fluxes. This methodology has been used before to assess the impact of metabolic reactions on ATP formation in the lactic acid bacterium Lactobacillus plantarum (Teusink et al., 2006). Reduced costs allow quantifying how much the objective flux could be improved by changing a capacity constraint. Scaled reduced costs represent the reduced cost normalized by the current biomass flux and the flux associated with the constraint. They allow computing the relative effect of a change in a parameter to the whole system.

The increase of sugar uptake rate—fructose or glucose—had a positive effect on the specific growth rate (scaled reduced costs of 2.59 and 1.67, respectively) as well as the transport of acids -malate or citrate uptake- (scaled reduced costs of 0.86 and 0.03, respectively), and the export of acetate, L- and D-lactate (0.47, 0.34, 0.09 respectively). The specific uptake rate of cysteine, serine and threonine also showed a positive effect on the specific growth rate (scaled reduced costs of 0.05, 0.01, and 0.06, respectively). On the contrary, the production rate of D-mannitol and D-erythritol had a negative effect on the specific growth rate, which showed scaled reduced cost of −2.67 and −1.48, respectively.

To evaluate the robustness of our results, we employed random sampling and FVA to identify and explore the different phenotypes achieved at optimal specific growth rate.

FVA revealed some important differences between unconstrained and constrained networks (Figure 7). 485 and 419 reactions were able to carry flux in the unconstrained and constrained network, respectively. The larger differences were found for reactions with a narrow flux range (less than 1 mmol gDW⁻¹ h⁻¹, i.e., a negative value for the logarithm of the flux range in Figure 7). We found only 30 reactions with a flux range of less than 1 mmol gDW⁻¹ h⁻¹ in the unconstrained network; meanwhile, this value increased to 370 in the constrained network. Moreover, 99% of the reactions in the constrained network showed a flux range lower than 3 mmol gDW⁻¹ h⁻¹, suggesting that the constraints applied (uptake and production rates) strongly delimit the solution space. Thus, applying these constraints, the phenotype is well defined.

Additionally, random sampling in the constrained network revealed that the solution space was tight and alternative pathways were limited. Except in the case of reactions of isomer interconversions which could result in a futile cycle, none of the reaction rates analyzed changed more than 1.1 mmol gDW⁻¹ h⁻¹.

Interestingly, when constrained by experimental rates, FVA shows that O. oeni requires oxygen to achieve growth at all ethanol levels. Moreover, the oxygen consumption rates needed for growth increase as the concentration of ethanol increases, ranging from 0.8–1.2 to 4.1–4.2 mmol gDW⁻¹ h⁻¹ for 0 and 12% ethanol, respectively.

The essential reactions of O. oeni were predicted by in silico simulations of reaction knockouts—inhibiting the activity of the enzyme(s) carrying away the respective reaction. This was conducted by further constraining the model, i.e., fixing the flux of the corresponding reaction to zero.

132 essential reactions were found by reaction deletion analysis, which represent 20% of the 660 total reactions of the model; of these, 28% correspond to fatty acid biosynthesis, and 14% to unsaturated fatty acid biosynthesis. The other main essential pathways include the biosynthesis of peptidoglycan (10%), glycerolipids (8%), and amino acid biosynthesis (8%) (Figure 8).