2.1.1. Model refinement.

The model refinement was executed following manual curation and gap filling. Every GPR association created during the draft model stage was investigated and validated by comparing the predicted GPR association based on BLAST to the annotation of the genes from different bioinformatic databases (e.g., KEGG[62], BioCyc[63], BRENDA[64], UniProt[65], and MetaNetX[66]) as well as available biochemical information from the literature (See Methods). Manual curation was based on protein sequence homology and EC number annotations. Additionally, the new validated GPR associations were cross-checked using a secondary assessment based on NCBI and PATRIC annotations. For GPR associations with exogenous proteins, multiple protein sequences were aligned using R. palustris Bis A53 proteome and proteins from the BiGG database [67] associated with those reactions. R. palustris amino acid sequences (RPE) aligned under the selected BLASTp parameters (see Methods) were manually verified based on the information in the databases and assigned to the corresponding reactions. Reactions without associations to R. palustris were identified and evaluated using sink metabolite algorithms [56]. Reactions with exogenous GPR associations were classified according to their flux contribution to the BOF and their biological relevance. The tested reactions with no contribution to the BOF and without biological relevance were removed from the model. Reactions involved in the production of BOF components or with biological significance were kept in the M-model and designated as orphan reactions (S6 Material).

The preliminary draft model with optimized BLASTp parameters (only containing iJB785 and iJN678 as references) consisted of 956 reactions, 802 metabolites, and 725 RPE genes. The draft model increased to 1,732 reactions, 1,569 metabolites, and 1,398 genes (1,198 of which had amino acid sequence information, i.e., RPE genes) after adding the homologous metabolic properties from iML1515 using the RAVEN toolbox [60]. Subsequently, the R. palustris GEM from the first manual curation stage was complemented using another metabolic model from the CarveMe prokaryotic modeling database [68]. Both metabolic models were compared based on reactions, metabolites, and genes involved in each model. The reactions only present in the CarveMe version were manually verified (GPR associations) and added to the draft model, checking energy and mass balances. A total of 566 reactions, 349 metabolites, and 117 genes were added to the semi-curated M-model (Fig 2). Prior to the gap-filling stage, all transport reactions of the draft model were curated using TransportDB [69] and metabolites were assigned to three compartments (cytosol, periplasm, and extracellular space). Reactions and metabolites that were originally located in the thylakoid and carboxysome compartment of the template models (neither of them is present in R. palustris) were reassigned to the cytosol and periplasm to maintain the biological relevance of the model. Carboxysome and thylakoid reactions were reallocated based on KEGG and BioCyc annotations regarding R. palustris’ photosynthetic pathway distribution. This step resulted in a model comprised of 2,349 reactions, 1,955 metabolites, and 1,275 genes distributed across three compartments.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Model properties and prediction capabilities of iDT1294.

A A general comparison was executed among the three principal template models (Escherichia coli K-12 substr. MG1655, iML1515, Synechocystis sp. PCC 6803, iJN678 and Synechococcus elongatus PCC 7942, iJB785) and iDT1294 reactions. The four models share 341 core metabolic reactions. B Six principal model versions were generated from the initial reconstruction process to the final validated model. Across the different stages, the model increased the number of reactions, metabolites, and RPE genes, while the number of exogenous genes was reduced to zero. C iDT1294 accuracy and phenotypic predictions capabilities were evaluated under the four main metabolic states of this versatile organism, testing over 350 experimental conditions. D R. palustris Bis A53 model contains 549 unique reactions related to aromatic compounds metabolism, PHB production, carotenoids, pigments and bacteriochlorophylls biosynthesis specifically synthetized by PNSB as well as nitrogen fixation and denitrification. Reactions (2,721 in total) were distributed in 17 subsystems representing the entire metabolism of R. palustris Bis A53.

https://doi.org/10.1371/journal.pcbi.1011371.g002

2.1.2. Gap filling.

Gap filling was executed in two separate steps: (1) gap filling of disconnected metabolites already present in the model and (2) gap filling based on experimental results, which consisted of (2a) phenotyping using Biolog plates and (2b) previously published studies. Initially, disconnected metabolites of the manually curated draft model were determined using COBRA Toolbox algorithms [56]. The disconnected metabolites were classified into three groups (see Methods) and reconnected depending on available information in bioinformatics databases. Metabolites remaining from template models but without evidence being present in R. palustris were removed together with their associated reactions. During the second step of gap filling, we used experimental data from Biolog experiments (PM1 and PM2 plates for carbon sources and PM3 for nitrogen sources) and modified reactions in the model accordingly. Each reaction added to the model in this step was manually reviewed to maintain concordance in the GPR associations, and energy and mass balance curation was performed to preserve the quality of the model. Transport reactions involved in the assimilation of carbon and nitrogen sources from Biolog results were associated with RPE genes using specific transporter information of the NCBI and PATRIC annotations. Unspecific transporting proteins were assigned to carbon and nitrogen sources without detailed information on the proteins implicated in inter- or intracelluar transport (e.g., general ABC amino acids transporters, general symporters, antiporters, porins, etc.). Furthermore, metabolic reactions involved in the catabolism of carbon and nitrogen sources revealed by Biolog experiments were added to the model with the corresponding RPE genes in GPR associations. Reactions employed in the suitable assimilation of the tested compounds without RPE evidence were kept in the model as orphan reactions. The resultant gap filled model contained 2,655 reactions, 2,081 metabolites and 1,282 RPE genes. In the second subsection of the final gap filling stage, literature information was deployed using an iterative approach to determine which reactions were missing in the model for their proper consumption. The model was modified by adding pathways for aromatic compound degradation and their use as carbon source under photoheterotrophic conditions. Further additions included pathways for nitrogen fixation and denitrification (nitrate and nitrite). In Fig 1C, the new relevant pathways added to the model are summarized and classified according to the pathways, reactions, and metabolites involved. Overall, a total of 372 reactions and 168 metabolites across all compartments were added to the final model during the gap-filling process.

2.1.3. Model properties.

The R. palustris Bis A53 metabolic model (iDT1294) consists of 2,721 metabolites, 2,123 reactions, and 1,294 genes, representing ~27% of all annotated coding genes in the NCBI genome reference sequence. Specific details about the reactions and metabolites from the model are summarized in Fig 2. iDT1294 was validated using over 350 experimental growth results under photoheterotrophic, photoautotrophic, nitrogen fixing (diazotroph), and heterotrophic conditions (both aerobically and anaerobically). iDT1294 contains all reactions, metabolites, and genes involved in nitrogen fixation and denitrification (nitrogen metabolism), PHB production, carotenoid and bacteriochlorophyll biosynthesis (S4 Material), anoxygenic photosynthesis, and aromatic compound degradation (Fig 2B). The entirety of the metabolic pathways present in the M-model is organized in 17 subsystems (including transport and exchange reactions) depending on the biological role of the reactions and metabolites utilized (Fig 2).

Most of the reactions (80%) in iDT1294 belong to lipid metabolism, amino acid metabolism, transport and exchange of metabolites, alternate carbon and cofactor, and vitamin metabolism. Specific metabolic capabilities of R. palustris Bis A53, such as nitrogen fixation and denitrification (nitrogen metabolism), PHB production (others), carotenoids metabolism (including pigments and bacteriochlorophylls biosynthesis), anoxygenic photosynthesis and aromatics compound degradation represent less than 6% of the metabolic reactions. Through this metabolic pathways’ distribution, iDT1294 can predict all four metabolic modes of R. palustris accurately (Figs 3 and 4). The three principal template models used during the optimization and reconstruction steps share 341 reactions with iDT1294. Most of the reactions present in the four models are related to core metabolic pathways (TCA cycle, glycolysis, gluconeogenesis, amino acids, lipids, carbohydrates and cofactor metabolism). Photosynthesis, carotenoid and pigments biosynthesis, and some alternate carbon pathways were taken from the two photosynthetic reference models (S. sp. PCC 6803, iJN678 and S. elongatus PCC 7942, iJB785). Gram-negative properties and metabolic pathways were obtained from the template E. coli K-12 substr. MG1655 (iML1515). iDT1294 shares 1,100 reactions with iML1515 (Fig 2), which are related to transport reactions, lipid metabolism, and the production of BOF constituents. Table 1 compares the properties of the different metabolic models reconstructed for R. palustris. To our knowledge, iDT1294 represents the most comprehensive M-model of R. palustris available to date.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Comparison of the principal model properties (reactions, metabolites, and genes) available for R. palustris.

https://doi.org/10.1371/journal.pcbi.1011371.t001

2.1.4. Biomass objective function.

The BOF contains the principal constituents and the stoichiometry values of each metabolite involved in biomass production. The fraction of each metabolite participating in the BOF composition is defined per gram of biomass. iDT1294 BOF was designed employing three different sources: (1) constituents of iML1515 and iJN678 based on their physiological similarity (Gram-negative bacteria or photosynthetic organisms), (2) estimations based on the proteomic and genomic data of R. palustris Bis A53, and (3) literature evidence of the BOF composition obtained under different experimental conditions [21]. The stoichiometric coefficients of amino acids, DNA, and RNA precursors present in the BOF were calculated based on the theoretical amino acid abundance obtained from proteomic and genomic references. The mineral fractions were determined based on mineral requirements from the literature (see Methods). Additionally, the carotenoid composition and their stoichiometric representations in the BOF were estimated assuming the presence of these metabolites in dark and light conditions [71]. The remaining constituents were estimated based on requirements of the Gram-negative model iML1515. The metabolites involved in biomass production were grouped into nine principal clusters (i.e. amino acids, cell wall, pigments, cofactors, RNA, DNA, carbohydrates, minerals, and miscellaneous). iDT1294’s final BOF composition contains 98 constituents (excluding the theoretical metabolites representing the BOF clusters). The BOF was employed as the primary optimization objective to perform metabolic predicitions of growth and internal fluxes for the four different modes of metabolism in R. palustris.

2.2.1. Growth validation.

R. palustris Bis A53 was grown under aerobic chemoheterotroph conditions in the dark using Biolog plates with 190 carbon (PM1 and PM2) and 95 nitrogen (PM3) sources. Growth was evaluated for 96 hours to determine which substrates were suitable nutrients for this PSNB. The growth values were normalized according to the average growth over triplicates per Biolog plate and reduced to qualitative data (growth or non-growth values). From this high-throughput phenotypic data, carbon and nitrogen compounds were mapped to iDT1294. Out of 190 carbon sources, 143 metabolites were identified in the M-model; carbon substrates without defined pathways for assimilation or catabolism were not mapped. Model simulations of these conditions were performed under heterotrophic aerobic conditions with ammonium and N₂ as the preferred nitrogen sources. Mineral and O₂ constraints were adjusted to -1000 and the final fluxes were calculated according to the carbon source limitations and BOF requirements. The carbon and nitrogen substrates employed during the phenotypic experiments are displayed in S2 Material. The Biolog plate validation results were classified and analyzed in two specific groups: carbon (PM1 and PM2) and nitrogen (PM3) compounds. For the group of carbon sources, a total of 143 compounds were mapped and analyzed using the R. palustris Bis A53 metabolic model. Carbon sources with little metabolic evidence (e.g., laminarin, 3-0-b-D-galactopyranosyl-D-arabinose, b-methyl-D-xyloside, etc.) were not included in the validation of the model. Before the gap-filling stage, the manually curated model version had 51% accuracy and Mathews Correlation Coefficient (MCC) of = 0.34 predicting carbon compound utilization. Most of the erroneous growth predictions were related to false negatives (70) since the model could not consume or metabolize different amino acids, carbohydrates, and organic acids. After the gap-filling stage, iDT1294 increased the accuracy to 85% and MCC to 0.68 (close to 35% more for both statistical parameters) regarding the carbon compounds. However, the growth estimations of the model were still affected by false negative predictions (19), mainly distributed in amino acids (8) and amino-containing compounds (8). The same procedure employed for the carbon set was followed in PM3 experiments to estimate the growth with 89 different nitrogen compounds. In this case, simulations were performed using sucrose as the sole carbon source. Statistical parameters (accuracy, sensitivity, specificity, positive predicted, negative predicted, and MCC) were calculated for chemoheterotrophic non-diazotrophic conditions with O₂ uptake adjusted to nitrogen requirements. Based on the PM3 simulation results, iDT1294 showed the highest accuracy and modeling precision across all Biolog conditions (Table 2). Predicting the qualitative growth of nitrogen compounds, iDT1294 achieves an accuracy of 94%, with 100% positive predictions and MCC close to 90%. The number of false negative predictions decreased to 6 compounds distributed in amino acids (3) and amino compounds (3).

To accurately determine how iDT1294 modeling capabilities compare to previously published models, the same procedure was applied to the four existing GEMs (Table 1) calculating the model simulation statistical parameters. Models distinct from BiGG naming nomenclatures (ModelSEED automatic reconstruction, iRpa940, and iAN1128) were translated to BiGG IDs to map carbon and nitrogen sources previously tested in iDT1294. However, from the four presented M-models, only two (iRpa940 and iAN1128) were suitable to perform Biolog simulations under the experimental conditions employed (aerobic heterotroph dark environments). The ModelSEED and CarveMe versions did not contain the required pathways and metabolic features to grow using a sole carbon or nitrogen source under oxic, dark conditions (chemoheterotrophic). Table 2 summarizes the results for the three metabolic models.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Comparative table with the statistical parameters of the three principal R. palustris metabolic models for carbon and nitrogen sources from the Biolog plates experiments (PM1, PM2, and PM3).

https://doi.org/10.1371/journal.pcbi.1011371.t002

Based on Biolog validation for carbon and nitrogen sources, iDT1294 demonstrated better prediction capabilities for growth and substrates assimilation: a global accuracy close to 90% (50% more than iRpa940, and 40% above iAN1128), sensitivity above 82% (almost 60% more) and MCC achieving 0.75. Most of the incorrect predictions of the previously published GEMs are based on the lack of transport reactions (S7 Material). For instance, at least 50% of the substrates from the Biolog sets were present only in the cytosolic compartment of iRpa940. Regarding iAN1128, the M-model contains significantly more transport reactions than iRpa940, although it lacks the metabolic pathways needed to employ these carbon or nitrogen sources as nutrients. Both previously published M-models contain specific information about the metabolites employed in the carbon and nitrogen validation but in the cytosol compartment. Further steps of manual refinement of the reactions transporting metabolites across the compartments and connecting the substrates to core metabolism can significantly improve the accuracy of the two metabolic models. Fig 3 displays the complete comparison of the performance of each model and the experimental results obtained from phenotyping.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Comparison of the high-throughput growth phenotypic experimental and simulated results across the three principal metabolic models.

R. palustris Bis A53 was cultivated in 190 carbon and 95 nitrogen sources under monoculture conditions for 96 hours. Subsequently, the three GEMs (iDT1294, iRpa940, and iAN1128) were properly constrained to simulate all the Biolog conditions. Each heat map compares every carbon (PM1 and PM2) and nitrogen (PM3) source’s experimental result against the simulation growth output per metabolic model. Growth results were classified into two possible results: Growth (purple) and No Growth (light pink). Additionally, we determined growth by measuring optical density (OD600, see methods) after 96 hrs, as indicated in blue to the right of each Biolog plate set. The substrates studied were sorted in descending order based on growth values. Statistical parameters to determine the global accuracy and prediction capabilities of the three models were calculated.

https://doi.org/10.1371/journal.pcbi.1011371.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Comparison of dynamic flux balance analysis with experimental data.

The set of plots shows the dynamic prediction capabilities of iDT1294 using nine different growth conditions. For each plot, the experimental time points (blue dots) were adjusted to the exponential phase, and time points in the stationary phase were removed. Later, dFBA was executed for iDT1294 to determine the growth rates and consumption of substrates across time. R2 and log₁₀P parameters were determined for iDT1294 (orange and star) to identify how the model predictions fit to the experimental results.

https://doi.org/10.1371/journal.pcbi.1011371.g004

2.2.2. iDT1294 metabolic trade-offs in steady state and dynamic scenarios.

Next, we tested iDT1294’s performance using steady state and dynamic data from the literature for diverse carbon and nitrogen sources. For steady-state data, two different sets of conditions were employed to determine the accuracy of the simulations: (1) aromatic compound degradation under anaerobic conditions in light (photoheterotroph conditions) with soluble nitrogen sources in the medium and N₂ in the environment [7,72], and (2) consumption of organic acids and glycerol under anoxygenic photosynthesis [17].

R. palustris Bis A53 contains high-specialized enzymes capable of degrading complex aromatic compounds under anaerobic growth [2]. To exploit the growth precision of iDT1294 we evaluated 14 aromatic compounds (with two lignin precursors) commonly found in wastewater from chemical and food industries [7,72]. Based on the experimental growth conditions of the reference studies [7,72], the model was constraint using a single organic carbon source (i.e., aromatic substrate) for each experiment, ammonium and N₂, and all minerals required for BOF optimization under anaerobic conditions in the light. The growth condition details employed for all experiments and simulations from the literature are summarized in S3 Material. iDT1294 accurately predicted the growth for 14 aromatic compounds under anoxygenic photosynthesis with 12 true positives (benzoate, benzoylformate, caffeate, cinnamate, ferulate, 3-hydroxybenzoate, 4-hydroxybenzoate, 4-hydroxybenzoylformate, DL-Mandelate, p-coumaric acid, p-coumaroyl-CoA, and vanillin) and two true negatives (protocatechuate and vanillate). For most of the carbon sources analyzed, the mechanism employed by the M-model for anaerobic catabolism of the aromatic substrates uses oxidoreductase reactions with different electron acceptors (NADH, NADPH, and FMN). The resulting metabolites from the oxidoreductases’ activity are transformed into common intermediates and integrated into core metabolic pathways (e.g., benzoate, pyruvate, and fumarate). Once iDT1294 was tested with the aromatic compounds using qualitative growth results, the quantitative growth rates were calculated for 1 mM substrate concentration. The predicted growth rates were compared with experimental evidence obtained from Harwood [1988] [7], shown in Table 3. The M-model predicted with an average accuracy of 84% the growth rates for all the carbon sources using ammonium or N₂ as nitrogen sources, achieving a minimum accuracy of 74% and a maximum value above 98%.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Predicted and experimental growth rates reported by Harwood [1988] [7] for R. palustris using aromatic carbon sources under anaerobic light conditions.

https://doi.org/10.1371/journal.pcbi.1011371.t003

iDT1294 was further validated using growth with organic acids and glycerol under anaerobic conditions in the light, employing experimental data from Govindaraju (2019) [17] to evaluate the model’s accuracy for predicting steady state and dynamic growth. Initially, the model was tested using nine different conditions with one or two carbon sources since the study was performed to establish the effect of co-substrate utilization in lactate assimilation (S3 Material). After the metabolic model was proven to successfully consume these metabolites, growth and concentration timepoint datasets from the nine experimental conditions were normalized to growth rates and consumption percentages, respectively. Subsequently, dFBA [56] was applied constraining the model with the initial biomass and substrate concentrations established by Govindaraju (2019) [17]. To calibrate the dFBA algorithm, we used the final growth rate and time point as reference data for each experiment. For the experiments with co-substrate utilization, the uptake rates for the two carbon compounds were estimated based on the exponential rate consumption of the dFBA. Regulation and inhibition features were not included in the dynamic simulation constraints. Ultimately, the correlation coefficient and exponential logarithmic parameters (R2 and log₁₀P) were calculated comparing the experimental growth and consumption rates against the model estimations across the time.

iDT1294 demonstrates a strong correlation R2 coefficient (98.6%) for the growth rate samples. For most of the conditions analyzed, the model accurately predicted lag and exponential phase trends with clear mismatches of the starting point of the stationary phase. Additionally, substrate consumption across the conditions was predicted correctly according to the dynamic simulation time points generated through dFBA model analyses (R2>97%). In some specific cases (acetate-glycerol, butyrate and butyrate-lactate), the consumption of carbon sources was incorrectly estimated since the COBRA Toolbox algorithms assume constant uptake fluxes, ignoring the biological regulation and uptake restrictions of the organism. Surprisingly, iDT1294 was capable of accurately predicting growth rates after depletion of the available carbon sources. When the carbon source is depleted, R. palustris utilizes CO₂ as the carbon source (carbon fixation). For instance, under malate and malate-lactate conditions nutrients were depleted after 100 hours and CO₂ became the sole carbon source available; iDT1294 accurately predicted the change in growth rates for these scenarios. Based on the statistical correlation parameters for growth rate and substrate consumption values, our comprehensive model is established as a suitable tool to estimate and analyze the dynamic metabolic behaviors of R. palustris Bis A53 under photoheterotrophic conditions.

4.2.1. Manual curation.

In the first step of the manual curation process, reactions in the initial draft model with exogenous proteins in the GPR associations were identified and verified in R. palustris metabolism using data from bioinformatics databases (KEGG, BioCyc, BRENDA, UniProt, and MetaNetX). Exogenous proteins assigned during the reconstruction stage were replaced with homologous sequences of R. palustris Bis A53 using the Enzyme Commission (EC) number as a reference to identify the metabolic function of the protein. For every subsystem of the model with exogenous proteins, the GPR associations were manually reviewed using KEGG, BioCyc, NCBI and PATRIC. When the sources disagreed about which protein should be associated to a reaction, a unique GPR was chosen by following the consensus among the databases and annotations. Additionally, reactions lacking any R. palustris Bis A53 proteins were checked through BLASTp. Candidate proteins from the R. palustris proteome were identified via BLAST against the proteins assigned to the GPR associations for the reaction in other microorganisms in the BiGG database. The BLASTp criteria were an e-value ≤ 1e-10, query coverage ≥ 80% and identity percentage ≥ 40%. A second step of manual curation was executed to confirm the correct assignation of the GPR associations [26,54,55]. The proteins for each reaction of the semi-curated draft model were manually reviewed based on the type of metabolic reaction, protein function, and cell compartment. Protein complexes were adjusted correctly from the GPR associations of the template organisms to the specific protein complex conformations in R. palustris. All validated reactions, metabolites, and GPR associations were distributed in three different compartments (periplasm, cytoplasm, and extracellular compartment). Metabolites were labeled according to the corresponding position in the cell. Metabolites obtained from the photosynthetic model references were carefully renamed and properly allocated to the corresponding compartment of R. palustris Bis A53. Duplicated metabolites generated during the renaming and reallocation of photosynthetic metabolites were merged and integrated into a unique metabolite identifier. Transport reactions were added using the TransportDB database [71]. Metabolite transport between compartments was curated using BLASTp. Hypothetical and putative proteins were not included in the GPR associations of the curated model to avoid false positive calls. The remaining reactions with exogenous genes in the GPR associations of the model were identified and analyzed through Flux Balance Analysis (FBA) using the COBRA Toolbox [56]. Reactions with no flux and exogenous GPR associations were removed from the M-model. Reactions with exogenous proteins carrying fluxes were maintained in the model as orphan reactions.

4.2.2. Gap filling.

Gap filling was performed in two steps, (1) gap filling of metabolic pathways already present in the manually curated draft model, and (2) gap filling by adding new metabolic pathways into the model obtained from different bioinformatic databases or experimental evidence from the literature.

Gap analysis was performed to recognize which compounds were disconnected in the M-model and which reactions were lacking in the analyzed pathways. Initially, disconnected metabolites were determined through dead-end analysis algorithms. The dead-end metabolites were classified according to their disconnection motive (present in one reaction, present only as a substrate or as a product). Subsequently, the involved reactions were identified to establish suitable connections of dead-end metabolites. Disconnected reactions were manually added using different bioinformatic databases (e.g., KEGG, Biocyc). After these analyses, gap filling was employed to connect pathways (lipid metabolism, gluconeogenesis, TCA cycle, etc.) through the data retrieved. Finally, dead-ends were identified through KEGG and BioCyc pathway modules and synthesized using COBRA gap filling algorithms [56]. A final test was performed to verify the correct production of each dead-end metabolite using sink reaction algorithms from the COBRA Toolbox. Subsequently, the second round of gap filling was performed to connect metabolites from the medium retrieved using literature information [17,76–79] and experimental data generated in the present study. Gap filling algorithms identified the reactions involved in the assimilation of carbon and nitrogen sources under photoautotrophic, photoheterotrophic, diazotrophic, non-diazotrophic, and heterotrophic conditions. Both aerobic and anaerobic conditions were considered to perform these simulations and the second round of gap filling. The reactions added in the gap filling process with no GPR associations were annotated as orphan reactions. Ultimately, reaction fluxes were validated using FBA to verify the predicted internal fluxes to identify the production of each metabolite present in the model.

In the second step of gap filling, new pathways were added to the refined model using semiautomatic algorithms. Specifically, carbon (aerobic and anaerobic degradation of aromatic compounds and PHB production), nitrogen (denitrification and partial denitrification), and photosynthetic (carotenoids, bacteriochlorophylls, and bacterial pigments) metabolisms and their annotation. The names of new reactions and metabolites were assigned according to the BIGG database; reactions and metabolites with no information in BiGG were added to the model according to the EC Number information (through BRENDA [64]) and bioinformatic databases (KEGG, Biocyc, etc.). Detailed information for reactions and metabolites (charge, formula, reversibility, direction, etc.) were extracted from well-reviewed biochemical databases (PubChem, UniProt, ModelSEED, KBase [73], and MetaCyc) or from the metabolic models of R. palustris Bis A53 from Machado et al 2018 and Navid et al., 2019 [51,68]. For R. palustris Bis A53-specific features, like bacteriochlorophyll biosynthesis and pathways for aromatic compound degradation, literature information using experimental evidence was employed to reconstruct these specific metabolic subsystems. Pseudoreactions with unspecific intermediates or used to represent transformations with no known pathway were not included in the M-model to avoid inaccurate predictions. All pathways of aromatic compound utilization were curated based on experimental evidence from Harwood and Ma [7,72]. A detailed list of reactions, metabolites, and genes added to the model for these pathways is presented in S5 Material. The functionalities of all reactions were validated using FBA through the model to predict biomass production. Reactions added in the second step of gap filling were tested by adding specific constraints in the model and performing simulations to measure the reaction flux distributions. Sink reaction algorithms were employed to assess the production of the new metabolites in the manually curated model.

4.2.3. Final quality control and quality analysis.

The final quality check was accomplished to guarantee correct GPR associations and suitably balanced reactions and metabolites. We performed in-silico single gene-deletion simulations (in-silico gene knockouts) to verify if the GPR associations are properly assigned using the COBRA and RAVEN Toolboxes. Next, we performed Mass and Charge Balance simulations of the COBRA Toolbox on the model to check for unbalanced reactions added during the model refinement. Unbalanced reactions were fixed by adding the correct formula and charge of each metabolite. Stoichiometric values assigned in every reaction were carefully reviewed and corrected for unbalanced reactions. Ultimately, the final model was analyzed by looking for ATP, NADH, and NADPH free energy production, by removing media nutrients (exchange reactions) and checking that they had zero flux.

4.3.1. R. palustris Bis A53 culturing and growth conditions.

R. palustris Bis A53 was purchased from the American Type Culture Collection (ATCC BAA-1125). R. palustris Bis A53 was grown in mineral medium [80] in light conditions for 14 days at 26°C. Purple cultures were transferred to an aerobic medium with the following composition: yeast extract, 0.1 g; NH₄Cl, 2.7 g; KH₂PO₄, 0.5 g; MgCl₂ · 6H₂O, 0.33 g; CaCl₂ · 2H₂O, 0.05 g; NaCl, 0.4 g; 9.5 ml Pfennig and Lippert’s mineral solution; and 8.5 ml vitamin solution [81]. R. palustris was cultured in aerobic medium for 3 days at 30°C under dark conditions. Optical density (OD) measurements were taken at 600 nm (OD600). A solution based on OD600 readings was prepared for the Biolog plates experiments. After the inoculation, Biolog plates were incubated aerobically at 30°C in the dark (chemoheterotrophic condition).

4.3.2. Carbon source utilization assays.

Bacterial suspensions of R. palustris were tested for growth on 190 different carbon sources using Phenotypic Microarray (PM) plates 1 and 2A from Biolog, Inc. (Biolog, Inc., Hayward, California) following the company’s instructions. Briefly, cell suspensions from a 10-day old culture (see above) were washed and resuspended in Biolog’s IF-0a GN/GP Base (1.2X) inoculating fluid (#72268) up to an optical density of 0.025 at 600 nm (OD600, Molecular Devices SpectraMax M3 Multi-Mode Microplate Reader (VWR, cat # 89429–536)). The suspentions were further supplemented with ATCC trace minerals and vitamin supplements (ATCC-MDTMS and ATCC-MDVS). 100 uL of these washed cell suspensions were inoculated into each well of the Biolog plates PM1 and 2A (a full list of compounds including a blank well with no carbon source (negative control) can be found here). Plates were incubated at 30°C without shaking with lids coated with an aqueous solution of 20% ethanol and 0.01% Triton X-100 (Sigma) to prevent condensation [82]. We defined substrate utilization by R. palustris as an OD600 increase > 0.02 (readout of total cell biomass from substrate use) over a 96 hour incubation period.

4.3.3. Nitrogen source utilization assays.

The bacterial suspensions for the nitrogen use assays were prepared in the same way as for the carbon source but were washed in IF-0a GN/GP Base (1.2X) inoculating fluid (#72268) containing 5 mM potassium phosphate monohydrate, pH 6 (Millipore-Sigma), 2 mM sodium sulfate (Millipore-Sigma) and 20 mM sucrose as the carbon source. The rest of the steps and determination of substrate utilization as a nitrogen source were similar to the ones described above for the carbon source assay.