An experimentally validated manually curated genome-scale metabolic model of Enterococcus faecium TX0016 reproduces high-throughput phenotyping data

Enterococci are Gram-positive opportunistic pathogens responsible for a wide range of nosocomial infections. One enterococcal species, Enterococcus faecium , is steadily increasing in prevalence and has been listed among major multidrug-resistant ESKAPE pathogens. To gain systems-level insights into its metabolism and support discovery of potential therapeutic targets, we constructed iDR478, a comprehensive manually curated genome-scale metabolic model (GEM) for E. faecium TX0016 (strain DO). The reconstruction was based on extensive homology searches and literature evidence, and was further refined and gap-filled through experimental validation. Phenotypic profiling using Biolog microarrays enabled assessment of carbon source utilization, while amino acid leave-out growth assays allowed the evaluation of auxotrophies. The final model, after phenotypic-based refinement, successfully predicted 100% of amino acid auxotrophy and 85% of growth on sole carbon sources. Discrepancies between model predictions and experimental phenotypes identified specific knowledge gaps across metabolic pathways, including unresolved carbon source utilization phenotypes, e.g., psicose, sorbitol, and palatinose utilization. Those gaps will guide future experimental characterization. Additionally, we conducted gene essentiality analysis to evaluate the predictive capacity of the iDR478 model. Since no experimental gene essentiality data are currently available for E. faecium , model predictions were compared against Tn-seq experimental results from E. faecalis MMH594. Under simulated unconstrained uptake conditions, iDR478 achieved 84.2% positive predictive value of the experimental essentiality results of E. faecalis MMH594. iDR478 thus provides a framework for studying E. faecium , offers insights into its metabolic network, and serves as a source for generating hypotheses for future experimental investigation.

Enterococci are Gram-positive facultative anaerobic bacteria that naturally colonize the gastrointestinal tract of humans and animals (Murray, 1990). Enterococcus faecalis and Enterococcus faecium are the most prevalent species causing nosocomial infections (Fisher and Phillips, 2009). The most common enterococcal infections include urinary tract infections (UTI), infective endocarditis, and bacteremia; they also contribute to intra-abdominal infections and meningitis (Sood et al., 2008). Among enterococci, E. faecalis has been the primary focus of genomic and functional studies because of its higher predominance in clinical settings (He et al., 2018; Chaguza et al., 2023); however, E. faecium is increasingly being recognized for its strong association with multidrug resistance (Zhou et al., 2020). As a result, E. faecium is now classified among the ESKAPE family of antibiotic-resistant pathogens (Rice, 2008), emphasizing the urgent need for new treatment targets.

Exploring metabolic drug targets is a promising alternative to addressing multidrug-resistant bacteria. Building metabolic network models can be a powerful tool for identifying metabolic factors linked to pathogenicity and potential antibiotic targets (Chavali et al., 2012). Despite E. faecium ’s clinical significance, its metabolism has received limited attention, as research focused more often on antibiotic resistance mechanisms, epidemiological typing, and the characterization of virulence and colonization-associated proteins (Gao et al., 2018; Dyrkell et al., 2021; Coll et al., 2024). Consequently, aspects of E. faecium ’s metabolism remain poorly studied, and comprehensive systems-level studies of E. faecium metabolism are essential.

Computational modeling and in silico simulations are potent tools that allow deep understanding of bacterial phenotypes, including metabolism. Genome-scale metabolic models (GEMs) have been valuable in elucidating genotype–phenotype relationships and generating computational predictions of biological functions (Bordbar et al., 2014). They are a key step toward digital twins, intended to allow full in silico simulation experiments on virtual cells. GEMs are constraint-based modeling methods that translate genetic, genomic, and biochemical information into a mathematical framework for metabolic network reconstruction. They enable simulation and prediction of metabolic capabilities under various conditions (O'Brien et al., 2015). They also allow the prediction of gene essentiality by simulating the effects of multiple gene knockouts. Accordingly, GEMs have a wide range of applications, including the identification of antibiotic targets, the prediction of enzyme functions, and strain development to produce chemicals and materials (Oberhardt et al., 2009).

Thanks to a large set of well-curated models, automated metabolic reconstruction is rapidly becoming achievable, e.g., Model SEED (Devoid et al., 2013), KBase (Allen et al., 2022), and AGORA (Magnúsdóttir et al., 2017). However, manual curation and experimental validation of GEMs for at least a reference strain per species remain valuable. Numerous manually curated GEMs have been developed for lactic acid bacteria (LAB), to which E. faecium belongs, which is a group characterized by significant genetic and biochemical diversity that contributes to differences in metabolic capacities and pathogenicity among species (Douillard and de Vos, 2014). Published GEMs for this group include models of Lactococcus lactis (Flahaut et al., 2013), Lactiplantibacillus (Lactobacillus) plantarum (Teusink et al., 2006), and several Streptococcus species, such as S. thermophilus (Pastink et al., 2009), S. oralis (Jensen et al., 2020), and S. pyogenes (Levering et al., 2016). Despite this progress, only one manually curated GEM has been published for the Enterococcus genus, E. faecalis (Veith et al., 2015), leaving a gap in the modeling of the multidrug-resistant ESKAPE pathogen, E. faecium.

The rising availability of whole genome sequences has significantly accelerated the development of GEMs (Fang et al., 2020). For E. faecium, the first partially assembled draft genome sequence was published in 2000 for strain TX0016 (Fox, 2000), followed by the first complete genome sequence in 2012 for strain AUS004 (Lam et al., 2012) and a complete sequence for the TX0016 strain shortly thereafter (Qin et al., 2012). Today, over 33,953 genomes are available for E. faecium, 826 of which are described as complete (NCBI database accessed on September 30, 2025).

This study introduces iDR478, a manually curated and experimentally validated GEM for E. faecium. The model is reconstructed based on the genome sequence of the strain, E. faecium TX0016, also known as strain DO, isolated from an endocarditis patient in 1992. Our GEM enables a comprehensive exploration of the organism’s metabolic capabilities and allows for the study of its network behavior under diverse experimental conditions. We validated the model on experimental data from Biolog nutrient utilization phenotypic arrays and amino acid auxotrophy experiments. iDR478 serves as a starting point for future enhancements and further refinements in the metabolic modeling of E. faecium.

Strain

The strain whose genome was used in this reconstruction is E. faecium TX0016, also known as the DO strain (ATCC BAA-472), isolated from the blood of a patient with infective endocarditis. The DO genome sequence was downloaded from the National Center for Biotechnology Information (NCBI) under taxonomy ID: 3338491 in May 2023.

Growth conditions and media preparation

E. faecium TX0016 was initially cultured aerobically at 37 °C in brain heart infusion (BHI) broth, as recommended by ATCC. Glycerol stocks (25%) were prepared in 2 mL cryovials and stored at −80 °C for long-term preservation. For the validation experiments, the strain was grown at 37 °C in a chemically defined medium for lactic acid bacteria (CDM-LAB). The CDM-LAB formulation used in our study included M9 salts, buffers, trace elements, all 20 amino acids (AAs), nucleotide bases, multiple vitamins, and diverse carbon sources. The pH of the medium was adjusted to a neutral range (6.8–7.3) by HCl or NaOH. Concentrated stock solutions of most components were prepared in advance, except for components that needed to be freshly prepared (e.g., cysteine). All stock solutions were either syringe or vacuum filter sterilized through 0.22 μm pore-size membranes and stored under appropriate storage conditions. A detailed list of medium components, preparation methods, and storage requirements is provided in Supplementary Table S1. The CDM-LAB composition was adapted from previously developed formulations for E. faecium (Zhang et al., 2013) and for the closely related E. faecalis (Jönsson et al., 2009; Veith et al., 2015).

Metabolic reconstruction

To build the GEM for the E. faecium TX0016 strain, we used the protocol by Thiele and Palsson (2010). We began developing a metabolic network reconstruction based on genome annotation data obtained from NCBI. Candidate metabolic proteins, associated functions, potential reactions, and accession IDs were identified and curated using the information from the Bacterial and Viral Bioinformatics Resource Center, BV-BRC, formerly PATRIC (Antonopoulos et al., 2019; Olson et al., 2023) and comprehensive biochemical databases such as KEGG (Kanehisa et al., 2006), BiGG Models (King et al., 2016) and MetaNetx (Ganter et al., 2013). The curated information was compiled into a structured spreadsheet, forming the foundation of the metabolic network.

This draft reconstruction was subjected to extensive manual curation and refinement. Each metabolic reaction and function was verified against experimental data in literature, whenever possible. Because of limited specific data on E. faecium DO, literature from closely related species, most often E. faecalis, was used. Metabolic maps were constructed to help track and identify missing functions, using the pathway visualization tool Escher (King et al., 2015), guided by KEGG pathway maps. To fill the gaps in the network, we used BLASTp for homology searches through the NCBI web server (Boratyn et al., 2013) with default settings, relying on data from closely related species.

Each reaction or metabolite within the model was assigned a standard BiGG identifier. Metabolite identifiers were manually reviewed and standardized to ensure connectivity across the reactions in the network. For species-specific or missing pathways, not covered in BiGG databases, new metabolites and reactions were manually created and assigned unique identifiers.

Enzyme Commission (EC) numbers and genomic subsystem assignments were also added to all reactions, when applicable. Gene-protein-reaction (GPR) associations were assigned, linking each reaction to its corresponding gene (via NCBI accession ID), and, when available, its gene symbol. A function may be associated with a single gene, multiple adjacent genes acting together (AND relationship), or multiple genes capable of independently performing the same function (OR relationship), depending on the genomic context.

Additionally, each reaction was also given a confidence score, reflecting the strength of the supporting available evidence on the metabolic function. The confidence scoring system (Supplementary Table S2) was adapted from the Thiele and Palsson’s protocol (2010), and it ranges from 0 to 4, with 0 indicating the lowest level of evidence and 4 representing the highest. The scoring and evidence data were critical for identifying low-confidence reactions that could be the reason for discordances in subsequent simulations. An additional column specifying if the function is annotated in our genome or based on homology searches was also incorporated.

Transport reactions were added into the reconstruction based on genome annotation data from NCBI, KEGG, the Transporter Classification Database, TCDB (Saier et al., 2021), and supporting literature, where available. Exchange reactions were also included to allow for metabolite uptake and secretion, completing the reconstruction network.

Constraint-based modeling and simulations

The reconstructed metabolic network was converted into a mathematical model, with all simulations conducted using the Constraint-Based Metabolic Modeling toolbox in Python (COBRApy) version 0.25.0 (Ebrahim et al., 2013). These simulations enabled the investigation of the metabolic capabilities and adaptations through the metabolic network. Flux balance analysis (FBA) was performed using the default glpk COBRApy solver. Flux bounds were set according to the reversibility of each reaction, guided by information from the E. faecalis template model and the BiGG database. Stoichiometric mass and charge balance were maintained by ensuring elemental balance in each reaction.

Gap filling

To resolve network gaps and ensure model feasibility, we supplemented our curated model with reactions from a CarveMe model, which constructs a model by filtering a universal (Gram-positive) model to a specific one through solving a mixed-integer linear program (Machado et al., 2018). However, the reliance on a universal model in CarveMe would have limited the specificity of the model. Thus, the updated model underwent another round of intensive manual curation. Each gap-filled reaction, added by CarveMe, was thoroughly evaluated using BLASTp homology search and literature evidence, when available. A confidence score was also assigned to each added reaction to reflect the level of evidence supporting its inclusion. The model was retested for in silico biomass production. Whenever the model remained non-functional, we performed additional gap filling based on Reed et al. (2006) until the model produced a non-zero flux with biologically relevant objective function. Briefly, we curated the draft GEM using the gap filling algorithm implemented in COBRApy v. 0.25.0, and the BiGG Models database, downloaded on August 1, 2023, as the universal database of candidate biochemical reactions. The COBRApy gapfilling module uses mixed-integer linear programming to compare the draft model against the universal database. It constrains the model to a defined media input and sets biomass production as the biological objective function. Based on these constraints, the algorithm identifies the minimal set of external reactions required from the universal database to achieve simulated growth. Like with CarveMe, the proposed reaction sets were also reviewed and integrated into the network to produce a functionally complete model.

Biomass reaction

The biomass reaction includes all the essential components required for bacterial growth, including biomass precursors, such as amino acids, nucleotide triphosphates, and lipids. Due to the lack of a biomass composition specific to E. faecium, we relied on the biomass reaction used in E. faecalis GEM (Veith et al., 2015), guided by the model Gram-positive organism, Bacillus (default of the CarveMe model).

Amino acid leave-out experiments

Leave-out experiments were conducted to verify the essentiality of each AA in E. faecium and to validate our model. Pre-cultures of E. faecium DO cells were grown overnight at 37 °C in CDM-LAB and then washed twice with phosphate buffer saline (PBS) to minimize any residual components of the rich media. The omission experiment was conducted by systematic removal of one AA from the CDM-LAB to test whether E. faecium can grow without it. The washed cells were inoculated into the modified media lacking the specified AA, with an initial optical density at 600 nm (OD 600) of 0.05–0.075. The cultures were incubated overnight at 37 °C in 3 mL volumes into 5 mL capped tubes. A culture grown in rich CDM-LAB with the complete AA composition was used as a positive control. Cell growth was monitored by visual inspection and by OD 600 measurement the following day. If growth was observed, sequential inoculation steps were performed, in which cells that grew in the CDM-LAB were washed again and transferred to a fresh CDM-LAB to ensure complete removal of any residual AA from the growth medium. This step served to confirm the prototrophy of the respective AA. All omissions have been tested in at least two independently prepared media and at least three separate inoculations.

Biolog phenotyping experiments

Biolog’s Phenotype Microarrays (Biolog, CA, USA) were used for carbon utilization assays. Kinetic measurements were read by the Omnilog™ plate reader system. Carbon source utilization was tested in PM01 and PM02 plates containing 190 carbon sources. E. faecium cells were first cultured overnight in CDM-LAB, washed twice with PBS, and then inoculated into fresh CDM-LAB lacking glucose. This medium served as a negative control, as the strain does not grow under this condition. One hundred microliters of the inoculated cells was added into each well of the microplate and incubated at 37 °C for 48 h. Replicate experiments were conducted to ensure reproducibility and accurately determine the sole carbon utilization profiles of E. faecium.

Biolog plate data processing

Kinetic time-series data were collected from the Omnilog™ plate reader machine for PM01 and PM02. Each time-series signal was smoothed with a Savitzky Golay filter set to a 3rd-degree polynomial and a window size of 50. Maximum signal (respiration kinetics) values were extracted for each signal and compared to the control well maximum signal values (negative controls) with a z -test with multiple test correction to make growth calls on the carbon sources, as described under PMkbase (Krishnan et al., 2025). Carbon sources with both replicates showing growth were assigned as growth-usable carbon sources, and sources with no growth across replicates were assigned as unusable carbon sources.

Model simulations using COBRApy

Amino acid auxotrophy

In silico simulations were conducted to determine auxotrophic AA(s). CDM-LAB medium was defined in silico, and model exchanges were opened up for all 20 AAs. Each AA was then removed iteratively by setting its lower and upper bounds to zero to determine the biomass yield. Growth was considered feasible if the resulting objective value exceeded a threshold of 0.01 h −1. If growth was not yielded upon removal of a specific AA, the organism was auxotrophic to that AA; otherwise, the organism was considered prototrophic.

Carbon source utilization

The medium used for Biolog PM experiments (CDM-LAB without glucose) was defined in silico, and model exchanges were opened up for all carbon sources available in the Biolog phenotype microarrays. Since the model predicted in silico growth in the absence of glucose by catabolizing AAs—a phenotype not observed experimentally—a specific growth threshold was established corresponding to the in silico growth value (biomass production without added glucose). Model growth was tested on each carbon source, and only substrates generating biomass above this threshold were determined to be usable carbon sources. Discrepancies were visualized by Escher and revised against information from KEGG and BiGG. A confusion matrix was generated to express the concordance of the in silico growth outcomes with the experimental results. Experimental data on amino acid auxotrophies and carbon source utilization were incorporated to enhance the model’s accuracy.

Gene essentiality

To assess gene essentiality, we conducted single gene knockouts in the metabolic network of E. faecium by subsequently performing in silico single gene deletions under two conditions: with all exchange reactions open and in defined medium. For each gene, the flux through the associated reaction was constrained to zero flux. If the resulting objective function value dropped below a threshold objective value of 0.01, the gene was classified as essential; otherwise, the gene is not essential.

Gene essentiality analysis provides means of prioritizing candidate metabolic functions for subsequent experimental testing. We could not find established experimental data on gene essentiality for the E. faecium DO strain; therefore, we compared between iDR478 and experimental gene essentiality data available for E. faecalis MMH594 (Gilmore et al., 2020), as a cross-species consistency assessment. Orthologs were identified between E. faecium DO and E. faecalis V583 (a strain whose locus tags are mapped to those of MMH594). The two RefSeq proteomes were compared by BLASTp, and only bidirectional best hits were retained as orthologs. Protein accession numbers were mapped to locus tags through the NCBI feature tables for each assembly. Genes assigned by Gilmore et al. (2020) as “critical, important, potentially critical, or potentially important” were counted as experimentally required for fitness (essential), and those classified as non-essential as not required (non-essential).

Comparison with semi-automated AGORA2 E. faecium reconstructions

AGORA2 contains semi-automated reconstructions for 188 E. faecium strains (Heinken et al., 2023), including strain DO (TX0016). We compared iDR478 to the reconstruction of the same strain in AGORA2, as well as TX1330, used as a reference strain in that database.2

Amino acid auxotrophies were predicted for all three models by the same approach described in Section 2.10.1. All three were simulated on the complex medium distributed with the DEMETER pipeline, obtained from the repository of the AGORA2 reconstructions, because the AGORA2 models do not produce biomass under CDM-LAB. Additionally, no defined medium specific for any enterococcal strain was available in the DEMETER input data. Because reaction identifiers differ between the BiGG namespace used in iDR478 and the VMH namespace used in AGORA2, the 157 exchange reactions defining the complex medium and the 20 amino acid exchanges were mapped between the two conventions before simulation. Reaction-level overlap between the reconstructions was not computed, since not all identifiers were mappable.

We compared predicted carbon source utilization with the validation data reported for both strains in AGORA2, which derive from the NJC19 and BacDive datasets, and with the phenotypically refined final model (iDR478). To identify genes present in the AGORA2 TX0016 reconstruction but absent from iDR478, we mapped locus tags between the complete TX0016 genome (GCF_000174395.2) used here and the draft assembly used in AGORA2 (GCF_000166815.1) through their shared non-redundant protein accession numbers. We then assessed the reactions carried by those genes for inclusion.

iDR478: a manually curated genome-scale metabolic model for E. faecium

The new manually curated and experimentally validated metabolic model of E. faecium DO, iDR478, is based on its genome sequence (Qin et al., 2012) and includes all metabolic reactions required for cell growth (detailed reconstruction steps in Supplementary text, Figure S1). Because of scarcity of literature about E. faecium DO, we curated data from the closely related bacteria E. faecium AUS004 (Staerck et al., 2021), E. faecium E1162 (Zhang et al., 2013), and most often E. faecalis (Brown and Wittenberger, 1972; Hedl and Rodwell, 2004; Ramsey et al., 2014).

iDR478 comprises carbohydrate metabolism, including glycolysis, the pentose phosphate pathway, and galactose, fructose, mannose, and pyruvate metabolism. It also covers nucleotide metabolism, such as purine and pyrimidine, and amino acid metabolism, including the synthesis and metabolism of all 20 AAs and the synthesis and metabolism of other peptides, such as glutathione. Additionally, it includes lipid metabolism, such as fatty acid and glycerophospholipid synthesis, as well as the synthesis of vitamins and cofactors such as folate, riboflavin, and pantothenate (Figure 1A).

The GEM (Table 1) consists of 938 reactions, 638 metabolites, and 478 genes. All metabolic reactions, functions, and subsystems, as well as their confidence levels and available experimental evidence are provided in Supplementary Table S3. Metabolites and genes are provided as supplementary material as well (Supplementary Tables S4, S5, respectively). The model is also available in JSON and SBML formats (Supplementary Data Sheets 1, 2). Reaction confidence scores in iDR478 ranged from 1 to 3. Out of the 938 reactions, 499 (53%) were assigned a score of 2 as their inclusion evidence was only based on annotation or homology, owing to the sparsity of literature evidence available for metabolic reactions in this species. Eighty-four reactions were added based on homology searches following gap filling, as their associated enzymes were not initially annotated (Figure 1B). GPRs were defined for 98.8% of the metabolic reactions (575 out of 582), excluding the biomass, exchange, and transport reactions. Only seven non-gene–associated reactions were included in the model because they were computed as essential to allow simulation of growth but could not be associated with a gene in the model.

Many transporters in our model were not annotated as such in the reference E. faecium DO genome. As these were backed by little evidence, they were mostly (162 transporters) added to support model functionality and assigned a low confidence score of 1. The final model includes 201 transport reactions that represent uptake of nutrients and export of metabolic end products, as well as 154 exchange reactions that describe the exchange of metabolites with the environment. Among the 201 transport reactions, only 16 were assigned a confidence score of 3 and 23 were assigned a score of 2 (Figure 1B).

The network reconstruction provided insight into the metabolic capabilities of E. faecium, including the following points of interest: Despite the limited direct experimental evidence about fatty acid biosynthesis in E. faecium, and given the close genetic similarities between E. faecium and E. faecalis, insights from studies on E. faecalis mediated by the type II fatty acid synthase (FASII) system are included in the E. faecium model. Fatty acids, including mainly palmitic acid (hexadecanoic acid), stearic acid (octadecanoic acid), and myristic acid (tetradecanoic acid), are attached to ACP and used as substrates in phospholipid biosynthesis (Zou et al., 2024). Glycerol-3-phosphate is initially acylated by glycerol-3-phosphate acyltransferase (PlsY, HMPREF0351_11144) then by 1-acylglycerol-3-phosphate O-acyltransferase (PlsC, HMPREF0351_12429) to form phosphatidic acid (PA). PA is subsequently converted to CDP-diacylglycerol by CDP-diacylglycerol synthetase (HMPREF0351_11681). This intermediate serves as a precursor for phosphatidylglycerol (PG) via phosphatidylglycerophosphate intermediate through phosphatidylglycerol phosphate synthase (HMPREF0351_10392), and further condensed to cardiolipin (CL) through cardiolipin synthase (HMPREF0351_12025 or HMPREF0351_11068). Experimental studies in E. faecalis have validated the essential roles of PlsC and acyl-ACP in this pathway, demonstrating that disruption of these enzymes leads to defective phospholipid biosynthesis and growth impairment—phenotypes that can be rescued by restoring acyl-ACP availability (Zhu et al., 2019). Fatty acid and phospholipid biosynthesis are illustrated in the Escher map (Figure 2).

E. faecium also synthesizes wall teichoic acid (WTA) and lipoteichoic acid (LTA), which are critical components of its cell wall. These anionic polymers play essential roles in maintaining cell shape and division, making their representation important in GEMs for Gram-positive bacteria. The WTA biosynthesis reaction in the model reflects the experimentally defined repeating unit of E. faecium WTA, which consists of two N-acetyl galactosamine (GalNAc) residues linked to a glycerol-phosphate backbone (Bychowska et al., 2011; van der Es et al., 2016). These polymers are covalently attached to the peptidoglycan layer, contributing to cell wall integrity and function. Likewise, the LTA biosynthesis pathway was reconstructed based on the presence of annotated homologs of known LTA-related enzymes, enabling its incorporation into the GEM (Figure 2). LTA, a key component of the Gram-positive cell envelope, consists of polyglycerol phosphate chains anchored to the cytoplasmic membrane via a glycolipid. The inclusion of the LTA biosynthetic pathway enhances the model’s ability to capture the essential features of Gram-positive envelope biosynthesis and provides a basis for exploring cell wall-targeting antimicrobials (Paganelli et al., 2017).

E. faecium TX0016 possesses an incomplete tricarboxylic acid (TCA) cycle, a common feature among LABs. Genomic analysis confirmed the absence of key oxidative TCA enzymes, including 2-oxoglutarate dehydrogenase and succinyl-CoA synthetase, effectively blocking the full conversion of citrate to oxaloacetate. As a result, E. faecium cannot generate NADH or FADH 2 efficiently through the TCA cycle and must rely predominantly on glycolysis and substrate-level phosphorylation for ATP production.

Homology searches exposed the absence of many enzymes in the folate biosynthesis pathway, including dihydropteroate synthase (folP), 2-amino-4-hydroxy-6-hydroxymethyldihydropteridine diphosphokinase (folK), and 7,8-dihydroneopterin aldolase (folB), enzymes present in E. faecalis, which can synthesize folate in vivo. Based on these findings, we hypothesize that folate is essential for E. faecium TX0016 and needs to be supplemented in the medium—a hypothesis that is consistent with previous studies (Ramsey et al., 2014). Furthermore, E. faecium TX0016 was found to be auxotrophic for eight AAs, a finding that has been supported by leave-out experiments and visualized as Escher maps.

Biomass reaction

The biomass objective function of iDR478 was assembled from two published formulations for Gram-positive organisms, comprising 45 components. Coefficients for the 20 amino acids, the growth-associated energy requirement, the ribonucleotides and deoxyribonucleotides and the cofactor set (thiamine, coenzyme A, pyridoxal 5 phosphate, tetrahydrofolate, S-adenosylmethionine) were taken from the Bacillus subtilis formulation distributed with CarveMe (Machado et al., 2018). Coefficients for the components that distinguish the enterococcal cell envelope, peptidoglycan, wall teichoic acid, lipoteichoic acid, cardiolipin, and undecaprenyl pyrophosphate, and for biotin were taken from the E. faecalis V583 reconstruction (Veith et al., 2015). We incorporated the experimentally derived parameters from E. faecalis V583 to approximately simulate the energy requirements of E. faecium. We added a separate ATPM reaction to simulate basal maintenance energy demand independent of biomass formation. This reaction was constrained with a lower bound of 0.81 mmol ATP/gDW/h, based on the experimentally determined mATP value reported for E. faecalis (Veith et al., 2015).

No coefficient was measured in E. faecium either, and the resulting formulation should be regarded as a composite intended to give a biologically reasonable objective function for a Gram-positive lactic acid bacterium rather than as a determination of the exact biomass composition of this strain.

Sole carbon source utilization profiles in E. faecium DO

We used the Biolog microarray system to investigate E. faecium ’s utilization of 190 different carbon sources. For model simulations, we used the CDM-LAB medium components as the one used in vitro and allowed model exchanges for all carbon sources available tested in the Biolog phenotype microarrays. Out of the 190 carbon sources tested in Biolog, we studied 80 carbon sources (42.1%) that were either mapped in the metabolic model or supported growth in the Biolog experiments.

Although the CDM-LAB without glucose could not sustain growth experimentally, the model growing on an in silico CDM-LAB without glucose predicted a biomass yield of 0.078 h −1. We adopted this baseline as a pragmatic threshold for classifying carbon sources as supporting growth or not. This discrepancy reflects the lack of transcriptional and allosteric regulatory constraints, which can artificially route amino acids into central carbon metabolism to generate biomass in the absence of a primary sugar. To identify which amino acids were driving this false-positive growth, we systematically removed them in silico. We found that the model could only maintain this glucose-free baseline growth if L-aspartate, L-glutamate, and L-asparagine were present. When these three amino acids were collectively removed from the medium, omitting glucose finally rendered all growth infeasible (0 h −1), perfectly aligning with experimental observation.

Experimentally, the strain utilized 46 carbon sources in the Biolog system. Based on these positive results, Escher maps of the corresponding pathways were generated. This allowed us to identify missing or incomplete pathways and thus reveal the gaps. Missing pathways included those for dextrin, cyclodextrins, psicose, lactulose, amygdalin, and gentiobiose. New pathways were thus reconstructed from a combination of KEGG pathway data, genome annotation, homology-based evidence and literature data. We also completed several incomplete pathways in the model, such as 5-keto-D-gluconic acid, sorbitol, and palatinose, by adding either missing reactions or transport mechanisms or both. All added reactions were incorporated in the refined final model.

Together, these additions extended the metabolic scope of the model. Before they were included, the reconstruction reproduced 49 of the 80 phenotypes (61.2%), with 21 true positives, 28 true negatives, six false positives, and 25 false negatives. Nineteen substrates changed from false negative to true positive in the refined model (Figure 3). After refinement, and with a selected growing threshold of 0.078 h −1, corresponding to the biomass yield of the model in absence of glucose, the model successfully predicted growth on 40 of the 46 Biolog substrates scored as utilized, thus achieving an accuracy of 86.9%. The remaining six carbon sources, L-lyxose, D-arabinose, D-lactitol, sorbic acid, 2,3-butanedione, and L-leucine, failed to support growth in silico and thus remained false negatives, offering opportunity for discovery. Conversely, no activity was detected with 34 mapped carbon sources, of which 28 were consistent with the model predictions, representing true negatives. However, six carbon sources, sucrose, 2`-deoxyadenosine, D-gluconic acid, thymidine, D-raffinose, and L-malic acid, were false positives, as the model predicted growth that was not observed experimentally.

Overall, the agreement of the final curated iDR478 model with experimental carbon source utilization profiles was 85%, with 68 out of 80 carbon sources, 40 true positives, and 28 true negatives (Figure 3B). A detailed table representing the Biolog results and validations is provided in Supplementary Table S6.

Amino acid essentiality in E. faecium DO

Given the limited characterization of amino acid metabolism in E. faecium, amino acid leave-one-out experiments were conducted to assess the amino acid essentiality and validate the newly built GEM (Figure 4A). Based on experimental results, omission of L-methionine, L-histidine, L-leucine, L-isoleucine, L-valine, L-arginine, L-threonine, and L-tryptophan resulted in a marked decrease in the growth, with final OD 600 values consistently below 0.1 across all replicates. Thus, these eight AAs are classified as essential for growth. The strain initially failed to grow in the absence of L-lysine, L-phenylalanine, and L-tyrosine. However, after repeated re-inoculation, the cultures adapted and demonstrated significantly improved growth, indicating a potential for adaptation upon passaging (Figure 4B). An OD 600 above 0.21 was observed across all experiments in cultures where L-alanine, L-proline, L-glutamate, L-glutamine, L-aspartate, and L-asparagine were omitted. The omission of L-serine, L-glycine, and L-cysteine resulted in a weak growth (OD 600 valued above 0.1 but below 0.21).

With an OD 600 threshold of 0.1 selected based on sensitivity analysis (Supplementary Table S7), the model’s essentiality prediction fully agreed with leave-out experiments, identifying eight essential AAs, L-methionine, L-histidine, L-leucine, L-isoleucine, L-valine, L-arginine, L-threonine, and L-tryptophan and 12 non-essential AAs, L-alanine, L-proline, L-glutamate, L-glutamine, L-aspartate, L-asparagine, L-lysine, L-phenylalanine, and L-tyrosine, L-serine, L-glycine, and L-cysteine (Figure 4C).

Gene essentiality

Gene essentiality was assessed under two conditions: (i) with all exchange reactions open (representing unconstrained uptake) and (ii) in CDM.

With all the exchange reactions open, the model predicted 19 essential genes, including genes from nucleotide metabolism (n = 4) and amino acid metabolism (n = 1). The largest proportion of essential genes, however, was assigned to the metabolism of cofactors and vitamins (n = 12), such as pantothenate and CoA biosynthesis and nicotinate and nicotinamide metabolism. In CDM, 88 essential genes were predicted in the model, including genes from nucleotide metabolism (n = 17), lipid metabolism (n = 12), amino acid metabolism (n = 18), peptidoglycan and glycan biosynthesis (n = 16), metabolism of cofactors and vitamins (n = 13), and the terpenoid backbone pathway (n = 7). All 19 genes that are essential under unconstrained uptake are also essential in CDM-LAB (Figure 5A; Supplementary Figure S2).

Since no experimental essentiality data were available for E. faecium DO, the model’s predictions were compared with a Tn-seq study of MDR E. faecalis MMH594 (Gilmore et al., 2020) in a cross-species consistency assessment. Orthologs were identified by reciprocal best BLASTp hits between the E. faecium DO and E. faecalis and 352 of the 478 genes of the model have an ortholog, and 349 of these carry a fitness classification in Gilmore’s study. Genes classified in the experimental data as critical, important, potentially critical, or potentially important are counted as required for fitness, and those classified as non-essential as not required. The analysis was repeated, counting only the critical and important categories and excluding potentially important categories, under which agreement is marginally higher (Matthews correlation coefficient, MCC, of 0.31 vs. 0.27 in defined medium), so the inclusive mapping reported here is the more conservative.

In the defined medium, 85 of the 88 predicted essential genes have an ortholog in the E. faecalis experimental dataset. Among these, 53 genes (62.4%) agreed with the experimental classification, giving true positive results. Across the 349 assessable genes, the whole comparison indicates 181 true negatives, 32 false positives, and 83 false negatives, corresponding to a sensitivity of 0.39, a specificity of 0.85, a positive predictive value (PPV) of 0.624, an accuracy of 0.67, a balanced accuracy of 0.62, and an MCC of 0.27. Under unconstrained uptake, 16 of the 19 predicted essential genes (84.2%) agreed with the experimental classification, with 210 true negatives, three false positives, and 120 false negatives, a PPV of 0.842, a specificity of 0.986 but accuracy of 0.648, a balanced accuracy of 0.552, and an MCC of 0.223 (Figures 5B,C). The full comparison, including the ortholog assignment and the experimental classification for every gene, is provided in Supplementary Table S8.

The model also predicted 31 essential reactions, when all the exchanges are open, including reactions from the pentose phosphate pathway (n = 1, e.g., phosphoribosylpyrophosphate synthetase), nucleotide metabolism (n = 4, e.g., UMP kinase), and cofactor and vitamin metabolism (n = 9), such as pantothenate and CoA biosynthesis and vitamin B6 metabolism pathways. It also included transporters (n = 3) and exchange reactions (n = 9). In the CDM, 140 essential reactions were predicted, including 15 exchanges (eight essential AAs and six cofactors) and their transporters (n = 6), as well as reactions from carbohydrate metabolism (n = 4), amino acid metabolism (n = 14), metabolism of cofactors and vitamins (n = 11), nucleotide metabolism (n = 16), and the terpenoid backbone pathway (n = 8) with the largest proportion of essential reactions assigned to lipid metabolism (n = 43), followed by cell wall and glycan biosynthesis (n = 19). A complete list of predicted essential reactions is provided in Supplementary Table S9.

Comparison with available semi-automated E. faecium reconstructions

Comparison of model characteristics

iDR478 was compared with the two AGORA2 reconstructions of E. faecium, those of strain DO and strain TX1330 (Heinken et al., 2023). AGORA2 reconstructions had more genes, reactions and metabolites than iDR478. Blocked reactions represent 10.6% of all iDR478 reactions vs. 30.0 and 38.0% blocked reactions in the AGORA2 reconstructions. In iDR478, GPRs are assigned to 98.8% of metabolic reactions, but to 81.5 and 94.5% in AGORA2 DO and TX1330 reconstructions, respectively (Supplementary Table S10). Although iDR478 has more GPR assignments for metabolic reactions, fewer transport reactions are connected to genes in iDR478 (39 out of 201, or ~19.4%). On the other hand, the AGORA2 reconstructions have a much larger proportion of assigned GPRs (95 of 149 for DO, 63.7%, and 122 of 167, 73%, for TX1330). A more comprehensive model comparison is provided in Supplementary text.

Comparison of predicted amino acid auxotrophies

iDR478 precisely predicts eight auxotrophies, which were experimentally confirmed by AA leave-out experiments in strain DO. Having applied the same prediction methodology to AGORA2 reconstructions, we made auxotrophy/prototrophy assignments to only 17 out of 20 AAs. DO predicts 13 auxotrophic AAs, of which only seven agree with experimental results, while TX1330 predicts nine auxotrophic AAs, but we do not have experimental data for this strain. In AGORA2 models, three AAs cannot be assessed because of missing corresponding exchanges (aspartate, cysteine, and lysine in DO, and aspartate, cysteine and asparagine in TX1330, Supplementary Table S11).

Gene overlap with AGORA2 DO reconstruction

We examined the 273 reactions carried by genes present in the AGORA2 DO reconstruction but absent in iDR478. The reactions were assessed for inclusion (Supplementary Table S12). Among these reactions, 104 were excluded as they were blocked in AGORA2 reconstruction itself, and 14 because their genes had no equivalent in the complete DO genome. Another 65 genes were excluded because their corresponding reactions were already assigned to different genes in iDR478, mostly because the AGORA2 model assigns broader GPR annotations than the iDR478 model. For example, for ASPTA, the aspartate transaminase enzyme, iDR478 assigned the HMPREF0351_11895 gene, annotated in different databases as aspartate transaminase, while AGORA2 also assigned to that reaction two genes, EFAEDRAFT_RS02760 (equivalent to HMPREF0351_10558) and EFAEDRAFT_RS03865 (equivalent to HMPREF0351_12085), annotated as aminotransferase and GntR family transcriptional regulator, respectively.

We aimed to only add reactions that would connect to metabolites and increase network functionality rather than introduce additional blocked reactions. The remaining 90 reactions were then further assessed. Most of them were not appropriate for inclusion in our model: thirty-two are transport reactions, several of which are transporters for metals present in the AGORA2 biomass but not in iDR478, such as copper and cobalt. Reactions representing DNA replication, RNA transcription, and protein biosynthesis were also excluded because iDR478 is a metabolic reconstruction rather than a metabolic and expression (ME) model. Of the remainder, several were functional but redundant with routes already present in our model, and several would have required importing further reactions lacking GPRs. One reaction gave a clear improvement to our model, methylglyoxal synthase (MGSA) (HMPREF0351_12249), which converts dihydroxyacetone phosphate to methylglyoxal, a metabolite consumed by three reactions already in iDR478 but produced by none. Its addition allows flux through LALDO2x, LCARR, 12PPDRte, EX_12ppd__R_e, ALR2, ALR4x, LGTHL, GLYOX, and LDH_D, and reduces the number of blocked reactions without introducing new model errors.

We here present iDR478, a comprehensive manually curated genome-scale metabolic model of E. faecium DO, a major multidrug-resistant member of the ESKAPE pathogens. The model serves as an established platform for conducting systems-level analyses of an organism’s metabolism by integrating data from genome annotation, biochemical characterizations, and literature. Model curation integrated existing knowledge of E. faecium annotations and metabolic pathways with insights from our experimental analyses Reconstruction was guided and refined through experimental validation, using Biolog phenotypic microarrays for carbon source utilization and amino acid leave-out growth assays for amino acid auxotrophies. After refinement, the model fully predicted amino acid auxotrophies in agreement with leave-out experiments, and achieved 85% agreement with Biolog’s sole carbon source utilization.

The metabolic potential of E. faecium was evaluated across 80 carbon sources. Model validation for carbon source utilization was conducted via growth simulation in a CDM-LAB containing all AAs but lacking glucose, which exactly mirrors the conditions of experimental Biolog carbon utilization assays. The model’s ability to sustain growth in glucose-free conditions, despite the absence of experimental growth, was likely driven by false-positive utilization of L-aspartate, L-glutamate, and L-asparagine. Removing all three together with glucose abolished predicted growth, whereas removing any two did not. This finding suggests that the model catabolizes these AAs for energy generation, leading to false-positive growth predictions under the Biolog-like conditions. Mapping and pathways analysis revealed that these AAs are interconverted through transamination and deamidation reactions, ultimately supplying carbon to the oxaloacetate node and allowing the model to synthesize essential biomass precursors. This behavior is unlikely to occur under the experimental conditions, suggesting that the current reconstruction overestimates the capacity of these AAs pathways to sustain growth.

Out of the 80 carbon sources tested, 40 were predicted by the model as true positives and 28 as true negatives. However, the predictions of six substrates were false negative, as they failed to support carbon utilization in the model although the substrates were effectively utilized experimentally. These discrepancies likely result from gaps in our understanding of specific metabolic pathways. For instance, sorbic acid and D-lactitol utilization failed in the model because of incomplete knowledge of the metabolic mechanisms of these substrates, as recovered from literature searches. As for D-arabinose, it could not be utilized in silico because of the absence of arabinose isomerase in the genome, with neither literature nor genomic evidence supporting its presence. Similarly, L-lyxose and 2,3-butanedione (diacetyl) metabolism remain unsupported owing to missing key enzymes, such as ATP: L-xylulose 5-phosphotransferase and diacetyl reductase, respectively. There is neither genomic evidence nor literature support for the presence of these enzymes or any homolog thereof in any Enterococcus species. Another false negative result is L-leucine, transaminated to 4-methyl-2-oxopentanoate, which is usually converted into isovaleryl-CoA through an intermediate step, but this needs a specific enzyme complex (BCKDH) that also could not be found in the genome sequence.

Conversely, six carbon sources represented false positives predictions of the model, as they were predicted to support growth in silico but were not experimentally validated. Sucrose was predicted as a utilized carbon source in the model based on two annotated sucrase enzyme-encoding genes in the model (HMPREF0351_11791 or HMPREF0351_11800, EC: 3.2.1.26, BiGG ID SUCR), yet no respiration was observed in Biolog experiments. Literature evidence also confirmed the presence of universal sugar transport system components in E. faecium (Hallenbeck et al., 2024). Thus, this discordance may be due to inducible expression requirements that Biolog conditions did not fulfill, since sucrase has previously been studied as an inducible enzyme in Gram-positive bacteria, e.g., in a previous Staphylococcus study (Gering and Brückner, 1996). Likewise, raffinose, thymidine, and deoxyadenosine utilization in the model was based on nonspecific broad-spectrum enzymes, notably alpha-galactosidase (HMPREF0351_11597, EC: 3.2.1.22, BiGG ID: RAFGH) in the case of raffinose; pyrimidine-nucleoside phosphorylase (HMPREF0351_10038, EC: 2.4.2.4, BiGG ID: TMDPP) in the case of thymidine; and purine-nucleoside phosphorylase (PNP) (HMPREF0351_10046, EC: 2.4.2.1, BiGG ID: PUNP2) in the case of deoxyadenosine. These observations emphasize how annotation-based assumptions may overestimate metabolic capabilities. In the case of deoxyadenosine, PNP may possibly prefer to act on other nucleosides such as inosine (which was indeed utilized both in silico and experimentally). The inability of raffinose to support the strain growth in real life may be also interpretable by the non-functionality of the enzyme in all E. faecium strains, as studied before. This capability was shown to be linked to a specific gene cluster located on a megaplasmid, encoding two alpha galactosidase enzymes, AgaA and AgaB, present uniquely in some strains and not in others (Zhang et al., 2011). Similar issues were noted in gluconate and L-malic acid utilization. Gluconate utilization was also predicted by the model but not observed experimentally. The model includes key enzymes for its catabolism: gluconokinase (HMPREF0351_11707, EC 2.7.1.12, BiGG ID: GLUKA_1) and 6-phosphogluconate dehydrogenase (EC 1.1.1.44). A likely explanation is the absence of a gluconate transporter in E. faecium DO, as the gluconate transporter included in our model was supported by a low-confidence score of 1 (added primarily just to ensure model connectivity).

For L-malic acid, the false prediction likely depends on transport reactions included at a confidence score of 1 without a gene association (BiGG IDs: MALt2r and MALt2_2), as was the case for gluconate. Alternatively, adaptation or multiple passages might be necessary to activate the pathway, highlighting the difference between genomic potential and observable phenotype. Similar considerations can be applied to all the previously mentioned false positive cases, where transporters or relevant metabolic enzymes may be conditionally expressed or not induced under Biolog assay conditions. Thus, it is possible that, with further passaging, the strain might eventually activate the required pathways for alternative carbon source utilization.

The difference between genomic potential and observable phenotype in all false positive cases highlights a common limitation of GEMs. GEM predictions generally assume that all genetically encoded reactions are available, whereas in vivo enzyme activity is strictly regulated by gene expression and environmental context. Thus, incorporating transcriptomic and regulatory data could significantly improve predictive accuracy (Rychel et al., 2021; Catoiu et al., 2025).

Another potential source of discordance is the different endpoints of the computational and phenotypic approaches. While the computational approach uses a growth threshold (0.078) as an endpoint for the biomass function, Biolog microarrays use a redox-dye to primarily cellular monitor respiration/metabolic activity as a proxy for growth/survival.

The metabolic network was further validated against experimentally determined amino acid auxotrophies. Amino acid leave-out experiments demonstrated strong agreement between in silico predictions and experimental outcomes. Eight AAs, L-methionine, L-histidine, L-leucine, L-isoleucine, L-valine, L-arginine, L-threonine, and L-tryptophan, were deemed essential, as the average optical density for cultures lacking them remained below 0.1 across all experiments after 24 h. These findings align with the model predictions as true negative results. It is worth noting, however, that these amino acid-negative conditions were not subjected to repeated passaging. Therefore, the possibility of adaptation over time cannot be excluded. In contrast, the remaining 12 AAs, L-alanine, L-proline, L-glutamate, L-glutamine, L-aspartate, L-asparagine, L-lysine, L-phenylalanine, and L-tyrosine, L-serine, L-glycine, and L-cysteine were considered as nonessential, merely following the selected threshold of growth and the model predicted prototrophy.

Digging deeper in the experimental results, beyond just the sensitivity analysis approach and the selected threshold, we further analyzed the exact OD 600 values for each condition and assessed the growth patterns across experiments and biological triplicates. The six AAs L-alanine, L-proline, L-glutamate, L-glutamine, L-aspartate, and L-asparagine are surely classified as non-essential for growth, as cultures consistently exhibited high OD 600 values across all experiments, and the model also predicted prototrophy, representing true positive results. However, upon initial omission of any of L-lysine, L-phenylalanine, and L-tyrosine, the strain grew marginally. Interestingly, following repeated passaging, the final optical density was well above the threshold, suggesting that a regulatory or metabolic adaptation may have occurred during passaging, perhaps through the activation of an alternative pathway or the repurposing of a generalist enzyme. These cases ultimately align with model predictions of prototrophy. Of course we cannot rule out the possibility of the selection and expansion of a random mutant member of the bacterial population; however, testing for this hypothesis is beyond the scope of this study.

Experimental growth in the absence of L-serine, L-glycine, and L-cysteine was consistently marginal, and the model predicted nonessentiality. The uncertainty surrounding L-glycine and L-serine essentiality may be linked to the metabolic constraints of glycine hydroxymethyltransferase (HMPREF0351_12288, BiGG ID: GHMT2r, glyA), which catalyzes the reversible conversion of L-serine and tetrahydrofolate (THF) to glycine and 5,10-methylenetetrahydrofolate (MTHF). Since E. faecium cannot synthesize folate de novo (Ramsey et al., 2014) and the CDM contains only traces of folate, limited folate availability may impair GHMT2r activity, restricting growth. The model does not account for kinetic constraints or real-world nutrient limitations, which may contribute to this discrepancy. The other source of glycine in the model is aminopeptidase, which is a broad-spectrum nonspecific enzyme with no direct evidence of its presence to produce glycine specifically. Key enzymes in the phosphoglycerate pathway of serine biosynthesis, the enzyme phosphoserine aminotransferase (EC: 2.6.1.52) that converts 3-phosphohydroxypyruvate to 3-phosphoserine and the enzyme phosphoserine phosphatase (EC: 3.1.3.3) that converts 3-phosphoserine to L-serine, are also lacking, which also confirms that L-serine production primarily relies on the GHMT2r reaction as well.

Similar observations were made in GEM of E. faecalis V583, although their explanation evolved around transcriptomic regulation rather than folate insufficiency, given its ability to synthesize folate (Veith et al., 2015). In Lactococcus lactis, the glyA gene that codes for GHMT, and other folate metabolism genes is regulated by the PurR repressor, whose activity is influenced by hypoxanthine and PRPP. A PurR repressor protein was also identified in the genome of E. faecalis (purR, EF0058), displaying notable sequence similarity to homologs in Bacillus subtilis (55% identity) and Lactococcus lactis MG1363 (53% identity). Furthermore, a PurR binding motif closely resembling the consensus sequences characterized in B. subtilis and L. lactis was detected upstream of the glyA gene in E. faecalis, implying a potentially conserved PurR-dependent regulatory mechanism within this species. For L-cysteine, the observed growth in its absence was also minimally above the threshold, with an average OD 600 of approximately 0.25. One possible explanation is that cysteine biosynthesis requires hydrogen sulfide (H 2 S) as a sulfur donor for the cysteine synthase reaction (HMPREF0351_10472 or HMPREF0351_11313, EC: 2.5.1.47, BiGG ID: CYSS). However, only trace amounts of H 2 S are present under the ambient air conditions used in the experiments, conditions that differ from the ideal assumptions of the in silico model (there is an exchange reaction for H 2 S as in E. faecalis GEM too). Furthermore, this strain lacks key genes involved in the reduction of sulfate to H 2 S, specifically sat (sulfate adenylyltransferase) and cysH (phosphoadenosine phosphosulfate reductase). These genetic limitations, combined with the low environmental H 2 S availability, likely explain the limited growth observed in the absence of L-cysteine and the discordance from the predictions of the in silico model. These findings highlight the need for integrating more kinetic constraints into GEM simulations.

Additionally, essential gene predictions were performed in the model under two simulated conditions: rich medium with unrestricted exchanges and in CDM-defined growth. Owing to the lack of experimental essentiality data for E. faecium, we used Tn-seq data from E. faecalis (Gilmore et al., 2020) for cross-species consistency assessment. As a result, 84.2% of the predicted essential genes in case of the unrestricted exchanges and 62.4% in case of the CDM growth aligned with experimentally determined essential genes, representing the PPV. A full confusion matrix represents moderate agreement between the species regarding essentiality.

These differences likely reflect high species-specific variations between Enterococcus species, differences in growth conditions, as well as limitations of the model as the absence of regulatory and adaptive constraints. This highlights the need for generating E. faecium specific experimental gene essentiality datasets to improve model accuracy in this aspect. Experimental approaches such as targeted gene knockout or CRISPRi studies, as demonstrated in previous studies for different genes including acpH, treA, and lacL (de Maat et al., 2019; Chua and Collins, 2022) will be critical in validating gene essentiality and identifying potential antibiotic targets in the future. Predictions of gene essentiality also depend on the composition of the biomass reaction, since a gene is predicted essential only where its product is required, directly or indirectly, for a biomass component. As no experimentally determined biomass composition is available for E. faecium, the limitation of the formulation noted above apply here as well.

Having finalized and refined our manually curated iDR478 to the best of our capacity, we compared the model to existing semi-automated reconstructions in the AGORA2 database. We selected the reconstruction of the same strain (DO or TX0016), and we also included strain TX1330, whose model is publicly deposited. AGORA2 reconstructions had more genes, reactions and metabolites than iDR478 and the reconstructions differed radically: iDR478 was built up reaction by reaction, with a recorded confidence score and supporting evidence for each reaction, when possible whereas the AGORA2 reconstructions were automatically generated from KBase drafts by the DEMETER pipeline. As a result, blocked reactions represent 10.6% of all iDR478 reactions, as opposed to 30.0 and 38.0% blocked reactions in the AGORA2 reconstructions (Supplementary Table S10).

The discrepancy in the proportion of transporter GPR between iDR478 and the AGORA2 reconstructions is due to our conservative approach in assigning transporter GPRs when there is only genomic evidence, but not experimental evidence: We only assigned GPRs in iDR478 when the NCBI or KEGG annotation assigned a transporter to the substrate in question; otherwise, transport reactions were given a confidence score of 1 and left without a gene association. In the AGORA2 semi-automated reconstructions, substrate-specific transport reactions are frequently associated to genes annotated at the level of transporter family or of a shared component. Glucosamine transport, for example, is assigned to EFAEDRAFT_RS05720 (equivalent HMPREF0351_10962 in the complete genome), annotated as a PTS enzyme I component, which appears in 13 separate transport reactions in that model; glucose ABC transport is assigned to seven genes annotated as sugar or general carbohydrate ABC transporter permease, and D-alanine symport to EFAEDRAFT_RS05575 (HMPREF0351_10346) annotated as an APC family amino acid transporter.

To investigate the basis of the differences in predicted amino acid auxotrophies between the manually curated iDR478 reconstruction and the AGORA2 reconstructions, we traced each discrepancy to specific reactions absent from one reconstruction and present in the other. Several differences could be traced to biosynthetic reactions absent from the AGORA2 DO reconstruction but supported by the genome annotation of strain DO and in iDR478 model. AGORA2 DO lacked biosynthetic routes for alanine, asparagine, phenylalanine, serine, and tyrosine. For example, AGORA2 DO predicts alanine and asparagine auxotrophy, but we demonstrated experimentally that these amino acids are nonessential in strain DO. In iDR478, alanine can be synthesized through alanine transaminase (HMPREF0351_11177), while asparagine can be synthesized through the glutamine-dependent asparagine synthase (HMPREF0351_11718). Both genes are annotated in the DO genome but are not represented in the AGORA2 DO reconstruction. Similarly, AGORA2 DO reconstruction lacks chorismate mutase, so chorismate cannot be converted to prephenate. Although prephenate dehydratase, prephenate dehydrogenase, and the corresponding aminotransferases are present, the pathway cannot proceed toward phenylalanine and tyrosine without the chorismate mutase step, resulting in predicted auxotrophy for both amino acids. In contrast, iDR478 includes the annotated chorismate mutase (HMPREF0351_10156) and can produce both amino acids. These examples illustrate how differences in strain-specific gene annotation and reaction-level curation can have direct consequences for predicted nutritional requirements. This highlights the value of phenotype-refined, genome-based manual curation when developing strain-specific GEMs.

Limitations and future directions

It should be noted that the quality and completeness of a metabolic reconstruction are highly dependent on the availability of accurate organism-specific data. For E. faecium, this remains a significant limitation, as the organism is much less characterized than related species, and comprehensive metabolic experimental datasets are not widely available. This limitation is especially critical for E. faecium, given its multidrug resistance; in such case, accurate metabolic models can help prioritize essential pathways for future drug target screening. This limitation is also evident in existing reconstruction resources. In AGORA2, several curation steps that distinguish the DEMETER pipeline from automated draft had limited input for E. faecium and particularly strain TX0016. The growth-requirements table records no positive entries for any of the 188 E. faecium strains; the carbon source dataset includes only 14 substrates drawn from two literature sources and identical across all strains; and the DO genome is recorded as not analyzed in 34 of the 35 subsystems considered for comparative genomic re-annotation.

Other limitations include the following:

In carbon utilization experiments, using a predicted baseline as the classification threshold accommodates several observed disagreements instead of resolving them. Constraining amino acid catabolism would require flux measurements not available for this strain and remains for future refinement.

In gene essentiality analysis, opening every exchange permits uptake of metabolites that would not be simultaneously available to the cell, and the resulting predictions differ from those made in a defined medium, not because the medium is richer but because the constraint structure is removed. This matters for the comparison with the transposon data, which were generated in Mueller-Hinton broth, a rich but defined-composition medium. Neither of the conditions simulated here corresponds to it: CDM-LAB is chemically defined and considerably more restrictive, while unconstrained uptake is more permissive than any real medium. The comparison is therefore made across conditions as well as across species, and both differences constrain what can be concluded from it.

In conclusion, iDR478 represents a manually curated GEM that provides a foundation for understanding E. faecium’s metabolism and predicting nutrient dependencies and essential pathways, which can guide targeted Tn-seq or CRISPR-based screens and thus inform the design of targeted antimicrobial strategies in future studies.