Untargeted metabolomics reveals organism specific biomarkers of carbapenem resistance in Klebsiella pneumoniae and Escherichia coli

Abstract

Antimicrobial resistance (AMR) is one of the greatest global concerns due to the increase in the rate of AMR infections and the lack of development of antimicrobial agents to combat AMR. The development of resistance to carbapenems among common pathogenic bacteria, including Klebsiella pneumoniae and Escherichia coli, is particularly concerning because carbapenems are relied upon for the treatment of Gram-negative infections. Metabolomics offers an approach to identify biomarkers associated with metabolic mechanisms of carbapenem resistance. We analyzed 512 annotated metabolites from a total of 297 bacterial isolates spanning eight organisms. We discovered that the metabolome has greater variation between organisms compared with the resistance phenotype. As a result, we conducted organism-specific statistical analyses to discriminate carbapenem-resistant from carbapenem-sensitive isolates, and supervised learning models resulted in test set mean and standard deviation of area under the receiver operating characteristic curve of 0.822 ± 0.092 and 0.670 ± 0.110, and area under the precision-recall curve of 0.973 ± 0.016 and 0.653 ± 0.121, for Klebsiella pneumoniae and Escherichia coli, respectively. Feature selection using lasso logistic regression identified four metabolite biomarkers of carbapenem resistance in Klebsiella pneumoniae as 6-hydroxyphenazine-1-carboxamide (6-OH-Phz-1-Cam), N-Lactoyl tyrosine (N-Lac tyrosine), flavin adenine dinucleotide (FAD), and amalorin, and seven metabolite biomarkers in Escherichia coli as ibha#30, Cyclo(Leu-Lys), Asukamycin A-II, LPA 16:0, sphinganine, phytosphingosine, and PA 14:0/4:1;O. These unique metabolic signatures provide a vital foundation for exploring the emergence of AMR, warranting follow-up studies to clarify their role in carbapenem resistance and inform improved diagnostics and treatments for AMR.

Introduction

Antimicrobial resistance (AMR) occurs when the treatment of bacteria with antibiotics leads to the death or inhibition of susceptible bacteria, while naturally resistant bacteria survive and multiply¹. AMR has numerous factors influencing its development and proliferation: inappropriate prescriptions for patients, greater numbers of susceptible patients, such as the elderly and immunocompromised, persistent use and overuse of antimicrobials in clinical settings, antimicrobial usage in animals and plants, and antimicrobial release into wastewater². AMR has now emerged as one of the greatest global concerns due to the rapid growth of the AMR infection rate and the lack of new antimicrobial medications being introduced to combat this issue^3,4. AMR was associated with 550,000 deaths globally in 2021, and estimates suggest 8.22 million AMR-associated deaths throughout the world by 2050⁵. Additionally, carbapenem-resistant Enterobacteriaceae (CRE), strains of different bacterial organisms that have acquired resistance mechanisms to carbapenems, have been growing in prevalence worldwide. CRE infections result in longer hospital stays, increased healthcare costs, and higher mortality rates compared to carbapenem-susceptible Enterobacteriaceae (CSE) infections^6–8. Microbiology laboratories must identify CRE not just for patient care but also for infection control purposes. Moreover, accurate antimicrobial susceptibility testing (AST) against various antimicrobial agents is necessary to offer clinicians alternative treatment options⁸. In addition to AST, phenotypic carbapenemase detection methods, such as the Carbapenemase Nordmann-Poirel (Carba NP) test and carbapenem inactivation method (mCIM), as well as molecular testing, are commonly used to identify CRE in clinical laboratories^9,10. These methods are widely used in clinical laboratories and are generally rapid and cost-effective; however, they have limitations, including variability in sensitivity for certain carbapenemases¹¹. Moreover, CarbaNP and mCIM are not used in current clinical practice in the U.S. Additional limitations of these methods include the requirement of unique reagents and restricted detection capabilities⁸. Thus, the identification of the phenotype of CRE using alternative methods may offer future opportunities for the development of new methods of CRE detection that resolve these issues.

Metabolomics, the comprehensive analysis of metabolites in biological systems, offers a promising alternative for elucidating resistance-associated mechanisms used by CRE by providing a detailed metabolic profile of bacterial isolates. Recent studies have demonstrated the potential of metabolomics to identify molecules, biomarkers, and pathways associated with CRE. For example, Wen et al. measured 33 metabolites and found that transcriptional and translational activity-related pathways, as well as purine metabolism, were affected in cefepime/avibactam-resistant CRE Klebsiella pneumoniae¹². Similarly, Li et al. used gas chromatography-mass spectrometry (GC-MS)-based metabolomics to investigate the metabolic mechanisms of antibiotic resistance in clinically isolated carbapenem-resistant Acinetobacter baumannii (CRAB). Through the measurement of 109 metabolites, these researchers found that the inactivation of the pyruvate cycle and purine metabolism are characteristic features of CRAB¹³. Additional studies have been conducted, which demonstrated impressive discrimination between susceptible and drug-resistant samples. These studies reported the measurement of 59 metabolites in multidrug-resistant Pseudomonas aeruginosa¹⁴, which included resistance to carbapenems, 203 metabolites in carbapenem-resistant Escherichia coli¹⁵, and 273 metabolites in multidrug-resistant Escherichia coli¹⁶, which included resistance to carbapenems.

Recently, machine learning models have been utilized to identify metabolite biomarkers of carbapenem-resistant Klebsiella pneumoniae. Wen et al. constructed a random forest (RF) model to classify patients with carbapenem-resistant Klebsiella pneumoniae compared with patients negative for carbapenem-resistant Klebsiella pneumoniae and determined metabolites indicative of infection¹⁷. Limitations include a low sample size of bacteria cultured from 38 total patients and the use of a small number of metabolites, totaling 58 for model building. In another study, surface-enhanced Raman spectroscopy (SERS) from 8 samples with carbapenem-resistant Klebsiella pneumoniae and 7 carbapenem-susceptible samples was used to discriminate samples, achieving perfect separation using a convolutional neural network¹⁸. However, the very small sample size of 15 and the small peak number of 19 used in model training may limit the ability to identify a new patient infected with carbapenem-resistant Klebsiella pneumoniae. Rees et al. measured 169 metabolites of 60 Klebsiella pneumoniae, 37 Escherichia coli, and 20 Enterobacter cloacae samples and fit random forest, linear support vector machine, and partial least squares discriminant analysis models to discriminate between CRE and CSE samples¹⁹. This investigation reports a very high predictive ability of the metabolome on resistance phenotype. However, these results may be inflated due to the process used by the authors to identify metabolomic biomarkers that were included for evaluation of the validation data. The feature selection method used, which averaged the feature ranks of 100 splits of the data into a training and validation set, may have led to the selection of features that were already influenced by the validation set. This would lead to an overinflated ability to discriminate between CRE and CSE samples, as information from the validation set was already used in the initial model training. Consistent with the other studies, the sample sizes are small for Escherichia coli and Enterobacter cloacae, and a small number of metabolites were analyzed in the comparison of susceptible and resistant samples.

As these studies suggested the metabolome offers high discriminatory ability between CRE and CSE isolates, we hypothesized that comparing a larger number of CRE and CSE isolates across a greater number of organisms, in combination with the measurement of a larger number of metabolites, could identify robust resistance-associated metabolites. This study aims to utilize 512 metabolomic features to identify distinct biomarkers that differentiate CRE from CSE across eight organisms of Enterobacterales, including Klebsiella pneumoniae and Escherichia coli. We discovered that the metabolome has a much greater variation between organisms than between CRE and CSE, which led us to investigate resistance-associated metabolites by organism. We identified seven metabolites in Escherichia coli and four metabolites in Klebsiella pneumoniae that were associated with the development of carbapenem resistance. By identifying specific metabolic signatures associated with carbapenem resistance, this research could enhance diagnostic capabilities, inform treatment decisions, and advance our understanding of bacterial resistance mechanisms. As noted in prior metabolomics studies²⁰, future work evaluating the diagnostic potential of these metabolite biomarkers could support the development of rapid, cost-effective tools for detecting carbapenem-resistant bacteria in patients. Such advances may improve clinical outcomes and contribute to more effective containment of resistant infections.

Results

Metabolite annotation and data analysis pipeline developed for the discovery of carbapenem-resistant biomarkers

To determine metabolites predictive of carbapenem resistance, we measured untargeted metabolomics for 325 bacterial isolates, composed of 223 CRE (69%) and 102 carbapenem-susceptible (CSE) (31%) isolates. The bacterial isolates were made up of 17 different organisms (Supplementary Table 1), with CRE isolates mostly from Klebsiella pneumoniae and CSE isolates mostly from Escherichia coli. CRE isolates produced different carbapenemases or no carbapenemases at all (Supplementary Table 2). To ensure statistical testing between CRE and CSE was possible, we kept only organisms with three or more CRE and CSE isolates. This resulted in 307 bacterial isolates encompassing eight remaining organisms. We filtered all m/z values to keep only those with a corresponding mapping to the RefMet²¹ or Metabolomics Workbench²² databases with a tolerance less than 0.01, resulting in a total of 512 remaining metabolomic features. The data preprocessing step excluded ten additional isolates, lowering the overall total to 297, due to these isolates having greater than 10% missing values (Fig. 1a). The remaining isolates were evaluated using different statistical testing techniques (Fig. 1b).

Fig. 1. Data analysis pipeline from isolated measurements to statistical analysis.

a Outer circle: distribution of isolates among all organisms. Inner circle: proportion of carbapenem resistance after removal of organisms with less than three CRE and CSE isolates and isolates with greater than 10% missing metabolite values. b Experimental pipeline for the analysis of CRE and CSE isolates, including isolate selection, metabolite annotation, data pre-processing, and all statistical analyses.

Metabolic profiles show an organism-specific pattern rather than AMR phenotype

To determine if our data were separable by resistance type, we plotted the first two principal components of the principal component analysis (PCA), labeling the PCA plot by organism and phenotype (Fig. 2a). The resulting graph showed a clear separation into two groups that were not separated due to resistance phenotype. To determine if the source of this separation could be due to the acquisition of metabolomics from our isolates, which took place across three different days, we labeled the PCA by measurement date. The resulting graph showed a clear separation into two batches between April 21 and April 25/April 27 (Fig. 2b). To measure the batch effect, we conducted a univariate analysis of variance (ANOVA) for each metabolite between the two batches. ANOVA testing revealed 305/512 (59.6%) of metabolites were significantly different after FDR correction between batches (Supplementary Data 1). To remove batch effects due to the date of measurement, we performed ComBat batch correction²³. The PCA plot after batch correction was characterized by isolates from different organisms overlapping in the plot (Fig. 2c), and coloring by the two batches illustrated the removal of the batch effects demonstrated by overlapping isolates between batches (Fig. 2d). To statistically validate that the batch effect had been alleviated, we again performed univariate ANOVA testing on the data post-batch correction. ANOVA testing resulted in zero metabolites being significantly different after FDR correction between batches (Supplementary Data 1).

Fig. 2. Correction of batch effects due to the date of measurement.

a PCA plot before batch correction, colored by organism, with shapes representing the resistance type of each isolate demonstrates no grouping by organism. b PCA plot before batch, correction colored by measurement date, showing a batch is present due to different isolate measurement dates. c PCA plot after batch correction shows bacterial isolates are still not completely separable by organism. d PCA plot after batch correction shows overlap between isolates measured across different batches.

Uniform Manifold Approximation and Projection (UMAP) is characterized by the preservation of local relationships between data points, while PCA maximizes the variance in the data. Therefore, to visualize potential clustering by organism within the dataset, we plotted the first two components of a UMAP for all isolates (Fig. 3a). The UMAP visualization suggested the isolates do not cluster by resistance type, CRE versus CSE, but rather by organism. As a result, we conducted univariate two-way ANOVA testing to identify if there was a strong influence of resistance phenotype on organism (Supplementary Data 2). The univariate two-way ANOVA tests revealed a great majority of significantly differential metabolites associated with organism (m = 435), fewer metabolites significantly associated with resistance phenotype (m = 63), and very few metabolites significantly different corresponding to the interaction between organism and phenotype (m = 5) (Fig. 3b). Therefore, the organism is the source of the greatest variation in the metabolomic data. Also, since the interaction terms are not significant, the variation due to the organism is not dependent on the resistance phenotype.

Fig. 3. Metabolomics identifies a greater isolate variability between organisms than carbapenem resistance.

a UMAP plot illustrates clustering of isolates by organism with little overlap between different organisms. The UMAP is colored by organism, and the circles represent CRE isolates and triangles represent CSE isolates. b Bar chart showing the number of significantly differential metabolites resulting from univariate two-way ANOVA testing concludes a much greater number of metabolites are significantly different due to organism than due to resistance type. The number of metabolites significantly associated with the organism–resistance interaction was low, suggesting that resistance type contributes minimally to the variability driven by organism. c Bar chart showing the number of significantly differential metabolites from univariate one-vs.-rest linear models comparing metabolite abundances between organisms identifies many metabolites that are significantly different between organisms.

To determine how well we could separate bacterial organisms, we conducted univariate one-vs-rest statistical testing of linear models for each metabolite to determine differential metabolites for each organism (Fig. 3c). For each organism, we performed hierarchical clustering using only the metabolites that were determined to be differential and displayed the corresponding organisms alongside the heatmaps (Supplementary Fig. S1). Hierarchical clustering showed that the organism whose significant metabolites were input into the algorithm had high similarity with other isolates of the same organism compared to those of other organisms (Supplementary Fig. S1). As a result of these analyses, we concluded that the identification of carbapenem resistance biomarkers required us to first subset the data by organism before testing for differences between phenotypes.

Multiple resistance-associated biomarkers were discovered for Klebsiella pneumoniae with high predictive power and Escherichia coli with moderate predictive power

As our analyses showed greater variation between organisms than between resistance phenotypes in our isolates, we subset our data by organism to determine organism-specific biomarkers of carbapenem resistance. Due to variability in the number of isolates for each organism, we separated our biomarkers into low-sample-size and high-sample-size groups. All organisms were subset, and univariate firth logistic regression was performed, resulting in zero significantly differential metabolites between CRE and CSE for low sample size groups, four significantly different metabolites for Klebsiella pneumoniae, and nine significantly different metabolites for Escherichia coli (Supplementary Data 3). For high sample size groups, supervised learning was performed using a nested cross-validation framework. For 300 random iterations, the data was split into an 80% train set and a 20% test set. Within each train set, optimal hyperparameters for the LASSO logistic regression model were selected using 5-fold cross-validation. The final model for each random iteration, trained with these optimal hyperparameters on the entire train set, was then evaluated on the corresponding test set for that iteration to assess performance (Supplementary Fig. S2).

For Klebsiella pneumoniae, the average and standard deviation (Std. Dev.) of all 300 iterations produced an area under the receiver operating characteristic (AUROC) of 0.822 ± 0.092 (Fig. 4a) and an area under the precision recall curve (AUPRC) of 0.973 ± 0.016 (Fig. 4b). This illustrates that the metabolome can predict carbapenem resistance in Klebsiella pneumoniae with high predictive power. For Escherichia coli, the result of the 300 iterations produced an AUROC of 0.670 ± 0.110 (Fig. 4c) and an AUPRC of 0.653 ± 0.121 (Fig. 4d). These results demonstrate a moderate ability to discriminate between CRE and CSE for Escherichia coli. Additional metrics for these models include accuracy, sensitivity, specificity, positive predictive value, and negative predictive value (Supplementary Table 3).

Fig. 4. Identification of metabolite biomarkers of AMR.

a Mean receiver operating characteristic (ROC) curve for 300 random iterations of lasso logistic regression for Klebsiella pneumoniae, including mean and standard deviation (Std. Dev.) (shown as blue shade around mean ROC curve), area under the ROC (AUC) score. b Mean precision-recall (PR) curve for 300 random iterations of lasso logistic regression for Klebsiella pneumoniae, including mean and standard deviation (shown as blue shade around mean ROC curve), AUPRC score. c Mean ROC curve for 300 random iterations of lasso logistic regression for Escherichia coli, including mean and standard deviation (shown as blue shade around mean ROC curve), area under the ROC (AUC) score. d Mean precision-recall curve for 300 random iterations of lasso logistic regression for Escherichia coli, including mean and standard deviation (shown as blue shade around mean ROC curve), AUPRC score. e Boxplots comparing the distributions of four metabolite biomarkers identified by fitting a lasso logistic regression, optimized using cross-validation, on all the data for Klebsiella pneumoniae. f Boxplots comparing the distributions of seven metabolite biomarkers identified by fitting a lasso logistic regression, optimized using cross-validation, on all the data for Escherichia coli.

We next identified biomarkers of carbapenem resistance for each high sample size group using a multivariable approach to account for correlations among predictors. Independent 5-fold cross-validated LASSO models were applied to the Klebsiella pneumoniae and Escherichia coli datasets to determine the optimal hyperparameters for each organism. Using these parameters, a final model for each organism was subsequently fit. This procedure resulted in four remaining metabolites after a fit of the final model for Klebsiella pneumoniae (Fig. 4e) and seven metabolites for Escherichia coli (Fig. 4f). For m/z values with more than one possible annotation for Klebsiella pneumoniae (Supplementary Table 4) and Escherichia coli (Supplementary Table 5), we searched the literature for each annotation. For Escherichia coli, the annotations from Metabolomics Workbench and RefMet for the metabolite biomarker represented by 449.2049286 were attributable to metabolites that are not naturally produced by this organism. Therefore, we conducted an additional search on LIPID MAPS²⁴ with a tolerance of 0.01 for lipids with similar m/z values (Supplementary Table 5). The labeled m/z values represent metabolite annotations that had an association with AMR. For Klebsiella pneumoniae, the metabolite biomarkers included 6-hydroxyphenazine-1-carboxamide²⁵ (6-OH-Phz-1-Cam), N-Lactoyl tyrosine²⁶ (N-Lac tyrosine), flavin adenine dinucleotide²⁷ (FAD), and amalorin²⁸. The Escherichia coli metabolites found to be in accordance with AMR literature were labeled as ibha#30²⁹, Cyclo(Leu-Lys)³⁰, Asukamycin A-II³¹, LPA 16:0³², sphinganine³³, phytosphingosine³⁴, and PA 14:0/4:1;O³⁵. Three (6-OH-Phz-1-Cam, N-Lac tyrosine and amalorin) and four (Cyclo(Leu-Lys), LPA 16:0, sphinganine, and PA 14:0/4:1;O) of the metabolites were significant in the univariate analysis for Klebsiella pneumoniae and Escherichia coli, respectively (Supplementary Data 3). Our models’ ability to identify metabolomic biomarkers associated with AMR demonstrates the reliability of our results.

Discussion

In the present study, we used metabolomics to measure 4936 metabolites from 297 isolates of eight bacterial organisms to identify metabolomic biomarkers of carbapenem resistance. Annotation of metabolites using Metabolomics Workbench²² and RefMet²¹ resulted in a finalized set of 512 metabolites used for statistical analysis. We found that the metabolome has greater variation between organisms compared with the resistance phenotype. This conclusion is consistent with other analyses of the bacterial metabolome^19,36 and proteome³⁷. Organism-specific supervised learning models constructed for 300 random splits of the data into training and testing sets yielded mean test set AUROC of 0.822 and 0.670, and AUPRC of 0.973 and 0.653, for Klebsiella pneumoniae and Escherichia coli, respectively. Fitting one lasso model on the complete dataset to maximize the number of isolates used to determine biomarkers of carbapenem resistance identified four metabolites in Klebsiella pneumoniae and seven metabolites in Escherichia coli that were associated with AMR. These distinctive metabolic signatures serve as a crucial starting point for investigating AMR emergence, evolution, and persistence, and will enable the development of targeted biomarkers for rapid, precise, and effective AMR diagnosis and treatment.

As the discrimination of bacterial organisms is necessary to properly identify and eliminate pathogenic bacteria, our finding that there is greater variation in the metabolome between organisms than between resistance types within organisms is of importance. This suggests that to optimally identify the most discriminative resistance mechanisms of CRE using metabolomics, research must be focused separately for each organism. For example, Rees et al. constructed machine learning models to distinguish carbapenem-resistant from carbapenem-sensitive isolates for three bacterial organisms together, as well as each organism separately¹⁹. The predictive power for the model using all three bacterial organisms together was much lower than that of two of the organism-specific models and much higher than the remaining organism-specific model. The use of a pan-organism model reduces the confidence to identify robust metabolic resistance mechanisms used across all CRE. In addition, increasing interest in the role of the gut microbiome in disease has led to the utilization of metabolomics to discriminate between multiple bacterial organisms with high sensitivity³⁸. Recently, small molecular metabolites and lipids were used to identify 359 taxon-specific markers that were able to distinguish between numerous bacterial organisms³⁹. Therefore, as there is clear heterogeneity in the metabolome across bacterial organisms, we believe there are unique metabolic mechanisms developed by different organisms that influence carbapenem resistance, and researchers should focus on identifying these resistance mechanisms through organism-specific investigations. Importantly, this highlights that translating metabolic biomarkers into routine clinical practice may require organism-specific diagnostics, which could present a challenge given the large number of organisms known to exhibit carbapenem resistance mechanisms.

The biomarkers of AMR in Klebsiella pneumoniae were annotated as 6-OH-Phz-1-Cam, N-Lac tyrosine, FAD, and amalorin (Fig. 4e). 6-OH-Phz-1-Cam is a phenazine metabolite, a class of molecules in which many have antimicrobial properties⁴⁰ as a result of their ability to undergo redox cycling in the presence of various reducing agents and molecular oxygen. This process leads to the accumulation of toxic superoxide (O²⁻) and hydrogen peroxide (H₂O₂) and eventually to oxidative cell injury or death⁴¹. Specifically, 6-OH-Phz-1-Cam has been shown to exhibit antimicrobial activity against Staphylococcus aureus, Streptomyces albus, and Escherichia coli with minimum inhibitory concentrations of 64, 32, 128 μg/mL, respectively²⁵. We observed a higher abundance of 6-OH-Phz-1-Cam in carbapenem-resistant Klebsiella pneumoniae, suggesting the bacteria have developed an inhibitory mechanism against the reactive oxygen species (ROS) produced by 6-OH-Phz-1-Cam. N-Lac tyrosine was produced in high abundance as a result of the antimicrobial activity of plantaricin Q7 on Listeria monocytogenes biofilm²⁶. This metabolite is produced from lactate and tyrosine by reversed action of the protease cytosolic nonspecific dipeptidase 2⁴². Lactate has antimicrobial properties against carbapenem-resistant Klebsiella pneumoniae⁴³. As we reported N-Lac tyrosine to have a greater abundance in resistant isolates than susceptible isolates, this result may suggest a resistance-associated alteration is utilized by resistant Klebsiella pneumoniae such that lactate metabolism is upregulated to remove potentially harmful lactate, thus producing a large abundance of N-Lac tyrosine. FAD is one of the cofactors produced upon the breakdown of riboflavin and plays a role in redox homeostasis, protein folding, DNA repair, fatty acid β-oxidation, amino acid oxidation, and choline metabolism⁴⁴. FAD is also an essential coenzyme within oxidative-reduction processes and significantly impacts the regulation of energy metabolism⁴⁵. Our study found a lower abundance of FAD in carbapenem-resistant isolates, suggesting there could be an upregulation of energy metabolism associated with resistance that results in the increased metabolism of FAD. Amalorin belongs to the canthin-6-one alkaloids, a subclass of beta-carboline alkaloids (https://pubchem.ncbi.nlm.nih.gov/compound/11-Hydroxycanthin-6-one). Canthin-6-one and 8-hydroxy-canthin-6-one displayed minimum inhibitory concentrations of 8–64 µg/mL against multidrug-resistant and methicillin-resistant strains of Staphylococcus aureus²⁸. Impressively, the canthin-6-one alkaloids achieved lower minimum inhibitory concentrations than the control drug norfloxacin for the Staphylococcus aureus 1199B and Staphylococcus aureus XU212 strains. The greater abundance of amalorin in resistant Klebsiella pneumoniae isolates may indicate the presence of an unknown inhibition mechanism against canthin-6-one alkaloids. Multiple resistance mechanisms we report here are consistent with a recent publication that conducted a proteomics analysis of carbapenem-resistant and carbapenem-sensitive Klebsiella pneumoniae⁴⁶ isolates. The authors observed increases in energy production and conversion, carbohydrate transport and metabolism, and amino acid transport and metabolism. These metabolic alterations are consistent with an overabundance of N-Lac tyrosine due to increased production and metabolism of lactate and tyrosine, as well as lower FAD abundance due to consumption of FAD to meet increased energy demands. In conclusion, the Klebsiella pneumoniae metabolomic biomarkers suggest resistance-associated alterations in lactate and energy metabolism, as well as inhibitory mechanisms against canthin-6-one alkaloids and ROS produced by 6-OH-Phz-1-Cam (Fig. 5).

Fig. 5. Proposed metabolic carbapenem resistance mechanisms.

The proposed metabolic resistance mechanism towards carbapenems for each metabolite in the LASSO logistic regression models for Klebsiella pneumoniae and Escherichia coli. Each resistance mechanism is given its own box. Red arrows pointing up illustrate that the metabolite is overabundant in the carbapenem-resistant bacteria. Blue arrows pointing down illustrate that the metabolite is underabundant in the carbapenem-resistant bacteria. Bacteria illustration from NIAID NIH BioArt source⁶⁶.

Escherichia coli had corresponding AMR-associated biomarker annotations in either Metabolomics Workbench, RefMet, or LIPID MAPS databases of ibha#30, cyclo(Leu-Lys), asukamycin A-II, LPA 16:0, sphinganine, phytosphingosine, and PA 14:0/4:1;O (Fig. 4f). The full name for ibha#30 is (3R,16R)-16-{[3,6-dideoxy-4-O-(1H-indol-3-ylcarbonyl)-α-L-arabino-hexopyranosyl]oxy}-3-hydroxyheptadecanoic acid. This compound is characterized by a heptadecanoic acid (C17) backbone, an intermediate in lipid metabolism, which is metabolized to produce propionyl-CoA. It is suggested that propionyl-CoA metabolism is integral to the biogenesis of cell walls in Mycobacterium tuberculosis, which can enhance the bacteria’s ability to evade macrophage antimicrobial mechanisms²⁹. The underabundance of ibha#30 in carbapenem-resistant Escherichia coli suggests that increased β-oxidation may contribute to resistance by elevating propionyl-CoA production, which could enhance protection against immune responses and environmental stressors. Previous research has focused on cyclic dipeptides as antimicrobial and antifungal compounds^30,47,48. The cyclic dipeptide containing leucine and proline (cyclo(Leu-Pro)) has specifically demonstrated significant antibacterial activity. If cyclo(Leu-Lys) produces a similar effect to cyclo(Leu-Pro), the lower abundance of this cyclic dipeptide in resistant Escherichia coli isolates in our results suggests that non-ribosomal peptide synthetases, one of the major catalyzers responsible for producing dipeptides in bacteria⁴⁹, is deregulated. Asukamycin A-II is a manumycin-type metabolite originally discovered in cultures of Streptomyces nodosus subspecies asukaensis through the conversion of asukamycin A-I⁵⁰. Asukamycin has been demonstrated to inhibit the growth of Gram-positive bacteria⁵¹. In our experiment, the carbapenem-resistant isolates are characterized by a lower abundance of asukamycin A-II, proposing a metabolic alteration may exist in resistant isolates that reduces the production of asukamycin A-II. Phytosphingosine is synthesized from sphinganine⁵², and the different chemical structures of these sphingoid bases affect cell permeability⁵³. Previous literature has illustrated the ineffectiveness of both sphinganine^34,54 and phytosphingosine³⁴ on Escherichia coli. We report a pattern in which there is a greater abundance of sphinganine and phytosphingosine in resistant Escherichia coli isolates. This may indicate that resistant isolates not only are unaffected by these metabolites but also deregulate sphingolipid metabolism. Lastly, PA 14:0/4:1;O and LPA 16:0 are both lipids that are phosphatidic acids. LPA is primarily produced from PA through phospholipase A⁵⁵. As an anionic lipid, PA 14:0/4:1;O can produce an antimicrobial effect by targeting the lipid matrix of the cytoplasmic bacterial membrane, potentially destabilizing bacteria and decreasing their ability to develop resistance mechanisms³⁵. As a greater abundance of PA 14:0/4:1;O was observed in resistant Escherichia coli isolates, one possible mechanism of resistance is the ability of resistant isolates to inhibit phosphatidic acids from interacting with the bacterial membrane. A study conducting a transcriptomic analysis of carbapenem-resistant Escherichia coli after treatment with a carbapenem reported underexpressed genes involved in catabolizing fatty acyl-CoAs into acetyl-CoA units for the TCA cycle, namely fadE, fadB, and fadH⁵⁶. When underexpressed, accumulation of fatty acyl-CoAs can lead to an increased shift towards PA synthesis as substrate supply increases, consistent with our finding of an overabundance of PA 14:0/4:1;O. In addition, glpA, glpB, glpT, glpF, and glpK, genes involved in glycerol metabolism, were significantly overexpressed. As glycerol-3-P is a precursor for phospholipids, its increased metabolism may relate to our finding of overabundant PA 14:0/4:1;O. On the contrary, our experiments found LPA 16:0 to be underabundant in carbapenem-resistant Escherichia coli. Previous research has demonstrated that LPA enhances the susceptibility of several Pseudomonas aeruginosa strains to the anionic β-lactams ampicillin, piperacillin, and ceftazidime³². Therefore, the lower abundance of LPA 16:0 is consistent with a carbapenem resistance phenotype. Taken altogether, a mechanism of carbapenem resistance by Escherichia coli may be through the deregulation of phospholipase A, resulting in a greater abundance of PA species and a lower abundance of LPA species. The results of our statistical modeling for Escherichia coli suggest resistance-associated metabolic alterations include increased metabolism of compounds with a heptadecanoic acid backbone, reduced non-ribosomal peptide synthetases activity, decreased production of asukamycin A-II, downregulation of sphingolipid metabolism, inhibition of phosphatidic acids from interacting with the cell membrane, and decreased enzymatic activity of phospholipase A (Fig. 5).

The design of our study included many important considerations. The Escherichia coli Metabolome Database (ECMDB) contains over 3755 metabolites⁵⁷, providing an example that there are many metabolites in bacteria that may influence the development of a resistance phenotype. To our knowledge, the studies we cited that directly compared the metabolome of susceptible and drug-resistant isolates reported the analysis of a maximum of 273 annotated metabolites¹⁶, a total coverage less than 10% of the entire measured metabolome of one bacterial organism. Our inclusion of 512 metabolites provides an 87% increase in the number of metabolomic features analyzed for an association with carbapenem resistance. Additionally, differences between the metabolites within sample materials such as serum and feces within the same subject have been shown⁵⁸. Therefore, it is likely that bacterial cultures will produce different metabolite abundances compared to direct patient samples. We determined biomarkers of resistance through the measurement of bacterial cultures of susceptible and resistant isolates because this design provides a controlled environment to identify metabolites produced solely by the bacteria without confounding due to other factors, such as diet. To our knowledge, only one prior clinical metabolomics study, referenced in the Introduction¹², has analyzed patient serum samples to distinguish carbapenem-resistant from carbapenem-sensitive Klebsiella pneumoniae. Although that study identified 58 significantly altered metabolites associated with resistance status, none corresponded to the four metabolites identified in our model. This lack of overlap suggests that resistance-associated metabolic profiles may vary across studies, potentially due to differences in clinical specimen type, cohort characteristics, sample handling, and metabolomics methodologies.

The present study has limitations that must be considered. First, biological functional validation was not conducted in this study. Future studies should include targeted metabolite supplementation and gene knockdown experiments to clarify the causal roles of the identified biomarkers in carbapenem resistance. Supplementing bacterial cultures with metabolites that were found to be depleted in resistant isolates, or reducing the expression of genes associated with overabundant metabolites in resistant isolates, would allow for an assessment of whether these metabolite biomarkers restore or enhance drug susceptibility. Following these manipulations, antibiotic response assays—such as minimum inhibitory concentration measurements or growth curve analyses—could be used to determine whether the altered metabolite abundances directly modulate the bacterial response to treatment. Together, these functional studies would help distinguish whether the proposed biomarkers are merely correlated with resistance phenotypes or are mechanistically involved in mediating them. We also have a small proportion of CSE isolates for Klebsiella pneumoniae. This sampling of Klebsiella pneumoniae isolates does not represent the proportions of the resistance phenotype in the US, which is approximately 3% CRE and 97% CSE, and influenced our machine learning design to use cross-validation on all data instead of using a held-out testing set for evaluation. In addition, further validation of metabolite identification using MS/MS is necessary for the establishment of exact molecular identification. Also, the bacterial isolates used in this study were from the United States. It is possible that isolates from different geographical locations throughout the world may have different metabolic mechanisms of carbapenem resistance due to environmental factors. Follow-up studies must be conducted with isolates from a wide variety of geographic locations to understand the universality of these metabolite biomarkers. Another limitation is the lack of stratification by carbapenemase production. CRE encompasses heterogeneous resistance mechanisms, and future studies integrating genomic or phenotypic characterization of carbapenemase status may reveal more specific metabolomic signatures associated with distinct resistance pathways. Lastly, the final predictive models are evaluated on one dataset due to the constraint we imposed, in which our biomarker detection is performed on the entire dataset to maintain the optimal power. It is important that future studies confirm the predictive power of our biomarkers on new isolates.

Materials and methods

Bacterial isolates/reagents and materials

All isolates were obtained from the CDC & FDA Antibiotic Resistance Isolate Bank (AR BANK) (Atlanta (GA): CDC 2022). The isolates had minimum inhibitory concentrations obtained by broth dilution available for reference on the AR Bank Webpage. AR Bank isolates were utilized from the following panels: Aminoglycoside/Tetracycline Resistance (ATR), Enterobacterales Carbapenem Breakpoint (BIT), Enterobacterales Carbapenemase Diversity (CRE), Gram Negative Carbapenemase Detection (CarbaNP), and Plazomicin (PLZ). All isolates were plated to tryptic soy agar with 5% sheep blood (BD BBL, Sparks, Maryland, USA) and were incubated at 37 °C in 5% CO₂ for 18–24 h. Isolates were subcultured to F2 on tryptic soy agar with 5% sheep blood (BD BBL, Sparks, Maryland, USA) for day of use testing. All isolates were maintained and cultured under standardized laboratory conditions to minimize variability related to environmental factors such as host milieu or nutrient sources.

Isolate preparation

Bacterial isolates were collected and subsequently washed using Phosphate-buffered saline. Following this, the isolates were vortexed and lysed via ultrasonication in the presence of ice. Lastly, the isolates were processed using a protein precipitation method. Briefly, 800 µL extraction buffer solution (Methanol: Ethanol: H₂O = 4:4:2) and 40 µL internal standard (L-carnitine:HCl, O-Decanoyl, D3, 500 ng/mL) were pipetted into a centrifuge tube with 100 µL bacterial lysis solution. Then, the mixture was vortexed for 1 min, followed by centrifugation at 14,000 × g for 10 min at 4 °C. Then, 800 µL of supernatant was transferred into a new centrifuge tube and dried under vacuum conditions. The residues were reconstituted into 100 µL mobile phase, followed by high-speed centrifugation before analysis. The quality control (QC) isolates were prepared by mixing all bacterial isolates: 10 µL was taken from each isolate and added into a centrifuge tube, vortexed and aliquoted into 100 µL each.

Instrumental conditions

HPLC-Q-TOF-MS/MS (Waters, Xevo G2-XS QTof quadrupole time-of-flight Mass Spectrometry, USA) was applied as a metabolite separation and detection platform to investigate the metabolite differences between the carbapenem-resistant and carbapenem-sensitive isolates. The data was collected under the positive ion modes (ESI⁺) of mass spectrometry.

HPLC conditions: Acquity UPLC BEH C₁₈ column (2.1 mm × 100 mm, 1.7 µm) at 45 °C with 5 µL injection volume; mobile phase was acetonitrile and water with 0.1% formic acid. The gradient was set as follows: 5% acetonitrile for 0–1 min, 35% acetonitrile for 1–8 min, 98% acetonitrile for 8–18 min, and 5% acetonitrile for 18–20 min. The flow rate was maintained at 0.4 mL/min.

MS conditions: Electrospray ion source was detected using positive ion mode; high purity N₂-assisted spray ionization and solvent removal were used; the flow rate of cone gas and desolvation gas were 20 L/h and 800 L/h, respectively. The mass scanning range was 50–1200 m/z, and the drying temperature was 120 °C. In ESI positive mode, the capillary voltage was 3000 V. Quality control isolates (QC isolates) were analyzed three times at the beginning of the run and injected once after every 30 injections of the random sequenced isolates.

Alignment and peak picking

Data pre-processing was performed using Progenesis QI (Nonlinear Dynamics, Waters, Newcastle upon Tyne, UK). LC-MS data acquired were imported to Progenesis QI for alignment (Aligned with QC isolates) and peak picking. The peak picking was using the default parameters, including data filter strength (1.0) and peak picking sensitivity (automatic, default strength). Then, the raw data, including all picked features, was exported as a CSV file.

Metabolite selection

To ensure our analysis would be able to produce biologically interpretable metabolite biomarkers, we first performed a mapping between each m/z value output from the mass spectrometry experiment with Metabolomics Workbench²² and Reference Set of Metabolite Names (RefMet)²¹. We selected only m/z values that had a corresponding annotation to either of these databases with a mapped m/z value less than 0.01, giving us confidence that the annotations were truly representative of the measured m/z value. This filtering resulted in 512 total metabolite features, which were further processed. Given this list of metabolite annotations within 0.01 m/z from our measured value, we manually searched through the literature for each possible match and assigned the final metabolite annotation reported within the manuscript to those with a reported association with antimicrobial resistance. For metabolite biomarker 449.2049286, an additional search on LIPID MAPS²⁴ was conducted with a tolerance of 0.01 for lipids with similar m/z values to find annotations for this compound.

Data preprocessing

For all data preprocessing, we used R version 4.2.1 and the Maplet⁵⁹ R package version 1.1.2. First, we kept only organisms that had three or more CRE and CSE samples measured. For these samples, we set all values equal to zero to missing. We used the mt_pre_filter_missingness function to remove metabolites missing greater than 20% of sample measurements and samples missing greater than 10% of metabolite measurements. We then performed quotient normalization using the mt_pre_norm_quot function, log2 transformed the data using the mt_pre_trans_log function, and z-scored the data using the mt_pre_trans_scale function. Visual inspection of metabolite distributions determined that most were approximately normally distributed. To remove outlier metabolites, we used the mt_pre_outlier_to_na function, which assumes each metabolite follows a normal distribution, and takes into consideration the sample size and alpha value, which we set to 0.05, to set metabolite values outside the range of a particular percentage of points to missing. We impute the remaining missing values using K nearest neighbors (KNN) imputation using the mt_pre_impute_knn function. Finally, we performed batch correction between the two batches (Fig. 2) using the ComBat method²³. ComBat was used due to its widespread use in metabolomics research^60–62 and its suitability as a batch correction method compared to other methods^63–65. As with any batch correction method, we cannot be certain that the use of ComBat did not remove biologically relevant information.

Statistical analysis

Data visualizations were performed using both R version 4.2.1 and Python version 3.11. In R, the function prcomp() was used to create the PCA plots for batch correction visualization (Fig. 2). The remaining data visualization was performed using Python. UMAPs were constructed using the umap-learn package version 0.5.4. UMAPs and box plots were plotted using seaborn version 0.12.2. PCA was calculated using sklearn version 1.4.2.

Univariate ANOVA testing to compare metabolites between batches was conducted using the aov() function in R. Univariate two-way ANOVA was performed using Python’s statsmodels package version 0.14.0 to test if there was a significant interaction between organisms and phenotype of all bacteria isolates for each metabolite. To determine metabolomic differences between organisms, we performed univariate one-vs-rest ordinary least squares comparing every metabolite for each organism to the rest of the organisms. The conclusion that 435 significant metabolite differences between organisms resulted, and only five significant interactions between organism and phenotype were found, suggested it was necessary to split the data by organism to properly determine metabolomic biomarkers of resistance.

Therefore, we separated the isolates by organism. We further separated the organisms into those with low sample sizes (Citrobacter freundii, Enterobacter cloacae, Klebsiella aerogenes, Klebsiella oxytoca, Proteus mirabilis, and Serratia marcescens) and those with high sample sizes (Klebsiella pneumoniae and Escherichia coli). For all organisms, we conducted univariate firth logistic regression on all isolates to determine differential metabolites between CRE and CSE isolates using the logistf package in R. Firth logistic regression was used because this test is suitable for models fit to a low number of samples. For organisms with high sample sizes, we split the isolates into 300 random iterations of an 80% train and a 20% test set. We then used the LogisticRegressionCV() function in Python to conduct 5-fold cross-validation on the 80% train data to identify the optimal regularization strength. For each train/test split, a final model with the optimal regularization strength was fitted to the full 80% train set, and its performance was evaluated on the remaining 20% test set. To attempt to handle class imbalance between carbapenem-resistant and sensitive isolates, we set the class_weight parameter to balanced. We report the mean and standard deviation of the AUROC, AUPRC, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value to measure the model’s ability to discriminate CRE and CSE isolates across all 300 iterations. Taken together, these metrics demonstrate the model’s ability to discriminate classes overall, as well as provide insight into the class-specific predictive power of the model. All metrics must be investigated to understand the impact of class imbalance between CRE and CSE isolates. For identifying metabolite biomarkers for organisms with high sample sizes, a 5-fold LogisticRegressionCV() model was fit to all samples belonging to the organism. This model begins with 5-fold cross-validation to identify the optimal hyperparameters for the dataset, then fits a model with the optimal hyperparameters to the entire dataset. All analyses for organisms with high sample sizes were performed using sklearn version 1.4.2.

ANOVA, two-way ANOVA, ordinary least squares, and firth logistic regression were all corrected for multiple testing using the false discovery rate with an adjusted significance threshold of 0.05.

Author contributions

Z.Z. initiated and designed the study. Y.L., J.M., and Y.H. handled the isolates and measured the metabolomics of the isolates. N.B., Y.L., and V.S. conducted all analyses of these data. Z.Z., M.J.S., S.Y., F.W., J.K., V.S., and H.S.Y. supervised the research. N.B., Y.L., J.M., and V.S. wrote the manuscript. All authors contributed to the editing of the manuscript and approved the manuscript.

Data availability

All data is publicly available at (https://github.com/Nicholas-Bartelo/Zhao-Lab-AMR-Paper/tree/main).

Code availability

All code is publicly available at (https://github.com/Nicholas-Bartelo/Zhao-Lab-AMR-Paper/tree/main).

Competing interests

J.K. is co-founder and holds equity in iollo and ExactRx (Celeste Health, Inc) and is an advisor to Everfur (Strand Health, Inc) but declares no non-financial competing interests. M.J.S. previously acted as a paid consultant for Shionogi and has received research funding from Merck, SNIPRBiome, bioMéreiux, and Selux Diagnostics, but declares no non-financial competing interests. Z.Z. previously acted as a paid consultant/advisor for ET Healthcare, Radiometer, Intelligenome, Roche and Siemens and has received sponsored research/seed instrument from Roche, Siemens, ET Healthcare, Gator Bio, Novartis, Polymedco, and Waters, but declares no non-financial competing interests. H.S.Y., F.W., N.B., Y.L., Y.H., V.S., and J.M. declare no financial or non-financial competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41540-026-00747-7.