Single-bacterium metabolome revealing heterogeneous cellular states in bacterial populations

Abstract

Bacterial metabolism is a complex system of interwoven pathways coordinated by an intricate, multilayered regulatory network. Dissecting the metabolism at single bacterial level is highly challenging. Herein, we greatly increase the sensitivity of the bacterial measurement with our home-made single-bacterium metabolome (SinBactM) mass spectrometry platform, which combines micro-extraction with induced nanoelectrospray ionization mass spectrometry, for mapping metabolites of the primary bacterium. The universal applicability of the SinBactM platform is demonstrated by successfully discriminating distinct metabolic profiles across different bacterial species. By measuring the bacterial uptake of antibiotics and observing the dynamic metabolic changes in response to external stimuli, the accuracy and reliability of SinBactM platform are further validated. We apply SinBactM to characterize clinical heteroresistant K. pneumoniae strains metabolic alteration under antibiotic perturbation. Several different subpopulations resulting from antibiotic perturbation are classified based on their metabolome, while pseudotemporal ordering reveals a trajectory of metabolic change across subpopulations. Thus, SinBactM platform paves the way for profiling metabolomics of individual bacterial cells, which might provide an exciting dimension to a thorough understanding of antibiotic-resistant evolution for in-depth interrogation of resistant mechanisms.

Here the authors present a single-bacterium metabolomics platform that enables sensitive metabolic profiling of individual bacteria, revealing metabolic heterogeneity in heteroresistant K. pneumoniae strains and providing insights into bacterial survival and resistance.

Introduction

Bacterial metabolism is essential to bacterial life, and it fuels all cellular activities with building blocks and energy, which is a large and complex connected network^1–5. Furthermore, within bacterial colonies, metabolite concentrations are highly heterogeneous and closely correlated with growth and stress response states, such as bacterial heteroresistance when exposed to antibiotics^6–10. Thus, it is necessary to analyze the intertwined metabolism comprehensively at single-bacterium level for a more thorough understanding of bacteria.

Mass spectrometry has served as the gold standard for bacterial metabolic analysis at population level, in which the information on bacterium-to-bacterium heterogeneity is masked^3,11–13. In recent years, single-cell RNA-seq has pioneered the study of metabolic heterogeneity in individual bacterial cells based on metabolite-related mRNA levels^14–20. However, inherent challenges, such as low mRNA content and the existence of cell walls, result in a low detectable amount of metabolism-related RNAs, which are insufficient for metabolite analysis^21–24. Due to excellent sensitivity, fluorescence and Raman spectroscopy are widely used for targeted/known single-bacterium metabolic analysis. Similarly, fluorescence-based methods can only measure a limited number of metabolites at a time with unavoidable band overlap and background interference^25–28. Meanwhile, establishing reliable correlations between specific metabolites and Raman spectra remains challenging²⁸. Consequently, a comprehensive view is still lacking for single-bacterium metabolism.

Despite the recognized complexity and importance of bacterial metabolic states in influencing their survival, virulence, and response to antibiotics^{6–9,29–31}, metabolic analysis at single-bacterium level remains a significant challenge. Herein, we report a single-bacterium metabolome (SinBactM) platform, by coupling micro-extraction with induced nanoelectrospray ionization mass spectrometry (InESI-MS), for the metabolomic profiling at single-bacterium level. Such a combination enables sensitive and untargeted metabolic profiling of individual bacteria. To establish its broad applicability, SinBactM successfully resolved species-specific metabolic profiles across diverse bacteria. We further applied SinBactM platform to characterize metabolic heterogeneity of clinical colistin heteroresistant Klebsiella pneumoniae populations. The complex and diverse bacterial metabolism was observed in different subpopulations rather than a single or uniform metabolic response in all bacteria. Notably, one rare subpopulation exhibited a metabolic state closely resembling that of resistant strains, and pseudotemporal ordering revealed a trajectory of metabolic change across subpopulations. While the functional implications of these metabolic states require further experimental validation, the SinBactM platform offers a culture-independent platform for mapping metabolic heterogeneity and tracking phenotypic adaptation under antibiotic pressure, thereby opening avenues for understanding bacterial survival strategies.

Results

Development of SinBactM platform for fL-level single bacterium metabolome analysis

The schematic of SinBactM platform is shown in Fig. 1a. Briefly, individual bacterial cell was immobilized on poly(L-lysine)-coated glass by electrostatic adsorption, and placed under a microscope. A pulled glass capillary containing ~25 pL extraction solvent was placed over the bacteria through a three-dimensional manipulation platform (Supplementary Fig. 1-2). Then, extraction solvent was extruded to form a droplet over the individual bacterial cell for metabolites extraction. After approximately 1 s of extraction, the droplet was sucked back by the syringe and subsequently analyzed by InESI-MS. Along with the volatilization of the solution in the extraction process, the metabolites were finally concentrated in an extraction solution of 4.8–5.8 pL (Supplementary Figs. 1-2). In this way, the metabolites from individual bacterial cells were extracted by picoliter droplet micro-extraction, enabling metabolites enrichment in a small volume during the extraction process. Then SinBactM platform was performed for individual B. subtilis metabolites analysis (Fig. 1b, Supplementary Fig. 8a). The average signal lasted about 53 scans for individual B. subtilis analysis by InESI, which was 40-fold longer than nanoESI (Supplementary Fig. 8b, e). For m/z ranging from 60 to 300, over 500 detected features met the criteria with SNR > 3, signal intensity greater than 10³, occurring at a frequency of > 20% in all B. subtilis events. Of the detected features, 141 ions, such as amino acids (proline, arginine and ornithine) and polyamines (agmatine, spermidine and putrescine), were further putatively annotated by matching the accurate mass with data from the HMDB (Supplementary Data 1).

Fig. 1. Single-bacterium metabolites analysis by SinBactM platform.

a Flow chart of single bacterial location, extraction and MS measurement. InESI-MS: induced nanoelectrospray ionization mass spectrometry. b Mass spectra of individual B. subtilis. c Heatmap of four different single-bacterium metabolite profiles. A total of 285 bacteria from two batches were measured, containing 90 E. coli, 74 P. aeruginosa, 50 S. aureus, 71 B. subtilis. Color indicates z scores of each metabolite ions in each single bacterium. Yellow represents high expression of the metabolites. Blue represents low expression of the metabolites. Each row represents a metabolite and each column represents a bacterial cell. d The average number of metabolites detected in individual bacterium for four different bacterial species. n = 90, 74, 50 and 71 for E. coli, P. aeruginosa, S. aureus and B. subtilis, respectively. The half violin plot displays the mean ± SD of each group as well as the data points. e t-SNE analysis of four different bacterial species. E.coli and its replicate (E.C.1 and E.C.2, red points, n = 33 and 57, respectively), P. aeruginosa and its replicate (P.A.1 and P.A.2, yellow points, n = 26 and 48, respectively), S. aureus and its replicate (S.A.1 and S.A.2, blue points, n = 30 and 20, respectively), B. subtilis and its replicate (B.S.1 and B.S.2, green points, n = 20 and 51, respectively). Data were acquired from two batches of experiments. E.C. represents E. coli, P.A. represents P. aeruginosa, S.A. represents S. aureus, B.S. represents B. subtilis. Source data are provided as a Source Data file.

In order to test the broad applicability of SinBactM platform across bacterial species and evaluate its ability to differentiate multiple species, we applied SinBactM platform to the Gram-negative E. coli, P. aeruginosa and the Gram-positive S. aureus, B. subtilis. Different metabolites were successfully detected in individual bacteria of different types. As expected, metabolic profiles were found to be similar within the same bacterial type and significantly different among different bacterial types (Fig. 1c, Supplementary Figs. 9-10, Supplementary Fig. 43). An average of 141, 159, 138 and 126 ions were further putatively annotated by matching the accurate mass with the HMDB in individual B. subtilis, E. coli, P. aeruginosa and S. aureus, respectively, which covered most groups of important functional metabolites, such as organic acids and derivatives, lipids and lipid-like molecules, organic oxygen compounds and organic nitrogen compounds (Fig. 1d, Supplementary Fig. 11, Supplementary Data 2-4).

To visualize the bacterial metabolic profiles clearly, individual bacterial cells were clustered based on metabolic fingerprints using Seurat. As shown in Fig. 1e, the separation of E. coli (n = 90), P. aeruginosa (n = 74), S. aureus (n = 50) and B. subtilis (n = 71) was clearly observed, demonstrating distinct metabolic profiles among four bacterial populations. For instance, carnitine and spermidine were significantly accumulated in B. subtilis. Acetylspermidine and agmatine were enriched in E. coli. Dopa and phenylalanine were enriched in P. aeruginosa. Aspartic acid and glutamic acid were accumulated in S. aureus (Supplementary Fig. 12). No obvious batch effects were observed among different batches (Supplementary Fig. 13-16). In addition, within-species metabolic subtypes and potential key underlying metabolites for each subtype were also identified for E. coli, P. aeruginosa, S. aureus and B. subtilis, respectively (Supplementary Figs. 44-47). Thus, we successfully characterized the metabolic profiles across different bacterial species, confirming the universal applicability of SinBactM platform.

SinBactM reveals distinct metabolomic between heteroresistant K. pneumoniae and resistant K. pneumoniae strain

Here, we applied SinBactM platform to explore the distinct metabolomic profiles of the clinical isolate heteroresistant K. pneumoniae (HR-S) strain and resistant K. pneumoniae (R) strains. HR-S strain was a colistin heteroresistant strain classified as susceptible confirmed by population analysis profiling (PAP) methods, and R strain was isolated from the above HR-S strain (Supplementary Fig. 24)³².

The metabolic profiling of the HR-S and R strains was shown in Fig. 2a, about 150 ions were putatively annotated by matching the accurate mass to the HMDB in clinical K. pneumoniae strain (Supplementary Fig. 26a, Supplementary Data 5). The complex metabolomic information between the HR-S and R strains was further visualized by the t-SNE algorithm in the two-dimensional plane (Fig. 2b). Metabolomic-based separation of HR-S and R strains was clearly observed, indicating that metabolites in HR-S and R strains were substantially different (Supplementary Fig. 48). For instance, lysine and cadaverine were enriched in HR-S strain, while glutamine and glutamic acid were enriched in R strain, which was consistent with previous reports that glutamate metabolism was affected in multidrug-resistant pathogenic bacteria (Supplementary Fig. 48). For further verification, we performed a supervised learning k-nearest neighbor (KNN) model based on single-bacterium metabolomic data, and found that the model had an excellent ability to distinguish between HR-S and R strains (Supplementary Fig. 49).

Fig. 2. Metabolomic profiling of heteroresistant K. pneumoniae strain and resistant K. pneumoniae strain.

a Heatmap of characteristic metabolites and their relative intensities in each HR-S (n = 159) and R (n = 223). Color indicates z scores of metabolites in each sample. Red represents high expression of the metabolites. Blue represents low expression of the metabolites. b t-SNE plot based on metabolic profiles of R (yellow dots, n = 223) and HR-S (green dots, n = 159). c t-SNE clustering of three samples: HR-S strain (green dots, n = 159), R strain (yellow dots, n = 223) and mixed sample containing 10% HR-S and 90% R (1:9) (gray dots, n = 151). d Bacterial composition prediction of the mixed (1:9) sample based on the output of (b). e Stacked bar graph showed percentage of  predicted and actual bacterial composition of the mixed (1:9) sample. f t-SNE clustering of three samples: HR-S strain (green dots, n = 159), R strain (yellow dots, n = 223) and mixed sample containing 50% HR-S and 50% R (5:5) (gray dots, n = 113). g Bacteria composition prediction of the mixed (5:5) sample based on the output of (b). h Stacked bar graph showed percentage of predicted and actual bacterial composition of the mixed (5:5) sample. HR-S: colistin heteroresistant K. pneumoniae. R: colistin resistant K. pneumoniae. Source data are provided as a Source Data file.

Next, to further demonstrate the ability of the SinBactM platform to discriminate potential clinical heteroresistant strains, we analyzed mixed samples containing pre-mixed HR-S and R strains at ratios of 1:9 (10% HR-S, 90% R) and 5:5 (50% HR-S, 50% R). The mixed bacteria could be divided into two groups (Fig. 2c, f, gray dots), which were assigned to either HR-S or R groups according to t-SNE analysis based on single-bacterium metabolic fingerprints. Bacterial compositions of two mixed samples were predicted using a supervised learning k-nearest neighbor (KNN) model based on the output of t-SNE according to their metabolic fingerprints. The prediction results showed that when the pre-mixed ratio was 1:9, it was predicted with 15% for HR-S group and 85% for R group. Meanwhile, when the pre-mixed ratio was 5:5, it was predicted with 45% for HR-S group and 55% for R group, which were similar to the pre-mixed ratio (Fig. 2d, e, 2g, h, Supplementary Fig. 50). All bacteria were randomly detected without any pretreatment which made it possible to indiscriminately identify heterogeneous bacteria from a mixture. The above results showed the potential of applying SinBactM platform to discriminate clinical heterogeneous strains.

SinBactM platform captured metabolomic responses of heteroresistant K. pneumoniae and resistant K. pneumoniae upon antibiotic stress

SinBactM platform was next applied to explore the metabolomic alterations of the HR and R strains in the presence of colistin perturbations. The HR-S and R strains were treated with 2 μg/mL colistin, respectively. Single-bacterium metabolomic responses to colistin were detected and compared after 1 h culture by SinBactM platform (Supplementary Fig. 25).

As expected, extensive metabolic alterations were observed between HR-S and R strains upon antibiotic stress (Fig. 3a). Among these identified metabolites, levels of 82 metabolites were found to be significantly different (p < 0.05, |FC| > 1.2), among which levels of 69 metabolites decreased and 13 metabolites increased in the presence of colistin perturbations (Fig. 3b, Supplementary Fig. 26, Supplementary Data 6). Namely, the polyamine levels, such as N-acetylputrescine, cadaverine and spermidine were found to decrease in HR-S strain, whereas levels of TCA-related metabolites were found to increase in HR-S strain, which was consistent with previous reports (Supplementary Fig. 27)^33–36. Above results indicated that bacterial metabolism would undergo a broad and complex change upon antibiotic stress.

Fig. 3. SinBactM platform captured metabolomic responses of K. pneumoniae upon antibiotic stress.

a Heatmap of characteristic metabolites and their relative intensities in each HR-S (n = 1275) and R (n = 280) treated with 2 μg/mL colistin. Color indicates z scores of metabolites in each sample. Red represents high expression of the metabolites. Blue represents low expression of the metabolites. Data were acquired from three days of experiments. bVolcano plot showed correlations between p values (p, evaluated by two-sided Student’s t-tests) and relative intensities (FC) between HR-S and R strains treated with 2 μg/mL colistin. Red dots and blue dots represent characteristic metabolites with a significant difference (|FC | >1.2, p < 0.05). Gray dots represent characteristic metabolites with no significance (p > 0.05, or |FC | <1.2, p < 0.05).c t-SNE plot based on metabolic profiles of HR-S (red dots, n = 1275) and R (blue dots, n = 280) strains treated with colistin. Data were acquired from three days of experiments. HR-S and its replicates (HR-treated.1, HR-treated.2, HR-treated.3, n = 333, 338 and 604, respectively). R and its replicates (R-treated.1, R-treated.2, R-treated.3, n = 133, 87 and 60, respectively). d Expression of representative metabolites from HR-S and R strains treated with colistin was highlighted on the t-SNE. Color bar indicates z scores of metabolites in each sample. Blue represents high expression of the metabolites. Gray represents low expression of the metabolites. e Pathway enriched (p < 0.05, one-sided hypergeometric test, not adjusted) by significant differential metabolites between HR-S and R strains treated with colistin. Color indicates p value of enriched pathway. Green represents high p value. White represents low p value. HR-S-treated: colistin heteroresistant K. pneumoniae treated with colistin. R-treated: colistin resistant K. pneumoniae treated with colistin. Sorce data are provided as a Source Data file.

The complex metabolomic perturbations between HR-S and R strains upon antibiotic stress were further visualized by the t-SNE algorithm in the two-dimensional plane. Metabolomic-based separation between HR-S and R strains upon antibiotic stress was clearly observed with no obvious batch effects, which further illustrated the complex metabolic changes in the presence of antibiotic perturbations (Fig. 3c, d, Supplementary Fig. 28-29, Supplementary Fig. 64). Furthermore, KEGG metabolic pathways were enriched based on differential metabolites between HR-S and R strains under colistin stress. Significant disturbance was found in a number of metabolic pathways, including i) arginine and proline metabolism, ii) tryptophan metabolism, and iii) tyrosine metabolism (Fig. 3e, Supplementary Fig. 30). Above results suggested that SinBactM platform could capture metabolic alterations upon antibiotic stress at single-bacterium level.

In addition, the differential metabolites between HR-S and R strains with or without colistin treatment were further compared. Based on our findings, we observed that the presence or absence of antibiotic treatment led to both different and similar metabolic changes between R and HR-S strains, highlighting the complex interplay between antibiotic pressure and bacterial metabolism (Supplementary Fig. 51-52). While the functional implications of these metabolic differences identified here require further mechanistic validation, the data generated by SinBactM platform provide a single-bacterium level phenotypic map that establishes critical targets for future investigation into heteroresistance.

Discovery of a resistance-related subpopulation in heteroresistant K. pneumoniae strain

Bacterial heteroresistance enables small fractions of the population to survive in the presence of antibiotics, which is a major reason for anti-infective therapy failure, especially after cyclic antibiotic administration. More and more evidence imply the correlation between this heteroresistant phenotype and the metabolic adaptation. However, screening the small fractions of surviving bacteria and tracing their dynamic developmental trajectory from the large populations during resistant evolution remains a significant challenge, mainly due to the lack of proper tools that can monitor the phenotypic response of primary individual bacterium. Herein, the metabolic atlas of HR-S strain was elaborately examined to discover heterogeneous dynamic metabolic responses of the above clinical HR-S strain upon antibiotic stress.

Eight subpopulations were found for HR-S and R strains through unsupervised clustering based on their metabolomic features (Fig. 4a, Supplementary Fig. 31). Unsupervised clustering was performed through t-SNE using Seurat, confirmed by KNN with an excellent reliability of clustering results (Supplementary Fig. 32). Among these eight subpopulations, HR-S strain was found to be distributed in all eight subpopulations (Fig. 4b). In contrast to HR-S strain, R strain was mainly distributed in cluster 1, showing a significant overlap with cluster 1 of HR-S strain (Fig. 4c, d, Supplementary Fig. 33).

Fig. 4. SinBactM reveals heterogeneous responses of K. pneumoniae to colistin exposure based on single-bacterium metabolomics.

a t-SNE clustering results of eight subpopulations of HR-S and R upon colistin stress based on their metabolomes. Different colored dotted lines represent different subpopulations. Gray dots represent HR-S (n = 1275) and R strain (n = 280). b t-SNE clustering results of HR-S treated with colistin based on their metabolomes. n = 67, 243, 236, 225,125, 165, 155 and 59 for cluster 1, 2, 3, 4, 5, 6, 7 and 8, respectively. c t-SNE clustering results of R treated with colistin based on their metabolomes. n = 210, 4, 0, 1, 57, 8, 0 and 0 for cluster 1, 2, 3, 4, 5, 6, 7 and 8, respectively. d Strip chart showed the number of HR-S and R treated with colistin in each cluster. e A circos plot showed the overlap of characteristic metabolite markers between R and eight HR-S clusters treated with colistin. The same characteristic metabolite markers in HR-S cluster and the R are connected by lines. f Stacked bar graph showed the overlap proportion of characteristic metabolites between R and eight HR-S subpopulations treated with colistin. The blue bar represents the proportion of characteristic metabolite markers of eight HR-S subpopulations overlap with R. g Pseudotime trajectory of eight subpopulations of HR-S was generated by Monocle2. HR-S plotted on the trajectory were colored by subpopulations (upper left), and pseudotime (lower left). The location of eight HR-S subpopulations projected on the trajectory (right). Pseudotime (lower left) was colored in a gradient from dark to light blue. The start of pseudotime is indicated by dark blue, the end of pseudotime by light blue. h Top 50 metabolite levels among the pseudotime trajectory were divided into 3 clusters, and the representative processes along the pseudotime of each cluster were shown. Color from green to purple indicates relative express levels of top 50 metabolites from low to high. HR-S-treated: colistin heteroresistant K. pneumoniae treated with colistin. R-treated: colistin resistant K. pneumoniae treated with colistin. Source data are provided as a Source Data file.

Furthermore, we ranked characteristic metabolites (p < 0.05, FC  >  20%) of HR-S strain in each group (Supplementary Fig. 34). Different metabolic characteristics were observed for each subpopulation (Supplementary Data 7-14). Interestingly, by comparing the overall characteristic metabolites of R strain upon antibiotic stress with the eight subpopulations of HR-S strain treated with colistin, the characteristic metabolites in R strain were found to be more similar to cluster 1 in HR-S strain than the other seven clusters. This indicated cluster 1 in HR-S strain might represent the small fraction of the resistant K. pneumoniae in HR-S strain, which survived the antibiotic stimulation (Fig. 4e, f).

To further characterize metabolic heterogeneity in HR-S under colistin exposure, we performed pseudotemporal ordering of HR-S subpopulations based on the metabolomics of each bacterium. The pseudotime trajectory revealed a distinct linear pattern that oriented from right to left (Fig. 4g, Supplementary Fig. 35). Among them, cluster 2 and cluster 4 were mainly situated at the beginning of the pseudotime path, which had lower metabolites overlap with R. Meanwhile, cluster 1 was concentrated at the end of the pseudotime path, which showed strong correlation with R. And cluster 3, cluster 5, cluster 6, cluster 7 and cluster 8 were dispersed across the entire stage prior to reaching the end of the pseudotime path, revealing graded metabolic heterogeneity among HR-S subpopulations. We further analyzed the dynamic metabolite changes along the pseudotime path. Based on metabolites expression along the branches, top 50 metabolites were divided into different expression trends with pseudotemporal expression pattern (Fig. 4h). For example, malic acid and lysine displayed decreasing expression as pseudotime advanced, while acetylputrescine and cadaverine increased along the trajectory. The identified metabolites may play a critical role in the development of resistance of K. pneumoniae (Supplementary Fig. 36). Together, the metabolic heterogeneity delineated by SinBactM platform, while not yet functionally decoded, may open an avenue for investigating heteroresistance.

Discussion

To reveal the complex and heterogeneous metabolism among individual bacteria, it is vital to dissect bacterial metabolism comprehensively at single-bacterium level. However, tools to address this issue are still lacking. To solve this problem, we developed SinBactM, a platform combined InESI-MS measurement with pipette micro-extraction. The picoliter droplet micro-extraction enables metabolites to be enriched in a small volume. Meanwhile, InESI allows stable mass spectrometry analysis of picoliter-level solution, offering advantages of direct analysis without any liquid-phase separation, microelectrophoresis effect to reduce salts interference and ultra-low spray flow rates, thereby enabling metabolome profiling at single-bacterium level^37–41. However, although our SinBactM platform employs InESI-MS without liquid-phase separation, which is great to boost throughput and simplify the workflow, abundant salts in samples may still affect metabolite detection. Electrophoretic separation represents the most effective method for salt removal^42,43. While InESI-MS has been demonstrated to possess some inherent electrophoretic separation capability, integrating additional electrophoretic techniques specifically optimized for trace samples within the electrospray ionization emitter could further improve measurement sensitivity.

Using SinBactM platform, tracking metabolic heterogeneity in HR-S strain with colistin perturbations was performed, and differential metabolism was found in a variety of subpopulations. In particular, we identified an unexpected metabolic state that was likely associated with bacterial resistance in a rare subpopulation of bacteria. Notably, pseudotemporal ordering of HR-S subpopulations was clearly depicted based on single-bacterium metabolome. However, it is important to note that while this study provides a foundational dataset on single bacterial metabolomics and reveals previously unappreciated metabolic diversity within heterogeneous bacterial populations, the functional linkage between specific metabolic changes and resistant phenotypes remains to be experimentally established. We anticipate that the SinBactM platform will play a significant role in such future investigations, enabling researchers to track metabolic dynamics at the single-bacterium level and thereby deepen our understanding of how metabolic regulation influences bacterial adaptation and survival under antibiotic stress.

To date, metabolic heterogeneity was discovered in part nodes of bacterial metabolism by individual bacterial RNA-seq^{14,15,17–19}. However, the total number of the metabolite-related mRNAs was greatly limited by inherent challenges, such as low mRNA content and the existence of cell walls. As an excellent single-bacterium technique, MATQ-seq achieved an average detection of 102 genes for single Pseudomonas aeruginosa. Among which, 26 metabolite-related mRNAs were captured (Supplementary Fig. 37a)¹⁴. By our SinBactM platform, a total of 104 metabolites were putatively annotated for individual Pseudomonas aeruginosa, with much higher metabolite coverage and wider metabolic pathways range compared with MATQ-seq (Supplementary Fig. 37b-d). Furthermore, among the resistance-related subpopulation we discovered in heteroresistant K. pneumoniae strains, high correlation with arginine and proline metabolic pathway were revealed (Supplementary Fig. 30). However, from data obtained for RNA-seq based techniques, these metabolic pathways were covered due to their lower metabolite-related mRNAs coverage (Supplementary Fig. 37a), which may lead to overlooking of this rare subgroup. Thus, considering the rapid outbreak of single bacterial RNA-seq technologies in recent years, single bacterial metabolism analysis platform may provide some complementary insights in single bacterium analysis.

Despite the excellent performance of this platform, several severe challenges remain. Firstly, the low concentration of metabolites within such small volume poses significant limitations in single bacterial metabolomics. While the SinBactM platform allows for the detection of more metabolites than current single bacterial transcriptomics, the metabolite coverage is still incomplete due to the inherent constraints of individual bacteria. The low concentration of metabolites in individual bacteria makes it difficult to achieve comprehensive metabolite profiling, underscoring the need to further enhance the platform’s sensitivity and expand the detectable metabolite range. Looking forward, further improvements in ionization efficiency, extraction strategies, and complementary positive/negative ion detection are expected to substantially expand metabolite coverages. Such advances will be essential for enabling more comprehensive and biologically informative single bacterial metabolomic analyses. Secondly, unlike traditional metabolomics, the lack of chromatographic separation and MS/MS in single cell analysis fundamentally limits our ability to discern noise from true biological signal. The inherent trade-off between metabolite coverages and data certainty, stemming from the analytical limitations of single cell MS, underscores the pressing need for innovative strategies to enhance the credibility of single cell metabolic data measurement in the future. Thirdly, higher throughput is clearly necessary, particularly for studying highly complex and diverse natural communities^15,20. Although SinBactM platform has the higher metabolite coverages than current individual bacterial RNA-seq techniques, it is far inferior in analytical throughput. High-throughput interrogation of individual bacterial RNA-seq has been developed to deal with complex and large bacterial communities^15–20. Thus, single bacterial metabolite measurements with high-throughput are essential for further in-deepth studies of microbial systems. What’s more, due to the complexity of the biological process, multi-omics analysis as a new field of single cell has been reported to provide a more comprehensive understanding of cellular processes^44–47. It is exciting that SinBactM platform is a morphology-preserving measurement method in which the complete structure of the bacteria (bacterial wall, bacterial morphology) remains after micro-extraction. This makes it possible for the SinBactM platform to be combined with other superior technology (such as individual bacterial RNA-seq) for single-bacterium multi-omics analysis. Still, we expect SinBactM platform has the potential to emerge as a potent technique across diverse bacterial research domains at single-bacterium level, such as host-pathogen interactions and bacterial interactions, expanding our knowledge of the microbiome.

Methods

Ethical statement

This study was approved by the Medical Ethics Committee of the First Affiliated Hospital of University of Science and Technology of China (2025KY-209). The K. pneumoniae strains (SD 159) were isolated from a clinical sample submitted by a patient from the First Affiliated Hospital of University of Science and Technology of China and informed consent was obtained from patient before sample collection.

Materials and reagents

Tryptone, yeast extract, formic acid (FA), ammonium hydroxide (NH₄OH), glucose, ammonium chloride (NH₄Cl), ammonium chloride-¹⁵N (¹⁵NH₄Cl), disodium hydrogen phosphate (Na₂HPO₄) and potassium dihydrogen phosphate (KH₂PO₄) were obtained from Sigma-Aldrich (St Louis, MO, USA). Sodium chloride (NaCl), potassium chloride (KCl), magnesium chloride (MgCl₂), magnesium sulfate (MgSO₄), ammonium bicarbonate (NH₄HCO₃) and 4-(2-Hydroxyethyl)piperazine-1-ethanesulfonic acid (HEPES) were obtained from Sinopharm Chemical Reagent Co., Ltd (Shanghai, China). Deionized water (18.2 MΩ/cm) was obtained from a Milli-Q direct water purification system (Millipore, Inc., Bedford, MA, U.S.A.). Ciprofloxacin and agarose A were obtained from Sangon Biotech (Shanghai, China). Colistin was obtained from National Institute for Food and Drug Control (Beijing, China). Poly(L-lysine)-coated glass was obtained from Xingrui Experimental Equipment Factory (Jiangsu, China).

For Luria-Bertani (LB) medium: 1 g tryptone, 0.5 g yeast extract and 0.5 g NaCl were dissolved in 100 mL deionized water. Then, the above solution was sterilized for 20 min at 120 °C. After cooling, the sterilized LB medium was stored at 4 °C for further use.

For LB agar medium: 1 g tryptone, 0.5 g yeast extract, 0.5 g NaCl and 1 g agarose were dissolved in 100 mL deionized water, followed by sterilization for 20 min at 120 °C. The medium was poured into the Petri dish and sterilized by UV for 20 min. After sterilization and solidification, the LB agar medium was stored at 4 °C for future use.

For M9 minimal medium: 0.5 g NH₄Cl, 4 g glucose, 6.8 g Na₂HPO₄, 3 g KH₂PO₄, 0.5 g NaCl and 2 g MgSO₄ were dissolved in 1 L deionized water. For ¹⁵NH₄Cl-sourced M9 minimal medium, the 0.5 g NH₄Cl was replaced with 0.5 g ¹⁵NH₄Cl. If not stated otherwise, M9 minimal medium means ¹⁴NH₄Cl-sourced M9 minimal medium.

For ciprofloxacin and colistin stock solutions: 1 g ciprofloxacin (CIP) or colistin was dissolved in 1 L sterile PBS. Then, the above solution was filtered with a 0.22 μm filter membrane. After filtering, the sterilized antibiotic solution was stored at 4 °C for future use.

For artificial intracellular solution: 10.3 g KCl, 0.34 g NaCl, 5.9 g MgCl₂, 0.79 g NH₄HCO₃, 1.98 g glucose and 2.4 g HEPES were dissolved in 1 L deionized water (pH was adjusted to 7.4).

Strains and growth conditions

Escherichia coli BL21 (E. coli), Pseudomonas aeruginosa PAO1 (P. aeruginosa), Staphylococcus aureus ATCC25923 (S. aureus), Bacillus subtilis (B. subtilis), Klebsiella pneumoniae (K. pneumoniae) and Salmonella (12023 S. Typhimurium wild-type) were used for this study. The K. pneumoniae strain (SD 159) was isolated from patients. The K. pneumoniae strains (SD 159) is a colistin heteroresistant strain, which refers to the presence of a resistant subpopulation in an overall susceptible strain (HR-S).

For metabolic profiling of individual bacterium: Bacteria were streaked onto an LB agar plate for colony growth. Then, a single colony was cultured in LB medium with 220 rpm shaking at 37 °C for 6–8 h. The harvested bacteria were centrifuged at 3445 ×g for 3 min and washed twice with deionized water, followed by single bacterial metabolic analysis.

For ciprofloxacin accumulation measurement in individual bacterium: E. coli was streaked onto an LB agar plate for colony growth. Then, a single colony was cultured in LB medium with 220 rpm shaking at 37 °C for 6–8 h. The harvested E. coli were centrifuged at 3445 ×g for 3 min and washed twice with deionized water, followed by dilution to OD600 ~ 0.2 in deionized water. E. coli was pre-incubated with or without carbonyl cyanide m-chlorophenylhydrazone (CCCP) for 10 min at 37 °C with 220 rpm shaking. Then, 10 μM or 50 μM ciprofloxacin was added to bacterial solution for 10 min at 37 °C with 220 rpm shaking. After incubation, E. coli were centrifuged at 3445 ×g for 3 min and washed twice with deionized water, followed by single bacterial metabolic analysis. The relationship between the accumulation of CIP and the coverage of metabolites in the same single bacterium was also analyzed (Supplementary Fig. 54-55).

For isotopic labeling metabolic analysis in individual bacteria: E. coli was streaked onto an LB agar plate for colony growth. Then, a single colony was cultured in LB medium with 220 rpm shaking at 37 °C for 6–8 h. The above E. coli were centrifuged at 3445 ×g for 3 min and washed twice with deionized water to remove the LB medium. Then, the harvested bacteria were cultured in M9 minimal medium or ¹⁵NH₄Cl-sourced M9 minimal medium for 1 h, 6 h and 24 h, respectively. After incubation, bacteria were centrifuged at 3445 ×g for 3 min and washed twice with deionized water, followed by single bacterial metabolic analysis.

For PAP assay of heteroresistant K. pneumoniae strain: Antimicrobial heteroresistance was determined by population analysis profiling (PAP). Generally, overnight cultured bacteria were serially diluted with Mueller-Hinton (MH) broth and were dropped onto MH agar plates containing 0–64 μg/mL colistin. Colonies on the plates were counted after incubation at 37 °C for 24 h. Colistin heteroresistance was defined as the growth of a resistant subpopulation at antibiotic concentrations at least eightfold higher than the MIC of the parental isolate. As shown in Supplementary Fig. 24a, with increasing colistin concentration, most of the bacteria were killed. The resistant subpopulation was detected at a frequency of about 10⁻⁷, when the colistin concentration reached 16 μg/mL. Thus, the K. pneumoniae strain (SD 159) was identified as a colistin-heteroresistant strain, in which exist a small group of resistant subpopulations in an overall susceptible strain, designated as HR-S strain.

For resistant K. pneumoniae strain: Colistin-resistant K. pneumoniae was isolated from the above clinical colistin heteroresistant K. pneumoniae (SD 159). Colistin-resistant K. pneumoniae were selected from the plate with the highest concentration of 32 μg/mL colistin MH agar plates where growth occurred, named as R strain. The MIC values of colistin were determined by broth microdilution (BMD) (Supplementary Fig. 24b)³². The minimum inhibitory concentration (MIC) experiment for R strain was performed before individual bacterial metabolism detection to ensure that the resistance of R has not been lost.

For clinical isolate K. pneumoniae treated with colistin: Bacteria were streaked onto an LB agar plate for colony growth. A single colony was cultured in LB medium with 220 rpm shaking at 37 °C overnight. Subsequently, the overnight culture was diluted 100-fold in LB medium and treated with 2 μg/mL colistin. Then, it was cultured in LB medium with 220 rpm shaking at 37 °C for 1 h. After incubation, bacteria were centrifuged at 3445 ×g for 3 min and washed twice with deionized water, followed by single bacterium metabolic analysis. 1 h antibiotic exposure is just a quick treatment to perturb cell metabolism which does not kill the bacteria (Supplementary Fig. 62). Samples were collected in batches, where each batch was prepared and collected within a strict 30 min interval. In our setup, after the antibiotic treatment for 1 h, bacteria were centrifuged and washed twice with deionized water, followed by 30 min single bacterial sampling. After 30 min sampling, a new sample was prepared for the next round of collection. All the data are derived from three days of sampling, with samples collected from different batches each day. The data have no obvious batch effect according to t-SNE analysis (Supplementary Fig. 25, Supplementary Fig. 56).

Imaging

The photos of different bacteria were taken under an inverted microscope (Nikon, Eclipse, Ti2). The bacterial length and width were measured using ImageJ software. The bacteria used in this study exhibit the following size ranges (length × width): Escherichia coli: 1.2–2 µm × 0.5–0.8 µm; Staphylococcus aureus: 1.2–1.3 µm (diameter, spherical); Pseudomonas aeruginosa: 1.2–2 µm × 0.6–0.8 µm; Bacillus subtilis: 2.5–4.1 µm × 0.6–0.8 µm; Klebsiella pneumoniae: 1.3–1.8 µm × 0.6–0.9 µm (Supplementary Fig. 39). As revealed by t-SNE and PCA analysis based on their metabolome, bacteria with different sizes (small and big) were distributed without size-guided clustering after TIC normalization (Supplementary Fig. 40).

Sampling individual bacteria metabolites

All single-bacterium sampling experiments were performed through a three-dimensional manipulation platform (Burleigh micromanipulator, PCS-5000). The extraction procedure was performed under an inverted microscope (Nikon, Eclipse, Ti2). In all spray experiments, the capillary micropipettes were fabricated by pulling from borosilicate glass capillaries (1.2 mm o.d., 0.9 mm i.d.) using a P-2000 laser-based micropipette puller (Sutter Instruments, Novato, CA). The resulting capillary micropipettes had a tip size of approximately 1 μm.

For single-bacterium sampling: 5 μL bacterial solution was deposited on the poly(L-lysine)-coated glass and left for 5 min, to allow bacterial adhesion via electrostatic adsorption. Then, excess solution was removed and the glass was washed twice with deionized water, followed by drying for 2 min at room temperature. The glass with bacteria immobilized was placed under a microscope. A pulled capillary micropipette (with 1 μm opening) was inserted into 1% FA aqueous solution approximately 1 s for solvent aspiration by capillary action. Then, capillary micropipettes containing ~25 pL extraction solvent were placed over the bacteria through a three-dimensional manipulation platform. The positive and negative pressures were both all achieved by a 5 mL syringe. By pushing the syringe inner tube 200 μL forward, the extraction solution in the capillary micropipette was extruded to form a droplet over the individual bacterium for metabolites extraction. After extraction approximately 1 s, that droplet was subsequently sucked back by moving the syringe inner tube 200 μL backward, then subjected to InESI-MS analysis. 1 s extraction time is enough for most metabolite extraction (Supplementary Fig. 53). During extraction, a steady negative pressure was created by always moving the syringe inner tube 200 μL backward to ensure a consistent extraction volume each time (Supplementary Fig. 41). In the positive mode, H₂O containing 1% FA was used as extraction solution. In the negative mode, H₂O containing 1% NH₄OH was used as extraction solution (Supplementary Data 15-16). Due to the lack of rapid metabolic quenching compatible with single-bacterium MS, the metabolite profiles reported here represent an operational metabolic state under defined handling conditions. More effective quenching strategies to further minimize the loss of metabolites for individual bacteria during the quenching process, such as using nanowell to prevent metabolite diffusion will be explore in the future.

For different picoliter-level solution sampling: The capillary micropipettes were dipped into the solution for 1 s, and ~5 pL solution was loaded into the capillary micropipettes. For 1 pL solution sampling, the capillary micropipettes were dipped into the solution for 1 s, followed by volatilization for about 5 s, so that the volume of the solution was reduced to ~1 pL as observed under the microscope. The MS signal produced by InESI remained stable for ~5–45 s (around 10 to 100 scans) for various solutions at 1–5 pL level, leading to an average flow rate ranging from 6.7 to 13 pL/min, which was sufficient for picoliter level solution measurement (Supplementary Figs. 3–7 and Supplementary Table 1).

Mass spectrometry settings

After capillary micropipettes had extracted metabolites of the single bacterium, the micropipettes were placed about 5 mm from the front of the MS inlet. To initiate the electrospray, an alternating current (AC) or direct current (DC) voltage was applied to the micropipette. The AC voltage was set at an approximate amplitude of 3 kV, with a frequency of 1000 Hz. The mass spectrometry analysis was performed using an Exactive Plus mass spectrometer (Thermo Scientific, San Jose, CA, USA). Typical parameter settings: mass resolution was 140 000, microscan was 1, maximum ion injection time was 10 ms, capillary temperature was 275 °C, AGC was 1 × 10⁶, ion mode was positive, and scan range was 60–300. All mass spectra were acquired via Xcalibur software (version 3.0.63).

Data preprocessing

A homemade MATLAB program (version 2016b) was developed to select reliable metabolite signals. The average signal lasted about 50 scans for single bacterium analysis, and about 50 scans were used to measure ion counts. Only ions with the signal-to-noise ratio (SNR) greater than 3, signal intensity greater than 10³ and the detected frequency more than 20% in all single bacteria, were selected to be reliable metabolite signals. The selected signals were further identified by the Human Metabolome Database (HMDB, http://www.hmdb.ca/). Metabolite adduct types [M + H]⁺, [M+Na]⁺, and [M + K]⁺ were chosen for positive mode. Metabolite adduct types [M-H]^-, [M-H₂O-H]^-, [M + K-2H]^-and [M+Na-2H]^- were chosen for negative mode. Metabolite annotation was performed in a tiered manner to assign differing levels of confidence. The selected ion signals were initially assigned putative molecular formulae based on accurate mass matches against the Human Metabolome Database (HMDB) within a mass error of <|10 ppm | . This is the lowest level of confident identification for metabolites. To increase confidence, some metabolites were further identified by comparing MS/MS spectra obtained from population bacteria with data from the standards or the HMDB (Supplementary Fig. 57-61). This is the second-lowest level of confident identification for metabolites. Metabolites with MS/MS spectra matching with the standards and the HMDB library were marked with double asterisks (**) or a single asterisk (*), respectively. Pathway analysis was performed on MBROLE 2.0 (https://csbg.cnb.csic.es/mbrole2/analysis.php). 0/NaN values were replaced with zero. We normalized the intensity to a ratio of the total ion current (TIC). We further compared TIC normalization with internal standard normalization for quantifying metabolic differences between HR and R strains, demonstrating no significant divergence between the two approaches (Supplementary Fig. 66).

Validation and technical performance of SinBactM platform

To validate the reliability of the SinBactM platform for detecting exogenous compounds at the single-bacterium level, we measured ciprofloxacin accumulation in individual E. coli⁴⁸. We observed that intracellular antibiotic signals increased with rising antibiotic concentrations (Supplementary Fig. 17-18). In addition to the accumulation of exogenous antibiotics in single bacterium, the reliability of SinBactM platform was further validated by tracking endogenous metabolite transformation in bacteria by isotopic labeling at single-bacterium level. The time-dependent metabolic isotope transformation was observed to increase with the prolongation of the ¹⁵N-NH₄Cl incubation time (Supplementary Fig. 19-23, Supplementary Fig. 42). Also, our platform resolved clear, dose-dependent metabolic remodeling with the increasing antibiotics concentrations, demonstrating its capacity to robustly capture true biological heterogeneity at single-bacterium resolution (Supplementary Fig. 65). In addition, using a carbon starvation time-course model, our platform captured progressive, literature-consistent metabolic dynamics (e.g., amino acid accumulation, nucleotide degradation) with single-cell resolution and strong time-metabolite correlations, demonstrating its capacity to resolve true biological heterogeneity from technical noise (Supplementary Fig. 67).

Calibration with external standards and reproducibility of SinBactM platform

Calibration curves of spermidine, leucine and tryptophan were performed in artificial intracellular fluid by SinBactM platform (Supplementary Fig. 38a–c). Nine repeated measurements (three times each day for 3 consecutive days) were performed to illustrate the reproducibility of the SinBactM platform for spermidine (1 mg/L), leucine (1 mg/L) and tryptophan (1 mg/L), with interday RSDs ranging from 2.7% to 12.2% and intraday RSDs ranging from 7.1% to 10.9% (Supplementary Fig. 38d-f).

Also, the correlation between the rhodamine B (RhB) fluorescence intensity and the mass spectrometry signal in single B. subtilis was compared to further illustrate the reproducibility of the SinBactM platform. Bacteria were stained with the 0.5 μg/mL fluorescent dye RhB for 5 min at 37 °C. After incubation, bacteria were centrifuged at 3445 ×g for 3 min and washed twice with deionized water. Then, bacteria were adsorbed on the glass followed by fluorescence imaging. Fluorescence intensity of individual bacteria was analyzed by RGB measurement in ImageJ software. The bacteria with measured fluorescence intensity were used for subsequent mass spectrometry detection. The RhB signal detected by mass spectrometry increased with increasing fluorescence intensity. The fluorescence signal of RhB for low, medium and high fluorescence was less than 15%, and the RhB signal measured by our method was also less than 15%, indicating that our method has good reproducibility (Supplementary Fig. 39).

In addition, we further conducted a PCR analysis independent of mass spectrometry to illustrate the reproducibility of the SinBactM platform. The Salmonella strain had its speG gene and speE gene knocked out (Supplementary Fig. 63). The speG gene encodes N-acetyltransferase, which is responsible for the acetylation modification of polyamines. After the speG gene was knocked out, the level of acetylspermidine decreased significantly. Meanwhile, spermidine could not be converted into acetylspermidine, resulting in an increase in spermidine levels (Supplementary Fig. 63a). The SpeE gene encodes spermidine synthase, which is responsible for the biosynthesis of spermidine. After the speE gene was knocked out, the level of spermidine decreased significantly. Meanwhile, due to the decrease in spermidine, the content of downstream acetylspermidine also decreased (Supplementary Fig. 63b, c). The above results demonstrated that the PCR analysis was consistent with the mass spectrometry measurements. After gene knockout, as the gene content decreased, the expression level of the corresponding metabolites also decreased.

Construction of the Salmonella ΔspeE and ΔspeG mutant strain

The Salmonella strain, Salmonella ΔspeE mutant strain and Salmonella ΔspeG mutant strain were a kind gift from Professor Yongqun Zhu in Zhejiang University⁴⁹. Briefly, the in-frame deletion of speE or speG in Salmonella strain was constructed individually by homologous recombination using the pDM4 suicide vector. For each gene, DNA fragments containing approximately 1000 base pairs of the upstream and downstream sequences were ligated into the pDM4 vector. The resulting plasmids were separately transformed into the donor Escherichia coli SM10 λπ strain. The transformed E. coli SM10 λπ strains and the recipient wild-type Salmonella enterica serovar Typhimurium strain 12023/14028 were cultured in 2YT medium at 37 °C until the optical density at 600 nm (OD600) reached 0.6. For each knockout, the donor and recipient strains were mixed at a ratio of 1:1 and transferred onto a sterilized 0.22 μm membrane filter placed on 2YT agar plates without antibiotics, followed by incubation at 30 °C overnight for conjugation. Transconjugants were selected on 2YT agar plates containing 25 μg/mL chloramphenicol and 50 μg/mL streptomycin. The first recombinant colonies were confirmed by standard PCR using primers targeting the sacB gene and the flanking regions of the respective target gene, and then grown on 2YT agar plates (without NaCl) supplemented with 6% sucrose to isolate the ΔspeE and ΔspeG mutant strains, respectively. Finally, each mutant strain was confirmed by standard PCR using at least two different pairs of primers, and further validated by DNA sequencing of the PCR products.

Identification of HR-S and R-KP subpopulations

We used the single-bacterium metabolome to cluster individual bacteria through a t-distributed stochastic neighbor embedding (t-SNE) algorithm using Seurat R package (version 3.0). The data matrix was imported into Seurat (version 3.0) to identify HR-S and R subpopulations. The metabolite intensity was normalized by a ratio of the total ion current and then the data were scaled. Next, PCA dimensional reduction was performed on the scaled data and then appropriate principal components were selected for subsequent analysis. The clusters were identified using a shared nearest-neighbor modularity optimization-based clustering algorithm, which can effectively determine clusters. Finally, clusters were visualized by t-SNE.

Identification of characteristic metabolites

According to the above HR-S and R subpopulations classification information, Wilcoxon rank-sum tests were carried out for one cluster to all other clusters. Ions with significant differences (p values < 0.05, FC greater than 20%) were considered as characteristic metabolites. The top 10 metabolites (based on p value rank) were selected for visualization of each cluster, and z score was used to plot a heatmap, which highlights the most distinctive metabolic features of each bacterial subpopulation. The z score is a numerical measurement that describes a value’s relationship to the mean of a group of values. It is calculated by subtracting the mean intensity of all bacterial cells from an individual raw intensity and then dividing the difference by the standard deviation of all bacterial cells. Circos plot was performed by TBtools, representing the characteristic metabolites overlap between different HR-S subpopulations and R strain.

Constructing trajectories of clinical isolate K. pneumoniae

We used the above screened characteristic metabolites for K. pneumoniae trajectories analysis, by Monocle2 R package (version 2.16). The metabolic data matrix was imported into Monocle and constructed the pseudotime trajectory. “DDRTree” was used to perform the dimension reduction, and the function “plot_cell_trajectory” was performed to visualize the trajectory of HR-S.

Supervised machine learning methods

Unsupervised clustering (Seurat) first identifies bacterial subpopulations based on metabolomic similarity, without prior labels. T-SNE was used to visualize the above subpopulation results. Subsequently, supervised validation (kNN) was employed solely as a post hoc validation tool to assess the robustness of the unsupervised clusters, with prior subpopulation label resulting from the aforementioned unsupervised clustering (Seurat). A supervised learning method based on k-nearest neighbor (KNN) were trained and tested using classification learner in MATLAB program (version 2016b) was performed.

Statistical analysis

Statistical analysis was performed by GraphPad Prism 6 and the R programming language. The statistical significance between two groups was compared by unpaired two-sided Student’s t-tests, and p values were plotted on the figures. Data show the mean ± SD. p < 0.05 was considered a statistically significant difference. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

G.H., Z.H., L.Z. and H.L. conceived and conducted the project. G.H. and Z.H. supervised the project. L.Z. and H.L. wrote the paper. L.Z., H.L., W.C., K.C., Y.G., S.Z. and B.C. performed the experiments and data analysis. W.C., K.C. and X.M. contributed to bacteria culture. X.M., R.Z. H.H. and H.Y. contributed to the interpretation of the data.

Peer review

Peer review information

Nature Communications thanks Bo Zhang and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The metabolomic data generated in this study have been deposited in the MassIVE database under accession code MSV000101304 (https://massive.ucsd.edu/ProteoSAFe/static/massive.jsp?redirect=auth). All data supporting the present study are available within the paper and the Supplementary Information. Source data are provided in this paper. Source data are provided with this paper.

Code availability

The code developed in this manuscript has been submitted to GitHub (https://github.com/GMHuang/SinBactMeta).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-72373-0.