Untargeted Metabolomics Sensitively Differentiates Gut Bacterial Species in Single Culture and Co-Culture Systems.

Gut microbiome plays a vital role in human health, and its characteristic has been widely identified through next-generation sequencing techniques. Although with great genomic insights into gut microbiome, its functional information is not clearly elaborated through metagenomic techniques. On the other hand, it is suggested that fecal metabolome can be used as a functional readout of the microbiome composition; therefore, we designed a proof-of-concept study to first characterize the metabolome of different gut microbes and then investigate the relationship between bacterial metabolomes and their compositions in co-culture systems. We selected eight representative bacteria species from Bifidobacterium (2), Bacteroides (1), Lactobacillus (4), and Akkermansia (1) genera as our model microbes. Liquid chromatography coupled mass spectrometry-based untargeted metabolomics was utilized to explore the microbial metabolome of bacteria single cultures and co-culture systems. Through spectral comparisons, our results showed that untargeted metabolomics could capture the similarity and differences in metabolic profiles from eight representative gut bacteria. Also, untargeted metabolomics could sensitively differentiate gut bacterial species based on our statistical analyses. For example, citrulline and histamine levels were significantly different among four Lactobacillus species. In addition, in the co-culture systems with different bacteria population ratios, gut bacterial metabolomes can be used to quantitatively reflect bacterial population in a mixed culture. For instance, the relative abundance of 2-hydroxybutyric acid changed proportionately with the changed population ratio of Lactobacillus reuteri in the co-culture system. In summary, we proposed a workflow that could demonstrate the capability of untargeted metabolomics in differentiating gut bacterial species and detecting their characteristic metabolites proportionally to the microbial population in co-culture systems.

Introduction

Human gut microbiome plays an important role in human health, where gut dysbiosis leads to various diseases from metabolic syndrome to cancer.[1,2] Meanwhile, the current advanced next-generation sequencing platform[3,4] for gut microbiome gives us great genomic insights into gut bacteria with their taxonomic identification (ID) and gut microbial composition. For example, from the gut microbiome sequencing information, bacteria genera such asBacteroides, Bifidobacterium, and Lactobacillus are predominant.[5] However, metagenomic techniques do not offer immediate functional information. Meanwhile, a previous study proposed that fecal metabolome could serve as a functional readout of microbial activity,[6] in which they reported that the fecal metabolome largely reflected the gut microbial characteristics by explaining on average 67.7% (±18.8%) of its variance from the large clinical study data set of TwinsUK cohort. Furthermore, more bioinformatics-based research studies have contributed to the development of statistical models to predict microbe abundance from metabolite abundance with multi-omics data.[7−9] These research studies have supported that certain statistical models have the ability to predict microbiome composition from their metabolomics profiles. Taken together, metabolome data have the potential to reveal the metabolic capabilities and functional characterizations of gut microbes and reflect bacterial compositions to some extent.

Mass spectrometry (MS)-based metabolomics analysis, as a system biology technique, has been applied to the microbiology field and helps to investigate the microbial metabolome in the large intestine.[10] It has advantages of high sensitivity and high-throughput discovery of a wide range of metabolite classes especially. Detection of microbial metabolites (e.g., short-chain fatty acids, amino acids, and bile acids) in both clinical and preclinical studies has been widely reported and utilized to many applications ranging from disease diagnosis to treatment evaluation.[11,12] In addition, MS-based metabolomics techniques have been developed to not only illustrate gut microbial metabolic profile but also help to differentiate between closely related pathogenic and nonpathogenic bacteria species.[13] Therefore, MS-based metabolomics would serve as a suitable tool to explore the gut microbial metabolome.

However, despite the existing efforts, there are still limited studies to unveil the complicated gut microbiome–metabolite relationship in human large intestine, and part of the challenges were due to the large number of gut microbes and it is not easy to untangle their complex interactions. Therefore, here we conducted a proof-of-concept study utilizing eight representative gut bacterial species from generaBifidobacterium, Lactobacillus, Akkermansia, and Bacteroidetes to map out their metabolome and identify their characteristic metabolic profiles. The aim of this study to first characterize the metabolome of different gut microbes and then investigate the relationship between bacterial metabolomes and their compositions in co-culture systems. We hypothesized that untargeted metabolomics could capture similarity and differences in metabolic profiles from eight representative gut bacteria and sensitively differentiate gut bacterial species. In addition, we hypothesized that gut bacterial metabolomes from the mixed culture can be used to reflect the bacterial composition in a simplified co-culture system.

Materials and Methods

Chemicals and Biological Reagents

HPLC-MS-grade acetonitrile, ammonium acetate, formic acid, and acetic acid were purchased from Fisher Scientific (Pittsburgh, PA, USA). Eight bacterial species, Bacteroides adolescentis (BA, ATCC 15703), Bacteroides longum subsp (BL, ATCC 15707), Limosilactobacillus reuteri (LR, ATCC 23272), Lactobacillus delbrueckii subsp. (LD, ATCC 11842), Lactobacillus acidophilus (LA, ATCC 4356), Lactobacillus fermentum (LF, ATCC 9338), Akkermansia muciniphila (AM, ATCC BAA-835), and Bacteroides thetaiotaomicron (BT, ATCC 29148), were purchased from American Type Culture Center (ATCC, Manassas, VA, USA).

Bacteria Culturing Conditions

For a single culture experiment, 15 mL of sterile tubes (ThermoFisher, Waltham, MA, USA) were utilized for bacterial cultures. The culturing procedure was modified based on the previous study.[14] Overnight cultures were prepared by inoculating 100 μL of bacterial stocks in 10 mL of sterilized Gifu anaerobic broth (GAM) and incubating at 37 °C in the anaerobic chamber (Coy Laboratory Products Inc., MI, USA) for 16 h. 100 μL of overnight cultures were then inoculated into 10 mL of fresh and sterilized GAM at 37 °C in an anaerobic chamber as testing cultures. Uninoculated sterilized GAM broths were also incubated as a control group and each group has three biological replicates. Eight bacterial species were cultured for 24.5 h and optical density (OD) measurements were taken for data normalization.

For the co-culture experiment, the testing cultures of eight representative bacteria species were prepared similarly as described in the single culture experiment for the first 24 h. Then, OD measurements were taken to determine the population ratio for different bacterial co-cultures. Eight bacteria species had similar OD values between 0.5 and 0.6 after 24 h of incubation. Three groups of two bacteria species were mixed at a population ratio of 100:0, 80:20, 50:50, 20:80, and 0:100. To mimic simplified bacterial model systems, BL and LR were mixed for genus level comparison, LR and LD were mixed for species level comparison, and LR and AM were mixed for lactic acid producing bacteria (LAB) versus non-LAB comparison. Each co-culture group has three biological replicates and these bacteria co-cultures were further incubated for half an hour before the sampling process.

Sample Preparation and Liquid-Chromatography MS (LC/MS)

A Vanquish UHPLC and a Hybrid Quadrupole Orbitrap Q Exactive mass spectrometers (ThermoFisher, Waltham, MA, USA) were used in this study. Metabolite extraction of the bacterial cultures was performed following a previously reported protocol.[15−17] Briefly, 1 mL of bacteria sample from each biological replicate was centrifuged. All supernatants were removed from the pellet and 100 μL of supernatant was used for extracellular polar metabolite extraction. The pellet was washed with three rounds of ∼500 μL of phosphate buffer saline and centrifuging for intracellular polar metabolite extraction. After adding 250 μL of methanol and 50 μL of 13C15N-labeled internal standard mixture to both intracellular and extracellular metabolites, the mix was vigorously vortexed and cooled in a fridge at −20 °C for 20 min. After incubation, 150 μL of supernatant was collected and dried via a SpeedVac system. 250 μL of 50% (v/v) aqueous ACN was then added to dissolve the extracted metabolites and then centrifuged. The 150 μL of reconstituted supernatant samples were transferred into LC vials and loaded onto a sample tray in the UHPLC System. The Xbridge BEH Amide (2.5 μm, 2.1 × 150 mm, Waters, Milford, MA, USA) column was used for metabolite separation. The detailed LC/MS parameters setup for polar metabolite detection was reported in our previous study.[18]

Data Processing and Statistical Analysis

For quality control purpose, pooled quality control (QC) samples, processed GAM samples as processed blanks (PBLs), and 50% ACN in water as instrument blanks (IBLs) were also used in the LC/MS-based metabolomics analysis. MS-generated raw files for polar metabolites were processed by MSDIAL (http://prime.psc.riken.jp/compms/msdial/main.html).[19] Exported features from both positive and negative modes were manually filtered through several data filtering steps: (1) exclude features with signal-to-noise (S/N) ratio smaller than 3 comparing biological samples (BSs) to PBL; (2) exclude features (i) having peak areas in IBL and (ii) having peak area = 0 in more than one group; (3) exclude features with S/N ratio smaller than 3 comparing BS to IBL; (4) exclude features with coefficient variance (CV) of QCs larger than 30%; and (5) BS peak area of all remained features were subtracted by their corresponding PBL peak area. Some tentative IDs were provided by MSDIAL. Addition ID process was accomplished by cross-referencing features with the human metabolome database (HMDB, https://hmdb.ca),[20] METLIN metabolite and chemical entity database (http://metlin.scripps.edu),[21] and in-house database with over 200 metabolites validated through the LC/MS system. All data visualization and statistical analysis were based on a list of features left after the abovementioned data filtering steps. Spectra of eight bacteria species were generated in ORIGIN (https://www.originlab.com, Northampton, MA, USA). Principal component analysis (PCA), analysis of variance (ANOVA), t-test, and partial least-squares discriminant analysis (PLS-DA) with variable importance in the projection (VIP) score were performed in MetaboAnalyst (https://www.metaboanalyst.ca).[22]

Results and Discussion

Untargeted Metabolomics Can Sensitively Detect Characteristic Metabolic Features of Gut Microbes

We first hypothesized that untargeted metabolomics can capture similarities and differences in metabolic profiles/features from eight representative gut bacteria. It is suggested that fecal metabolome can be used as a functional readout of the microbiome composition; therefore, we developed a proof-of-concept study utilizing untargeted metabolomics workflow to study the gut bacterial metabolome for potential IDs/differentiations of these important gut bacteria. As shown in Figure , eight representative bacterial species from different taxonomies were selected as model gut bacteria. Based on MSDIAL data processing, there were 14,746 metabolic features exported from the positive mode and 16,462 features exported from the negative mode, in which 6855 (46.5%) and 4318 (26.2%) features, respectively, have tentative IDs. Five steps of data filtering and data processing procedure were performed as described in the data processing section and features left from each step are shown in the Supporting Information (Figure S1). Finally, 354 features in the negative mode(Table S1) and 425 features (Table S2) in the positive mode were left by the end of data processing, in which 52 (14.7%) and 55 (12.9%) features, respectively, have tentative IDs after cross-referencing with a microbial metabolite database based on the incorporated databases in MSDIAL. Examples of identified metabolites matched with ms/ms data, such as adenine, l-glutamine from the positive mode, and oleic acid, uridine 5′-monophosphate from the negative mode, are demonstrated in Figure S2. These compounds were also identified by cross-referencing our in-house database. Although we analyzed both intracellular and extracellular metabolites from the positive and negative mode for the analysis of comprehensive gut bacteria metabolome, extracellular metabolites from the negative mode provided us the most abundant and comparable information relating to gut bacterial metabolic profile and therefore were primarily used for the following discussions, while other spectral data were also made available in the Supporting Information (Figures S3 and S4).

Figure 1

Schematic of the experiment workflow from experimental design and MS analysis to data analysis.

To comparatively visualize the metabolome of eight representative gut bacteria at species level, the profiles of 425 extracellular metabolites for these bacteria from the negative mode after data filtering were shown individually in Figure . Because of the presence of some dominant features in each bacteria species, normal scale on the y-axis would not provide enough detailed information for gut bacteria species differentiation in the spectra. For example, gamma-aminobutyric acid (GABA) (m/z = 102.0563, identified from our in-house database) was highly abundant in all four gut bacteria species from Lactobacillus genus, which was consistent with previous finding of high GABA production by culturable Lactobacillus from human intestine.[23,24] As a result, features with relative abundance between 1 and 10% were also zoomed in on the y-axis in the spectra to highlight the differences in features of bacterial metabolome at the species level. For example, for the same feature at m/z = 187.0421, the relative abundance of this feature in eight gut bacteria was ranging from 27.76% in BT to 100% in AK and BA. Besides, for features with relative abundance that fall within 1 to 10% in some gut bacteria species such as N-acetyl-dl-glutamic acid (m/z = 188.0562, identified from in-house database), relative abundance in AK was 9.55%, while relative abundance in LF was 2.34%. N-Acetyl-dl-glutamic acid has been reported as an intermediate in the arginine production pathway in bacteria produced by N-acetylglutamate synthase.[25] Also, there were certain peaks that were absent in certain bacteria but present in other bacteria species which also led to the difference in their spectra. For example, for feature with m/z at 308.2202, it was absent in BA and AK; however, in other six bacteria species, this feature was present. Additional spectra of these bacteria from the positive mode and intracellular extraction can be found in the Supporting Information (Figure S3). As a result, it is clearly visualized that metabolome from eight representative gut bacteria species were different from each other. In particular, the presence and absence of certain unique features as well as different peak intensities of certain shared features in different bacteria species primarily contributed to the differences in their metabolic profiles. Therefore, spectra generated from untargeted metabolomics analysis could show different polar metabolomes from different bacteria species, resulting from the absence or presence of metabolic features with different relative abundances.

Figure 2

Spectrum of eight bacteria strains, (A) B. adolescentis, (B) B. longum subsp., (C) L. reuteri, (D) L. delbrueckii subsp., (E) L. acidophilus, (F) L. fermentum, (G) A. muciniphila, and (H) Bacteroides thetaiotaomicron. All features from extracellular polar metabolites in the negative mode after data filtering were presented as an example. y-axis represented the relative abundance for each feature. 1 to 10% on y-axis were zoomed in order to differentiate among different bacteria spectrums. x-axis represented m/z of the features.

Untargeted Metabolomics Can Differentiate Gut Bacteria via Their Metabolome

Next, we hypothesized that untargeted metabolomics can differentiate gut bacteria at different levels based on their metabolic profiles. We first investigated the genus level comparison, and eight representative bacteria species were divided into four genera: Bifidobacterium (BA and BL), Lactobacillus (LR, LD, LA, and LF), Akkermansia (AM), and Bacteroides (BT). Three hundred and fifty-four extracellular metabolites from negative modes were analyzed and the PCA plot of Figure A indicated that there was a distinct separation among these four genera, and the color-coded circles for each genus were added for visualization purpose. Principal component (PC) 1, 2, and 3 accounted for 22.6, 20.3, and 7.8% of total metabolic variations among the four genera in this analysis, respectively. Furthermore, ANOVA t-test and VIP score were calculated for each feature, which identified 94 metabolites that were significantly contributed to this genus level separation with p-value < 0.05 and VIP score > 1 (Table S3). Some metabolites such as methylacetate (m/z = 73.0296, tentatively identified) and proline (m/z = 114.0564, identified from in-house database) were detected at significantly different levels among four gut bacterial genera. Particularly, proline was significantly higher in Akkermansia compared to Lactobacillus, while genera Bacteroides and Bifidobacterium were in the middle. It is well-known that proline is one of building block amino acids for protein synthesis in biological systems,[26] which indicated that among four bacteria genera, they may have a significant difference in amino acid metabolism and protein synthesis relating to proline. Additionally, there was a study that reported a higher level of proline from fecal sample in children with celiac disease under gluten-free diet compared to non-celiac children. Meanwhile, the levels of Lactobacillus and Bifidobacterium were significantly lower in celiac children, which indicated that proline production could be negatively associated with the level of Lactobacillus and Bifidobacterium.[27] For further taxonomic level comparison, we looked at the species level comparison of LR, LD, LA, and LF under the Lactobacillus genus. The PCA plot of Figure B suggested that even within the same gut bacterial genus, there were clear separations highlighted by color-coded circles among these four different species, which indicated the capability of untargeted metabolomics in capturing subtle metabolic differences and differentiating closely related bacterial species. PC1, 2, and 3 of Figure B accounted for 28.6, 23.2, and 9.5% of the total metabolic variance among four groups, respectively. Furthermore, ANOVA t-test and VIP score indicated that 45 metabolites were significantly different among four Lactobacillus species (p-value < 0.05 and VIP score > 1) (Table S4). Several metabolites were identified that significantly contributed to the differentiation, for example, dihydroorotic acid (an amino acid derivative) was most abundant in LR and least abundant in LA, and it is also related to pyridine metabolism in gut bacteria.[28] Other than extracellular metabolites from the negative mode, from other modes, significantly different metabolites such as citrulline and histamine were also detected among the four Lactobacillus species (Tables S7, S10, and S13). This finding was consistent with our previous study reporting the rapid differentiation of Lactobacillus species via metabolic profiling.[14] Although two different types of media were used (GAM vs MRS broth), multivariate analysis (PCA and PLS-DA) in the previous study also showed a pattern of differentiation among four Lactobacillus bacteria species. In addition, metabolites that differentiated four Lactobacillus bacteria species, such as citrulline, and histamine, which were involved in bacterial arginine biosynthesis, were commonly reported in both studies. Additionally, a previous study reported that the abundant Lactobacillus genus associated with the highest abundance of citrulline and arginine in fecal samples from breastfed Chinese infants compared to formula-fed, complementary food-fed, and mixed-fed infants.[29] Arginine and its precursor citrulline have been reported to protect intestinal cell tight junction from hypoxia in an animal model.[30] As for histamine, as a Lactobacillus metabolite in the gut microbiome, it has shown to act as an immunoregulatory signal to suppress pro-inflammatory TNF production[31] and modulating the host immune system.[32]

Figure 3

3D PCA plot of (A) bacteria at genus level comparison (Akkermansia vs Bacteroides vs Bifidobacterium vs Lactobacillus), (B) Lactobacillus bacteria at species level comparison (L. acidophilus vs L. delbrueckii subsp. vs L. fermentum vs L. reuteri), and (C) LAB and non-LAB comparison. Each group was circled by corresponding colors for better visualization. All features from extracellular polar metabolites in the negative mode after data filtering were presented as an example.

Our results indicated that untargeted metabolomics could sensitively detect many important microbial metabolites and use this metabolic information to differentiate gut bacteria at taxonomic levels (e.g., genus and species level).

Meanwhile, we also investigated if untargeted metabolomics could differentiate gut bacteria with their reported metabolic capabilities (such as lactic acid producing ability). Therefore, we divided eight bacteria species into two groups: LAB which containing BA, BL, LR, LD, LA, and LF, as well as non-LAB-containing AM and BT. LAB is recognized as gut bacteria that produce lactic acid as the primary metabolic end product of carbohydrate fermentation.[33] Therefore, other than lactic acid, whole gut bacteria metabolomes acquired from untargeted LC/MS metabolomics were analyzed to see if we can observe the differentiation between LAB and non-LAB based on their metabolic profiles. Similarly, PCA plot pointed out that metabolic profiles obtained from LAB could be separated from the metabolic profiles of non-LAB (highlighted by red and green circles in Figure C). PC 1, 2, and 3 accounted for 22.6, 20.3, and 7.8% of total metabolic variance, respectively. Fourteen metabolites, such as methylacetate, were identified as significantly different metabolites that drive the separation (p-value < 0.05 and VIP score > 1) between LAB and non-LAB (Table S3), including tentative IDs for some metabolites. The relative abundance of methylacetate was detected at a significantly lower (p < 0.001) level in the LAB group, and this compound has been reported to be at a low level from the gut microbiome of cystic fibrosis patients.[34]

Metabolome of the Mixed Culture Can Be Used to Quantitatively Estimate the Bacterial Membership in Simplified Co-Culture Models

In the human large intestine, gut bacteria grow in a complicated environment with interactions with other bacteria species. To mimic the complex gut microbial community, we hypothesized that, using simplified co-culture systems, gut bacterial metabolome from mixed culture can be used to quantitatively estimate the bacterial composition. To test this hypothesis, we utilized a co-culture system of incubating two bacterial species at different ratios (100:0, 80:20, 50:50, 20:80, and 0:100) for half an hour in order to investigate the relationship between metabolic changes and the changes of mixed gut bacterial population. Consistent with the single gut bacterial comparisons reported in the previous section, we mixed LR with BL to represent the genus level co-culture test, mixed LR with LD to represent the species level co-culture test, and mixed LR with AK to represent the LAB and non-LAB co-culture test. From the analyzed co-culture metabolome, the trend of gut bacterial ratio change in the mixed bacterial culture can be visualized based on PCA plots (Figure ). The arrows in each figure represented the trajectory of the metabolic shifts or the co-culture system that is corresponding to the changes of mixed gut bacterial ratio. Gut bacteria metabolome showed the pattern of change based on population ratio increase, which is consistent in three comparison groups. For example, in Figure A, the changing trend of the co-culture metabolome was primarily following PC1 direction as the population ratio of BL increases, but took a sharp turn when the culture was 100% BL, suggesting that metabolome from 100% BL was distinctively different on metabolic features summarized by PC2 compared to the other co-culture metabolome. Meanwhile, in Figure B, the changing trend of metabolic profile-based co-cultures was primarily following PC1 direction as the population ratio of LR increases. Similarly, Figure A,C shows a similar pattern of first following the direction of PC2 when the LR population ratio increases and then it took a sharp turn along the PC1 direction when 100% LR was tested, indicating that metabolome from 100% LR was distinctively different on PC1 compared to the rest metabolomes from other co-cultures.

Figure 4

2D PCA plot of the bacteria mixture at different population ratios for different biological comparison, (A) B. longum subsp. and L. reuteri mixture for genus level comparison, (B) L. reuteri and L. delbrueckii subsp. for species level comparison, and (C) L. reuteri and A. muciniphila mixture for LAB and not LAB comparison. Features from extracellular metabolites in the negative mode after data filtering were presented. Arrows indicated the increasing trend of population ratio of a certain bacteria species in the mixture.

In addition, to analyze the pattern of specific metabolites that changed when we mixed the bacteria at different ratios, we first used a simple equation of multiplexing peak area from single bacteria with their weighted population ratio in the co-culture system to predict the profiles of these metabolic features and then compared peak areas of bacterial co-cultures acquired from the experiment to our predictions. Correlation analysis was performed to calculate the degree of correlations between experimental and predictive peak areas. Top 10 most abundant metabolites with correlation coefficient larger than 0.7 between predicted and experimental results are shown in Figure . In heatmaps, the log transformation of the peak area was done for better visualization. Missing values in Figure C referenced to the negative peak area after PBL subtraction. Left half in the heatmaps were experimental peak areas and right half were predicted peak areas. Overall, a similar trend of changes on the left half and right half was observed, indicating that with changes in the population ratio in the co-cultures, peak areas of many metabolic features would alter accordingly in all three co-culture systems. Specifically, in LR and BL co-cultures, the synchrony of metabolic and bacterial population changes was particularly obvious in top four features at m/z = 82.9544, 84.9513, thioglycolate (m/z = 90.9858, tentatively identified), and 2-hydroxyisobutyric acid (m/z = 103.0406, tentatively identified) with correlation coefficient of 0.765, 0.787, 0.936, and 0.903, respectively (Figure A). While in LR and LD co-cultures, 2-hydroxyisobutyric acid (m/z = 103.0406, tentatively identified), propyl propionate (m/z = 115.0770, tentatively identified), and feature at m/z = 149.0614 with a correlation coefficient of 0.971, 0.900, and 0.910, respectively, were more responsive to the changes in the gut bacteria population ratio (Figure B). In the LAB and non-LAB co-culture, 2-hydroxyisobutyric acid (m/z = 103.0406, tentatively identified), features at m/z = 206.9975, and 298.0694 with correlation coefficients of 0.704, 0.979, and 0.709, respectively, were also having similar changing trends between experimental data and predicted values. Among all annotated features, 2-hydroxybutyric acid was commonly reported in all three co-culture groups. 2-Hydroxybutyric acid has been reported as a microbial metabolite derived from butyric acid and was associated with lactic acidosis and ketoacidosis in humans and diabetes in animals.[35,36] Lactic acidosis is a condition referred to the high level of lactic acid circulating in the blood leading to decreased blood pH.[37] As mentioned above, one of the primary fermentation products from LAB is lactic acid, while non-LABs are not able to produce lactic acid. Moreover, our finding was consistent with the literature supported by the lowest log transformation of peak area of 2-hydroxybutyric acid in AK, a non-LAB in eight selected bacteria species.

Figure 5

Heatmaps of selected predicted and experimental features in different biological comparison groups, (A) B. longum subsp. and L. reuteri mixture for genus level comparison, (B) L. reuteri and L. delbrueckii subsp. for species level comparison, and (C) L. reuteri and A. muciniphila mixture for LAB and not LAB comparison. Features from extracellular metabolites in the negative mode after data filtering were selected. Selected features were further filtered according to Spearman correlation analysis (correlation coefficient > 0.7 or < −0.7) and their absolute peak areas (peak area > 106). Top 15 most abundant features were shown in the heatmaps. Log transformation of the peak area was performed. Missing data points in Figure C were negative values after blank subtraction.

Although utilizing gut bacterial metabolome to estimate bacterial composition in a mixed environment has been complicated by the large number of gut bacteria in the gut and the compositional nature of the metagenomics and metabolomics data, there were studies using linear and nonlinear statistical models to explain the gut microbiome-metabolites relationship. For example, Morton et al. proposed a neural network architecture called MMvec to predict the metabolic profile from a single microbe sequence using multi-omics data sets such as Pseudomonas aeruginosa and its metabolites related to cystic fibrosis. However, because MMvec is based on the co-occurrence probability of each metabolite’s presence given presence of a specific bacteria, the suitability of assessing the statistical significance of interactions using co-occurrence probability remained unclear.[7] In this study, we focused on the experiment-driven estimation of the microbial composition using microbial metabolomes. By utilizing a simplified co-culture system, we were able to understand, from the basics, how microbial metabolomes changed proportionally to the change of bacterial compositions in co-cultures. Similarly, there were other studies that reported microbial metabolite changes after the co-culture system as well. For example, a previous publication reported the increase in carboxylic acids after mixing propionic acid producing bacteria (PAB) with LAB compared to a single LAB culture.[38] Taken together, we believe our study showcased that gut bacteria metabolome has the potential to reflect bacterial composition in the mixed culture.

In conclusion, the data from our analysis of metabolic profiles produced by BA, BL, LR, LD, LA, LF, AM, and BT demonstrated that untargeted metabolomics combined with data processing workflow and statistical analysis method has a great potential for gut bacterial ID and functional understanding in an in vitro system. Using untargeted LC/MS metabolomics, we are able to differentiate gut bacteria at the taxonomy level (genus and species) and of different metabolic capabilities. As a proof-of-concept study, we also acknowledge that our study has limitations. For example, the brief co-culture systems of two bacterial species in a simplified system may not fully capture the long-term gut bacterial interactions, which might lead to more complicated shifts in the mixed bacterial metabolome. Also, there were literature suggesting that the growth medium and growth condition would affect metabolic profiles of bacteria,[39] but we have not fully explored this perspective in our study. Meanwhile, due to the availability of some chemical standards, we were only able to validate a fraction of the potentially useful microbial specific metabolite biomarkers in this study. Moving forward, we will keep these challenges/opportunities in mind and develop more integrated complex culture systems to study the application of metabolome on quantitatively estimating gut bacteria membership in a mixed culture. Also, we will continue to validate the IDs of additional metabolites with commercial or synthetic standards when resources become available and to explore metabolomes acquired from untargeted metabolomics for additional gut bacteria species and a variety of growth conditions in an in vitro system.