Integrated metagenomics identifies a crucial role for trimethylamine-producing Lachnoclostridium in promoting atherosclerosis

The gut microbiota is known to convert carnitine, betaine, and choline to TMA by utilizing three TMA-lyase complexes, cntA/B, yeaW/X, and cutC/D, respectively. CntA, yeaW, and cutC are TMA-lyases, while cntB, yeaX, and cutD are their activators (Fig. 1a). We identified the TMA-lyase sequences of the TMA-producing bacteria based on the three reference resources.

The final 885 TMA-lyase reference sequences were identified from three resources, 451 from the Human Microbiome Project database, 216 from the Unified Human Gastrointestinal Genome, and 218 reference sequences from 4930 species-level genome bins by taking the initial 983 sequences as query (Supplementary Table 1). These sequences were distributed in 216 bacteria species from 102 genera as shown in Supplementary Table 2. Specifically, cntA existed in 15 genera, cntB in 20 genera, yeaW in 15 genera, and yeaX in 17 genera (Fig. 1b). All these genera belong to the phylum of Proteobacteria (Fig. 1b). CutC was distributed in 3 genera from Actinobacteria, 27 from Firmicutes, 18 from Proteobacteria and 4 from other phyla like Bacteroidetes, Desulfobacterota_A, and Fusobacteria. CutD was distributed in 4 genera from Actinobacteria, 39 from Firmicutes, 30 from Proteobacteria and 10 from other phyla like Bacteroidetes, Desulfobacterota_A, Dictyoglomi, Fusobacteria, Spirochaetes, Thermotogae, and Verrucomicrobia (Fig. 1b). Furthermore, cntA/B and yeaW/X coexisted in 15 genera, while cutC/D coexisted in 48 genera (Fig. 1c).

It is worth noting that cntA/B and yeaW/X coexisted in 49 species (Supplementary Table 2). For sequence similarity, the average sequence identity was 81.64% for cntA and yeaW, and 65.63% for cntB and yeaX (Fig. 1d). Compared to cntA/B and yeaW/X, the sequence identity of cutC/D was relatively low; 32.27% for cutC and cntA/yeaW, and 33.60% for cutD and cntB/yeaX (Fig. 1d). Since the TMA-lyases cntA/yeaW/cutC usually coexisted with their corresponding activators cntB/yeaX/cutD and there was high homologous similarity between cntA/B and yeaW/X rather than cutC/D, we focused our attention on cntA and cutC in subsequent analysis.

An integrated reference metagenomic sequence dataset was acquired from several public cohorts (Supplementary Table 3). Multivariable-adjusted analysis showed that population played a major role in the relative abundance (R.A.) of cntA (27.15% contribution to total variations) and cutC (21.02% contribution to total variations, Supplementary Fig. 1). A total of 645 healthy individuals worldwide were included to figure out the R.A. of TMA-producing bacteria with sequence alignment. Five genera containing cntA were found in the human intestine, and Escherichia occupied 66.5% followed by Klebsiella with 17.7% and then Shigella with 15.4% (Fig. 2a). The cumulative R.A. of cntA varied widely among populations (p_org = 0.013). Swedish exhibited the highest cumulative R.A. of cntA (4.55 × 10⁻⁵), while Americans exhibited the lowest cumulative R.A. (2.48 × 10⁻⁶) (Fig. 2a). For cutC, the top five most prevalent cutC-carrying genera were Lachnoclostridium (22.8%), Desulfovibrio (13.5%), Klebsiella (12.5%), Clostridium (11.2%) and Escherichia (7%) (Fig. 2b). The cumulative R.A. of cutC also varied largely among populations (p_org = 6.1e−06). Japanese showed the highest cutC cumulative R.A. (2.66 × 10⁻⁵), while Swedish showed the lowest cumulative R.A. (2.02 × 10⁻⁶).

Subsequence analysis of the TMA-producing bacterial strains was based on the Human Microbiome Project which has defined strain taxonomic information among the three reference sources. We conducted phylogenetic analysis of cntA and cutC in candidate TMA-producing bacterial strains. Totally, 13 strains were found to have detectable R.A. containing cntA, 8 of which belonged to the dominant genera Escherichia (Fig. 2c, left panel). In the phylogenetic classification, the cntA from Shigella was close to Escherichia. Two strains from Klebsiella and one strain from Pseudomonas were farther on the phylogenetic tree, suggesting that their genetic sequences differ widely. The R.A. of cntA in the 13 strains ranged from 1.62 × 10⁻⁶ to 2.55 × 10⁻⁵ and the prevalence ranged from 1.90 to 6.72% (Fig. 2c, right panel).

CutC existed in 30 strains with detectable R.A. in three phyla, Actinobacteria, Firmicutes, and Proteobacteria. There were three main branches in the phylogenetic tree of cutC. The first branch was mainly from Clostridium (R.A., 3.71 × 10⁻⁸–7.54 × 10⁻⁶ and prevalence, 2.20–7.80%). The next branch was mainly from Lachnoclostridium (R.A., 4.54 × 10⁻⁶–1.72 × 10⁻⁵ and prevalence, 3.18–5.53%) and Desulfovibrio (R.A., 3.80 × 10⁻⁷–4.61 × 10⁻⁷ and prevalence, 2.24–4.66%). The third branch was mainly from Escherichia (R.A., 1.06 × 10⁻⁶–3.99 × 10⁻⁶ and prevalence, 1.30–3.41%) and Klebsiella (R.A., 2.25 × 10⁻⁶–1.30 × 10⁻⁵ and prevalence, 1.73–4.97%). Among them, the highest genus containing cutC was Lachnoclostridium (Fig. 2d).

We then explored the correlation of cntA and cutC with various diseases such as, adenoma, colorectal cancer, impaired glucose tolerance, type 2 diabetes, cardiovascular disease, hypertension, and obesity using several public case-control gut metagenomic datasets (Supplementary Table 3). Interestingly, the R.A. of cntA showed no significant differences in the various disease groups compared with healthy individuals. There was an elevated trend but no statistically significant difference between persons with atherosclerosis and the healthy individuals (Supplementary Fig. 2). Surprisingly, cutC showed significantly increased R.A. in the patients with atherosclerosis (Mann–Whitney U test p^a = 0.033) and a significant difference in prevalence of same (Chi-squared test p^b = 0.0091) compared with healthy individuals. Moreover, cutC showed significantly increased R.A. in the patients with obesity (Mann–Whitney U test p^a = 0.013) and a significant difference in prevalence (Chi-squared test p^b = 0.033) compared with healthy individuals (Fig. 3a). No significant differences in the R.A. of cutC were observed in other diseases compared with the healthy controls. The prevalence of cutC was significant higher in patients with colorectal cancer compared with healthy individuals (Chi-squared test p^b = 0.045).

In further analysis of the correlation between the risk of atherosclerosis and cutC-containing genera, we observed that the R.A. of Lachnoclostridium (Mann–Whitney U test p^a = 2.9e−05), Clostridium (p^a = 5.8e−04) and Olsenella (p^a = 0.018) were significantly increased in patients with atherosclerosis compared with the healthy persons (Fig. 3b). Other genera showed no significant differences in the R.A. between the atherosclerotic patients and healthy individuals (Fig. 3b). Lachnoclostridium and Clostridium showed highest prevalence in this cohort. The prevalence of Escherichia (Chi-squared test p^b = 1.6e−06), Klebsiella (p^b = 0.00046) and Desulfovibrio (p^b = 0.024) were significant different. The genera Anaerococcus, Hungatella, and Proteus were not detected in this dataset.

Since cutC but not cntA was significantly increased in patients with atherosclerosis, we then compared the TMA-producing ability of 5 candidate cutC-containing genera with high R.A. We performed homologous modeling, molecular docking, and dynamic simulation analysis between choline and cutC protein from different strains (specific sequence is shown in Supplementary Table 4). The homologous modeled structure of cutC from L. saccharolyticum WM1 gave the lowest binding energy (-4.97 kJ/mol) with choline compared with L. asparagiforme DSM 15981 (−4.49 KJ/mol) and L. phytofermentans ISDg (−4.41 KJ/mol). In addition, the binding energy between choline and cutC from L. saccharolyticum WM1 was lower than the other four cutC-containing genera (from −4.26 to −4.82 KJ/mol) (Fig. 4a). The seven active site residues of cutC were all conserved among these cutC-containing genera (Fig. 4b)²⁹. However, the site residues close to the seven active sites, such as V225, M344, V404, and M406, appeared to differ among these genera, possibly accounting for the differences in the cutC enzyme activity. The binding conformation between choline and cutC from L. saccharolyticum WM1 was relatively stable within 100 ns by the dynamic simulation (Fig. 4c). In this binding conformation, choline exhibited three hydrogen bonds with two key residues, i.e., Cys497 and Glu499 (Fig. 4d). These findings demonstrated that L. saccharolyticum WM1 exhibited favorable binding affinity to substrate choline. The binding conformation of cutC from other candidate cutC-containing genera are showed in Supplementary Fig. 3.

L. saccharolyticum WM1 was chosen for subsequent in vitro- and in vivo- experiments. The TMA-producing activities of several cutC-containing strains were then investigated by incubating each strain with the substrate choline. The concentration of TMA was accurately quantified using an isotope-labeled internal standard method by liquid chromatography-triple quadrupole mass spectrometry. L. saccharolyticum WM1, effectively converted choline to TMA in a time-dependent manner at 12, 24, 36, 48, and 96 h (Fig. 4e). The conversion rates of L. saccharolyticum WM1 for various concentrations of choline (2 mM, 4 mM and 6 mM) were higher than 98.68% (Fig. 4f). C. sporogenes BNCC 104015 showed moderate conversion rates for choline (63.79–67.51%). B. fragilis BNCC 352061 and P. intermedia BNCC336948, two negative control strains without cutC enzymes, showed no capability in transforming choline to TMA (less than 0.1%) (Fig. 4f).

After showing that L. saccharolyticum effectively converted choline to TMA by in vitro incubation, we then investigated the possible implication of L. saccharolyticum in atherosclerosis in ApoE^−/− mice. We compared the atherosclerotic phenotypes of mice in the following four groups: normal group, normal + L. saccharolyticum group, choline group, and choline + L. saccharolyticum group (Fig. 5a). L. saccharolyticum administration notably elevated the level of L. saccharolyticum in fecal samples (Fig. 5b). The serum levels of TMAO, the downstream metabolite of TMA, significantly increased in the choline + L. saccharolyticum group compared with the normal group (p = 9.9e−04), normal + L. saccharolyticum group (p = 3.4e−05) and choline group (p = 5.3e−04) (Fig. 5c). Compared with the normal control, the plaques and the stenosis of the blood vessels showed an increased trend in the L. saccharolyticum group and the choline group. The plaques and the stenosis of aortic arch were greatly pronounced in the choline + L. saccharolyticum group (Fig. 5d). More importantly, the level of L. saccharolyticum in the fecal samples was positively correlated with TMAO level in serum (R = 0.31, p = 0.0052), aortic lesion area (R = 0.54, p = 0.00024) and vessel lesion area (R = 0.48, p = 0.0085, Supplementary Fig. 4a). Administration of choline increased the thickness of the vessel wall and inflammatory cells infiltration and lipid deposition (Fig. 5e). These changes were more obvious in the choline + L. saccharolyticum-treated mice (Fig. 5e). L. saccharolyticum and choline administration triggered inflammation in the vessel, as indicated by gene inductions of Il-1β, Tnf-α, Icam-1 and Vcam-1, Mcp-1, Cd68, and F480 (Fig. 5f). Similarly, L. saccharolyticum administration increased circulating contents of FFAs, while other metabolic parameters remained unaffected (Supplementary Fig. 4b). These results indicated that L. saccharolyticum abundance impaired the aorta and favored the formation of atherosclerosis, probably due to the increased production of TMAO.

The first reference genomes resource was from the Human Microbiome Project in September 2014 which contained 1751 bacterial strains covering 1253 species. The second reference genomes resource, thus, the unified Human Gastrointestinal Genome, comprised of 204,938 nonredundant genomes from 4644 gut prokaryotes³⁶, and the third reference genomes resource contained recapitulated 4930 species-level genome bins³⁷ (Supplementary Fig. 5). The genes and related proteins of these bacterial genomes were predicted by MetaGeneMark (v2.8)³⁸, and the taxonomic information of these genes/proteins were directly extracted from their respective strains.

The query sequences of choline-TMA-lyase (cutC/D) were collected from the RefSeq database of the National Center for Biotechnology Information database using the keywords “choline trimethylamine-lyase” (575 sequences) and “choline TMA-lyase-activating enzyme” (347 sequences). Carnitine-TMA-lyase (6 sequences for cntA; 6 for cntB)¹² were collected from the original CaiT protein sequence (CAA52110↗)³⁹. γ-butyrobetaine-TMA-lyase (30 sequences for yeaW; 19 for yeaX) were collected with the original sequences (yeaW: dioxygenase, GeneID 6060925; yeaX: oxidoreductase, GeneID 6060982)¹³. An initial total of 983 TMA-lyase sequences were taken as query and BLASTP was used to search for candidates with parameters of e-value of 1e−5 and sequence identity of 45% (Supplementary Fig. 6) as cutoff in the reference genomes resources.

The public metagenomic sequence data of individuals were collected from 11 populations from six continents, including Hadza^40,41 (PRJNA278393, pre-agricultural communities in Tanzania), Chinese⁴² (PRJNA422434), South Korean⁴³ (PRJEB1690), Japanese⁴⁴ (PRJDB3601), Australian (PRJEB6092), Austrian⁴⁵ (PRJEB7774), Dane⁴⁶ (PRJEB2054), French⁴⁷ (PRJEB6070), Swedish⁴⁸ (PRJEB1786), Peruvian⁴⁹ (PRJNA268964), and American²⁷ (PRJNA275349 and PRJNA48479). A restriction criterion was used to construct metagenomic datasets of healthy individuals from each country. The individuals with age >3 and BMI < 30 were included and subjects were excluded if they had definite diseases like inflammatory bowel disease, liver cirrhosis, and colorectal cancer.

Public metagenomic datasets of persons with different diseases were included to study the differences in the relative abundance (R.A.) of the TMA-lyase between healthy persons and the persons with diseases, such as colorectal cancer, adenoma, type 2 diabetes, impaired glucose tolerance, hypertension, obesity and atherosclerosis (ERP023788 from the European Bioinformatics Institute database)⁶.

A total of 2134 individuals were downloaded with their accession codes and the R.A. at different taxonomic levels in each cohort calculated (Supplementary Table). 3

All raw reads were assessed and filtered using the FASTX-Toolkit (v 0.0.13) with parameters: “-Q33 -q 20 -p 80”, and high-quality microbiome sequencing reads were assembled with SOAPdenovo2 (v 2.04)⁵⁰ using the parameters: “-avg ins 250 -K 63 -k 45 -R Y -M 3”. After assembling process, contigs with at least 500 bp were further used to predict the genes using MetaGeneMark (v2.8)³⁸ with the parameters: “gmhmmp -a -d -f G -m MetaGeneMark_v1.mod”. A non-redundant protein set was then constructed by pair-wise comparison of all protein sequences within populations using BLAT (v 35×1)⁵¹ at 95% identity and 90% overlapping thresholds.

Protein sequences were aligned to the NCBI-NR database using BLASTP (v 2.2.29) with the parameters: “-e-value 1e−5”. The taxonomic assignments and functional annotations were constructed in MEGAN (v 5.2.3) with lowest common ancestor algorithms. The high-quality reads from each individual were aligned against the gene catalog using SOAPdenovo2 (v 2.04) with the parameters: “-r 2 -m 150 × 350 -v 5”. The quantification of each protein sequence of each individual was based on two steps: (i) Calculation of the entire copies of each gene with correct insert-size; (ii) Calculation of the R.A. of each gene in each individual, only if the other reads mapped outside the genic region, as previously described in detail⁴². The cumulative R.A. was calculated by the sum of R.A. of each genus in each population. TMA-lyase was identified in the above population datasets using BLASTP with the parameters: “-evalue 1e-5 -outfmt 0 -max_target_seqs 3” and 885 query sequences of e-value at 1e−5 and 80% sequence identity as the cutoff values. Persons without zero R.A. were employed for the comparisons within the various cohorts.

The protein sequences of cntA and cutC were aligned using mafft v7.455⁵², and the resultant multiple sequences were trimmed for poorly aligned positions with Gblock 0.91b⁵³. RAxML v8.2.12⁵⁴ was used to build the most likely phylogenetic tree of genomes with the parameters “-m GTRCAT” and protein with the parameters “-m PROTGAMMAILGX”. R package “ggtree”⁵⁵ was used to construct the phylogenetic tree.

Protein Homology/analogY Recognition Engine V 2.04⁵⁶ based on intensive mode was used to predict the homologous structures of cutC in strains from Lachnoclostridium, Clostridium, and other abundant cutC-containing genera. The details of cutC sequence used for modeling are provided in Supplementary Table 4. Choline ligand was downloaded from ZINC database⁵⁷. The AutoDock (v.4.2.6)⁵⁸ was employed to generate an ensemble of docked conformations for each ligand bound to its target. We used the genetic algorithm for conformational search and performed 100 individual GA runs to generate docked conformations for each ligand.

For each docked conformation, the top-ranked docking pose was optimized in the binding pocket and used as the initial geometry in molecular dynamics. We performed 100 ns molecular dynamics simulation for choline-bound states. All of the molecular dynamics simulations were performed with Amber16⁵⁹.

To verify the in vitro activity of TMA-producing bacteria, Lachnoclostridium saccharolyticum WM1 (ATCC 35040) was cultured anaerobically on ATCC medium 1118 at 37 °C. Clostridium sporogenes (BNCC 104015) was cultured anaerobically in Trypticase soy agar/broth with defibrinated sheep blood at 37 °C. Two representative intestinal strains, Prevotella intermedia (BNCC 352061) and Bacteroides fragilis (BNCC336948), were selected as negative controls and were cultured in anaerobically sterilized Trypticase soy agar/broth with defibrinated sheep blood at 37 °C.

For TMA assessment, bacterial cultures were inoculated with choline chloride (Sigma-Aldrich, USA) in sterile tubes and incubated anaerobically (80% N₂ and 20% CO₂) at 37 °C (typically OD_600nm ~ 1.0). The cultures were centrifuged at 16,200 × g for 10 min at 4 °C.

Quantification of TMA was performed on a liquid chromatography-20A system coupled to a triple quadrupole mass spectrometer (Shimadzu 8050, Japan) using an internal standard method¹⁵. Briefly, 10 μL of bacterial culture medium was first mixed with 10 μL of isotope-labeled d9-TMA internal standard (100 μg/mL). Then, 960 μL of n-hexane/n-butyl alcohol (2:1, v/v) and 20 μL of 1 M NaOH were added and vortexed for 3 min to fully extract TMA. After centrifugation (16200 × g, 4 °C, 1 min), 500 μL of the organic layer was transferred to a sealed container and acidified with 200 μL of 0.2 N formic acid. Finally, the mixture was vortexed for 3 min and centrifuged at 4 °C for 1 min. An aliquot of 1 μL of the aqueous layer was injected for LC-MS/MS analysis.

LC separation was carried out on a HILIC column (100 × 2.1 mm, 1.7 μm; WATERS). Pure water and acetonitrile both containing 0.1% formic acid were used as mobile phases A and B, respectively, delivered at a flow rate of 0.4 mL/min. An isocratic elution with phase B/phase A (80:20, v/v) was employed and the detection was operated in the positive ion mode. Multiple reaction monitoring (MRM) transitions were performed at m/z 60.08 → 44.04 for quantitative detection of TMA, and m/z 69 → 49 for d9-TMA. The ESI source parameters were as follows: nebulizing gas flow, 3 L/min; heating gas flow, 10 L/min; drying gas flow, 10 L/min; interface temperature, 300 °C; DL temperature, 250 °C; and heat block temperature, 400 °C. Multiple reaction monitoring transitions were performed at m/z 60.08 → 44.04 for quantitative detection of TMA, and m/z 69 → 49 for d9-TMA.

The care and treatment of mice were performed in accordance with the Provisions and General Recommendation of Chinese Experimental Animals Administration Legislation and the study was approved by the Animal Ethics Committee of China Pharmaceutical University (No. 2021-04-002). All procedures conformed to the European Parliament directive on the protection of animals used for scientific purposes (Directive 2010/63/EU) or the NIH Guide for the Care and Use of Laboratory Animals.

Apolipoprotein E knockout mice (C57BL/6J ApoE^−/−) were housed in a pathogen-free environment at a temperature of 22–24 °C, humidity 40–60%, and a strict 12 h light cycle. After acclimatization for a week, both female and male ApoE^−/− mice were randomized into 4 groups: normal group, mice fed standard chow and gavaged with the sterile medium; normal + L. saccharolyticum group, mice fed standard chow and gavaged live L. saccharolyticum at a dose of 5 × 10⁸ CFUs/100 μl; choline group, mice fed chow supplemented with 1.0% choline (Sigma-Aldrich, USA) and gavaged with the sterile medium; and choline + L. saccharolyticum group, mice fed standard chow supplemented with 1.0% choline and gavaged with live L. saccharolyticum at a dose of 5 × 10⁸ CFUs/100 μl. The mice number was 20 (half male and half female) for each group. The group sizes were selected according to the minimum experimental requirements and natural factors such as fight-related injury.

After 20 weeks, overnight-fasted mice were anesthetized with 4% isoflurane and their blood was collected from the ventriculus dexter for the assay of glucose (Solarbio, BC2505), non-esterified free fatty acids (Solarbio, BC0595), triglycerides, total cholesterol, low-density lipoprotein, and high-density lipoprotein using commercial Kits (Jiancheng Biotechnology Co., Ltd., Nanjing, China). Their stool samples were collected and immediately stored at −80 °C until further analysis. After that, the mice were euthanized by exsanguination following the guidelines of the Institutional Animal Care, the cardiac and aortic tissues were immediately removed, rinsed with ice-cold physiological saline, and stored at −80 °C or fixed in 4% paraformaldehyde until further analysis.

Cardiac and aortic tissues were fixed with 4% (w/v) paraformaldehyde overnight. Then the tissues were embedded in paraffin, cut into 5 μM slices, and stained with hematoxylin and eosin or oil red O. Images were captured using NanoZoomer 2.0 (Hamamatsu, Japan).

Perivascular adipose tissue and adventitious blood vessels were carefully removed and the prepared aorta was stained with Sudan IV solution (0.1% Sudan in 50% aceton) for 6 min, followed by de-staining in 80% ethanol for 5 min. The images were captured with a Canon EOS 80D Digital Camera. Areas stained red were considered atherosclerotic lesions and were quantified using ImageJ software.

Fecal samples of each mouse were used for metagenomic DNA extraction using TIANamp Stool DNA Kit according to the manufacturer’s instructions. DNA concentration was measured using NanoDrop 2000 spectrophotometer (Biotek, Germany). Quantitative PCR was carried out using the LC480 detection system (Roche Diagnostics, Basel, Switzerland) and HiScript Q RT SuperMix (Vazyme biotech, Nanjing, China). Primers used in this work are listed in Supplementary Table. 5

An isotope-labeled internal standard method was employed for the quantitative analysis of TMAO in serum. Targeted detection of TMAO was performed on an Agilent 1290 infinity liquid chromatography system coupled to a triple quadrupole mass spectrometer (Agilent 6470, USA) operated in the positive ion mode. An aliquot of 10 μL serum was precipitated by adding 980 μL of acetonitrile/ water/formic acid (94:5:1, v/v/v) solution and 10 μL of internal standard working solution (2 μg/ml of d9-TMAO), followed by vortex-mixing for 30 s. Precipitated protein was subsequently removed by centrifugation at a speed of 16200 × g at 4 °C for 10 min. Then, 5 μL of supernatant was analyzed by ultra-performance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS).

Chromatographic evaluation was achieved on an ACQUITY UPLC^® BEH HILIC column (2.1 × 100 mm, 1.7 μm) maintained at 40 °C. The mobile phase consisted of (A) 10 mM ammonium acetate aqueous solution and (B) 10 mM ammonium acetate water/acetonitrile (1:9) solution delivered at a flow rate of 0.4 mL/min. The gradient elution program was 5–80% B at 0–7 min, 80–100% B at 7–12 min, 100% B at 12–13 min, and then back to initial conditions, with 2 min for equilibration. Targeted detection of TMAO was performed on an Agilent 1290 infinity LC system coupled to a triple quadrupole mass spectrometer (Agilent, LC-MS/MS 6470) operating in the positive ion mode. Multiple reaction monitoring (MRM) mode was performed and the ion transitions monitored were m/z 76 → 58 for TMAO and 85 → 66 for d9-TMAO. The detailed MS parameters were set as follows: fragmental voltage, 100 V; capillary voltage, 3500 V; nebulizer gas, 35 psig; drying gas flow rate, 10 L/min; drying gas temperature, 300 °C.

Statistical differences between the two groups were calculated using the Wilcoxon rank-sum test. The multivariable-adjusted analysis using linear regression model was performed to evaluate the R.A. of TMA-lyase against other confounding factors. The Kruskal–Wallis test was for multi-group comparison followed by Holm–Bonferroni correction. The differences among groups in animal experiments were analyzed by single-factor ANOVA with Tukey HSD test. All analyses were performed using R 3.6.1, and p < 0.05 were considered statistically significant.

All public metagenomic data used in this manuscript were provided their web links or the accession codes in “Methods” section.

All code used for statistical analysis is available from the corresponding author on reasonable request.