The fecal samples were collected from 20 healthy adult individuals (median age 30 years) who had not received antibiotics in the previous 3 months. Informed written consent was obtained from each donor, and a donor number was randomly assigned for blinding. Fecal samples were collected by donors immediately after defecation using a sterile container. An atmosphere generator (AnaeroGen™; Thermo Fisher Scientific, Pratteln, Switzerland) was added into the container to produce anaerobic conditions during transportation. The samples were transferred within 2 hours into an anaerobic chamber (10% CO₂, 5% H₂, and 85% N₂) (Coy Laboratories. Grass Lake, MI, USA), aliquots were stored at −80°C for DNA isolation, and the samples were used to immediately inoculate batch fermentations, as outlined below. The Ethics Committee of ETH Zürich exempted this study from review because the sample collection procedure was not performed under conditions of intervention and was done anonymously.

A 10% w/v (weight/volume) fecal suspension was prepared inside an anaerobic chamber by resuspending 1 g of fresh feces in 10 mL of sterile peptone water in a 50 mL sterile Falcon™ tube with four sterile glass beads (5 mm). The tube was vortexed for 5 min and the mixture was left to stand for 5 min to allow the heavy particles to settle. The resulting supernatant was used to inoculate (0.1% w/v) serum flasks containing 20 mL of Macfarlane medium prepared according to Bircher et al. (2018b) and supplemented with 200 nM of PhIP (Chemie Brunschwig AG, Basel, Switzerland) and 100 mM of glycerol (Sigma-Aldrich, Buchs, Switzerland). This concentration of glycerol was chosen based on previous reports (from De Weirdt et al., 2010; Fekry et al., 2016; and Engels et al., 2016a), and was the minimum dose at which PhIP transformation was detected. The fermentations in the Macfarlane medium without added glycerol served as a control. The additional fermentations were supplemented with 1% of a 16-h overnight culture of A. hallii DSM 3353 [final concentration in the medium measured by quantitative PCR (qPCR) as 10⁶ cells/mL] grown in a yeast extract—casein hydrolysate—fatty acid (YCFA) medium prepared as previously described (Bircher et al., 2018a). Triplicate fermentations were incubated at 37°C with magnetic stirring (120 rpm) and sampled and analyzed as previously described (Ramirez Garcia et al., 2021). Briefly, 1-mL samples were taken after 0 h, 6 h, and 24 h of fermentation using sterile syringes flushed with CO₂ and centrifuged in sterile Eppendorf™ Tubes™ (14,000 × g for 10 min at 4°C).

DNA was isolated from the pure culture of A. hallii inoculum, from the fecal samples, and from the fermentation samples using the FastDNA™ SPIN Kit for Soil (MP Biomedicals, Zurich, Switzerland), following the manufacturer’s protocol. The concentration of the isolated fecal DNA was assessed with a NanoDrop^® 1000 instrument and a Qubit™ 4 fluorometer (Thermo Fisher Scientific), whereas the DNA integrity was verified via gel electrophoresis. A previously established qPCR assay was used to quantify the abundance of A. hallii in the overnight culture used for A. hallii supplementation, using specific primers for gdh (forward primer 5′-CGTTATGCTCCATTTAATGCT-3′ and reverse primer 5′-CCAAGGAGTATCATCACCATC-3′) (Ramirez Garcia et al., 2021). The reactions were performed using a Roche LightCycler^® II device (Roche Diagnostics, Rotkreuz, Switzerland). For shotgun metagenome sequencing, all fecal samples were diluted to 25 ng/µL⁻¹ and sent for sequencing (Novogene Co. Ltd., Cambridge, UK), with the number of reads/samples ranging from 405,268,068 to 620,549,190. The sequencing was performed using the Illumina NovaSeq 6000 platform in 150-bp paired-end mode. The number of generated raw read pairs ranged between 405 million and 620 million, with an average of 466 million.

The quality control of raw reads was performed with BBMap (v.38.71) by removing adapters, reads mapping to the human genome, and low-quality reads (Bushnell et al., 2017). The filtered reads were merged with a minimum overlap of 16 bases and were assembled using the SPAdes assembler (v3.14) in metagenomic mode (Bankevich et al., 2012). Gene calling on the assembled scaffolds was performed using Prodigal (Hyatt et al., 2010), and the genes were clustered using CD-HIT-EST (Li and Godzik, 2006) with a cutoff value of 95% identity and 90% overlap of the shorter gene. The reads from all samples were mapped against all representative genes. To account for the differences in gene size, the number of inserts mapped were normalized to represent 1,000-bp genes. The abundance of each gene in each sample was normalized using the median abundance of the 10 single-copy marker genes described by Milanese et al. (2019) to estimate gene copies per bacterial cells, which was expressed per 10¹⁰ cells in the manuscript.

The taxonomic profiles for each sample were generated by profiling quality-controlled metagenomic reads using the metagenomic-based operational taxonomic units (mOTUs) profiler version 3.0.1 using default parameters (Milanese et al., 2019).

The contigs harboring genes belonging to the pduC/gdh-annotated clusters from the gene catalog were identified. The collections of metagenome-assembled genomes (MAGs) that were specific for each donor (i.e., list of contigs belonging to a particular MAG) were then used to link pduC/gdh genes to bacterial populations and their Genome Taxonomy Database (GDTB) assignments. When the identified contigs were not assigned to any MAG, the taxonomy of the gene belonging to pduC-annotated clusters was determined by aligning the corresponding contig sequence against the National Center for Biotechnology Information (NCBI) non-redundant nucleotide database (nt).

The merged mOTU profiles were loaded into R (The R Foundation for Statistical Computing, Vienna, Austria) as phyloseq objects (McMurdie and Holmes, 2013), and data analysis was conducted using DivComAnalyses R functions (Constancias and Sneha-Sundar, 2022). The taxonomic structures were clustered to determine the enterotype, as previously defined (Arumugam et al., 2011). The principal coordinate analysis was used to depict similarity between donors’ microbiota. The correlation between family-level proportion and ordination pattern was evaluated, and the top 10 significantly correlated families were displayed (p < 0.05).

The association between microbiota and metadata (i.e., PhIP transformation, age, diet) was investigated at the alpha/beta and mOTU abundance levels. A permutational multivariate analysis was used to identify marginal associations between community structure and covariates. In a second step, distance-based partial redundancy analysis was used to evaluate the association between microbiota structure and PhIP transformation while controlling for differences in the microbiota associated with enterotypes.

The assembled scaffolds were length filtered (≥ 1,000 bp), and quality-controlled metagenomic reads from all samples were individually mapped against the scaffolds of each sample using the Burrows–Wheeler aligner (v0.7.17-r1188) (Li and Durbin, 2009), allowing for secondary alignments. The alignments were filtered to > 45 bases, with an identity of ≥ 97% and covering ≥ 80% of the read sequence. To provide within- and between-sample coverages for each scaffold, the resulting binary alignment map (BAM) files from the previous step were processed using MetaBAT2 (v2.12.1) (Kang et al., 2019). The scaffolds on all samples were finally binned with MetaBAT, and the quality of the resulting bins was assessed using CheckM (v1.0.13) (Parks et al., 2015). Only the bins with a completeness level of ≥ 50% and a contamination level of < 10%, and which belonged to the domains of Bacteria or Archaea as reported by Anvi’o (v5.5.0) (Eren et al., 2015) were used as MAGs for further analysis. The taxonomic annotation of MAGs was performed using GTDB-Tk (v1.5.0, database R06-RS202) (Chaumeil et al., 2019). The MAGs assigned to the Anaerobutyricum genus were subjected to the Anvi’o pangenomic workflow, which identified two fragmentated MAGs in donors 2 and 15. The FASTA files corresponding to these MAGs were concatenated and subjected to CheckM and GToTree together with the other MAGs for final completeness and contamination estimation.

To identify the presence of the pdu-cob-cbi-hem operon in the MAGs assigned to the genus Anaerobutyricum, the MAGs were annotated using Prokka (Seemann, 2014) and the resulting annotation was manually screened for the pdu-cob-cbi-hem operon genes. Publicly available genomes of the strains A. hallii DSM 3353 and A. soehngenii L2-7 (DSM 17630) were annotated and used as reference strains.

The consumption of glycerol and the production of SCFA and 1,3-PD was monitored using high-pressure liquid chromatography—refractive index (HPLC-RI), as described before (Zhang et al., 2019b). In brief, culture supernatant was filtered with a 0.45-µm nylon filter (Phenomenex, Basel, Switzerland), diluted 1: 1 in H₂O, and transferred into an HPLC vial with a 250-µL conical glass insert (BGB Analytik, Boeckten, Switzerland). The samples were eluted isocratically with 10 mM H₂SO₄ at 40°C and at a flow rate of 0.4 mL/min⁻¹ using an Aminex HPX-87H column (300 mm × 7.8 mm, with a 9-µm particle size; Bio-Rad Laboratories, Reinach, Switzerland). Glycerol, 1,3-PD, and SCFA were quantified by independent calibration with external standards (all obtained from Sigma-Aldrich). The detection limits were 0.7 mM for acetate, propionate, and butyrate, 0.9 mM for glycerol, and 0.2 mM for 1,3-PD.

The transformation of PhIP into PhIP-M1 was determined via nanospray liquid chromatography coupled to tandem mass spectrometry (nanoLC-MS)/MS using a nanoACQUITY Ultra Performance LC™ (UPLC^®) System (Waters Corporation, Milford, MA, USA) attached to a TSQ Vantage™ triple-stage quadrupole mass spectrometer (Thermo Fisher Scientific), as described previously (Ramirez Garcia et al., 2021).

The calibration curves were generated using the ratio between the peak area of 200 nM of the internal standard 5-amino-6-methylbenzamidazolone (AMBI) and of 5–250 nM PhIP and PhIP-M1. The transformation of PhIP was reported as a percentage of the initial PhIP transformed to PhIP-M1.

R (version 4.2.2) was used for the statistical analyses. Based on the Shapiro–Wilk test, the data displayed non-normal distribution. Therefore, the differences between the groups for PhIP transformation, glycerol consumption, SCFAs, 1,3-PD, and A. hallii abundance (low vs. high) were assessed using a two-sided non-parametric Wilcoxon test. Adjusted p-values of ≤ 0.05 were considered to be significant. Spearman-ranked correlations were used to investigate associations between (i) observed species (i.e., number of mOTUs) and Shannon and Simpson diversity indices (i.e., diversity of mOTUs in a community), or (ii) mOTU proportions of the top 200 mOTUs and PhIP transformation for which only correlations with adjusted p-values of ≤ 0.05 and |Spearman’s rho| ≥ 0.6 were considered. Computed p-values were false discovery rate (FDR)-adjusted for multiple comparisons.

Raw sequence data were deposited with NCBI (in the Sequence Read Archive) and are available under BioProject PRJNA720271.

To understand the diversity of the microbiomes used for this study, the microbial composition of fecal samples collected from 20 young healthy adults was assessed by shotgun metagenomics ( Figure 1 ). Collectively, we identified 1,839 individual mOTUs within all 20 fecal samples. In the samples, taxa belonging to the Firmicutes phylum were the most abundant of all microbiota, with relative abundances of between 48.4% and 90.5%. This was followed by taxa belonging to the Bacteroidetes phylum, which ranged between 1.9% and 38.4% ( Figure 1A ). With regard to community diversity values, these ranged from 2.9 to 4.9 by the Shannon index, and 0.91 to 0.98 by the Simpson index, and individual fecal microbiotas harbored from 162 (D.13) to 701 (D.08) mOTUs ( Figure 1B ). To identify enterotype-associated clusters of the microbiota, a beta-diversity analysis (i.e., comparison of samples based on overall microbiota similarity) based on the Jensen–Shannon divergence was computed at the genus level. The analysis assigned six microbiota to the Bacteroides enterotype, nine to Prevotella, and five to Ruminococcus ( Figure 1C ).

As a next step, the proportion of taxa contributing to the gdh prevalence in each fecal microbiota was evaluated by creating a gene catalog from the shotgun metagenomes to quantify the proportion and distribution of gdh-harboring taxa in the fecal microbiota. In the 20 fecal microbiota, the total number of gdh copies expressed per 10¹⁰ bacterial cells ranged from 1.32 × 10⁸ to 1.15 × 10⁹ ( Figure 2A ). When looking into specific gdh-harboring taxa, a total of 17 taxa harboring gdh [i.e., the propanediol dehydratase large subunit (pduC)—functional annotation from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database K01699] were identified ( Figure 2B ). Specifically, the dominant gdh-harboring taxa in the microbiota of 18 donors was A. hallii, representing between 33% (D.20) and 94% (D.11) of the total gdh relative abundance of the samples. When the most abundant gdh-harboring taxon was not A. hallii, it was Faecalicatena gnavus (78%) for D.14 or a taxon that could not be assigned in D.18, accounting for 56% of the total gdh community ( Figure 2B ). To further explore such unassigned gdh-harboring taxa in D.18, it was BLASTed and could be identified with Alkalimarinus sp. SCSIO (84% identity and 13% electable coverage). Besides A. hallii and F. gnavus, other taxa that contributed significantly to the gdh communities in the fecal microbiomes were Blautia obeum (present in the feces of 18 donors, representing between 3% and 40% of the total gdh), an unidentified species of the Blautia genus (present in the feces of nine donors at between 0.6% and 9% of the total gdh), and Flavonifractor plautii (present in the feces of eight donors at between 1% and 11% of the total gdh) ( Figure 2B ).

Given that A. hallii was the most abundant gdh-harboring taxa in the fecal microbiome, we further estimated its proportion on the basis of both a taxonomic marker (i.e., 16S rRNA) and the proportion of a functional gene of A. hallii (i.e., gdh) in the fecal samples. At first, we evaluated the coherence of these two approaches and found a significant correlation (p = 0.00043), thus validating the results previously obtained for A. hallii and gdh for all the fecal samples ( Figure 3A ). Based on the A. hallii proportion, 1% was arbitrarily chosen to stratify the fecal microbiota in two groups defined as “low” (A. hallii < 1%) and “high” (A. hallii > 1%), and 14 microbiota were assigned to the “low” group and six microbiota to the “high” group ( Figure 3B ). Importantly, A. hallii identified by (Eubacterium hallii) mOTU_v3_03632 was the only taxa significantly associated with PhIP transformation in the 20 fecal metagenomes ( Supplementary Table 1 ).

To further investigate A. hallii populations, we used genome-resolved MAGs to compare the composition of the cbi-cob-hem-pdu operon in each fecal sample ( Figure S1 ). Overall, binning of the obtained assembled contigs from all samples resulted in 2,891 MAGs. These MAGs were clustered at 95% similarity for species identification, resulting in a collection of 850 species-level MAGs. In all microbiomes, we recovered one MAG belonging to the A. hallii species-level MAG cluster. The taxonomy of those MAGs matched with A. hallii, with the only exceptions being D.14 and D.03, in which one MAG assigned to an unidentified species of the genus Anaerobutyricum was recovered ( Figure S1 ). Despite the methodological limitations of MAG reconstruction related to fragmented MAGs, we identified donor-specific A. hallii populations that differed in the content of the genes required for acrolein production. Specifically, nine A. hallii MAGs harbored a pdu-cob-cbi-hem operon with highly similar gene content and organization to A. hallii DSM 3553 and A. soehngenii L2-7. The cbi and hem genes were not recovered in MAGs from D.17, D.20, D.10, D.06, or D.19, whereas the same MAGs harbored only two cob genes (D.19 only one). ( Figure S1 ).

Once we identified A. hallii as being the main contributor of gdh activity for 18 fecal microbiota and confirmed that the 20 diverse fecal microbiota of healthy adults could be stratified by proportion of A. hallii mOTUs, we hypothesized that the groups would differ in their actual capacity to transform PhIP to PhIP-M1. Thus, fecal batch fermentations were conducted anaerobically in the presence of PhIP (200 mM) and glycerol (100 mM), with or without supplementing additional A. hallii, and samples were analyzed after 24 h and 48 h for PhIP transformation ( Figure 4A ). In the presence of glycerol, the microbiota with high A. hallii transformed more of the PhIP than the microbiota with low A. hallii did [after 6 h, median of 2% versus 0% (p = 0.17), and after 24 h, 9% versus 4% (p = 0.0508), respectively]. In the same low A. hallii group, when fecal microbiota were supplemented with A. hallii at a final concentration of 10⁶ cells per mL a significant increase in PhIP transformation was observed after 6 h compared with glycerol alone [median of 4% compared with 0% transformation in the presence of glycerol only (p = 0.007)] ( Figure 4A ). Finally, when fecal microbiota were incubated for 24 h in controlled batch fermentations without glycerol, no PhIP transformation was detected (data not shown). In line with the batch fermentation results, PhIP was better depleted with increasing proportions of resident A. hallii in all 20 fecal microbiota, at both 6 h and 24 h ( Figure S2A ). This association was not observed for any other taxa. On the other hand, PhIP transformation could not be associated with either alpha- or beta-diversity metrics ( Figure S3 and Table S1 ). Furthermore, no significant correlation between PhIP transformation and community structure was detected when we controlled for variations associated with enterotype (p = 0.45; Table S2 ).

For PhIP transformation to occur, acrolein has to be produced from glycerol via the production of the intermediate 3-HPA, leading to the generation of 1,3-PD as the final metabolic product (Vollenweider and Lacroix, 2004). Therefore, besides PhIP transformation, the impact of A. hallii on the consumption of glycerol and the production of 1,3-PD was also assessed in the same batch fermentations after 6 h and 24 h. 1,3-PD production was used as a validation for the production of HPA in equilibrium with acrolein, both not detectable in complex cultures (Ramirez Garcia et al., 2021). When stratifying the microbiota in low and high resident A. hallii, consumption of added glycerol after 6 h was comparable for both groups. However, aligning with the previous observations, after 24 h, an increase in glycerol consumption was observed in both groups when the microbiota were supplemented with glycerol alone (median of 18% and 14% (p = 0.0495) in the low and high groups, respectively) ( Figure 4B ). On supplementing microbiomes with A. hallii, a significantly reduced glycerol consumption was observed in only the low A. hallii group compared with glycerol alone after 24 h (median of 8% consumed versus 18% consumed in the presence of glycerol alone, Figure 4B ). In line with this, glycerol consumption significantly positively correlated with resident A. hallii proportion after 24 h (p = 0.59 and p = 0.006; Figure S2B ). Concurrently, a decrease in 1,3PD production was observed ( Figure 4C ), which negatively correlated with the abundance of resident A. hallii in the microbiota ( Figure S2C ).

To assess whether the abundance of resident A. hallii and supplementing with A. hallii played roles in the fermentation activity in the presence of glycerol, total SCFAs (as sums of acetate, propionate, and butyrate) and individual metabolites produced during the fecal fermentations were compared with control fermentations without glycerol. After 6 h, a significant decrease in total SCFAs relative to the control group was observed in both the low (median of −18% versus −39% without and with A. hallii, respectively) and high (median of −33% and −50% without and with A. hallii, respectively) groups when A. hallii was supplemented, as compared with glycerol alone ( Figure 4D ). The same trend was observed after 24 h, when a significant decrease of total SCFA production relative to the control was observed in the presence of glycerol and A. hallii in the high group (median of −75%), as compared with glycerol alone (median of −57%). Specifically, acetate production decreased in glycerol-supplemented microbiota of both low- and high-resident A. hallii groups after 6 h (median −40% and −46%, respectively) and 24 h (median of −67% and −65%, respectively), as compared with the control. However, when A. hallii was further supplemented, a significantly higher decrease in acetate was observed in the low group after 24 h (median of −75% versus 65%) as compared with glycerol alone ( Figure 4E ). Butyrate production increased in both the low and the high groups when glycerol was supplemented after both 6 h (+45% and +16% in the low and high groups, respectively) and 24 h (+53% and +12%, respectively), as compared with the control. However, on supplementing microbiomes with A. hallii, butyrate production relative to the control significantly decreased, as compared with glycerol alone, in the low group after 6 h (–2%) and 24 h (–45%) ( Figure 4F ). In alignment with previous observations, a negative correlation was observed between the proportion of resident A. hallii and total SCFA production relative to the control after 6 h and 24 h when glycerol was supplemented in the 20 fecal microbiota and the trend was confirmed when A. hallii was additionally supplemented after 24 h ( Figure S2D ).