Studying the interplay between coffee and the microbiome

A total of 35,214 metagenomes from five ZOE PREDICT cohorts (PREDICT1, PREDICT2, PREDICT3 US21, PREDICT3 US22A and PREDICT3 UK22A)27, the MBS28, and the MLVS29 were included in this study, together with 18,984 metagenomes from public sources including healthy individuals (n = 8,728), non-Westernized individuals (n = 1,195; Methods), newborns and infants (n = 1,623), ancient microbiome samples (n = 37), non-human primates (n = 201) and 7,423 samples from individuals with a specific disease (35 diseases). Overall, these sets lead to a total number of 54,198 metagenomic samples considered (Fig. 1a), all profiled using MetaPhlAn 4 (ref. 30). In addition, the metagenomic profiles of 23,115 of these samples were coupled with highly detailed food frequency questionnaires (FFQs) covering over 150 food items from the PREDICT1, PREDICT2, PREDICT3 US22A, PREDICT3 UK22A, MBS and MLVS cohorts (Methods). We also analysed 438 plasma metabolomes and 364 faecal metatranscriptomes from the MBS and MLVS cohorts (Fig. 1b).

Fig. 1: Consistent global links between coffee consumption and the human gut microbiome.

a, Five UK and/or US PREDICT cohorts (n = 975, 11,798, 8,470, 1,098 and 12,353), the MBS and the MLVS (n = 213 and n = 307, respectively) were used to assess diet–microbiome relationships (total n = 35,214). For later comparisons of microbiome distributions across different populations, we retrieved n = 18,984 metagenomic samples from public sources, including healthy adult individuals, newborns, non-Westernized (non-West.) individuals, ancient samples and non-human primates (NHP). P1, PREDICT1; P2, PREDICT2; P3, PREDICT3. b, We combined faecal metagenomics (n = 54,198), faecal metatranscriptomics (n = 364) and plasma metabolomics (n = 438), with the latter two from the MBS and MLVS cohorts. FFQs surveyed nutritional habits of the participants from four PREDICT cohorts, MBS and MLVS (n = 22,867 after removing individuals above the 99th percentile of coffee intake in the PREDICT cohorts as outliers). Participants were categorized as ‘high’, ‘moderate’ and ‘never’ coffee drinkers as previously established25. c, Median Spearman’s correlation and median AUCs from a random forest regressor and a random forest classifier trained on the microbiome composition estimated by MetaPhlAn 4 (ref. 30). d, The number of never (light green), moderate (dark cyan) and high-coffee drinkers (brown). e, ROC and AUC of random forest classifiers discriminating participants between pairs of the three coffee drinker classes, assessed in a tenfold, ten times repeated cross-validations (CV) that benefited from the other cohorts during the training phase as in the leave-one-dataset-out approach (LODO; Methods). The shaded areas represent the 95% confidence intervals (CIs) of a linear interpolation over all the folds of the test. Machine learning results using either only a CV or a LODO approach are reported in Extended Data Fig. 2a,b.

Taking advantage of the four PREDICT cohorts with full dietary information (n = 22,595), we first used a random forest algorithm to associate single food items with microbiome. Microbiome species-level genome bin (SGB) composition was highly predictive of total coffee intake from FFQs in a cross-validation setting. Spearman’s correlation between predicted versus real coffee consumption reached values above 0.5 for the PREDICT1, PREDICT3 US22A and PREDICT3 UK22A studies. When using random forests to distinguish the top versus bottom quartiles of food intake, coffee was again the one most accurately predicted by microbiome composition (area under the curve (AUC) of >0.8; Fig. 1c). Importantly, milk, dairy cream, sugar and honey, which are commonly added to coffee and are present in the PREDICT FFQs, achieved much lower prediction performances (Spearman’s ρ < 0.1, AUC ~0.55–0.58), with the only partial exception of milk and soy/rice milk in the PREDICT3 UK22A cohort (ρ = 0.26 and 0.21, AUC of 0.76 and 0.7, respectively). Coffee is therefore not only a food with ascertained health effects, but also most strongly associated with the human microbiome among ~150 food items.

Coffee consumption is strongly linked to the gut microbiome

To better characterize the link between the gut microbiome and coffee, we classified individuals from PREDICT1, PREDICT2, PREDICT3 US22A, PREDICT3 UK22A and MBS–MLVS into three coffee-drinking levels: ‘never’, ‘moderate’ and ‘high’ (Fig. 1b,d). This categorization used the same thresholds previously applied by a panel of nutritionists in the PREDICT1 study25. Individuals with a coffee intake up to 20 g a day (less than three cups a month) were classified into the group ‘never’, individuals with an intake ≥600 g of coffee per day (more than three cups a day) were classified as ‘high’ and coffee drinkers between these two values were categorized as ‘moderate’. These thresholds, representing the 24.9th and 88.95th percentiles of coffee intake in the PREDICT1 cohort (Supplementary Table 1), were applied to all other cohorts (Fig. 1d and Extended Data Fig. 1). In addition, samples from individuals with a coffee intake above the 99th percentile were excluded as outliers potentially due to data collection issues. In total, 5,730 individuals were classified as ‘never’, 14,647 as ‘moderate’ and 2,490 as ‘high’ coffee drinkers (Fig. 1d and Supplementary Table 2).

We next used machine learning to evaluate the strength of the link between microbiome composition and coffee consumption levels31,32. We employed random forest classifiers trained on SGB-level abundances to distinguish between three pairs of conditions: never versus moderate, moderate versus high and never versus high coffee drinkers. The high and moderate categories were both highly separable from the never category (tenfold, ten times repeated cross-validation, median AUC across cohorts of 0.92 and 0.86, respectively; Extended Data Fig. 2a and Supplementary Table 3). Weaker predictions were achieved for moderate versus high coffee drinkers (median AUC of 0.63), suggesting a limited dose-dependent association of coffee intake with the microbiome.

Cross-validation performances in distinguishing coffee consumption levels were consistent across datasets, but to further test the cross-population reproducibility of the microbiome signature, we applied the leave-one-dataset-out (LODO)32 as well as a ‘cross-LODO’ hybrid approach (that is, augmenting the training folds of a specific dataset with external datasets; Methods). The ‘cross-LODO’ analysis yielded a median AUC of 0.93 for the never versus high comparison, 0.87 for the never versus moderate comparison and 0.65 for the moderate versus high comparison (Fig. 1e, Extended Data Fig. 2b and Supplementary Table 3). These findings suggest that the gut microbiome has distinct compositions in coffee drinkers compared with non-coffee drinkers, with a modest effect on differentiating the dose of coffee drinking.

L. asaccharolyticus is strongly linked with coffee

To investigate which gut microbiome features are associated with coffee intake, we correlated the ranked abundances of SGBs with participant’s coffee intake in each cohort (Spearman’s correlation) using raw and partial correlations (Supplementary Table 4 and Methods). Correlations were then meta-analysed across all cohorts (Supplementary Table 5 and Methods). In total, we found 291 correlations at q < 0.001 with 132 SGBs having ρ ≥ 0.05, and 298 when taking into account the effect of sex, age and body mass index (BMI) of the participants (q < 0.001, 130 with ρ ≥ 0.05; Supplementary Tables 6 and 7 and Extended Data Fig. 3).

While most of the correlations were positive, 46 partial correlations were negative and significant (q < 0.001), although none exceeded ρ values lower than −0.1. This suggests more stimulatory rather than inhibitory effects of coffee and its components on microbial species relative abundances (Extended Data Fig. 4a,b). Among the positive correlations, the strongest one involved the species L. asaccharolyticus (SGB15154), reaching ρ = 0.43 (0.41–0.45) and q < 10−10. In contrast, correlation between coffee and ɑ-diversity was much lower (ρ = 0.1 (0.05–0.15), p = 1.8 × 10−4; Extended Data Fig. 3). L. asaccharolyticus was first isolated in 201826,33, and as this strain remains the only one deposited in public biobanks, here we consider it to be representative for the species although other genomes (but not isolates) with conflicting taxonomic labels have been deposited (that is, Clostridium phoceensis34; Methods). The species responsible for the next two strongest associations were Massilioclostridium coli (SGB29305, ρ = 0.31 (0.27–0.35), q < 10−10) and the so-far uncharacterized Clostridium species ‘12CBH8’ (SGB7259, ρ = 0.3 (0.28–0.32), q < 10−10; Fig. 2a).

Fig. 2: L. asaccharolyticus drives the association between the gut microbiome and coffee intake.

a, The top ten SGBs from a meta-analysis of partial correlations between SGB-ranked abundances and total per-individual coffee intake considering the five cohorts analysed in this study (q < 0.001). The black markers show the per-cohort partial correlations and the light blue markers indicate the average Spearman’s correlations adjusted by sex, age and BMI. b, The same SGBs are meta-analysed with Spearman’s partial correlations (par. corr.) between SGB abundances and decaffeinated (decaf.) coffee intake in the PREDICT1 and PREDICT3 UK22A cohorts, excluding individuals who consumed caffeinated coffee only (n = 262 and 4,055). The black markers show the per-cohort correlations and dark blue symbols refer to average correlations adjusted by sex, age, BMI and caffeinated coffee. c, The prevalence of the ten SGBs in the five cohorts analysed. d, The prevalence of L. asaccharolyticus across never, moderate and high coffee drinkers and nine US regions in the PREDICT2 and PREDICT3 US22A cohorts (n = 9,210).

We further investigated whether these associations were driven by caffeine by performing two meta-analyses on the PREDICT1 and PREDICT3 22UKA samples for which the intake of decaffeinated and caffeinated coffee was available. Partial correlations between SGB-ranked abundances and decaffeinated versus caffeinated coffee were run independently (excluding individuals who drank exclusively caffeinated or decaffeinated coffee, respectively). In addition, partial correlations were also adjusted for the other type of coffee consumed by the individuals in case their record reported both kinds. We identified 150 correlations, which remained highly significant after controlling for the decaffeinated coffee intake (q < 0.001, nparticipants = 12,089; Extended Data Fig. 4c and Supplementary Table 8). This indicated a substantial independence on caffeine of the observed impact on the microbiome. Next, we analysed the decaffeinated coffee association with the microbiome in individuals consuming decaffeinated coffee and adjusting by caffeinated coffee as well as by sex, age and BMI. In this reduced set of samples (nparticipants = 6,089) we identified 22 correlations at q < 0.001 and 66 at q < 0.1 (Supplementary Table 9). The top three correlations identified were L. asaccharolyticus (ρ = 0.27 (0.21–0.33), q < 10−10), the Lachnospiraceae SGB4777 (ρ = 0.18 (0.16–0.21), q < 10−10), and M. coli (SGB29305, ρ = 0.17 (0.13–0.2), q < 10−10; Fig. 2b).

As expected, several coffee-associated SGBs were also L. asaccharolyticus co-abundant SGBs, possibly indicating similar independent stimulatory effects of coffee rather than ecological relations (Extended Data Fig. 5 and Supplementary Table 10). The top-five SGBs associated with L. asaccharolyticus abundance were, however, not among the strongest associations with coffee. In particular, the two SGBs with the strongest co-abundance pattern with L. asaccharolyticus were Dysosmobacter welbionis (SGB15078) and the Clostridiales bacterium SGB15143 (ρ = 0.57 and 0.51, respectively, q < 1 × 10−10; Extended Data Fig. 5), which were both only weakly associated with total coffee (ρ ≤ 0.05; Supplementary Table 7). Overall, these results indicate that a panel of species, and in particular L. asaccharolyticus is robustly associated with total and decaffeinated coffee consumption, suggesting that the association is not purely due to caffeine.

Effect of coffee on L. asaccharolyticus is supported in vitro

Among the top coffee-associated SGBs, L. asaccharolyticus showed the highest and the most uniform prevalence across all the cohorts (93.5%; Fig. 2c). In the ‘never’ group from the USA, its prevalence was uniformly high (average prevalence of 87.8 ± 2%) across nine different regions (samples from PREDICT2 and PREDICT3 US22A, n = 9,210). Over and above this, however, it was uniformly increased in all regions when considering coffee consumption; it increased from 87.8% to 95.6% in moderate drinkers and from 95.6% to 97.7% in high drinkers (Fig. 2d and Supplementary Table 11). Degree of urbanization (rural versus urban living context) was not associated with L. asaccharolyticus in the microbiome (Extended Data Fig. 6 and Supplementary Table 12). Overall, the median abundance of L. asaccharolyticus ranged from 4.5- to 8-fold higher in the high compared with the never group (in the PREDICT3 US22A and MBS–MLVS cohorts), and 3.4- to 6.4-fold higher in the moderate versus the never group (in the PREDICT2 and MBS–MLVS cohorts; Supplementary Table 13). By contrast, the highest median fold change between moderate and high drinkers was only 1.4 and did not reach statistical significance in three out of five cohorts (Fig. 3a and Supplementary Table 14).

Fig. 3: L. asaccharolyticus is highly prevalent with about fourfold higher average abundance in coffee drinkers, and its growth is stimulated by coffee supplementation in vitro.

a, The relative abundance of L. asaccharolyticus in each cohort by coffee consumption category (never, moderate or high). The boxes represent the median and interquartile range (IQR) of the distributions, and top and bottom whiskers mark the point at 1.5 IQR. The median fold change of the high versus never comparison is reported on top if post hoc Dunn q < 0.01, and median fold change (FC) of the other two comparisons are reported on the top of each combination. n.s. (not significant) refers to post hoc Dunn q > 0.01. Total sample sizes are presented in Extended Data Fig. 1. b, L. asaccharolyticus growth on agar plates supplemented with increasing concentrations of coffee and measured by plate count (c.f.u. per ml). P values refer to one-sample t-tests compared with the control (ctrl) experiment value. c–e, Bacterial growth of L. asaccharolyiticus (c), E. coli (d) and B. fragilis (e) in liquid medium supplemented with increasing coffee concentrations and measured by changes in optical density (OD650). Percentage growth is relative to the culture medium control not supplemented with coffee (100%). Absolute OD650 values are reported in Supplementary Tables 15 and 16. The bars and error lines indicate the mean ± s.d. of five technical replicates, except for E. coli control (n = 3 and n = 4) and B. fragilis instant 5 g l−1 (n = 4). The minus and plus signs refer to significant tests (Dunnett q < 0.01) that overcome specific thresholds of fold increase (incr.) or decrease (decr.).

To test whether these associations are at least partially due to a direct effect of coffee on L. asaccharolyticus, we performed in vitro experiments by supplementing coffee on L. asaccharolyticus cultures. To this end, the type strain L. asaccharolyticus DSM106493 was separately cultured with two selected common coffee preparations, that is, moka brewed and instant coffee (Methods). For both coffee preparations, we also tested the commercially available decaffeinated variants. The growth of L. asaccharolyticus was stimulated on agar plates supplemented with coffee at concentrations of 5 and 10 g l−1, regardless of coffee preparation (moka versus instant) and caffeine presence (one-sample t-test P = 0.02 and 0.01 for 5 and 10 g, respectively; Fig. 3b). We further tested the growth of L. asaccharolyticus associated with coffee supplementation in liquid media as assessed by optical density measurements. Despite the inherently low growth levels of L. asaccharolyticus (OD650 range of 0.0138–0.217), this experiment confirmed the stimulatory effect of coffee (average increase of 350% (3.5 median fold change), Dunnet’s q < 0.01 in ten out of 16 preparations; Fig. 3c and Supplementary Tables 15 and 16). As comparative controls, we applied the same experimental conditions to two isolates that we obtained from faecal samples of healthy donors, namely Escherichia coli (SGB10068) and Bacteroides fragilis (SGB1855) (Methods), representing both facultative and obligate anaerobes of the gut microbiome but unrelated with coffee intake in our study participants (meta-analysis ρ = −0.02 for E. coli and 0.01 for B. fragilis). While the B. fragilis isolate showed a slight significant (but much weaker, 8% on average) growth at 1 and 2 g l−1, no-clear trend was detected at 5  g l−1 in any of the species. A greater growth decrease was instead observed in both species at 10 g l−1 (−42% and −30%), suggesting an inhibitory action at higher concentrations (Fig. 3d,e and Supplementary Table 16). These results suggest that the increased abundance of L. asaccharolyticus in the gut of coffee drinkers can be due to direct fitness stimulatory effects of coffee on the bacterium.

L. asaccharolyticus is ubiquitous in Western populations

We then aimed to survey the prevalence of L. asaccharolyticus across more diverse populations, by exploiting the availability of metagenomes and curated sample metadata in curatedMetagenomicData35. We analysed 11 categories of hosts differing in age group, health status, lifestyle and species (N = 54,198) for the presence of L. asaccharolyticus in 43 countries (Methods and Supplementary Table 17). L. asaccharolyticus prevalence was above 60% in 52 out of 74 cohorts (70%; Fig. 4a), with a median prevalence of 75%, mostly representing adult populations in urbanized Western-lifestyle environments (Fig. 4a). In contrast, its prevalence in individuals belonging to rural societies with non-typically Western lifestyles (20 cohorts) was much lower (median prevalence 2.4%), and L. asaccharolyticus was also only rarely found in newborns and children. Considering the available metagenomic data from non-human primates (201 samples; Methods), L. asaccharolyticus was detected in only one sample. All these lines of evidence point at a dependency of the population-level prevalence of L. asaccharolyticus with the broad availability of coffee in the diet in the population, a hypothesis further strengthened by the detection of this species in only two of the 37 samples from ancient populations we had access to (Fig. 4a and Supplementary Table 17).

Fig. 4: L. asaccharolyticus is ubiquitous in modern, Westernized, adult populations and almost absent elsewhere.

a, The prevalence of L. asaccharolyticus in 11 different types of host (219 subpopulations, N = 54,198) including children and adults; healthy and diseased participants; from Westernized and non-Westernized communities; non-human primates and ancient samples, compared with the ZOE PREDICT and MBS–MLVS cohorts. Human, modern samples and participant records were obtained from a development version of curatedMetagenomicData35 (Supplementary Table 17). b, The per capita coffee consumption (kg per year, estimated by https://worldpopulationreview.com) for 25 countries (AUT, Austria; CHE, Switzerland; DEU, Germany; DNK, Denmark; ESP, Spain; FIN, Finland; FRA, France, GBR, UK; IRL, Ireland; ITA, Italy; LUX, Luxembourg; NLD, Netherlands; SWE, Sweden; CHN, China; IND, India; ISR, Israel; JPN, Japan; KAZ, Kazakhstan; KOR, Korea; MNG, Mongolia; MYS, Malaysia; ARG, Argentina; CAN, Canada; AUS, Australia) correlates with the prevalence of L. asaccharolyticus in healthy and diseased populations. The shaded areas around the regression line represent the 95% confidence interval estimated by bootstrapping.

Analysing microbiome samples from previous studies of various diseases (Methods), L. asaccharolyticus prevalence was comparably high in 7,004 samples from non-healthy adults and children from Westernized populations as well as 411 from non-Westernized populations (Fig. 4a and Supplementary Table 17). Prevalences ≥80% were found in all but one cancer type tested and in cardiometabolic diseases, and also no differences were found when meta-analysing studies with case–control metagenomic information across 25 diseases (7,154 controls and 5,670 cases; Extended Data Fig. 7a,b and Supplementary Table 18 and 19).

L. asaccharolyticus correlates with worldwide coffee intake

To investigate whether different average coffee consumption rates are the potential drivers of the variable prevalence of L. asaccharolyticus in the Westernized populations (95% confidence interval, 40–97%), we took advantage of the available data to directly correlate estimated annual coffee consumption with L. asaccharolyticus prevalence (Methods). Country-level coffee consumption was strongly correlated with L. asaccharolyticus prevalence (ρ = 0.62 and 0.70, P = 1.9 × 10−3 and 2.3 × 10−3 in healthy and diseased cohorts, respectively; Fig. 4b). This finding strongly reinforces the hypothesis that not only is L. asaccharolyticus abundance in a person stimulated by their coffee intake but also the overall prevalence in a population is driven by the population-level average coffee consumption.

Coffee plasma metabolites linked with L. asaccharolyticus

We next analysed 235 MLVS and 203 MBS plasma metabolomes spanning a total of ~14,000 metabolic features each (Methods). These included six metabolites in the caffeine metabolism pathway36, including caffeine, 1-methyluric acid, 1,7-dimethyluric acid, 1-methylxanthine, 3-methylxanthine and 5-acetylamino-6-amino-3-methyluracil, as well as quinic acid and trigonelline. We observed a significant positive correlation between these eight coffee metabolites and coffee intake in both cohorts (Fig. 5a). Next, we used MACARRoN37, a bioactive metabolite prioritization workflow to find modules (clusters) of covarying metabolic features associated with L. asaccharolyticus in each cohort (Methods). In the MLVS metabolomes, known coffee metabolites were prioritized to be associated with L. asaccharolyticus (Fig. 5b). Additionally, they covaried with several unannotated features in two modules: one (module 125) containing caffeine and its derivatives, while the other (module 33) containing trigonelline and quinic acid (Fig. 5b). Similarly, in the MBS cohort, both caffeine- and trigonelline-related metabolites (quinic acid was not measured in this cohort) were found to be separated into different modules (Extended Data Fig. 8a,b).

Fig. 5: Unannotated metabolites covarying with quinic acid are associated with L. asaccharolyticus.

a, The correlation of coffee intake versus abundances of six known coffee metabolites in plasma metabolomics samples from the MLVS (blue) and MBS (red). The highest rank correlation is reported in each plot. Three metabolites were not measured in MBS. b, Left, a heat map showing standardized abundances of the 14 unannotated and 8 previously annotated metabolites in the MLVS cohort (n = 307) with the highest MACARRoN priority score with respect to the presence of L. asaccharolyticus. QA, quinic acid; Trig, trigonelline. Right, MACARRoN priority scores. Samples are reported by coffee intake category. c, The log2-transformed abundances of quinic acid and the top six quinic acid-correlated unannotated metabolites according to L. asaccharolyticus relative abundance (RA) categories (absent, RA <0.01%; low, 0.1%> RA ≥0.01%; high, RA >0.1%) in 190 coffee drinkers. The boxes represent the median and IQR of the distributions, and top and bottom whiskers mark the point at 1.5 IQR.

L. asaccharolyticus and coffee enrich quinic acid metabolites

In addition to the known metabolites, unannotated metabolic features that strongly correlated with caffeine and quinic acid in modules 125 and 33, respectively, were also prioritized by MACARRoN as associated with L. asaccharolyticus (Supplementary Table 20). Furthermore, similar to known coffee metabolites, these unannotated prioritized features were enriched in moderate and high coffee drinkers, thus corroborating their correlation with L. asaccharolyticus (Fig. 5b). To further confirm the uniqueness of the associations between coffee-associated SGBs and coffee-associated metabolites, we repeated the aforementioned analyses and prioritized for another coffee-linked species M. coli (Fig. 2a) and two other species, Roseburia hominis (SGB4936) and Dorea formicigenerans (SGB4575), that are not coffee-associated but have the same prevalence and abundance patterns as L. asaccharolyticus in the MLVS cohort. As expected, coffee-associated metabolites were prioritized only in M. coli (Supplementary Table 21 and Extended Data Fig. 8c).

Because an interaction model identified a significant effect between coffee intake and L. asaccharolyticus only for quinic acid (interaction P = 3.39 × 10−13; Supplementary Table 22), this raised the hypothesis that the association of the microorganism with coffee may be particularly related to the biochemistry of the six unannotated compounds in module 33 that strongly correlated with quinic acid (Fig. 5b and Extended Data Fig. 8d). Testing this among coffee drinkers in the MLVS cohort (n = 190), individuals with higher abundance of L. asaccharolyticus showed higher abundance of quinic acid and related prioritized compounds (Fig. 5c). Upon examination of their masses, we found that they all differed from quinic acid by a small mass difference, suggesting that they are potential derivatives of quinic acid, although we could not confidently assign their identities (Extended Data Fig. 8e). One of them, F1976 (neutral mass 174.9947), which correlated with quinic acid (ρ = 0.43), was similar in mass to shikimic acid (monoisotopic molecular weight of 174.0528 g mol−1), which is synthesized from quinic acid by gut microorganisms11,38 (Extended Data Fig. 8d,e). We also observed a feature in this module that potentially represents pyrogallol (monoisotopic molecular weight of 126.03 g mol−1; [M + H], 127.0389; neutral mass, 126.0311; feature, F787), another gut microbial derivative of dehydroshikimic acid and quinic acid11, although it was not differentially abundant in participants carrying L. asaccharolyticus. Together, this analysis suggested the presence of pathways responsive to coffee and, more specifically, quinic acid in L. asaccharolyticus.

Towards this, we identified L. asaccharolyticus transcripts in metatranscriptomes of 364 samples from the MLVS cohort for which corresponding species abundance data were available. Transcripts of only 1,225 total UniRef90 protein families (146 Enzyme Commission (EC) numbers) were attributed to L. asaccharolyticus and were detected in only 12 (3.29%) samples (Methods). Moreover, there was no correlation between L. asaccharolyticus abundance and number of transcripts detected per sample, and only one sample contained more than 50% (n = 715) of all detected L. asaccharolyticus UniRef90 families (Extended Data Fig. 9a–c). To further improve transcript detection, we carried out SGB-level functional profiling using a custom bowtie database (Methods), which increased the number of detected UniRef90s from 1,225 to 3,158 in 352 (96.7%) samples. Despite this significant improvement in the number of samples containing at least one L. asaccharolyticus transcript, only 14 (3.84%) samples contained >10% of the total transcripts and only two contained >25% transcripts. Our integrated metabolomic, metagenomic and meta-transcriptomic analysis thus allowed us to identify the pathways probably connected with coffee metabolism, although single-transcript analysis did not have enough resolution to confirm the increased activity of the single genes in such pathways, a common phenomenon for low-abundance species that are neglected by highly abundant and highly transcribing species39.