Multicohort gut metagenomics with detailed dietary data

The aim of this study was to elucidate how prolonged dietary preferences affect the structure and function of the human gut microbiome at both the global and single-species level. To do so, we capitalized on three cohorts from the ZOE PREDICT programme from the United Kingdom (P1 n = 1,062 individuals10,11, P3 UK22A n = 12,353) and from the United States (P3 US22A n = 7,931; Methods). We further included two additional, publicly available cohorts comprising Italian participants (Tarallo et al. (2022)12n = 118 individuals and De Filippis et al. (2019)13n = 97 individuals; Fig. 1a). Each participant of the five cohorts reported their nutritional habits as being either ‘omnivore’ (including meat, dairy and vegetables), ‘vegetarian’ (excluding meat) or ‘vegan’ (excluding both meat, dairy and other animal products) and donated stool samples that underwent shotgun metagenomic sequencing. In total, 656 vegans, 1,088 vegetarians and 19,817 omnivores were considered (Fig. 1a). In addition to participants’ overall dietary habits, the ZOE PREDICT cohorts included data on habitual consumption of over 150 single foods per individual, obtained from validated quantitative food frequency questionnaires (FFQs; Methods). Dietary patterns were partially confirmed by DNA-based detection of food in the stool microbiome14 which, however, would require greater sequencing depth to be used for this goal (Methods).

Fig. 1: A large, integrated, metagenomic dataset with detailed dietary information.

a, Sample size for each diet pattern across the five cohorts (logarithmic scale). b, Observed richness of each diet pattern's gut microbiome within each of the five cohorts (nP1 = 1,062 individuals, nP3 UK22A = 12,353, nP3 US22A = 7,931, n Tarallo et al. (2022)12 = 118, n De Filippis et al. (2019)13 = 97). Boxplots show the median, 25th and 75th percentiles, and whiskers extend to 1.5× the interquartile range. Asterisks denote significance level of Dunn’s tests with BH correction (Methods and Supplementary Table 6); *P < 0.05, **P ≤ 0.01, ***P ≤ 0.001. c, Distribution of hPDI for each diet pattern within each of the five cohorts (P1 with 841 omnivores, 49 vegetarians and 10 vegans; P3 UK22A with 11,289 omnivores, 610 vegetarians and 192 vegans; P3 US22A with 6,720 omnivores, 309 vegetarians and 346 vegans). Boxplot integrated into violin plots have the same parameters as in b. Asterisks denote the same significance as in b, but for Tukey contrasts for multiple comparisons following an ANOVA model (Methods and Supplementary Table 3). d–h, Beta diversity of gut microbial composition accounting for phylogenetic diversity using unweighted UniFrac distances. Each dot in the principal coordinates analysis (PCoA) plots represents an individual. Ellipses indicate 95% CIs. Statistical differences between diet patterns were assessed via PERMANOVA, correcting for sex, age and BMI with 999 permutations. There is one PCoA plot per cohort: P1 (d), P3 UK22A (e), P3 US22A (f), De Filippis et al. (2019)13 (g), Tarallo et al. (2022)12 (h).

To quantify the consumption of plant-based foods, we considered the plant-based diet index (hPDI), which gives higher scores to healthy plant foods and reverse scores to less healthy plant and animal foods15 (Methods). Within each of the three PREDICT cohorts, hPDI significantly differed between diet patterns as expected (analysis of variance (ANOVA), P < 0.001 across all PREDICT cohorts; Fig. 1c and Supplementary Table 1), with significantly higher hPDI in vegans compared with vegetarians and similarly for vegetarians compared with omnivores (Tukey P < 0.01; Supplementary Tables 2 and 3).

Gut microbial diversity and composition across diet patterns

Gut microbial richness differed significantly according to diet patterns in the PREDICT cohorts (Kruskal–Wallis, P < 0.05; Supplementary Table 4), with a lower observed richness in vegans (median between 209 and 266 species-level genome bins (SGBs)) and vegetarians (median 201–269) compared with omnivores (median 217–299; Fig. 1b and Supplementary Table 5), but no significant differences between vegans and vegetarians (Dunn’s test, P > 0.05; Supplementary Tables 6 and 7). This highlights that alpha diversity might correlate with diet patterns that are potentially more diverse.

Overall gut microbial composition also differed significantly according to diet pattern (permutational multivariate analysis of variance (PERMANOVA) on unweighted UniFrac distances, R2 = 0.002–0.028; P < 0.05 for all five cohorts; Fig. 1d–h and Supplementary Table 8 with additional distance metrics; Methods), with the variation in beta diversity explained by diet pattern aligning with previous studies16. In addition, diet patterns were highly distinguishable based on quantitative gut microbial profiles when using machine learning classifiers17. By evaluating the performance of the model trained in a variant of cross validation in which training folds are merged with external cohorts (cross-validation leave-one-dataset-out, that is, cross-LODO; Methods)18, we obtained a mean area under the receiver operating characteristic (ROC) curve (AUC) across all diet patterns and across all five cohorts of 0.85. The highest predictability was obtained when separating vegans from omnivores (mean cross-LODO AUC = 0.90), followed by separating vegetarians from vegans (0.84), and finally vegetarians from omnivores (0.82; Fig. 2 and Supplementary Table 9). Similar results were achieved when using the LODO approach that does not consider any training folds from the target cohort (Supplementary Table 9). Because we did not log when diet patterns may have been switched, we hypothesize that the non-perfect classification might be due to individuals who switched diet patterns recently, and some associations may actually be stronger than what we observed. Altogether, these results warranted further investigation into the specific microbiome components responsible for these differences.

Fig. 2: Highly accurate classification of individual diet patterns based on gut microbial features.

Average ROC curves and AUCs showing the discrimination between all pairs of the three diet patterns (omnivores vs vegetarians; omnivores vs vegans; and vegetarians vs vegans) per cohort using random forest classifiers in a hybrid cross-LODO (Methods) approach. Shaded areas correspond to 95% CIs.

Gut microbe signatures of vegans, vegetarians and omnivores

To explore which microbes are associated with the different gut microbial compositions between vegans, vegetarians and omnivores, we performed a meta-analysis across the five cohorts on the differential relative abundance of each SGB within each individual and their respective diet pattern (Methods). In total, 488 SGBs were significantly differentially abundant in omnivores compared with 112 SGBs in vegetarian microbiomes; 626 SGBs were significantly differentially abundant in omnivores compared with 98 SGBs in vegans; and 30 SGBs were significantly differentially abundant in vegetarins compared to 11 SGBs in vegans (Supplementary Tables 10–12). When focusing on the top 30 microbial markers, the majority of these strongest associations were linked to the least restrictive diet pattern (Figs. 3b,h and 4b).

Fig. 3: Gut microbial signatures of an omnivore vs vegetarian and vegan diets.

Top panels: omnivore vs vegetarian diet. Bottom panels: omnivore vs vegan diet. a, Prevalence of the top 30 signature SGBs (with their respective SGB IDs in parentheses) in omnivore (left) and vegetarian (right) gut microbiomes. b, Meta-analysed correlations between SGB relative abundance and diet pattern (omnivore n = 19,817 in pink vs vegetarian n = 1,088 in purple). The top 30 SGBs with the largest absolute SMD are reported, with upper and lower confidence intervals. Smaller shapes are per-cohort correlations (black indicates Wald q-value < 0.1, grey indicates Wald q-value ≥ 0.1). The black horizontal bar indicates the separation between the correlations with omnivores vs vegetarians for ease of visualization only. c, Meta-analysed pooled effect sizes with upper and lower confidence intervals from correlations between SGB relative abundance and consumption of five major food groups (meat: nP1 = 841 individuals, nP2 = 843, nP3 UK22A = 11,533, nP3 US22A = 7,228; dairy: nP1 = 890, nP2 = 843, nP3 UK22A = 12,156, nP3 US22A = 7,558; fruits/vegetables: nP1 = 900, nP2 = 843, nP3 UK22A = 12,353, nP3 US22A = 7,931). d, Meta-analysed pooled effect sizes with upper and lower confidence intervals from correlations between SGB relative abundance and hPDI within omnivores (nP1 = 841, nP2 = 843, nP3 UK22A = 11,289, nP3 US22A = 6,720) and vegetarians (nP1 = 49, nP3 UK22A = 610, nP3 US22A = 309). e, ZOE MB health ranks of each signature SGB. Values closer to zero indicate positive CMH outcomes, closer to one indicate negative CMH outcomes30. f, Machine learning predictions (random forest cross-LODO AUC; Methods) of the presence of each of the signature microbes between omnivores and vegetarians based on FFQs. g, Prevalence of the top 30 signature SGBs (with their respective SGB IDs in parentheses) in omnivore (left) and vegan (right) gut microbiomes. h, Same as b, except between omnivores (n = 19,817 in pink) and vegans (n = 656 in green). i, Same as c, except between omnivores and vegans. j, Same as d, except between omnivores and vegans (vegans: nP1 = 10, nP3 UK22A = 192, nP3 US22A = 346). k, Same as e, except between omnivores and vegans. l, Same as f, except between omnivores and vegans.

Fig. 4: Gut microbial signatures of a vegetarian vs vegan diet.

a, Prevalence of the top 30 signature SGBs (with their respective SGB IDs in parentheses) in vegetarian (left) and vegan (right) gut microbiomes. b, Meta-analysed correlations between SGB relative abundance and diet pattern (nvegetarian = 1,088 in purple vs nvegan = 656 in green). The top 30 SGBs with the largest absolute SMD are reported. Smaller shapes are per-cohort correlations (black indicates Wald q-value < 0.1, grey indicates Wald q-value ≥ 0.1). The black horizontal bar indicates the separation between the correlations with vegetarians vs vegans for ease of visualization only. c, Meta-analysed pooled effect sizes with upper and lower confidence intervals from correlations between SGB relative abundance and consumption of five major food groups (meat: nP1 = 841 individuals, nP2 = 843, nP3 UK22A = 11,533, nP3 US22A = 7,228; dairy: nP1 = 890, nP2 = 843, nP3 UK22A = 12,156, nP3 US22A = 7,558; fruits/vegetables: nP1 = 900, nP2 = 843, nP3 UK22A = 12,353, nP3 US22A = 7,931). d, Meta-analysed pooled effect size with upper and lower confidence intervals from correlations between SGB relative abundance and hPDI within vegetarians (nP1 = 49, nP3 UK22A = 610, nP3 US22A = 309) and vegans (nP1 = 10, nP3 UK22A = 192, nP3 US22A = 346). e, ZOE MB health ranks of each signature SGB. Values closer to zero indicate positive CMH outcomes, closer to one indicate negative CMH outcomes. f, Machine learning predictions (random forest cross-LODO AUC; Methods) of the presence of each of the signature microbes between vegetarians and vegans based on FFQs.

Knowledge of the predicted functions of the SGBs linked to the various diet patterns revealed potential dietary-specific niches. Several SGBs increased in omnivore microbiomes are linked to meat consumption by aiding in its digestion through for example, protein fermentation (Alistipes putredinis), utilizing amino acids and via bile-acid resistance (Bilophila wadsworthia19), or are mucolytic indicators of inflammation that have been linked to inflammatory bowel diseases (Ruminococcus torques20,21; Fig. 3b,h). In contrast, several SGBs overrepresented in vegan microbiomes are known butyrate producers (Lachnospiraceae22, Butyricicoccus sp.23,24 and Roseburia hominis22,25) and are highly specialized in fibre degradation (Lachnospiraceae26; Figs. 3h and 4b). In addition, Streptococcus thermophilus, a common dairy starter and component27, had the highest effect size in vegetarian versus vegan gut microbiomes with a standardized mean difference (SMD) of −0.67 and second highest effect size in omnivore versus vegan gut microbiomes (SMD = −0.62). Thus, when a major differentiating characteristic between diet patterns lies in dairy consumption, the SGB with the greatest ability to differentiate between those diets is abundantly found in cheese and yogurt products. This was supported by other dairy-linked SGBs associated more with omnivore and vegetarian than vegan diets such as Lactobacillus acidophilus, Lactobacillus delbrueckii, Lactococcus lactis, Lacticaseibacillus paracasei and Lacticaseibacillus rhamnosus27,28. On the basis of these findings, we next explored the links between these diet pattern-specific microbes and the major food groups that distinguish the diet patterns.

Gut microbial diet signatures are linked to major food groups

We further investigated the role of major food groups, such as red and white meat, dairy, fruits and vegetables, in differentiating the gut microbial profiles across diet patterns (Methods and Supplementary Table 13). The amount of meat (either red or white) ingested by omnivores was positively correlated with the vast majority of SGBs linked to an omnivorous diet versus a vegetarian (23 out of 25 SGBs; Fig. 3c) or vegan one (16 of out 19 SGBs; Fig. 3i). In addition, compared with omnivore gut microbiomes, meat negatively correlated with all 5 SGBs strongly associated with vegetarian gut microbiomes and with 10 out of the 11 SGBs strongly associated with vegan gut microbiomes. The SGBs strongly associated with omnivore gut microbiomes correlated more strongly with red than with white meat consumption. Red and white meat correlated with the same SGBs in all but one case: ‘Candidatus Avimicrobium caecorum’, found in human gut microbiomes and assembled from chicken caecum29, which positively correlated with white meat consumption in omnivore gut microbiomes versus vegetarian and vegan ones (Fig. 3c,i).

In contrast, fruits and vegetables were positively correlated with 3 of out 5 SGBs overrepresented in vegetarian (Fig. 3c) and in 10 out of 11 SGBs overrepresented in vegan versus omnivore gut microbiomes (Fig. 3i). The majority of these correlations were more greatly associated with vegetables than with fruits. There were no cases of negative correlations between fruits and vegetables and the SGBs most strongly associated with vegetarian or vegan gut microbiomes. Conversely, any SGB strongly linked to an omnivore gut microbiome that correlated with fruits or vegetables showed negative and not positive correlations.

When considering dairy, which differentiates vegans from vegetarians and contributes to the difference between a vegan and an omnivore diet, SGBs that differentiate vegetarian from vegan gut microbiomes showed positive correlations with dairy in vegetarians and negative ones in vegans (Fig. 4c). Similarly, SGBs that differentiate omnivore from vegan gut microbiomes showed positive associations with dairy in omnivores and negative ones in vegans (Fig. 3i). Thus, the gut microbial signatures of these three diet patterns are linked to the inclusion or exclusion of major food groups.

Plant-based food diversity shapes the microbiome across diets

While the three diet patterns differed significantly in their hPDI scores (Fig. 1c), we next aimed to understand whether their correlations with the SGB relative abundance were consistent across diet patterns using a meta-analytical approach (Methods). Regardless of which diet patterns were compared, there was concordance in the correlations between hPDI and the SGB signature of each diet pattern (Figs. 3d,j and 4d). This means that if hPDI was correlated (either positively or negatively) with a signature SGB in omnivore gut microbiomes, it would show similar correlations in vegetarians and vegans as well. Thus, overall dietary factors may transcend diet patterns, suggesting that omnivores could share beneficial gut microbial signatures with other diet patterns if they also incorporate similar diversity of plant-based food items in their diets. In practice, however, omnivores generally ingest significantly less healthy plant-based foods than vegetarians or vegans (Fig. 1c).

Cardiometabolic health is linked to gut microbial diet patterns

To investigate the gut microbial links between the three diet patterns and human health, we employed the ZOE Microbiome Ranking 2024 (Cardiometabolic Health), ZOE MB Health ranks for short30, which assigns a numeric ranking to SGBs found to significantly correlate with cardiometabolic markers (Methods). We found that rankings of SGB signatures of omnivore microbiomes were statistically less favourable (mean rank = 0.53 and 0.58) when compared with vegetarian (mean rank = 0.44, two-sample t-test, P = 0.040, t(197) = 2.07) and vegan ones (mean rank = 0.38, two-sample t-test, P < 0.001, t(230) = 5.59; note that values closer to zero indicate positive CMH outcomes, whereas values closer to one indicate negative CMH outcomes; Extended Data Fig. 1). When comparing rankings of SGB signatures of vegan versus vegetarian microbiomes, vegan-associated SGBs had more favourable rankings (mean rank = 0.33) than vegetarian-associated ones (mean rank = 0.54, two-sample t-test, P = 0.028, t(30) = 2.30; Extended Data Fig. 1). These patterns were reflected when considering the 30 SGBs most distinguishable between the diet patterns. The majority of the ranked SGB signatures of an omnivore gut microbiome were associated with worse cardiometabolic health (CMH) compared with both vegetarian and vegan gut microbiomes, with the opposite being true for vegetarian and vegan gut microbiomes (Fig. 3e,k). When comparing vegetarian with vegan gut microbiomes, the latter again showed a majority of signature SGBs to be associated with positive CMH, whereas the pattern for the former was more split, with just under half of the vegetarian signature SGBs linked with more favourable CMH (Fig. 4e). Thus, omnivore signature microbes are associated with less favourable CMH, whereas signature vegan microbes are associated with more favourable CMH.

Entire diet profiles can predict specific gut species

Moving from major food groups to the entire set of food items in the FFQs, we next tested the extent to which habitual-diet information is linked to the presence or absence of each SGB of relevance for the three diet patterns (Methods18). The most diet-linked SGBs were those that most differentiate between omnivore and vegan gut microbiomes, in particular S. thermophilus, predictable from whole FFQ items at AUC = 0.72, R. torques (0.63), several Lachnospiraceae SGBs (all 0.65) and Lawsonibacter asaccharolyticus (0.78; Figs. 3f,l and 4f, and Supplementary Tables 14–16), which is strongly tied to coffee consumption10,31. This demonstrates the role that other foods may play in influencing this analysis based on entire FFQs versus highlighting only major food groups of interest. When comparing vegetarian with vegan gut microbiomes, the signature vegetarian microbes with the highest predictability are those linked with dairy consumption, for example, S. thermophilus (AUC = 0.72), L. rhamnosus (0.66), L. delbrueckii (0.70), L. paracasei (0.62), L. lactis (0.65) and L. acidophilus (0.68; Fig. 4f), aligning with our results thus far. These AUCs show that there exists a non-random, albeit mild, link between ingested food and the presence of specific species, suggesting causal links and potential transfer of microbes from food to gut.

Diet-dependent gut microbiome contribution of food microbes

Until now, our results suggest the potential for diet patterns to select for gut microbes, but gut microbes might be derived directly from food itself32, as may be the case for S. thermophilus, a common dairy component27,33 that we found to be one of the most differentiating SGBs between diet patterns that differ in dairy consumption (Figs. 3h and 4b). To establish how many SGBs in each diet pattern’s gut microbiome may be derived from food, we searched for food SGBs collated in ‘curatedFoodMetagenomicData’ (cFMD)32 across our five cohorts and found 260 to be present (Methods). We found that the number of distinct food SGBs differed according to diet pattern, with significantly fewer food SGBs in vegan (zero-inflated negative binomial mixed model, β = −0.36, P < 0.001; Fig. 5b and Methods) compared with omnivore and vegetarian (β = 0.35, P < 0.001) microbiomes, but not between vegetarians and omnivores (β = −0.004, P < 0.686).

Fig. 5: Contribution of food microbes to the gut microbiome across diet patterns.

a, Cumulative relative abundance (log10) and b, number of food SGBs (either meat, dairy, or fruits and vegetable-derived SGBs) within each individual’s gut microbiome, coloured by diet pattern (omnivores: nP3 UK22A = 11,533, nP3 US22A = 7,228; vegetarians: nP3 UK22A = 623, nP3 US22A = 330; vegans: nP3 UK22A = 197, nP3 US22A = 373) and grouped by cohort (either P3 UK22A or P3 US22A; for all cohorts, see Extended Data Fig. 2). Asterisks denote significance level of BH-corrected Dunn’s tests; **P ≤ 0.01, ***P ≤ 0.001 (Supplementary Tables 18 and 19). Boxplot parameters the same as in Fig. 1b. c, Prevalence of the 20 most common food SGBs (with their respective SGB IDs in parentheses) per diet pattern across all n = 5 cohorts. Hashtags denote the number of cohorts (out of the three that were tested: P1, P3 UK22A, P3 US22A; Methods) in which two-sided chi-squared tests showed significant differences in SGB prevalence across all three diet patterns (Supplementary Table 17). d, Prevalence (log10) of the 20 most common food SGBs across three major food categories (meat, dairy, fruits and vegetables) to indicate which food group each SGB is likely a signature of. White/blank boxes indicate that SGB was not prevalent in that particular food category.

When labelling these food SGBs as signatures of meat, dairy, and/or fruits and vegetables if they had a prevalence >0.1% across these food groups, the effect of food group was significant and larger than the effect of diet pattern, with a greater number of food SGBs found for dairy (β = 0.73, P < 0.001), followed by fruits and vegetables (β = 0.47, P < 0.001). Thus, the largest factor impacting food-to-gut species sharing is the food group that SGBs are derived from, with greatest sharing from dairy products and lowest from meat. The number of food-associated SGBs was greatest in omnivores and vegetarians who both eat foods from the groups with the highest transmission (dairy, fruits and vegetables), and lowest in vegans who exclude meat and, more importantly here, dairy products, thus probably minimizing food-to-gut transmission rates.

We further found that the cumulative relative abundance of food SGBs in vegetarian gut microbiomes was significantly higher than in both omnivore (zero-inflated linear mixed-effects model, β = 0.38, P < 0.001) and vegan (β = 0.51, P < 0.001) gut microbiomes, and significantly lower in vegans compared with omnivores (β = −0.12, P = 0.026; Fig. 5a). This again highlights the minimal food-to-gut species sharing in vegans. The fact that vegetarians had a greater cumulative relative abundance of food SGBs than omnivores but a similar number of distinct food SGBs may reflect the similar richness of food SGBs ingested (especially since meat-derived SGBs play an inferior role compared with SGBs derived from dairy and fruits and vegetables), but also the greater amount of fruits and vegetables ingested by vegetarians (Fig. 1c) instead of meat, which in turn drives a higher cumulative relative abundance of food SGBs. To summarize, dietary choices are linked to changes in the gut microbiome via not only potential selection but also food-to-gut acquisition.

Food–gut shared microbes differ across diet patterns

We then identified 20 food SGBs with the highest prevalence across the five cohorts (Fig. 5c) and which major food groups these SGBs were signatures of (Fig. 5d). As expected, S. thermophilus was among them, showing the greatest prevalence in dairy and significantly lowest prevalence in vegans (chi-squared tests; Methods, Fig. 5c and Supplementary Table 17). Similar patterns were observed for common dairy SGBs, for example, L. acidophilus, L. delbrueckii, L. lactis, L. paracasei and L. rhamnosus27,28,33—all SGBs that we found to be most greatly differentiated between vegan and non-vegan diet patterns. To lend more support to this hypothesis, we assessed omnivore and vegetarian frequency of dairy consumption (milk, yogurt, cheese, butter, other dairy) according to FFQs. We found that 96% of omnivores and 90% of vegetarians consume dairy at least once per week (Extended Data Fig. 3) with similar fractions (90% and 84% respectively) when restricting to fermented dairy products (yogurt and cheeses; Extended Data Fig. 3). Thus, we conclude that, while some microbe signatures of diets that include dairy could be selected to help digest dairy, others could be present in the gut microbiome as transient members derived from dairy foods themselves.

Several food SGBs with a high prevalence in vegans, such as Enterobacter hormaechei34, Citrobacter freundi35, Raoultella ornithinolytica36 and Klebsiella pneumoniae37,38 are members of the soil microbiome and/or nitrogen-fixing bacteria. Among them, E. hormaechei promotes growth in tomato and sweet pepper plants34,39, while some strains of K. pneumoniae are nitrogen fixers and thus used as plant-growth promoters in wheat and soybeans37,38. This supports previous findings that, aside from more obvious possible sources of food-to-gut transmission such as cultured dairy products, agricultural practices could also play a role