Gut micro-organisms associated with health, nutrition and dietary interventions

Abstract

The incidence of cardiometabolic diseases is increasing globally, and both poor diet and the human gut microbiome have been implicated¹. However, the field lacks large-scale, comprehensive studies exploring these links in diverse populations². Here, in over 34,000 US and UK participants with metagenomic, diet, anthropometric and host health data, we identified known and yet-to-be-cultured gut microbiome species associated significantly with different diets and risk factors. We developed a ranking of species most favourably and unfavourably associated with human health markers, called the 'ZOE Microbiome Health Ranking 2025'. This system showed strong and reproducible associations between the ranking of microbial species and both body mass index and host disease conditions on more than 7,800 additional public samples. In an additional 746 people from two dietary interventional clinical trials, favourably ranked species increased in abundance and prevalence, and unfavourably ranked species reduced over time. In conclusion, these analyses provide strong support for the association of both diet and microbiome with health markers, and the summary system can be used to inform the basis for future causal and mechanistic studies. It should be emphasized, however, that causal inference is not possible without prospective cohort studies and interventional clinical trials.

Fig. 1. The ZOE PREDICT studies comprise over 34,000 healthy people from five cross-sectional studies from the UK and the USA with gut microbiome samples, detailed individual information and dietary habits.

a, In this study, we considered and harmonized five cross-sectional ZOE PREDICT cohorts with participants from the UK and the USA (Supplementary Fig. 1). For each cohort, sample size and the percentage of female participants (% F) are reported in the upper bar plots, with sequencing depth (left-hand columns, darker colour, average size in gigabases) and the total number of detected species (right-hand columns, lighter colour) are reported in the middle bar plots, showing that cohorts with lower sequencing depths do not have fewer total numbers of detected species. Bottom box plots, distributions of age (left-hand columns, darker colour) and BMI (right-hand columns, lighter colour) in the five PREDICT cohorts (the PREDICT 1 (P1) cohort is split into its UK and US parts, but considered as a single cohort). Box plots show first and third quartiles (boxes) and the median (middle line); whiskers extend up to 1.5 × interquartile range (IQR). b, Random forest classification (discriminating the first three quartiles against the fourth quartile) and regression machine learning models (Methods) trained on the whole microbiome SGB-level relative abundance values with a cross-validation approach, show moderately strong and consistent associations with different categories of clinical data available across the five cross-sectional ZOE PREDICT cohorts (full machine learning results are reported in Extended Data Fig. 1 and Supplementary Tables 2 and 3). HEI, healthy eating index; PDI, plant-based diet index; hPDI, healthful PDI; oPDI, overall PDI.

Fig. 2. The 15 top and bottom health-ranked SGBs show consistent associations across the five PREDICT cohorts.

a, Average percentiles for the 15 most favourably (top) and unfavourably (bottom) ranked SGBs (selected for visualization purposes) across all five PREDICT cohorts. Percentiles are computed from the ranking of the correlations between SGBs and the different markers in each clinical data category. Percentiles close to 0 reflect SGBs consistently correlated positively with positive markers and negatively with negative markers, and vice versa for percentiles close to 1. For each cohort, the average percentiles for three clinical data categories are shown (personal, fasting and postprandial) and the cohort-level average. The rightmost column of the heatmap reports the ZOE MB health-ranks with the distribution of their relative abundance values when present (derived from n = 34,694 participants spanning the five PREDICT cohorts). Box plots as in Fig. 1. b,c, Detailed percentiles for the 15 most favourably and unfavourably ranked SGBs against the markers of the three clinical data categories of the PREDICT 1 (UK) (b) and PREDICT 3 US22A (US) (c) cohorts. Detailed panels of the percentiles for the other three cohorts can be found in Extended Data Fig. 2. iAUC, incremental area under the curve; PUFA, polyunsaturated fatty acid; QUICKI, quantitative insulin sensitivity check index; THR, total-cholesterol-to-HDL ratio; VLDL, very-low-density lipoprotein.

Fig. 3. ZOE MB health- and diet-ranked species show significant and reproducible associations with BMI and diseases.

a, Concordance of ZOE MB health-ranks with partial Spearman’s correlations against BMI (corrected for sex and age) across PREDICT cohorts. Favourably ranked SGBs correlate negatively with BMI; unfavourably ranked SGBs correlate positively (ZOE MB diet-ranks in Extended Data Fig. 7a). Shading represents 95% confidence interval. b,c, Cumulative relative abundance of favourably (b) and unfavourably (c) ranked SGBs across BMI categories. As BMI increases, reflecting higher health risks, the abundance of favourable SGBs decreases whereas that of unfavourable SGBs increases. Similar patterns were seen for SGB richness (Extended Data Fig. 7b,c). Only non-significant (NS) false discovery rate (FDR)-corrected P values (Q > 0.01, two-sided Mann–Whitney U-test) are annotated. Box plots as in Fig. 1. d, Meta-analysis of the 50 most favourable and unfavourable SGBs comparing participants of healthy weight with those with obesity from public cohorts. Lower BMI is associated with more favourable SGBs; people with higher BMI carry more unfavourable SGBs. Meta-analysis on ranks defined on UK and US participants shows reproducibility across countries. Other comparisons are in Extended Data Fig. 8 and the diet-ranked SGBs meta-analysis in Extended Data Fig. 9. Country codes: ARG, Argentina; AUT, Austria; DEU, Germany; DNK, Denmark; FRA, France; GBR, United Kingdom of Great Britain and Northern Ireland; IRL, Ireland; ISR, Israel; KAZ, Kazakhstan; NLD, Netherlands; USA, United States of America. e, Meta-analysis of disease group (adjusted by sex, age and BMI) on standardized mean differences (SMD) of cumulative relative abundance of the 50 most favourable (left) and unfavourable (right) SGBs from both rankings (meta-analysis on SGB richness in Supplementary Fig. 5a; Methods). f, Meta-analysis of normalized ZOE MB health-ranks and diet-ranks, weighted by arcsin square-root of relative abundance values (right, weighted score sum; left, score sum (unweighted)). Horizontal lines in meta-analysis plots represent 95% confidence intervals. CRC, colorectal cancer; IBD, inflammatory bowel disease; IGT, impaired glucose tolerance.

Fig. 4. Dietary interventions have a large impact on microbiome composition.

a,b, Pre–post dietary intervention variations in prevalent gut microbial SGBs (at least 10% at both time points). The plots show the effect size (log₂-transformed ratio of mean relative SGB abundance at endpoint over baseline) against the significance (Q values, FDR–Benjamini–Hochberg-corrected P values). a, BIOME cohort (ClinicalTrials.gov NCT06231706) with n = 321 healthy adults from the UK (n = 106 prebiotic blend, n = 106 probiotic and n = 109 control), significance threshold set to Q < 0.01. b, METHOD cohort (ClinicalTrials.gov NCT05273268) with n = 347 US individuals (n = 177 PDP, n = 170 control), and significance threshold set to Q < 0.1. c, Change in relative abundance for the significant SGBs in the intervention arms of BIOME (prebiotic blend, n = 57). d, Change in relative abundance of METHOD (PDP, n = 46), separated into those that increase from those that decrease from baseline (B) to endpoint (E). Extended Data Fig. 10a–d reports the change in relative abundance and prevalence of the control and prebiotic arms. Two-sided Wilcoxon test; box plots as in Fig. 1).

Fig. 5. Gut microbial SGBs that increase after dietary interventions are linked to more favourable ZOE MB health- and diet-ranks.

a, The 20 most significant gut microbial SGBs with the greatest effect sizes following the BIOME dietary intervention from Fig. 4a (left), paired with their ZOE MB health-ranks and diet-ranks, if available (right). b, The 20 most significant gut microbial SGBs with the greatest effect sizes following the METHOD personalized dietary intervention programme from Fig. 4b (left), paired with their ZOE MB health-ranks and diet-ranks, if available (right). The x axis shows the log₂-transformed ratio of mean relative abundance SGB values at endpoint over baseline. All values are reported in Supplementary Table 25. c,d, The distributions of the ZOE MB health-ranks (c) and diet-ranks (d) for the prebiotic blend arm of the BIOME cohort (n = 57 of tested SGBs). e,f, The distributions of the ZOE MB health-ranks (e) and diet-ranks (f) for the PDP arm of the METHOD cohort (n = 46 of tested SGBs). The distributions show that SGBs increasing in relative abundance have significantly more favourable ranks, whereas decreasing SGBs have more unfavourable ranks (two-sided Mann–Whitney U-test, P = 7.78 × 10⁻³, P = 3.00 × 10⁻⁵, P = 5.20 × 10⁻⁵ and P = 2.03 × 10⁻⁵, respectively). Distributions of the ZOE MB health-ranks and diet-ranks for the other arms are reported in Extended Data Fig. 10e,f. Box plots as in Fig. 1.

Extended Data Fig. 1. Microbiome predictive potential for personal information, dietary indices, fasting, and postprandial metabolic markers, via classification and regression random forest models.

Distributions of the random forest median AUCs (a) and median Spearman’s correlation coefficients (b) (Methods) in the five cross-sectional PREDICT studies for the different clinical data divided into four categories: ‘Personal’, ‘Dietary’, ‘Fasting’, and ‘Postprandial’. The AUC and Spearman’s index thresholds of 0.7 and 0.3, respectively, are indicated with a dashed line. a) Each point represents the median AUC value obtained in cross-validation for each cohort when testing the first versus the fourth quartile of the corresponding clinical marker values on the x-axis. b) Each point represents the median Spearman’s correlation coefficient for the predicted values by the regressor and the true values in the cross-validation setting for each cohort.

Extended Data Fig. 2. Detailed associations of the 15 top and bottom cardiometabolic-ranked SGBs in the PREDICT3 UK22A, PREDICT2, and PREDICT3 US21 cohorts.

The single-marker percentiles, divided into the three categories (‘Personal’, ‘Fasting’, and ‘Postprandial’) for the 15 most favorable and unfavorable ZOE MB Health-ranked SGBs the other three PREDICT cohorts not reported in Fig. 2 (a, PREDICT3 UK22A; b, PREDICT2, and c, PREDICT3 US21). Heatmaps with the single Spearman’s partial correlations for all PREDICT cohorts are available in Supplementary Fig. 4.

Extended Data Fig. 3. Spearman’s partial correlations of the 15 top and bottom cardiometabolic-ranked SGBs.

a-e) Spearman’s partial correlations (corrected for age, sex, and BMI) between SGB relative abundance and single marker values show consistency across the five PREDICT cohorts. These partial correlations were ranked and averaged first within and then across the three data categories (‘Personal’, ‘Fasting’, and ‘Postprandial’, reported in Supplementary Fig. 3) separately in each cohort. The cohorts’ averages were then used to define the cardiometabolic rank (for those SGBs analyzed in at least two cohorts).

Extended Data Fig. 4. Diet associations of the 15 top and bottom diet-ranked SGBs.

a-e) For each PREDICT cohort, we computed Spearman’s partial correlation between the SGBs’ relative abundances and different diet indexes. Associations were ranked and averaged in each cohort separately. f) The ZOE MB Diet-ranking was computed for SGBs ranked in at least two PREDICT cohorts. The raw Spearman’s partial correlations are available in Supplementary Fig. 7.

Extended Data Fig. 5. Spearman’s partial correlations of the 15 top and bottom diet-ranked SGBs.

a-e) Study-wise Spearman’s partial correlation coefficients (corrected for sex, age, and BMI) for the 15 most favorable and unfavorable ZOE MB Diet-ranked SGBs in different diet indexes. The associations appear consistent across cohorts.

Extended Data Fig. 6. Comparison of the ZOE MB Health and Diet ranks and with geography.

a) The ZOE MB Health and Diet ranks are overall in agreement (Spearman’s correlation = 0.72), albeit some SGBs show discordant rankings (absolute difference between the two ranks ≥ 0.3). These SGBs are highlighted in orange, and their ranks and taxonomy assignment are reported in Supplementary Table 6. b,c) Comparison of the ZOE MB Health (b) and Diet (c) ranks computed only on the PREDICT UK and US cohorts (Spearman’s correlations of 0.61 and 0.26, respectively). The top and right-side histograms depict the x and y-axis marginal distributions in each plot.

Extended Data Fig. 7. ZOE DIET ranks and their associations with BMI.

a) Comparison of the ZOE MB Diet-ranks (x-axis) with the Spearman’s partial correlations (corrected for sex and age, y-axis) for the 661 ranked SGBs in the five PREDICT cohorts. b) The number of the 50 most-favorably ranked SGBs (ZOE MB Health-rank, Richness) detected in different BMI categories, showed that increasing BMI, linked with increasing health risks, is reflected by a lower presence of favorable SGBs. On the other hand, c) unfavorably-ranked SGBs show an increasing count in higher-risk BMI categories. d,e) The box plots report the number of the 50 most favorable and unfavorable ZOE MB Diet-ranked SGBs of individuals stratified into three BMI categories (healthy-weight, overweight, and obese) in each PREDICT cohort. f,g) Similarly, the box plots represent the cumulative relative abundance of the 50 most favorable and unfavorable ZOE MB Diet-ranked SGBs in individuals categorized into the three BMI categories in each cohort. h,i) The box plots report the number of the 50 most favorably and most unfavorably ranked SGBs, ranked using the same markers and categories as in the ZOE MB Health-ranks (Methods), but partial correlations were corrected only for sex and age. j,k) Similarly, the box plots report the count of the 50 most favorable and unfavorable SGBs in the three BMI categories, with SGBs ranked according to their partial correlation with BMI, adjusted by sex and age. Only non-significant FDR-corrected P values (ns, P value > 0.01) from the Mann-Whitney U test are reported.

Extended Data Fig. 8. Meta-analysis of the 50 most favorably and unfavorably ZOE MB Health-ranked SGBs in overweight vs obese and healthy-weight vs overweight individuals.

a) Overweight individuals tend to carry a higher number of the 50 most favorably ZOE MB Health-ranked SGBs than obese individuals (left); the 50 most unfavorably ranked SGBs are increased in obese individuals vs overweight individuals (Methods). b) Healthy-weight individuals tend to carry a higher number of the 50 most favorably ZOE MB Health-ranked SGBs than overweight individuals (left); the 50 most unfavorably ranked SGBs are found in similar amounts in healthy-weight and overweight individuals (Methods). Error bars represent the 95% confidence interval.

Extended Data Fig. 9. Meta-analysis of the 50 most favorably and unfavorably ZOE DIET-ranked SGBs comparing individuals from different BMI categories.

a) Comparison of the number of the 50 most favorable Diet-ranked SGBs in pairs of BMI categories. Healthy-weight and overweight individuals tend to have a higher number of favorably-ranked SGBs than obese individuals (Methods). b) Comparison of the number of the 50 most unfavorably Diet-ranked SGBs in pairs of BMI categories. Obese individuals tend to have a higher number of unfavorably-ranked SGBs (Methods). Error bars represent the 95% confidence interval.

Extended Data Fig. 10. Significantly changing SGBs after dietary interventions show consistent patterns across cohorts in terms of relative abundance, prevalence, and ZOE MB Diet-ranks.

a) Distributions of the mean relative abundance of the significant SGBs for the probiotic and control arms of the BIOME cohort (relative to Fig. 4a). b) Distributions of the mean relative abundance of the significant SGBs for the control arm of the METHOD cohort (relative to Fig. 4b). c) Distributions of the prevalence of the significant SGBs of the BIOME cohort (relative to Fig. 4a) and d) of the METHOD cohort (relative to Fig. 4b). SGBs are separated into “increasing” and “decreasing”, depending on their trend in relative abundance values, showing that SGBs found to be increased in relative abundance are also more prevalent, while the opposite is observed for SGBs decreasing in relative abundance. e) Distributions of the ZOE MB Health ranks for the significant SGBs in the Probiotic and Control arms of the BIOME cohort and METHOD cohorts. f) Distributions of the ZOE MB Diet ranks for the significant SGBs in the Probiotic and Control arms of the BIOME and METHOD cohorts.