Faecal metabolites as a readout of habitual diet capture dietary interactions with the gut microbiome

Abstract

The interplay between diet and gut microbiome composition is complex. Faecal metabolites, the end products of human and microbial metabolism, provide insights into these interactions. Here, we integrate faecal metabolomics, metagenomics, and habitual dietary data from 1810 individuals from the TwinsUK and 837 from the ZOE PREDICT1 cohorts. Using machine learning models, we find that faecal metabolites accurately predict reported intakes of 20 food groups (area under the curve (AUC) > 0.80 for meat, nuts and seeds, wholegrains, tea and coffee, and alcohol) and adherence to seven dietary patterns (AUC from 0.71 for the Plant-based Diet Index to 0.83 for the Dietary Approaches to Stop Hypertension score). Notably, the faecal metabolome is a stronger predictor of atherosclerotic cardiovascular disease risk (AUC = 0.86) than the Dietary Approaches to Stop Hypertension score (AUC = 0.66). We identify 414 associations between 19 food groups and 211 metabolites, that significantly correlate with microbial α-diversity and 217 species. Our findings reveal that faecal metabolites capture mediations between diet and the gut microbiome, advancing our understanding of diet-related disease risk and informing metabolite-based interventions.

Fig. 1. Faecal metabolites act as a readout of habitual diet and bridge the gap between diet and the gut microbiome.

The schematic describes the discovery and replication cohorts as well as the dietary, faecal metabolomics, and metagenomics datasets, highlighting the univariate and multivariate analyses carried out in response to the aims of the study. UHPLC-MS/MS ultra high-performance liquid chromatography-tandem mass spectrometry, FFQ food frequency questionnaire, LMER linear mixed-effects regression, PERMANOVA permutational multivariate analysis of variance. Created in BioRender. Pope, R. (2025) https://BioRender.com/c51b012.

Fig. 2. Random forest machine learning models trained on faecal metabolite profiles accurately predict diet quality, identifying a targeted panel of 54 dietary-predictive metabolites.

a Prediction of adherence to seven dietary patterns, habitual intakes of 20 food and beverage groups, and 10-year atherosclerotic cardiovascular disease (ASCVD) risk using faecal metabolites as predictors for RF binary classification (TwinsUK n = 905; ZOE PREDICT1 n = 159) and regression models (TwinsUK n = 1810). The left y-axis reports the area under the curve (AUC) values for RF binary classification models (TwinsUK cohort 20% as circles; ZOE PREDICT1 as diamonds). The right y-axis reports the distribution of the Spearman’s rank correlation coefficients between predicted and observed labels for the TwinsUK 20% testing set, computed by 1000 bootstrapped samples with replacement. Boxplots show the median (centre line) and interquartile range (box limits), with whiskers extending to 1.5 times the interquartile range. b, Receiver operating characteristic (ROC) curves for prediction of top and bottom quartiles of 10-year ASCVD risk using faecal metabolites, BMI, and the Dietary Approaches to Stop Hypertension (DASH) diet score. c Venn diagram of the 54 faecal metabolites identified as part of the metabolite panels for prediction of adherence to dietary patterns and habitual intake of nine food and beverage groups with an original AUC > 0.70. d Sankey plot of the corresponding metabolic pathways of the 54 faecal metabolites included in the panel. Created in BioRender. Pope, R. (2025) https://BioRender.com/9z4te71.

Fig. 3. Faecal metabolites represent a readout of habitual consumption of food and beverage groups.

Chord diagram of 222 significant positive associations between characterised faecal metabolites and food and beverage groups. Associations identified from fixed-effects meta-analysis of linear mixed-effects regression models, corrected for age, sex, BMI, and twin family structure in the TwinsUK (n = 1810) and ZOE PREDICT1 (n = 318) cohorts, were considered significant below a Bonferroni-derived threshold of 9.78 × 10⁻⁵. Faecal metabolites are organised and coloured according to their metabolic superpathways. To facilitate the interpretation of the results we show here the positive associations, as they reflect intake of the food and beverage groups instead of dietary choices between food groups. A figure with the full set of 414 associations is depicted in Supplementary Fig. 4. Created in BioRender. Pope, R. (2025) https://BioRender.com/06wl9fo.

Fig. 4. Dietary-associated faecal metabolites are strongly associated with gut microbiome composition and species, providing insights into diet-microbiome interactions.

a Clustered heatmap of 23 faecal metabolites (y-axis) significantly associated with ≥4 food and beverage groups for the PDI (x-axis). Associations were tested using linear mixed-effects regression models (two-sided) in TwinsUK (n = 1810) and ZOE PREDICT1 (n = 318), adjusted for age, sex, BMI, and twin family structure. Results from both cohorts were combined using fixed-effects meta-analysis. Significance was defined as meta-analysis-derived p values below the Bonferroni threshold (p < 9.78 × 10⁻⁵). The tile colour represents the direction of association (blue = negative, red = positive); * denotes significance. Hierarchical clustering used the ACC distance metric and defined clusters based on the optimal silhouette score. Exact regression coefficients (β), standard errors, and p values are reported in Supplementary Data 7. b Associations of faecal metabolites with gut microbiome α-diversity (forest plot) and food groups (heatmap) from fixed-effects meta-analysis of β estimates (TwinsUK = purple squares, ZOE PREDICT 1 = red squares, meta-analysis = blue diamonds). The whiskers represent the 95% confidence intervals around the β estimates. Associations were considered significant if the meta-analysis-derived p value was below the Bonferroni thresholds of p < 2.51 × 10⁻⁴ (α-diversity) and p < 9.78 × 10⁻⁵ (food groups). Exact regression coefficients (β), standard errors, and p values are reported in Supplementary Data 7 and 9. Mediation analysis in the TwinsUK cohort (n = 474) showing direct and indirect effects of microbial species (c) or faecal metabolite (d) mediators in response to habitual diet. Linear mixed-effects regression models (two-sided) adjusted for age, sex, BMI, and accounting for twin family structure, were used to define direct (solid arrows) and indirect (dotted arrows) paths. Significance: *p < 0.05, **p < 0.01, ***p < 0.001. The average causal effect (ACME) and average direct effect (ADE) are summed to give the total effect with Prop. Mediated showing the percentage of the effect attributed to the mediator. The metabolite isohyodeoxycholate is abbreviated to isoHDCA. Test statistics are reported in Supplementary Data 13. Created in BioRender. Pope, R. (2025) https://BioRender.com/ksv93ie.