Machine Learning Metabolomics Profiling of Dietary Interventions from a Six-Week Randomised Trial

Abstract

Metabolomics can uncover physiological responses to prebiotic fibre and omega-3 fatty acid supplements with known health benefits and identify response-specific metabolites. We profiled 534 stool and 799 serum metabolites in 64 healthy adults following a 6-week randomised trial comparing daily omega-3 versus inulin supplementation. Elastic net regressions were used to separately identify the serum and stool metabolites whose change in concentration discriminated between the two types of supplementations. Random forest was used to explore the gut microbiome's contribution to the levels of the identified metabolites from matching stool samples. Changes in serum 3-carboxy-4-methyl-5-propyl-2-furanpropanoate and indoleproprionate levels accurately discriminated between fibre and omega-3 (area under the curve (AUC) = 0.87 [95% confidence interval (CI): 0.63-0.99]), while stool eicosapentaenoate indicated omega-3 supplementation (AUC = 0.86 [95% CI: 0.64-0.98]). Univariate analysis also showed significant increases in indoleproprionate with fibre, 3-carboxy-4-methyl-5-propyl-2-furanpropanoate, and eicosapentaenoate with omega-3. Out of these, only the change in indoleproprionate was partly explained by changes in the gut microbiome composition (AUC = 0.61 [95% CI: 0.58-0.64] and Rho = 0.21 [95% CI: 0.08-0.34]) and positively correlated with the increase in the abundance of the genus Coprococcus (p = 0.005). Changes in three metabolites discriminated between fibre and omega-3 supplementation. The increase in indoleproprionate with fibre was partly explained by shifts in the gut microbiome, particularly Coprococcus, previously linked to better health.

Keywords: fibre; gut microbiome; indoleproprionate; machine learning; metabolomics; omega-3.

Figure 1

Schematic diagram of the study design.

Figure 2

Changes in serum metabolomic data in response to 6 week supplementation with inulin or omega-3. (A) Receiver operating characteristic (ROC) curve using the selected features from the five-fold cross-validated elastic net model repeated five times; (B) boxplots of the performance [area under the curve (AUC) and quantitative estimate (Spearman’s correlation coefficient)] of a random forest model incorporating the metabolites identified by elastic net. The mean and 95% confidence intervals (CIs) of both the AUC and the Spearman’s Rho across the repeated folds are shown; (C) variance importance plot of the elastic net coefficients of the metabolites whose changed levels in serum were identified by elastic net regression as discriminating between the two interventions; and (D) boxplots showing the change in the identified serum metabolites in both arms separately (univariate paired t-test). ns: not significant, **** p < 0.0001.

Figure 3

Changes in stool metabolomic data in response to 6 week supplementation with inulin or omega-3. (A) ROC curve using the selected feature from the five-fold cross-validated elastic net model repeated five times; (B) boxplots of the performance (AUC) and quantitative estimate (Spearman’s correlation coefficient) of a random forest model incorporating the metabolites identified by elastic net. The mean and 95% CIs of both the AUC and the Spearman’s Rho across the repeated folds are shown; (C) variance importance plot (elastic net coefficient) of the metabolites whose changed levels in serum were identified by elastic net regression as discriminating between the two interventions; and (D) boxplots showing the change in the identified serum metabolites in both arms (univariate paired t-test). ns: not significant, **** p < 0.0001.

Figure 4

Associations of differentiating metabolites with SCFAs and gut microbiome composition. (A) Lack of association between the change in discriminating metabolites and change in circulating SCFA levels; (B) Random forest model showing to what extent can changes in gut microbiome composition in response to the intervention explain changes in the metabolites that can discriminate between the two interventions; and (C) Spearman’s correlations between change in circulating IPA and changes in two key taxa which increased with the intervention (Coprococcus and Bifidobacterium).