Gut microbiota signatures predict host and microbiota responses to dietary interventions in obese individuals

Abstract

Background: Interactions between the diet and intestinal microbiota play a role in health and disease, including obesity and related metabolic complications. There is great interest to use dietary means to manipulate the microbiota to promote health. Currently, the impact of dietary change on the microbiota and the host metabolism is poorly predictable and highly individual. We propose that the responsiveness of the gut microbiota may depend on its composition, and associate with metabolic changes in the host.

Methodology: Our study involved three independent cohorts of obese adults (n = 78) from Belgium, Finland, and Britain, participating in different dietary interventions aiming to improve metabolic health. We used a phylogenetic microarray for comprehensive fecal microbiota analysis at baseline and after the intervention. Blood cholesterol, insulin and inflammation markers were analyzed as indicators of host response. The data were divided into four training set - test set pairs; each intervention acted both as a part of a training set and as an independent test set. We used linear models to predict the responsiveness of the microbiota and the host, and logistic regression to predict responder vs. non-responder status, or increase vs. decrease of the health parameters.

Principal findings: Our models, based on the abundance of several, mainly Firmicute species at baseline, predicted the responsiveness of the microbiota (AUC = 0.77-1; predicted vs. observed correlation = 0.67-0.88). Many of the predictive taxa showed a non-linear relationship with the responsiveness. The microbiota response associated with the change in serum cholesterol levels with an AUC of 0.96, highlighting the involvement of the intestinal microbiota in metabolic health.

Conclusion: This proof-of-principle study introduces the first potential microbial biomarkers for dietary responsiveness in obese individuals with impaired metabolic health, and reveals the potential of microbiota signatures for personalized nutrition.

Figure 1. Result of data normalization.

Principal co-ordinates plots (with Bray-Curtis distances) show that the microbiota profiles segregate based on the study before (panel A), but not after (panel B) data normalization.

Figure 2. Validation of the microbiota responsiveness model.

The model selection and validation were conducted four times, each time leaving out one study (marked with letters A-D). The resulting model was used to predict the stability values in the left-out study. The dashed line represents the ideal situation where observed  =  predicted.

Figure 3. Predicting cholesterol responses to dietary intervetions.

Panels A, B, C: Three cholesterol response models: cholesterol response predicted by the microbiota stability (panel A), by the baseline abundance of E. ruminantium and C. felsineum (B), and by the baseline abundance of C. sphenoides (C). The data were divided randomly into a training set (75% of the data) and test set (the remaining 25%), and the ROC curves represent the ability of the models, fitted to the training data, to predict the cholesterol response (increase vs. decrease) in the test data. The ROC curve shows the true positive rate ( = sensitivity) against the false positive rate ( = 1-specificity) for the different possible cut points of a diagnostic test. The perfect diagnostic test would have a sensitivity  =  1 and specificity  =  1, and therefore the area under the curve (AUC) would be 1. A random guess would have a ROC curve following the diagonal; curves above the diagonal indicate that the classifier works better than a random guess. Shaded areas represent 95% confidence intervals for the ROC curve. Panels D, E, F: Comparison of cholesterol response groups (increase vs. decrease), with respect to microbiota stability (D), E. ruminantium and C. felsineum abundance (E), and C. sphenoides abundance (F).

Figure 4. Validation of the HOMA (panel A) and CRP (panel B) response models.

In each case, one study was left out, while data from the other studies were fitted to the model, which was then used to predict the HOMA and CRP response for the independent data set (A–D). The dashed line represents the ideal situation where observed  =  predicted.