The intestinal microbiome is a co-determinant of the postprandial plasma glucose response

Abstract

Elevated postprandial plasma glucose is a risk factor for development of type 2 diabetes and cardiovascular disease. We hypothesized that the inter-individual postprandial plasma glucose response varies partly depending on the intestinal microbiome composition and function. We analyzed data from Danish adults (n = 106), who were self-reported healthy and attended the baseline visit of two previously reported randomized controlled cross-over trials within the Gut, Grain and Greens project. Plasma glucose concentrations at five time points were measured before and during three hours after a standardized breakfast. Based on these data, we devised machine learning algorithms integrating bio-clinical, as well as shotgun-sequencing-derived taxa and functional potentials of the intestinal microbiome to predict individual postprandial glucose excursions. In this post hoc study, we found microbial and clinical features, which predicted up to 48% of the inter-individual variance of postprandial plasma glucose responses (Pearson correlation coefficient of measured vs. predicted values, R = 0.69, 95% CI: 0.45 to 0.84, p<0.001). The features were age, fasting serum triglycerides, systolic blood pressure, BMI, fasting total serum cholesterol, abundance of Bifidobacterium genus, richness of metagenomics species and abundance of a metagenomic species annotated to Clostridiales at order level. A model based only on microbial features predicted up to 14% of the variance in postprandial plasma glucose excursions (R = 0.37, 95% CI: 0.02 to 0.64, p = 0.04). Adding fasting glycaemic measures to the model including microbial and bio-clinical features increased the predictive power to R = 0.78 (95% CI: 0.59 to 0.89, p<0.001), explaining more than 60% of the inter-individual variance of postprandial plasma glucose concentrations. The outcome of the study points to a potential role of the taxa and functional potentials of the intestinal microbiome. If validated in larger studies our findings may be included in future algorithms attempting to develop personalized nutrition, especially for prediction of individual blood glucose excursions in dys-glycaemic individuals.

Fig 1. Data overview and experimental design.

Host variables including phenomics, biochemical and lifestyle and gut microbiome variables were summarized as functional and species abundance profiles adjusted for bacterial cell counts of faecal samples. Data were randomly split into training (70%) and test (30%) sets. By two approaches, a hypothesis-driven and an explorative approach, features were selected and ranked according to the number of times selected in 200 rounds of top-down searches for important features applying wrapper algorithms build around random forest models predicting postprandial plasma glucose responses. In the hypothesis-based approach the number of features were reduced from 36 features selected from a priori knowledge to eight features, whereas in the explorative approach eight features were selected from all of the data comprising a total of 1368 features. For each approach, a random forest model predicting postprandial plasma glucose responses based on the eight selected features was trained and evaluated on the test set by correlating predicted with measured postprandial plasma glucose excursions. A combined random forest model predicting postprandial plasma glucose responses was trained based on the top features from each feature selection results thereby exploiting both approaches. The combined model was evaluated on the test set by correlating predicted with measured postprandial plasma glucose responses and the importance of the predictor features in terms of magnitudes and directions were evaluated individually and for the microbiome features combined. Finally, glycaemic features were added to the combined model to enable comparisons to previous findings in the literature.

Fig 2. Association between predicted and measured postprandial plasma glucose.

Predictions are based on the combined model including the following features: fasting serum concentration of triglycerides, age, fasting total serum cholesterol concentration, abundance of MGS.hg0341, abundance of bifidobacteria, BMI, systolic blood pressure and MGS richness. The black line represents the fitted regression line and the grey shaded area represents the 95% CIs. (A) The association in the training set (n = 75). Pearson R = 0.90, 95% CI: 0.84 to 0.93 and p<0.001. (B) The association in the test set (n = 31). Pearson R = 0.69, 95% CI: 0.45 to 0.84 and p<0.001.

Fig 3. Contributions of all features in the combined model.

Contributions are calculated by removing the feature of interest from the combined model, retrain, evaluate the model by correlating measured and predicted postprandial plasma glucose responses in the test set and subsequently subtract the Pearson correlation R from the one obtained from the combined model including all of the features (delta R). Circles are scaled according to prevalence (prevalence refers to the number of individuals harboring the microbiome feature in %) and coloured according to abundance (abundance is the relative representation of the individual microbiome feature in the gut microbiome in %) (n = 31).

Fig 4. Partial dependence plots showing the marginal contribution of each feature in the combined model to the predicted postprandial plasma glucose responses.

A positive trend indicates a non-beneficial effect of the feature whereas a negative trend indicates a beneficial effect. Dot plots indicate the distributions of each feature in the study population, with the median is indicated by a blue diamond (n = 31).

Fig 5. Comparison of the association strengths between the actual and predicted postprandial glucose responses in the test set based on the five different models and a perfect correlation as a comparison (exact); the combined model including glycaemic variables in the fasting state, the combined model, the bio-clinical features-only model, microbiome-only model and the null model of the permuted glucose responses.

The lines represent the fitted regression lines and the corresponding shaded area represent the 95% CIs for each model, respectively. Exact: Pearson R = 1. The combined model including glycaemic variables in the fasting state: Pearson R = 0.78, 95% CI: 0.59–0.89 and p<0.001. The combined model: Pearson R = 0.69, 95% CI: 0.45–0.84 and p<0.001. The bio-clinical features-only model: Pearson R = 0.69, 95% CI: 0.45–0.84 and p<0.001. The microbiome-only model: Pearson R = 0.37, 95% CI: 0.02–0.64 and p = 0.04. The null model of the permuted glucose responses: Pearson R = -0.13, 95% CI: -0.47.0.25 and p = 0.51.