Comparison of fecal and blood metabolome reveals inconsistent associations of the gut microbiota with cardiometabolic diseases

Abstract

Blood metabolome is commonly used in human studies to explore the associations of gut microbiota-derived metabolites with cardiometabolic diseases. Here, in a cohort of 1007 middle-aged and elderly adults with matched fecal metagenomic (149 species and 214 pathways) and paired fecal and blood targeted metabolomics data (132 metabolites), we find disparate associations with taxonomic composition and microbial pathways when using fecal or blood metabolites. For example, we observe that fecal, but not blood butyric acid significantly associates with both gut microbiota and prevalent type 2 diabetes. These findings are replicated in an independent validation cohort involving 103 adults. Our results suggest that caution should be taken when inferring microbiome-cardiometabolic disease associations from either blood or fecal metabolome data.

Fig. 1. Workflow of the present study.

A total of 1007 participants from the Guangzhou Nutrition and Health Study with matched gut metagenomic and fecal and blood metabolomics data and without taking antibiotics within two weeks are included in this study. Shotgun metagenomic sequencing is performed for fecal samples to obtain the metagenomic data, including taxonomic composition and microbial pathways. A targeted metabolome profiling is performed to obtain the fecal and blood metabolomics data. After removing metabolites with missing rate ≥ 0.2 or with relative standard deviation ≥ 0.3, 132 matched fecal and blood metabolites are remained for subsequent analysis. We estimate the gut microbiota-fecal/blood metabolite associations using the machine learning pipeline, and compare the associations of taxonomic composition/microbial pathways with paired fecal and blood metabolites. We then explore the associations of gut microbiota-related fecal and blood metabolites with cardiometabolic diseases. We further replicate the identified significant associations in an independent validation cohort. This figure has been designed using images from Flaticon.com. NAFLD nonalcoholic fatty liver disease.

Fig. 2. Phenotypic and genetic correlations between paired fecal and blood metabolites.

a Phenotypic and genetic correlations between paired fecal and blood metabolites for the top 30 metabolites that are ranked by phenotypic correlations. Phenotypic correlations between paired fecal and blood metabolites are estimated by partial Spearman correlation analysis, adjusted by age, sex and BMI. Genetic correlations are calculated using bivariate GREML analysis. Correlations with FDR < 0.05 and $\mid r\mid$ > 0.3 (red dashed lines) are considered significant. FDR is controlled by the Benjamini-Hochberg method. *FDR < 0.05, **FDR < 0.01, *** FDR < 0.005. The results for remaining metabolites are shown in Supplementary Fig. 2. For genetic correlation analysis, there are 38 metabolites that do not converged during the calculation process and thus have no results. b The overlap between significant phenotypic and genetic correlations. All statistical tests are two-sided. Source data are provided as a Source Data file.

Fig. 3. Comparisons between paired fecal and blood metabolites in their associations with taxonomic composition.

a The associations of taxonomic composition with fecal metabolites, and b with blood metabolites. The random forest model with five-fold cross-validation is used to predict the fecal or blood metabolite levels based on taxonomic composition. Spearman’s correlation between measured and predicted metabolite levels is used to measure the association of taxonomic composition with fecal or blood metabolites. *FDR < 0.05, **FDR < 0.01, *** FDR < 0.005. c The distributions of the associations of taxonomic composition with fecal and blood metabolites (n = 132). The difference between the distributions of taxonomic composition-fecal metabolite associations and taxonomic composition-blood metabolite associations is tested by Wilcoxon signed-rank test. Box plots indicate median and interquartile range (IQR). The upper and lower whiskers indicate 1.5 times the IQR from above the upper quartile and below the lower quartile. ***P < 0.0001. d Differences between the associations of taxonomic composition with paired fecal and blood metabolites that are well-predicted in both feces and blood (marked with red in y axis), metabolites that are only well-predicted in blood and not in feces (marked with green in y axis), and the top 30 metabolites that are only well-predicted in feces and not in blood (marked with blue in y axis). Metabolites are ranked by the predictability of fecal metabolites. Differences between the associations of taxonomic composition/microbial pathways with paired fecal and blood metabolites are tested by the method proposed by Hittner et al. (see ”Methods”). *FDR < 0.05, **FDR < 0.01, ***FDR < 0.005. The results for the top 31–90 metabolites that are only well-predicted in feces and not in blood are presented in Supplementary Fig. 7a. e The number of well-predicted fecal and blood metabolites based on taxonomic composition and the number of validated associations between taxonomic composition and well-predicted fecal/blood metabolites in the validation cohort. Well-predicted metabolites are defined as Spearman’s correlation coefficient > 0.3 and FDR < 0.05. FDR is controlled by the Benjamini–Hochberg method. Associations with Spearman’s correlation coefficient > 0.3 and FDR < 0.05 are considered as being validated in the validation cohort. All statistical tests are two-sided. Source data are provided as a Source Data file.

Fig. 4. The associations of well-predicted fecal and blood metabolites with cardiometabolic diseases.

a The associations of well-predicted fecal metabolites with cardiometabolic diseases. b The associations of well-predicted blood metabolites with cardiometabolic diseases. In a, b, the intensity of colors represents the partial regression coefficients that are computed by logistic models, adjusted by age, sex, smoking status, alcohol status, education, income, physical activity, and total energy intake for obesity, and by age, sex, BMI, smoking status, alcohol status, education, income, physical activity, and total energy intake for T2D, hypertension and NAFLD. FDR is controlled by the Benjamini–Hochberg method. ^#P < 0.05, ^##FDR < 0.05. c Replication of the significant associations of well-predicted fecal metabolites with cardiometabolic diseases in the validation cohort (n = 1007 in the discovery cohort; n = 103 in the validation cohort). Error bars are partial regression coefficients with 95% confidence intervals. d The scatter plot demonstrates the partial regression coefficients for the associations of well-predicted fecal metabolites with cardiometabolic diseases obtained in the discovery (x axis) and validation cohort (y axis). The correlation between the partial regression coefficients obtained in the discovery and validation cohort is computed by Pearson correlation. Error band is linear regression line with 95% confidence band. All statistical tests are two-sided. Source data are provided as a Source Data file. T2D, type 2 diabetes; NAFLD nonalcoholic fatty liver disease, FDR false discovery rate, CI confidence interval.

Fig. 5. Comparisons between paired fecal and blood SCFAs in associating taxonomic composition and type 2 diabetes.

a The association between taxonomic composition and fecal acetic acid. bThe association between taxonomic composition and blood acetic acid. c The association between taxonomic composition and fecal butyric acid. d The association between taxonomic composition and blood butyric acid. The random forest model with five-fold cross-validation is used to predict the fecal or blood metabolite levels based on taxonomic composition. The scatter plot is plotted by the predicted and measured metabolite values. Spearman’s correlation between measured and predicted metabolite values is used to measure the association of taxonomic composition with fecal or blood metabolites. Error bands in a–d are linear regression lines with 95% confidence bands. e The association of fecal butyrate acid with T2D. f The association of blood butyrate acid with T2D. In e, f, logistic regression model is used to assess the associations of fecal or blood butyrate acid with T2D, adjusted for age, sex, BMI, smoking status, alcohol status, education, income, physical activity, and total energy intake. Error bands in e, f are logistic regression curves with 95% confidence bands. FDR is controlled by the Benjamini-Hochberg method. All statistical tests are two-sided. Source data are provided as a Source Data file. T2D, type 2 diabetes; SCFA, short-chain fatty acids; FDR, false discovery rate.