Fig 1.
Cohort description and model prediction scheme.
(a) Illustration of cohorts and machine learning process. A training set of 27,075 individuals was randomly selected out of 30,083 Israeli individuals and was used for model parameter selection using 10-fold cross validation and microbiome, age and gender features. For each phenotype the selected model was trained on the 27,075 training samples and then tested on both the held out 3,008 samples of the Israeli population and a separate U.S. test cohort of 3,974 individuals. (b) Distribution of age in the 3 cohorts, training test1 and test2. (c)-(e) Same for HbA1C%, BMI and alpha diversity. (f) A scatter plot comparing the mean log RA of each species, in the Israeli training cohort vs. the Israeli test1 cohort. R value represents Spearman correlation. (g) Same as (f) in the Israeli training cohort vs. the US test2 cohort. R value represents Spearman correlation.
Fig 2.
Species level Shannon alpha diversity significantly associates with many phenotypes.
(a) A box-plot of the distribution of phenotype values, for each of 10 deciles of Shannon alpha diversity, on the IL (train + test1) population. Phenotype values in the first and last deciles of alpha diversity are compared using Mann-Whitney rank-sum test where *** signifies P value< 10−16 after FDR correction. Boxes correspond to 25–75 percentile of the distribution and whiskers bound percentiles 5–95. (b) Running average of alpha-diversity (y-axis) for the combined training and test1 cohort (green curve), and for the separate test2 cohort (purple curve), ordered by the phenotype values. For the larger Israeli cohort the average is on 1000 individuals with shift of 100 individuals; for the smaller U.S. cohort the group size and shift were chosen to obtain 10 points and the shift was 10% of the group size. The Spearman correlation and P-value shown are of the Israeli cohort, and are calculated on individual level data. Individual level data points are presented with grayscale colors. Data points with light colors have higher frequency.
Fig 3.
Explained variance of phenotypes based on microbiome features.
(a) The proportion of variance of various phenotypes that can be explained using Shannon alpha diversity in the Israeli (green) and U.S. cohorts (purple) based on a linear model with covariates for age and gender. Also shown is the 95% confidence interval. (b) The proportion of variance of various phenotypes that can be explained using species-level RAs in the Israeli (green) and U.S. (purple) microbiome composition based on a linear mixed model estimation with covariates for age and gender (microbiome association index [32]). Also shown is the 95% confidence interval. Estimates from the larger cohort have smaller confidence intervals.
Fig 4.
Prediction of phenotypes by the microbiome.
(a) Coefficient of determination (R2) of GBDT prediction of different phenotypes based only on species level gut microbiome abundance. Results are obtained in a 10-fold cross validation scheme on the training set. Predictions are shown for three models, a model using the whole cohort, and a model for each gender. (b) Same as (a), but shown is the area under the curve (AUC) for predicting binary phenotypes. (c)-(e) Scatter plot of the phenotype and 10-fold cross-validation predicted values of the phenotype, for age, HbA1C% and BMI when training on the Israeli train cohort using GBDT. R2 of prediction is reported. Black line represents regression, dashed black line is x = y. (f) Coefficient of determination (R2) of predictions of age, HbA1C% and BMI, for models trained with different sets of input features using GBDT, and tested on both the held-out Israel and U.S. test sets. Error bars of the test set are from bootstrapping. (g)-(i) Coefficient of determination (R2) and standard deviation error bars of predictions of age, HbA1C% and BMI obtained using GBDT (purple) or Ridge regression (green) models trained on sub-samples of the cohort train IL, of different sizes, and tested of the test IL cohort. For each cohort size k, 10 random sub-samples of k individuals were obtained and the mean and standard deviation of their predictions are shown.
Fig 5.
Correlations of single species with age, HbA1C% and BMI.
(a) Spearman correlation of each bacterial species with age in the Israeli training cohort (x-axis, N = 27,075) and the U.S. test cohort (y-axis, N = 3,974). The correlation and P-value between the correlation coefficients of each cohort are shown. Bacteria are colored according to the P-values of the Spearman correlation coefficient in the Israeli cohort. The top three bacteria by Israeli P-values that replicate in the U.S. cohort are highlighted. (b)-(c) Same as (a) for HbA1C% and BMI. (d) R2 of predictions on Test1-IL and Test2-US cohort, using a partial set of bacteria features for XGBoost. X-axis represents the number of bacteria (with the highest Spearman correlation to phenotype, on the Train-IL cohort) used to build the model. (e)-(f) Same as (d) for HbA1C% and BMI. (g) Spearman correlation between the correlation coefficients of the Israeli cohort and U.S. cohort as in (a) but for different sub-samples of cohort sizes. For each cohort size k, a sub-sample of k individuals was obtained from both the Israeli and U.S. cohorts and this procedure was repeated 10 times to obtain standard deviation error bars. (h)-(i) Same as (g) for HbA1C% and BMI.
Fig 6.
Functional analysis of modules and pathways.
(a)-(c) Spearman correlation of each KEGG KO gene with age, HbA1C% and BMI in the Israeli training cohort (x-axis, N = 27,075) and the U.S. test cohort (y-axis, N = 3,974). The correlation and P-value between the correlation coefficients of each cohort are shown. Uniref are colored according to the P-values of the Spearman correlation coefficients in the Israeli cohort. (d)-(e) Heat map displaying only significant positive (blue) or negative (red) association between 6 phenotypes and top associated KEGG KO modules (d) or pathways (e). M00060 is the KDO2-lipid A biosynthesis, Raetz pathway, LpxL-LpxM type. M00116 the Menaquinone biosynthesis, chorismate → menaquinol. M00126 the Tetrahydrofolate biosynthesis, GTP → TH. Ko00540—Lipopolysaccharide biosynthesis. Ko01110—Biosynthesis of secondary metabolites. Ko01100—Metabolic pathways. Ko03010- Ribosome. Ko00740—Riboflavin metabolism. Ko00020—Citrate cycle (TCA cycle). Ko00970—Aminoacyl-tRNA biosynthesis. Ko00790—Folate biosynthesis. Ko02024—Quorum sensing. Ko02010—ABC transporters. Ko01230—Biosynthesis of amino acids. Ko00310—Lysine degradation.