Fig. 2

Identification of metabolites well-predicted by the microbiome using machine learning. A Number of samples per dataset. Colored portions of each bar represent samples from the healthy/control study group. B The number of well-predicted metabolites in each dataset. Black diamonds represent the total number of metabolites analyzed per dataset (with the percentage labels indicating the percent of analyzed metabolites that were well-predicted in each dataset). C The number of datasets each metabolite was well-predicted in, stratified by the number of datasets each metabolite appeared in. D Examples of predictability results for 30 metabolites. Each heatmap row represents a dataset and each column denotes a specific metabolite. Cell colors and labels represent predictability levels (Spearman’s ρ on out-of-fold predictions and FDR-corrected P values: *FDR < 0.05; **FDR < 0.01; ***FDR < 0.001). White cells indicate that the metabolite was not available in that dataset. The examples presented here include metabolites never well-predicted (leftmost panel), well-predicted in only one dataset (center panel), and well-predicted in several datasets (rightmost panel)