Fig. 4

Comparisons of metabolite models between datasets. A An overview of pairwise model comparisons for each robustly well-predicted metabolite. For each metabolite, we compared the number of shared features (i.e., genera; grey points) and the number of shared significant contributors (purple points) for every pair of models trained on two different datasets. For each metabolite we additionally note the average number of significant contributors over all models trained for that metabolite (turquoise points). Error bars represent standard deviations. B A detailed comparison of models for predicting beta-alanine levels. Left panel: contributors comparison. Each row represents a dataset in which the metabolite was well-predicted by the microbiome and each column represents a genus feature (only features significantly contributing to at least one model are included). Purple-colored cells denote the significance of the specific feature in the specific model (P < 0.1). White cells indicate that the feature was not available in the specific dataset. Right panel: Cross-predictability analysis. Matrix columns indicate the dataset used for training and matrix rows indicate the dataset used for testing. Numbers in cells indicate Spearman’s correlation between predicted and actual metabolite levels in the test dataset. Red cells denote cases where the model was well transferred from the column-dataset to the row-dataset. C. D include similar plots as B, for sebacic acid and l-tyrosine metabolites, respectively. *P < 0.05; **P < 0.01; ***P < 0.001; CP cross predictability