Fig. 1

Analysis scheme. A We collected data from multiple studies that included both microbiome and metabolome profiles from human fecal samples. Data from case-control studies were split into two datasets: “healthy” and “disease”. B We implemented a processing pipeline for both the microbiome and the metabolome data. For the microbiome, we collapsed MetaPhlAn-based or 16S rRNA gene-based profiles into genus-level profiles with unified names. For the metabolome, we retained only metabolites for which HMDB-IDs were identified, imputed missing values, and scaled log values to zero-mean unit-variance (see “Methods” section). C For each metabolite in each dataset, we trained a random forest regression model (see “Methods” section). Models were only trained for metabolites that appeared in 3 or more datasets. We identified the well-predicted metabolites in each dataset, i.e., metabolites that can be successfully predicted by genus-level profiles of the gut microbiota (Spearman’s ρ > 0.3 and FDR-corrected P value < 0.1 on out-of-fold predictions). D We next applied meta-analysis random-effects models to compare metabolite predictability results across datasets and identify metabolites which are consistently well-predicted by the microbiota’s composition. E To further evaluate whether robustly well-predicted metabolites also demonstrate similar dynamics in relation to the microbiome across datasets, we analyzed how well metabolite models trained on one dataset perform on another dataset, using only shared genera features. F We additionally identified the main genera contributors to the model and again compared contributors across datasets to evaluate the similarity between models trained on different datasets for the same metabolite, and identify consistent contributors. G Lastly, we identified metabolites for which genera contributors change in disease