FIG 5.

The e-cig user periodontal microbiome resembles those of both conventional smokers and nonsmokers. Supervised learning sample classification can accurately predict sample inclusion in the CS and NS but struggles with the ES cohort (A). The model accuracy was tested on 34 samples that were excluded from the training data set (n = 134). All cohorts had areas under the curve (AUC) well above what would be expected by chance (B). A hierarchical-clustering mean-relative-abundance heat map of the 50 most important features (ASVs) shows important clusters with distinct abundance patterns and pathological relevance (C).