Fig. 4

Putative compound activity of LAB BGCs. a Performance of four machine learning classifiers [logistic regression, elastic net regression, support vector machines (SVM), and random forest] in determining compound activities using tenfold cross-validation. The receiver operating characteristic (ROC) curves were based on aggregated performances of tenfold cross-validation. Average AUROC was shown. b Chord diagram showing the predicted activity of 129,878 BGCs. The scale was the proportion of each BGC class or predicted activity. The number shown in brackets refers to the BGC count and percentage relative to overall BGCs. Antibacterial-antifungal and antibacterial-cytotoxic represent BGCs encoding bifunctional SMs. The connections between BGCs and their predicted activities are highlighted with different colors according to the activity types. c Proportion of antibacterial SMs encoded by BGCs from LAB and non-LAB species. The proportion of antibacterial activity was calculated from randomly selected 10,000 BGCs of LAB or non-LAB, with be resampled 1,000 times. Data are mean ± standard deviation. P value was given by Wilcoxon rank-sum test (two sided), with “***” denoting P < 0.001. d The proportion of putative activities of BGCs detected in six body sites. AN, anterior nares; St, stool; PF, posterior fornix; SP, supragingival plaque; BM, buccal mucosa; TD, tongue dorsum