Fig. 5

Performance of the random forest classifier. A The AUC and predicted accuracy of the farmed and wild groups under different dataset stratifications with 10 replications. B Effect of the abundances of OTUs within and below prediction (see Fig. 4C and D for the definition of the OTUs) on host source tracking. The training set: test set is 5:5 and the bootstrapping number is 100. The dashed line indicates the upper boundary of 90% confidence interval. Fisher’s exact test is applied to compare the samples above and within the confidence interval. C The predicted accuracy of random forest classifier based on leave-one-out validation for each batch (splitting training set: test set = 8:2, bootstrapping n = 20). D The top 15 most important OTUs identified by the random forest classifier. E The detected frequency of the top 15 most important OTUs in the four main batches, two-tailed Student’s t-test, FDR-corrected, *P < 0.05, **P < 0.01, ***P < 0.001. F The predicted probabilities of samples from batch S based on the optimized classifier established using the top 15 most important OTUs (top panel). Heatmap shows the relative abundance of these OTUs in batch S (bottom panel)