Figure 7.

Machine learning classification accuracy in a nested cross-validation random forest model built to predict jejunum occludin concentrations (in pg/ml) based on features of the cecal microbiome in mice from this study. Feature inputs from the raw feature table included in this analysis consisted of all bacteria matched to the Silva_132 ribosomal RNA gene reference database with 99% accuracy. (A) Receiver operating characteristic (ROC) plot shows macro-average precision (each jejunum occludin classification equally-weighted, light blue dashed line, AUC = 0.68) at several true-positive rate (TPR) against false-positive rate (FPR) thresholds; micro-average precision (averaged metrics across each sample, dark blue dashed line, AUC = 0.67) at several TPR against FPR thresholds; and classification error rate achieved by random chance (black dashed line). (B) Per-class ROC shows the prediction results for low (forest green, AUC = 0.66) and high (red, AUC = 0.66) jejunum occludin classifications in study mice. The overall accuracy in classifying low and high jejunum occludin in study mice was 66%. (C) The prediction matrix shows that all but 2 samples were correctly classified by the machine learning algorithm. More accurate assessments of the true group designation by the algorithm are correlated with darker shading. (D) Top 15 feature rankings in the random forest analysis. AUC area-under-the-curve, FPR false positive rate, ROC receiver operating characteristics, TPR true positive rate.