Multivariate model validation. (A) Z-scored values of predicted lung CFU and measured lung CFU for each of 10 rounds of cross-validation. Box-and-whisker plots depict bounds from the 25th to 75th percentile, median line, and whiskers extending to the largest value or a maximum of 150% the interquartile range from the hinge, with all points shown. (B and C) Spearman correlations from 10 rounds of fivefold cross-validation (cv) are shown for (B) predictive NCan lung CFU models compared with null models and (C) canonical lung CFU model compared with a null model. Mann–Whitney U tests compare fivefold cross-validation distributions to the null distributions or to models built without any one individual strain, with P values corrected via Benjamini–Hochberg (**P < 0.01, ***P < 0.001). Mean ±1 SD shown as crossbars. (D) Th17-associated features plotted against lung CFU values, with linear regression lines and 95% confidence intervals shown. (E) Average classification accuracy across all cells per mouse for each of 100 rounds of model prediction using LDA models in Fig. 9. (F) Ensemble LDA model accuracy of phenotype prediction for each mouse across 100 rounds of fivefold cross-validation, compared with null models built on random features or shuffled labels. Mann–Whitney U tests compare cross-validation distributions with P values corrected via Benjamini–Hochberg (***P < 0.001). Mean ±1 SD shown as crossbars.