Figure 5: Gut MLGs classify colorectal carcinoma and adenoma samples from healthy controls.

(a) Distribution of 5 trials of 10-fold cross-validation error in random forest classification of carcinoma as the number of MLGs increases. The model was trained using relative abundance of the MLGs (>100 genes) in the controls and carcinoma samples (n=55 and 41). The black curve indicates average of the five trials (grey lines). The pink line marks the number of MLGs in the optimal set (Supplementary Data 5). The same MLGs were selected if age and BMI were included along with the MLGs. (b) Box-and-whisker plot for the probability of carcinoma in the cross-validational training set according to the model in a. (c) Receiver operating curve (ROC) for the training set. The area under receiver operating curve (AUC) is 98.34% and 95% confidence interval (CI) is 96.29–100%. (d) Classification of the test set consisted of 8 controls (green), 47 advanced adenoma (blue) and 5 carcinoma (red), that is, 18 unused samples and 42 adenoma samples used in analyses in Figs 1, 2, 3, 4, 6 and 7. (e) ROC for the test set. The AUC is 96% and 95% CI is 87.88–100%. (f–j) Training and testing the model that classifies adenomas from controls, performed as in a–e. The AUC for the training set (n=55 controls, 42 adenomas) is 87.38% and 95% CI is 80.21–94.55%; the AUC for the test set (8 controls, 5 advanced adenomas and 46 carcinomas) is 59.56% and 95% CI is 37.51–81.61%. (k–o) Training and testing the model that classifies adenomas from controls, performed as in (f–j) except that age and BMI were included along with the MLGs. Age was selected by the model, making the optimal number of markers 11. The AUC for the training set is 89.74% and 95% CI is 83.32–96.16%; the AUC for the test set is 59.56% and 95% CI is 37.64–81.48%.