Fig. 4

The selected 13 DE DMP markers were effective at classifying adenomas and carcinomas. A Heat map and hierarchical clustering analysis of the discovery EPIC dataset based on the 13 identified DE DMP markers shows a block like structure with almost half of the markers being hypermethylated in carcinoma and hypomethylated in adenomas and vice versa for the other half. B MDS clustering of the discovery dataset using the 13 markers shows 2 distinct clusters. C tSNE clustering of the discovery dataset using the 13 markers could also resolve the two tumor types. D tSNE clustering of the validation dataset using the 13 markers shows a clear separation between adenomas and carcinomas, only 2 carcinomas are falsely classified. E ROC curves for the final 13 DE DMP classifier model for both discovery and validation datasets from EPIC arrays. Sensitivity and specificity, for distinguishing between adenomas and carcinomas, at various cut-off values for the datasets are plotted. The model yielded an AUC of 0.99 and reached a sensitivity and specificity of 96.33% and 95.28%, respectively, while overall model accuracy was 95.81% in the discovery dataset. In the validation dataset it had an AUC of 0.85, and reached a sensitivity and specificity of 89.36% and 69.78%, respectively. The diagonal dotted line represents the line of no discrimination between the two tumor types. DE DMP double evidenced differentially methylated probes, ROC receiver operating characteristic, MDS multidimensional scaling, tSNE t-distributed stochastic neighbor embedding, TPR true positive rate, FPR false positive rate, A adenoma, C carcinoma