Fig. 2

Simulation results. a The assumed underlying true DM patterns Q. Each row indicates one of the five patterns; each column corresponds to one cancer type. A darker color suggests a higher probability of being DM. b Detailed DM patterns of the five clusters \(\hat {Q}\) learned by PanDM, which almost exactly match the true patterns. c The BIC plot for PanDM from K=2 to K=20. PanDM achieves the minimal BIC at the assumed true K=5. d The five DM patterns learned by separate analysis from the simulation data, which are prone to noise and deviate from the underlying true patterns. e Pattern matching matrix of PanDM and the assumed true cluster labels. The color in cell (k,l) corresponds to the number of CpG sites identified as model-fitted pattern l by PanDM while actually belonging to pattern k normalized by the total number of true cluster k CpG sites. Thus, the diagonal suggests the proportion of matched cases. From the heatmap we can see that PanDM correctly groups most of the CpG sites. f Pattern matching matrix from the separate analysis and the assumed true cluster labels. The vague diagonal indicates the high misclassification rate of the separate analysis