Classifier construction for risk stratification of patients with JMML using DREAM. (A) Patients with JMML or NS/MPD (discovery cohort, n = 99; validation cohort, n = 38) were subjected to unsupervised hierarchical clustering using DREAM data. The heat map displays the methylation ratios calculated for a selected subset of 1000 CpG sites with high differential average methylation levels between HM_DREAM and LM_DREAM subgroups. The methylation ratios were color coded, with a gradual change from blue (0% methylation) to red (100% methylation). Known clinical and biological features were annotated for each patient. Each column indicates 1 patient. (B) Schematic outline of the SVM classifier construction using DREAM data. (C) Performance of SVM. This plot shows the performance of various models using color coding. Darker regions with lower mean square error (MSE) indicate better accuracy. (D) OS of patients with JMML in the discovery cohort using clustering with DREAM data. (E) OS of patients with JMML in the validation cohort using SVM with DREAM data. (F) Patients with JMML (n = 9) and NS/MPD (n = 1) were subjected to unsupervised hierarchical clustering using RRBS data. The heat map displays the methylation ratios calculated for a select 1000 CpG sites with high differential average methylation levels between HM (HM_RRBS) and LM (LM_RRBS) subgroups. HbF, fetal hemoglobin; LOH, loss of heterozygosity; PLT, platelet.