UCSF

Juvenile myelomonocytic leukemia (JMML) is a rare heterogeneous hematological malignancy of early childhood characterized by causative RAS pathway mutations. Classifying patients with JMML using global DNA methylation profiles is useful for risk stratification. We implemented machine learning algorithms (decision tree, support vector machine, and naïve Bayes) to produce a DNA methylation-based classification according to recent international consensus definitions using a well-characterized pooled cohort of patients with JMML (n = 128). DNA methylation was originally categorized into three subgroups: high methylation (HM), intermediate methylation (IM), and low methylation (LM), which is a trichotomized classification. We also dichotomized the subgroups as HM/IM and LM. The decision tree model showed high concordances with 450k-based methylation [82.3% (106/128) for the dichotomized and 83.6% (107/128) for the trichotomized subgroups, respectively]. With an independent cohort (n = 72), we confirmed that these models using both the dichotomized and trichotomized classifications were highly predictive of survival. Our study demonstrates that machine learning algorithms can generate clinical parameter-based models that predict the survival outcomes of patients with JMML and high accuracy. These models enabled us to rapidly and effectively identify candidates for augmented treatment following diagnosis.