Genome-Wide Methylation Analysis Using the Digital Restriction Enzyme Analysis of Methylation for Stratification of Patients with Juvenile Myelomonocytic Leukemia

Background

Juvenile myelomonocytic leukemia (JMML) is a rare myelodysplastic/ myeloproliferative neoplasm that occurs during infancy and early childhood. The clinical course of the disease varies widely. The majority of children require allogenic hematopoietic stem cell transplantation (HSCT) for long term survival, but the disease will eventually resolve spontaneously in ~15% of patients. Previous studies have identified clinical and molecular risk factors in JMML. More recently, three groups independently discovered that genome-wide methylation profiling using 450K Illumina array revealed that the high methylation (HM) subgroup was significantly associated with poor survival compared to the low methylation (LM) subgroup (Murakami 2018 Blood, Stieglitz 2017 Nat. Commun., Lipka 2017 Nat. Commun.). 450K could be a standard assay for stratification of JMML. However, it is now unavailable because the manufacture replaced it with EPIC array. Here, we developed a next-generation sequencing-based clinical test recapitulate 450K clustering results using the digital restriction enzyme analysis of methylation (DREAM) method (Jelinek 2012 Epigenetics).

Patients and Methods

We studied 99 children (67 boys and 32 girls) with JMML. All the patients were included in our previous publications. First, we assessed JMML samples with DREAM. Briefly, genomic DNA was sequentially cut with two enzymes SmaI and XmaI recognizing the same sequence, CCCGGG sites in DNA. Enzyme-treated DNA was then used to generate sequencing libraries according to the Illumina protocols, and run on an Illumina Hiseq 2500.

We assessed 10 JMML samples with reduced representation bisulfite sequencing (RRBS) (Meissner 2005 Nucleic Acids Res.). In brief, purified genomic DNA was digested by the methylation-insensitive restriction enzyme MspI to generate short fragments that contain CpG dinucleotides at the ends. The CpG-rich DNA fragments (40-220 bp) were size selected, subjected to bisulfite conversion, PCR amplified and end sequenced on an Illumina Genome analyzer.

Results

We analyzed 99 samples using the DREAM with 8.87 (4.09-16.35) million reads (median, [range]), and determined methylation level in 62,525 (52,356-75,185) CpG sites (median [range]). We observed a strong correlation between DREAM methylation ratio and 450K beta-value of overlapping CpG sites (Pearson r² = 0.95 [0.913-0.962], median [range]).

We performed unsupervised consensus clustering with DREAM methylation data of 7,704 CpG sites within ±1 kb from TSS on autosomal chromosomes detected in ≥95% of the samples with imputation of the missing data using the median of each CpG site methylation level. Clustering identified two distinct subgroups, the HM subgroup (n = 35) and the LM subgroup (n = 64), matching 95% (94 of 99) with the 450K clustering results. The HM subgroup patients showed significantly poorer 5-year OS than the LM subgroup patients (41.9% [95% confidence interval {CI}], 25.3%-57.6%) vs. 71.4% [95% CI, 56.2%-82.1%]; P = 0.00345). Discrepancies in the clustering results between DREAM and 450K were observed in only 5 patients (2 survived and 3 died); all 5 patients were reclassified as those with LM with DREAM from being HM with 450K.

We also performed RRBS methylation analysis on 10 patients. Unsupervised consensus clustering using promoter-associated 4,971 CpG sites measured with RRBS identified HM (n = 5) and LM (n = 5) subgroups and completely matched with the classification made using DREAM and 450K.

Then, we developed a prediction model of the methylation subgroups using a machine-learning program. We selected 85 CpG sites from 7,704 CpG sites used for unsupervised clustering of the DREAM assay that showed a distinct difference in the average methylation level (>0.3) between the HM and LM subgroups of the learning cohort (n = 70) and developed a support vector machine (SVM) model. As a validation cohort, we analyzed the remaining 29 JMML samples with a SVM model and confirmed a high matching rate with 450K clustering results (100%, 29 of 29).

Conclusions

We could develop a methylation test for JMML using the DREAM assay. Both the unsupervised clustering analysis and SVM model could repeat the result of 450K-based methylation classification, i.e., the HM and LM subgroups. The relatively lower cost of the DREAM assay (US$200/sample) enabled us to incorporate methylation classification in JMML in most settings.