Pediatric Acute Myeloid Leukemia (AML) is a heterogenous disease with a dismal outcome. Recently, transcriptomic-based leukemic stemness scores have highlighted the importance of DNA methyltransferase 3B in pediatric AML prognosis, which invites a deeper investigation of leukemic stem cells through the scope of their DNA methylome. Other studies have shown the prognostic impact of genome-wide methylation burden (PMID: 29928480, 33080932). As the World Health Organization has recognized methylation analysis as the most apt technology for determining lineage and cell population origins of tumors (PMID: 34921008), it is befitting to evaluate the DNA methylome as a prognostic tool in hematopoietic malignancies. Thus, we hypothesize that a DNA methylation-based signature has the potential to enhance the prognostic precision in pediatric AML. In this study, we applied machine learning approaches to DNA methylation data from 924 pediatric AML patients (only including data from samples with diagnostic bone marrow specimens) from three Children's Oncology Group clinical trials: AAML1031, AAML0531, and AAML03P1, available through GSE190931, GSE124413 and GDC_TARGET-AML, respectively. Clinical endpoints available were defined as :i) Minimal residual disease after induction 1 (MRD1): positive MRD1 if ≥ 1 leukemic cell per 1,000 mononuclear bone marrow cells (≥ 0.1%) determined by flow cytometry; ii) event-free survival (EFS) defined as the time from study enrollment to induction failure, relapse, secondary malignancy, death, or study withdrawal for any reason, with event-free patients censored on last follow-up; iii) overall survival (OS) defined as the time from study enrollment to death, with living patients censored on the date of last follow-up. Processing of the raw data followed best practices in the literature using SeSAMe (PMID: 30085201, 27924034). To independently validate the findings derived from the discovery cohort, we processed in parallel meDNA array data from bone marrow specimens at diagnosis from 201 AML patients treated on the multi-site clinical trials AML02 and AML08 (NCT00136084 and NCT00703820). As step 1, we performed Epigenome-Wide Association Study with Cox Proportional Hazards Regression (Cox-PH EWAS) to identify CpGs with methylation levels most predictive of OS, controlling for risk group categories. At p-value threshold of 1.0e-6, 167 CpGs were associated with OS, which were chosen as candidate CpGs for step 2 analysis using LASSO-based Cox-PH model, with 1,000 iterations of 10-fold cross-validation. 35 CpGs were consistently selected as non-zero coefficients in at least 95% of the models and were used to define the CpG based signature. Patients were further categorized into either Low MethylScoreAML Px (n=563, 60%) or High MethylScoreAML Px (n=361, 40%) groups using recursive partitioning statistics. In the discovery cohort, patients with High MethylScoreAML Px demonstrated a markedly poorer probability of OS (HR 3.91, p<0.0001) and EFS (HR 2.61, <0.0001) in comparison to the low MethylScore. These results were further tested in the validation cohort (n=201) for OS (HR 2.45, p=0.0006) and EFS (HR 2.15, p=0.0004). Notably, in subgroup analyses for both discovery and validation, MethylScoreAML Px remained statistically significant in multivariable models for EFS and OS with adjustments for MRD1 status, risk group, FLT3 status, white blood cell count at diagnosis, and patient age, indicating it as a robust, independent model supplementing currently known prognostic factors. In conclusion, this study unveils MethylScoreAML Px, a methylation-based prognostic risk score that robustly predicts clinical outcome in pediatric AML patients. To our knowledge, we are the first group to describe a Cox-PH-EWAS analysis in pediatric AML, as well as the first to devise and validate a prognostic risk score in >1000 patients from multiple pediatric-only trials. Future directions of research are focused on testing MethylScore in fresh peripheral blood samples using long read, real-time Oxford Nanopore sequencing technology, and comparing it with other risk scores.
Read full abstract