A Measure of DNA Methylation for Predicting Outcome in De Novo AML

Background: Accurate prediction of response to induction chemotherapy and survival is essential for improving care of AML patients. Established prognostic schemes in AML are based on 1) clinical features and 2) pre-treatment karyotype. More recently molecular markers have been shown to be of prognostic importance. Correlation between DNA methylation in leukemic cells and clinical prognosis has been described but is not assessed routinely in clinical practice due to lack of a clinical assay. Here we report the ability of a novel, clinically feasible assay of DNA methylation status to predict clinical outcomes in AML patients.

Methods: A novel microsphere-based assay for multiplex evaluation of DNA methylation (xMELP) was used to simultaneously measure methylation status at multiple prognostic loci in banked AML samples. A methylation risk score (M score) was assigned to each sample using a 17-locus random forest classifier developed using an independent cohort. Cytogenetic risk was assigned to each sample using the MRC schema (Grimwade et al. Blood 2010). The primary clinical endpoints were failure to achieve complete remission (CR) within 90 days of induction and death before 2-years (2-year death). The receiver operating characteristic (ROC) curve was used to evaluate the performance of the M-score as prognostic biomarker. Univariate and multivariate logistic models were used to assess the ability of the M-score to predict 2-year death in combination with co-variates.

Results: This study included 101 patients with de novo AML age < 60 years with diagnostic samples available for xMELP analysis who underwent induction therapy at the University of Pennsylvania (2001-2012). The median age was 48 years (range 19-59), the median WBC count at diagnosis was 34.5 K/uL (range .8 -283.5 K/uL; 20% ≥100K/uL), and 55% were male. The majority of patients had intermediate cytogenetic risk (13% favorable, 65% intermediate, 18% unfavorable, 4% unknown). Among the study patients, 25% did not achieve CR and 52% died within 2 years of diagnosis.

The mean M-scores for the entire population were 84.6 (SD 30.3) versus 104.9 (SD 29.3) for those who achieved and failed to achieve CR, respectively (p=.004) while the mean M-scores were 77.5 (SD 24.7) versus 100.5 (SD 32.7) for those alive and dead at 2 years, respectively (p=.002). Within the intermediate cytogenetic risk group, results were similar: the mean M-scores were 87.3 (SD 31.8) versus 105.1 (SD 29.6) for those who achieved and failed to achieve CR, respectively (p=.06) and the mean M-scores were 78.9 (SD 25) versus 101.2 (SD 33.9) for those alive and dead at 2 years, respectively (p=.004). The M-score was not associated with sex or WBC at diagnosis, but it differed significantly among the 4 cytogenetic risk groups (ANOVA p=.01) with a higher M-score in less favorable cytogenetic groups.

The area under the ROC curve (AUC) was .69 for failure to achieve CR (95% CI .58-.80) and .71 (95% CI .61-.81) for 2-year death. For AML patients with intermediate cytogenetic risk, the AUC was .66 (95% CI .52-.80) for failure to achieve CR and .70 (95% CI .57 to .82) for 2-year death. For the cohort overall, the odds ratios for failure to achieve CR and 2-year death were 1.02 (p=.007) and 1.03 (p=.001), respectively, indicating that every 1-point rise in the M-score resulted in a 2% higher chance of not achieving CR and a 3% greater chance of dying by 2 years.

We investigated the ability of the M-score to predict 2-year death in combination with other prognostic co-variates. Adding age, diagnostic WBC count, and cytogenetic risk to the prognostic model did not improve prediction of 2-year death. A model incorporating M-score and CR did predict 2-year death better than M-score alone (AUC .80 versus AUC .71, p=.005). Results were similar when restricting analysis to intermediate cytogenetic risk group (AUC .77 versus AUC .70, p=.048).

Conclusion: There is a clinical need for predictive biomarkers in AML. Here, we show that a novel method of measuring methylation in AML predicts failure to achieve CR and likelihood of death within 2 years. A model incorporating methylation and remission status was better for prediction of 2-year death than a model with only methylation information. Although validation in additional cohorts is necessary, these results demonstrate the feasibility and benefit of clinical application of a multiplex DNA methylation assay to survival prognostication in AML.