DNA Methylation Profiling Predicts Clinical Outcomes and Reveals Unique Insights Into the Molecular Complexity of Acute Myeloid Leukemia.

Abstract 707

Epigenetic deregulation of genes through aberrant DNA methylation has been widely reported in cancer. We hypothesized that in AML this aberrant DNA methylation does not occur randomly, but rather occurs in specific and distinct patterns. Therefore, large-scale genome-wide analysis of the DNA methylome could help explain and define the complexity underlying leukemia biology and reveal the existence of epigenetically defined variants of AML. Using the HELP microarray assay, which measures DNA methylation at 50,000 CpG sites annotated to ∼14,000 promoters, we obtained DNA methylation profiles for 344 AML patients seen at Erasmus University Medical Center. Median follow-up based on survivors was 18.2 months (7-215); median age: 48 years (15-77). Unsupervised analysis (hierarchical clustering, correlation distance with Ward's clustering method) demonstrated that based on their methylation profiles AML patients distributed into 16 cohorts. 11 of these groups were also defined by the presence of specific molecular lesions: inv(16) [cluster 1], t(8;21) [cluster 3], t(15;17) [cluster 6], CEBPA-mutant [clusters 4 and 9], CEBPA-silenced [cluster 10] NPM1-mutant [clusters 12, 13, 14 and 16] and 11q23 abnormalities [cluster 11]. Enrichment for cases harboring a specific molecular lesion within a given cluster was determined using Fisher's exact test (p<0.01). Additionally, 5 new AML subtypes were defined based on epigenetic profiling alone and had no other clinical or molecular feature in common. Kaplan-Meier survival analysis revealed a significant difference in overall survival (OS) between these novel AML subtypes: 2-year OS±SE; 58.8%±8.4% and 45.2%±8.9% for clusters 5 and 7, respectively, vs. 23.6%±5.7%, 26.4%±9.2% and 33.3%±13.6%, for clusters 2, 8 and 15, respectively (log rank test, p=0.04). After adjustment for age, cytogenetic risk, NPM1 and FLT3-ITD status in a multivariate Cox proportional hazards regression model including all the clusters with ≥ 10 patients, 4 of these 5 novel clusters presented a statistically significant increased hazard ratio compared to the favorable risk inv(16) cluster. In contrast, the clinical outcomes of patients in cluster 5 were not significantly different from favorable risk patients with inv(16). In order to identify the genes affected by aberrant DNA methylation for each cluster, we performed a supervised analysis comparing each of the 16 clusters to normal CD34+ bone marrow progenitors (n=8) using ANOVA followed by Dunnet post hoc test, and selected genes with adjusted p values <0.05 and a methylation change >30%. The DNA methylation signatures of each cluster featured involvement of distinct gene networks and DNA regulatory elements, and displayed distinct degrees of hyper or hypomethylation with respect to normal CD34+ bone marrow cells. Of note, in spite of the variation in methylation across the 16 clusters, we identified a set of 45 genes that were almost universally aberrantly methylated (in >70% cases and present in at least 10/16 cluster signatures). This common epigenetic signature included the tumor suppressor PDZD2, the nuclear import proteins IPO8 and TNPO3, PIAS2, a regulator of MAP kinase signaling, CDK8, and CSDA, a regulator of CSF2. Gene expression profiling of the same patients indicated that at least 50% of these genes were also aberrantly silenced compared to normal CD34+ cells. Finally, we randomly divided the 344-patient cohort into a training group of 200 patients, a test group (n=95) and an independent validation group (n=49), and using the Supervised Principal Components algorithm identified a 15-gene methylation classifier that was predictive of OS (p<0.009) and event free survival (p<0.013). Furthermore, after adjustment for age, cytogenetic risk, NPM1, FLT3 and CEBPA status in a multivariate analysis, this classifier remained an independent risk factor for OS (Hazard ratio 1.29, 95% CI: 1.11-1.49; p<0.001). In summary, we have i) demonstrated that unique and distinct DNA methylation patterns characterize distinct forms of AML; ii) identified novel, epigenetically defined subgroups of AML with distinct clinical behavior; iii) revealed the presence of a consistently aberrantly methylated signature across AML subtypes, with confirmed silencing of the genes involved; and iv) report a 15-gene methylation classifier predictive of OS, and confirmed as an independent risk factor when adjusted for known AML covariates.