Abstract PR13: DNA methylation mapping and computational modeling in a large Ewing sarcoma cohort identifies principles of tumor heterogeneity and their impact on clinical phenotypes

Ewing sarcoma is an excellent model for studying the role of epigenetic deregulation and tumor heterogeneity, given its low mutation rates and the well-defined oncogenic driver. We have recently shown that the fusion oncogene EWS-FLI1 induces widespread epigenetic rewiring in proximal and distal enhancers (Tomazou et al. Cell Reports 2015). In the current study, we validate the clinical relevance of our results in a large cohort of primary tumors, and we explore the prevalence, characteristics, and clinical impact of epigenetic tumor heterogeneity in Ewing sarcoma. We used reduced representation bisulfite sequencing (RRBS) to generate genome-wide profiles of DNA methylation in 141 Ewing sarcoma primary tumors, 17 Ewing sarcoma cell lines, and 32 primary mesenchymal stem cell (MSC) samples. Deep sequencing resulted in DNA methylation measurements for an average of 3.5 million unique CpGs per sample with excellent data quality (>98% bisulfite conversion efficiency). In addition, for three primary tumors we generated comprehensive reference epigenome maps using whole genome bisulfite sequencing (WGBS) and ChIP-seq for seven histone marks (H3K4me3, H3K4me1, H3K27me3, H3K27ac, H3K56ac, H3K36me3, and H3K9me3). We show that DNA methylation data can be used to infer enhancer activity differences among tumors, allowing us to exploit our large primary tumor dataset to systematically compare the regulation of EWS-FLI1 correlated and anticorrelated enhancers. We also identified Ewing-specific DNA methylation patterns. For example, Ewing sarcoma samples consistently show higher DNA methylation than MSCs at AP-1 binding sites, but lower DNA methylation at EWS-FLI1 binding sites. To explore epigenetic heterogeneity within individual tumors, we developed a bioinformatic algorithm that quantifies DNA methylation disorder. Using individual reads containing multiple DNA methylation measurements from single cells, we assign scores at single-nucleotide resolution. This method uses a probabilistic model to account for overall methylation rate and expected disorder levels. By evaluating the likelihood of the data in a model that assumes that the DNA methylation status of a CpG is independent of the methylation status of a nearby CpG, we identify extremely heterogeneous as well as highly epigenetically conserved genomic elements. These different region types show distinct patterns of enrichment for regulatory modes and transcription factor binding. We also compared the observed D