Epigenetic patterns and methylation-based models for robust outcome prediction in osteosarcoma

Osteosarcoma (OSA) is a relatively rare but aggressive bone malignancy that primarily affects children and adolescents and has not seen survival improvements in at least three decades. No clinical or pathological prognostic factor has been identified that can lead to improved outcomes. Further, the complex genomic architecture and genetic heterogeneity of OSA and the rare nature of the tumor have significantly hampered identification and validation of molecular predictors. Recent assessment of DNA methylation in OSA has associated patterns with clinically informative outcomes, yet models building on these findings that can be applied in the clinic have not been developed. To address this unmet need, we developed methylation-based models for individualized patient risk prognostication and treatment response prediction at the time of diagnosis. By applying a region-based method to assess methylation, we were able to greatly reduce dimensionality and soften CpG-level noise, yielding stable features for model construction. These models, based on variably methylated regions (VMRs), underwent comprehensive internal cross-validation, and when possible, external validation in a completely independent dataset that was generated using a different technology from the discovery dataset. Further, they were contextualized with an unsupervised exploration of the features, showing widespread global hypermethylation being associated with unfavorable outcomes. Additionally, survival-associated signal appeared to be concentrated within a smaller, more focal subset of features, while in contrast, the chemoresponse signal seemed to come from a larger, more diffuse feature set, potentially suggesting a shift in cell state. An examination of epigenetic age using a methylation-derived epigenetic clock was also shown to provide an orthogonal approach of investigating the methylation landscape with significant association to outcome. Finally, clustering phenotypes were compared between methylation and several other omics platforms (mRNA, miRNA, and copy number variation) and showed very little overlap with one another. Together, these findings establish a foundation for methylation-based outcome stratification at the time of diagnosis and underscore the need for continued investigation including validation in larger and prospective cohorts.