Mapping classifiability in the cancer DNA methylome: A data-learned disease hierarchy

DNA methylation-based tumor classification has been successfully applied in clinical settings. Traditional classifiers assume that diagnostic labels are fixed and mutually exclusive, yet many biological entities are inconsistently defined or intrinsically overlapping. We propose a general framework for quantifying and interpreting classifiability in DNA methylation-based disease classification. Here, we treat classifiability itself as an empirical property of data, measured through cross-validation across more than 13,698 harmonized methylome cohorts drawn from TCGA and GEO, spanning over 324 cancer types. This approach reveals which disease or cancer types are stably separable at the molecular level and which collapse across labels, providing a principled, data-driven view of biological boundaries. By analyzing cross-validation consistency and label confusability, we reconstruct a hierarchical taxonomy derived directly from the methylome, in which relationships between entities emerge without prior human definitions, recapitulating known lineage relationships and uncovering novel cross-entity proximities. We further integrate major public datasets into a pan-disease foundation classifier that reports both predictions and classifiability-aware confidence scores, reflecting separability along the learned hierarchy. Finally, we demonstrate that the same framework can evaluate new or rare cohorts, testing whether a proposed entity forms a distinct, classifiable unit or merges with established types. Together, these advances recast methylation classification from a task of prediction into one of discovering the structure of classifiability itself in the human epigenome, offering a data-driven foundation for refining tumor taxonomies and diagnostic criteria.