DNA methylation-based prediction of differentiation state at cell-type resolution reveals an age-associated epigenetic mesenchymal drift

Abstract Quantifying the differentiation state of cells is an important task in regenerative medicine, cellular reprogramming and in elucidating how differentiation hierarchies are altered in aging and disease. For DNA methylation data (DNAm) there is however an accute shortage of methods for quantifying differentiation state. Here we use machine learning to build EpiDiff, a predictor of differentiation that yields a continuous absolute differentiation index (DI) when applied to a DNAm sample. We extensively validate EpiDiff in single-cells, sorted cells and bulk tissue, as well as in in-vitro differentiation, organoid, reprogramming and cellular rejuvenation experiments. In mixed samples, EpiDiff measures the extent of differentiation across all cellular lineages. In sorted cell populations, we find that DI increases with age and, underpinning this increase, we discover that many cell-types undergo a developmental epigenetic drift that is most prominent towards the mesenchymal lineage. When applied to normal-tissues, EpiDiff is predictive of carcinogenic exposure and cancer risk. In summary, EpiDiff is a unique, general and versatile tool that can accurately quantify the global differentiation state of a DNAm sample with broad applicability to many disciplines.