A DNA methylation score reads out cell differentiation, revealing an age-linked mesenchymal drift
Synopsis
DNA methylation data lacked a general way to quantify differentiation state; EpiDiff trains a random forest on 1,298 differentiation-associated CpGs to output a continuous index from 0 to 1. It reached 0.97 classification accuracy in independent datasets and rose with age in many sorted cell types, underpinned by a developmental epigenetic drift toward the mesoderm lineage.
Interpretation
A DNA methylation sample can now yield a continuous differentiation index, near 0 for pluripotency and near 1 for terminal differentiation, comparable across datasets. DNA methylation data are abundant and measured stably, yet there was no general method for quantifying differentiation state, so such judgments relied on noisier modalities such as mRNA expression. On a model selection set of 180 samples from five independent datasets, the random forest reached a classification accuracy of 0.97, and performance was robust to parameter values.
The predictor treats differentiation as three classes, pluripotent, multipotent and terminally differentiated, and combines their probabilities into a continuous index. Features come from 1,298 differentiation-associated CpGs, of which 882 are methylated in hESCs and become progressively hypomethylated with differentiation, while 416 behave oppositely. Training used 146 Illumina 450k samples covering 25 hESC, 86 fetal tissue and 35 adult tissue samples; in cell lines undergoing population doublings, the index did not change monotonically with doubling number.
In sorted cells the differentiation index increases with age, and this increase aligns with methylation changes seen during early fetal development. It was unclear whether differentiation state itself changes with age or merely reflects shifts in cell composition. Across 55 hepatocyte samples the index was associated with age (linear P=1e-11 adjusted for sex), and it also rose in bronchial epithelial cells, naive CD4+ T cells and several glial populations, while neurons and myeloid cells showed no clear increase.
Perspective
The index applies to DNA methylation samples, and its training set is entirely Illumina 450k data, with validation covering 450k, EPIC, snmCT-Seq and snm3C-Seq3 data. Validation on single-nucleus data was done on a relative scale, and the authors note that the high sparsity of single-nucleus methylation data could compromise absolute-scale accuracy. They suggest improved versions could be built from future EPIC, whole-genome bisulfite sequencing or single-nucleus data.
The authors note the training set draws on different studies, which could introduce batch effects, but selected CpGs had to change significantly in both fetal and adult differentiation, with effect-size thresholds above what batch effects can explain. They also state there is no evidence that the methylation drift at germ-layer-specific CpGs is functionally significant, leaving that an open question.