Abstract
<title>Abstract</title> <p>Background. DNA-methylation age estimators are routinely applied to “blood” as though blood were a single tissue. Published studies assay whole blood, buffy coat, peripheral blood mononuclear cells, granulocytes, or sorted subsets, and frequently pool or compare across studies that differ in that choice. The magnitude of the resulting variation has not been quantified. Methods. We estimated the Horvath, Hannum, PhenoAge and Horvath skin‑and‑blood clocks in two independent datasets under two different designs. GSE35069 (HumanMethylation450) provides a complete within-donor grid: six healthy donors each profiled in ten preparations, so that within a donor the only variable is which cells were assayed; variance was partitioned between donor and preparation, and within-donor range computed across preparations. GSE110554 (EPIC) provides a between-donor design: 37 purified arrays from 36 healthy donors across six cell types, with chronological age recorded; age acceleration was computed as the residual of the clock estimate on age, and compared across cell types by one-way analysis of variance. The two datasets differ in platform, laboratory, donors, and cell-separation method, and the two analyses use different estimators. Results. In the within-donor design, estimated epigenetic age varied by a median of 15.6 years (Horvath), 36.7 years (Hannum), and 36.7 years (PhenoAge) within the same donor across preparations, reaching 54.0 years in one instance. Preparation accounted for 9.4% of total variance for Horvath, 32.3% for Hannum, and 33.8% for PhenoAge, against donor shares of 86.5%, 62.9%, and 59.6% respectively; leave-one-donor-out refitting placed those preparation shares between 7.8 and 17.2% (Horvath), 28.4 and 56.7% (Hannum), and 30.2 and 56.4% (PhenoAge). In the between-donor design, cell type explained 43.4% of age-adjusted variance for Horvath (permutation p = 2.6 × 10⁻³), 85.8% for Hannum and 89.5% for PhenoAge (both permutation p < 1 × 10⁻⁴), with acceleration spanning 8.5, 26.5, and 51.2 years respectively across cell types. Under PhenoAge, CD8 + T cells read 30.9 years younger and monocytes 20.4 years older than the donor’s chronological age; unadjusted PhenoAge estimates in CD8 + T cells were negative. Preparation accounted for 8.1% of total variance for the skin‑and‑blood clock and 9.4% for multi‑tissue Horvath, against 32.3% and 33.8% for the two clocks trained on whole blood; the two groups did not overlap. Substituting PBMC for whole blood, the commonest substitution in published work, shifted the Hannum estimate by a median of 5.6 years and PhenoAge by 4.5 years within the same donors, against 1.5 and 0.8 years for the two clocks trained across tissues; comparing PBMC against granulocytes shifted Hannum by 9.6 years. Robustness tracked training across more than one tissue rather than training on blood, and the two groups did not overlap under leave‑one‑donor‑out resampling; the between‑donor design of purified populations did not reproduce that grouping, where the skin‑and‑blood clock behaved more like the whole‑blood‑trained clocks. Hannum and PhenoAge were consistently more affected than Horvath but were not reliably separable from each other: they exchanged rank under BMIQ normalisation and their leave-one-donor-out intervals overlap almost entirely. Conclusions. Epigenetic age measured in blood is a joint property of the individual and the cell population assayed, and the effect is large enough to dominate the biological differences these instruments are typically used to detect. Blood preparation should be reported as a primary methodological variable, and comparisons across studies that differ in preparation are confounded. Blood-specific training does not confer robustness to blood cell composition; the multi-tissue clock was the most robust of the three, though not immune.</p>