Abstract
<title>Abstract</title> <p> Background. Blood-based DNA-methylation classifiers are commonly validated by scoring publicly deposited disease cohorts against a healthy reference. Where the classifier is a deviation metric — the distance of a sample from a healthy baseline — it inherits the hematopoietic compartment of that baseline. Publicly deposited hematologic-malignancy series are heterogeneous in compartment: some are peripheral whole blood, others bone marrow, sorted fractions, or mobilized apheresis product. The consequence for validation has not been quantified. Methods. We defined a fully disclosed burden metric — mean absolute β-deviation from a healthy peripheral whole-blood reference (GSE40279, <italic>n</italic> = 656) — over 230,213 probes common and missingness-free across all cohorts. Compartment effect at zero disease was calibrated on GSE35069, in which six healthy donors were each profiled as whole blood, peripheral blood mononuclear cells (PBMC), granulocytes, and seven purified leukocyte subsets. Three hematologic-malignancy series were then scored against the same reference: GSE105420 (chronic myelomonocytic leukemia, CMML), GSE106600 (chronic myeloid leukemia, CML), and GSE118241 (myelofibrosis). Discrimination was quantified as the area under the receiver-operating-characteristic curve (AUC) with 10,000-fold bootstrap confidence intervals and Mann–Whitney tests. Leukocyte composition was estimated by robust partial correlation. Reference choice was varied across three independent healthy cohorts. Results. In healthy donors, compartment alone generated burden spanning 0.0229 (whole blood) to 0.0573 (CD8 + T cells), a range of 0.0344. Discrimination in the disease series tracked compartment matching rather than disease. Where cases and controls were drawn from different compartments, discrimination was near-perfect: AUC 1.000 (95% CI 1.000–1.000) for CML cells versus apheresis product, and 0.972 (0.931–1.000) for CMML bone marrow versus mobilized CD34 + peripheral blood. In the single compartment-matched comparison available — myelofibrosis versus control, both peripheral blood cells (GSE118241) — discrimination was absent: AUC 0.535 (0.316–0.763), <italic>p</italic> = 0.83. Holding disease constant and varying compartment within the same series produced substantially greater separation than disease did within matched compartment: AUC 0.827 (0.669–0.951), <italic>p</italic> = 0.0016. Estimated composition explained 45.2% of burden variance across the disease cohorts ( <italic>p</italic> = 6.7 × 10⁻¹⁶); the mobilized CD34 + controls of GSE105420 carried a median neutrophil fraction of 0.909 against 0.450 in the marrow cases they were compared with. Diagnosis and compartment were completely collinear in GSE105420. Reference choice alone shifted whole-blood burden by 82.7% across three healthy cohorts and moved the metric 39% as much as compartment did. Conclusions. Compartment mismatch between reference and test cohort is sufficient to produce near-perfect apparent discrimination in the absence of any demonstrated disease effect. Blood-referenced classifiers validated on publicly deposited hematologic series without compartment matching, reported composition estimates, and a stated reference are not interpretable as evidence of disease discrimination. We propose three reporting requirements. </p>