Fig. 6: Fragment-level deconvolution using cell type-specific biomarkers.

a, Cell-type-specific markers achieved less than 0.1% resolution. In silico simulations for five cell types, in which held-out samples were computationally mixed within leukocytes then analysed using 1,246 atlas markers plus 25 additional megakaryocyte markers (red) or an array-based deconvolution of these mixes²⁸ (grey). Box plots show average contribution in ten simulations, with error bars representing 1 s.d. b,c, Cell type composition in leukocytes (b) and plasma samples (c) from healthy donors. Box plots show overall proportions of leukocytes, megakaryocytes and erythroblasts (MEP) and other cell types. d, Analysis of low-coverage plasma samples from 52 patients with SARS-CoV-2 (ref. ⁴⁴) identified endothelial-derived cfDNA in patients with WHO ordinal scale seven or higher (requiring admittance to intensive care unit). e–i, Fragment-level deconvolution of Roadmap/ENCODE samples^4,5 showing cell-type-specific contributions. e, Heart ventricle samples contained a mixture of cardiomyocytes, endothelial cells, fibroblasts and blood. f, Liver samples contained around 60% of hepatocyte DNA, plus blood and endothelial cells. g, Colon samples contained approximately 50% epithelium, plus fibroblasts and blood. h, Lung samples contained less than 30% of lung epithelial cells. i, Pancreatic islet samples contained beta, alpha, duct and acinar cells. Box plots denote median and IQR, with whiskers 1.5× IQR.