Fig. 2: Selection and validation of unbiased fluctuating CpG (fCpG) sites.

A Starting with all CpG sites (gray dots), sites with an average methylation value β between 0.4 and 0.6 in both breast cancers (N = 634, TCGA cohort) and normal samples (N = 79, Normal cohort) were labeled as Unbiased Set I (yellow shaded rectangle). B Starting with sites in Unbiased Set I (yellow dots), sites with an average methylation value β between 0.4 and 0.6 in both normal breast luminal epithelial tissue samples (N = 3, single-cell DNAm atlas) and normal breast basal epithelial tissue samples (N = 4, single-cell DNAm atlas) were labeled as Unbiased Set II (red shaded rectangle). C Sites in Unbiased Set II were ranked by their inter-tumor standard deviation, and the top 500 were included in the clock set \({{\mathscr{C}}}\). D Distribution of individual fCpG β-values across tumor and normal cohorts; each curve represents the kernel density estimate of the distribution of a single fCpG across the TCGA cohort (blue; N = 633) or the normal breast cohort (yellow; N = 79). E In a separate cohort of breast cancers (N = 146, Lund cohort), the inter-tumor standard deviation of \(\beta\) was higher for fCpG sites in the clock set \({{\mathscr{C}}}\) (median: 0.188) compared to CpGs not included in \({{\mathscr{C}}}\) (median: 0.102; P < 10⁻¹⁰; Wilcoxon rank-sum test). F In a small cohort of patients (N = 5) with multiple primary samples (3-5 per patient), the intra-tumor standard deviation of β was higher for CpG sites in the clock set \({{\mathscr{C}}}\) (median: 0.056), compared to CpGs not included in \({{\mathscr{C}}}\) (median: 0.026; P < 10⁻¹⁰, Wilcoxon rank-sum test).