A smoothing method for DNA methylome analysis to enhance epigenomic signature detection in epigenome-wide association studies

Epigenome-wide association studies (EWAS) are instrumental for mapping DNA methylation changes in human traits and diseases but often suffer from low statistical power and false positives, especially in small cohorts. We developed an EWAS smoothing method that exploits co-methylation of adjacent CpG probes within CpG islands via a sliding-window average and generalized it using Savitzky-Golay filtering. We applied the smoothing approach-with window widths of 1-3 CpGs and, for generalization, Savitzky-Golay filters of varying polynomial orders and window sizes-across five distinct EWAS settings. Performance was quantified by signal-to-noise ratio (SNR), noise-variance reduction, variance ratio (VR), Bayes factors, and sample-size sensitivity. In the MMACHC epimutation dataset, a 5-CpG window (width, w = 2) increased SNR by 90 %, reduced noise variance by 80 %, and elevated VR by 176 % at the target CpG island, with no genome-wide false positives. For MLH1, smoothing preserved the top association and suppressed background signals. In the aging EWAS, a "Polyepigenetic CpG aging score" was derived following smoothing. This score correlated strongly with chronological age in the discovery cohort (Spearman's ρ = 0.89; P = 3.0 × 10^-219) and was independently validated in a separate dataset, significantly distinguishing newborns from nonagenarians (P = 3.4 × 10^-8). Savitzky-Golay filtering of order 0 with a 5-CpG window yielded optimal SNR across bootstrap iterations, supporting this configuration as a robust choice for methylation array smoothing. As an extension of the Savitzky-Golay-based smoothing framework, reanalysis of a liver cancer dataset identified five top loci surpassing a smoothed P-value threshold of 1 × 10^-8. Among these, MIR10A within the HOXB3 locus was the only previously reported functionally relevant site. In conclusion, the smoothing method improves EWAS performance by enhancing SNR, enabling detection of meaningful associations even in small cohorts, and offers a valuable tool for reanalyzing existing Infinium methylation array datasets to uncover previously undetected epigenomic signatures.