Evaluating DNA methylation age on the Illumina MethylationEPIC Bead Chip

Abstract

DNA methylation age (DNAm age) has become a widely utilized epigenetic biomarker for the aging process. The Horvath method for determining DNAm age is perhaps the most widely utilized and validated DNA methylation age assessment measure. Horvath DNAm age is calculated based on methylation measurements at 353 loci, present on Illumina's 450k and 27k DNA methylation microarrays. With increasing use of the more recently developed Illumina MethylationEPIC (850k) microarray, it is worth revisiting this aging measure to evaluate estimation differences due to array design. Of the requisite 353 loci, 17 are missing from the 850k microarray. Similarly, an alternate, 71 loci DNA methylation age assessment measure created by Hannum et al. is missing 6 requisite loci. Using 17 datasets with 27k, 450k, and/or 850k methylation data, we compared each sample's epigenetic age estimated from all 353 loci required by the Horvath DNAm age calculator, and using only the 336 loci available on the 850k array. In 450k/27k data, removing loci not on the 850k array resulted in underestimation of Horvath's DNAm age. Underestimation of Horvath DNAm age increased from ages 0 to ~20, remaining stable thereafter (mean deviation = -3.46 y, SD = 1.13 for individuals ≥20 years). Underestimation of Horvath's DNAm age by the reduced 450k/27k data was similar to the underestimation observed in the 850k data indicating it is driven by missing probes. In analogous examination of Hannum's DNAm age, the magnitude and direction of epigenetic age misestimation varied with chronological age. In conclusion, inter-array deviations in DNAm age estimations may be largely driven by missing probes between arrays, despite default probe imputation procedures. Though correlations and associations based on Horvath's DNAm age may be unaffected, researchers should exercise caution when interpreting results based on absolute differences in DNAm age or when mixing samples assayed on different arrays.

Fig 1. Epigenetic age by chronological age in combinations of CATHGEN dataset and epigenetic clock.

The plot of DNA methylation by chronological age shows the impact of the missing probes, by applying the (a) Horvath and (b) Hannum epigenetic clocks to CATHGEN 450k ('full' and 'reduced') and 850k datasets.

Fig 2. Difference of 'full' and 'reduced' Horvath epigenetic Age by chronological age.

The difference of ‘full and reduced’ epigenetic ages calculated in the GEO (450k and 27k) and CATHGEN 450k data are presented as (a) boxplot by 5 year chronological age categories and (b) as a scatterplot.

Fig 3. Difference of 'full' and 'reduced' Hannum epigenetic Age by chronological age.

The difference of ‘full and reduced’ epigenetic ages calculated in the GEO (450k and 27k) and CATHGEN 450k data are presented as (a) boxplot by 5-year chronological age categories and (b) as a scatterplot.