Figure 2

Development of the epigenetic aging signature. (a) The most relevant AR-CpGs were selected by iterative division of 575 DNAm profiles into training and testing sets (different split ratios). Age predictions were made for training sets using either 51 AR-CpGs or subsets of 5 CpGs. The results indicated that subsets with five CpGs (selected by recursive feature elimination) can enable age predictions with a mean absolute deviation (MAD) from chronological age of less than 6 years. (b) The frequency of occurrence of individual AR-CpGs in the best performing subsets of five CpGs. Five specific CpGs occurred in more than 50% of these filtered subsets and hence seemed to provide the best complement for age predictions. (c) DNAm at relevant AR-CpG sites was subsequently analyzed by pyrosequencing after bisulfite conversion. The sequences surrounding three of the five AR-CpGs were particularly suitable for this approach (CpG sites represented on the HumanMethylation27 BeadChip platform are indicated). (d) DNAm levels at these AR-CpGs were analyzed in a training set from 82 blood samples. The results were in line with the microarray data and revealed a clear age-associated correlation for each of the three CpGs. For cg17861230 (PDE4C) this correlation was even better at a neighboring CpG locus, which was therefore preferred for further analysis. (e) Based on the results with these three AR-CpGs, we generated a multivariate model that enabled relatively precise age predictions (MAD of 5.4 years). (f) Notably, the precision was even slightly higher when we validated this method in an independent set of 69 samples (MAD of 4.5 years).