Fig. 1

Overview of analysis strategy in the present study. We tested whether subsets of probes showed similar predictive capacities to total DNAm array content (1) (‘all available probes’, n = 393,654). We first identified subsets of interest. We restricted primary analyses to probes without known genetic influences (i.e. non-mQTL probes) and those with mean Beta-values (β) between 10 and 90% (2). These probes were termed ‘variable non-mQTL probes’ (n = 115,746). We then extracted the 50,000, 20,000 and 10,000 probes with the highest standard deviations from the pool of 115,746 non-mQTL probes (3). In our primary analyses, we compared the predictive performances of these four probe subsets against that of the full set of probes used in our analyses (4). In further analyses, we tested the relative performances of subsets based on (i) probes without known mQTLs and with mean Beta-value between 10 and 90% (shown in green in (2), highlighted in (3)), (ii) probes with known mQTLs and with mean Beta-value between 10 and 90% (shown in red in (2)) and hypo- or hypermethylated probes (mean Beta-value ≤ 10% or ≥ 90%, also shown in red in (2)). DNAm, DNA methylation; mQTL, methylation quantitative trait locus; SD, standard deviation. Image created using Biorender.com