Fig. 4
DNAm-based prediction of complex traits using all available probes and four subsets of decreasing size. LASSO regression was used to build blood DNAm-based predictors of seventeen biochemical and complex traits (n ≤ 4450 training sample and n ≤ 2578 test sample). The four traits with the highest proportion of variance captured by DNAm predictors in the test sample are displayed (incremental R2 estimates above null model, see main text). The first set of probes included those that passed quality control in the training sample, were common to both the EPIC and 450K arrays and included both probes with known methylation QTLs (mQTLs) and probes without known mQTLs reported in the GoDMC consortium. The next four sets of probes included non-mQTL probes only and had decreasing numbers of probes but increasing mean variabilities. DNAm, DNA methylation; HDL, high-density lipoprotein; LASSO, least absolute shrinkage and selection operator; mQTL, methylation quantitative trait locus