Fig. 2

Distinction between two primary analysis methods in the present study. We employed both variance components and penalised regression models in order to examine the amount of phenotypic variance captured by each respective probe set (n = 18 in total, see Methods). Variance component estimates were obtained using the restricted maximum likelihood method in OSCA. Here, we were able to estimate the amount of phenotypic variance captured by all probes in a given probe set in the training sample (n ≤ 4450). We also employed penalised regression to build linear DNAm-based predictors of traits using probes in a given probe set in the training sample. We then applied the predictors to the test sample (n ≤ 2578) in order to estimate how much variance in a given trait the predictor could explain over basic covariates (such as age and sex). This coefficient reflected the incremental R² estimate and pertained to an out-of-sample setting as the predictor was applied to a sample outside of that in which it was derived. LASSO, least absolute shrinkage and selection operator; OSCA, OmicS data-based complex trait analysis. Image created using Biorender.com