Devising reliable and accurate epigenetic clocks: choosing the optimal computational solution

Abstract

Illumina DNA methylation arrays are frequently used for the discovery of methylation signatures associated with aging and disease. One of the major hurdles to overcome when training trait prediction models is the high dimensionality of the data, with the number of features (CpGs) greatly exceeding the typical number of samples assessed. In addition, most large-scale DNA methylation-based studies do not include replicate measurements for a given sample, making it impossible to estimate the degree of measurement uncertainty or the reliability of the prediction models. A recent study proposed that training penalized regression models on derived principal components (PCs) rather than on the original features (CpGs) results in more reliable age predictions, as estimated from technical replication. Moreover, the same method could be applied for predicting other phenotypes more reliably. Here, we aimed at validating the proposed PC method. We found that although dimension reduction with PCA consistently led to small improvements in the reliability of age prediction models, it severely compromised their accuracy. PC-based models needed far larger training set sizes to be similarly accurate as CpG-based models, whereas reliability did not depend on the sample size of the training set data for either approach. Finally, the PC version of a novel multiclass predictor for breast, ovarian and endometrial cancer we trained using weighted ensembles of deep-learning models also had a markedly lower predictive accuracy compared to a CpG version, suggesting limited applicability of the proposed PC method for predicting phenotypes beyond age.