Optimizing effect sizes and specificity trumps machine learning when building DNA methylation reference panels for cell-type deconvolution

Accurate cell-type deconvolution is critical for correct interpretation of Epigenome-Wide Association Studies. Such cell-type deconvolution involves estimating underlying cell-type fractions in a sample, which is accomplished using a DNA methylation reference panel built from sorted or single-cell DNAm data. Two competing approaches have emerged to build such reference panels, one which uses machine-learning, and another based on optimizing effect size and cell-type specificity. Here we demonstrate that the latter approach is preferable, because, owing to the relatively small number of sorted samples used in building panels, standard machine learning does not optimize effect size and cell-type specificity, causing the model to overfit and underperform when tested in independent data. Furthermore, adult blood panels built from cell-type specific hypomethylated markers improves inference when compared to panels built from hypermethylated ones. These insights provide important guidelines for optimizing the construction of future DNAm reference panels. To aid this task, we have added a function for building an optimized DNAm reference panel to our EpiDISH R-package.