UC Berkeley

Biological age estimation from DNA methylation and the identification of relevant biomarkers is an active research problem that has predominantly been tackled using penalized regression (e.g., elastic net). Such models are commonly used to select small subsets of CpG probes from hundreds of thousands of candidates and to regularize training. Here, I show that feature sets discovered by these approaches can lack biological interpretability and relevance in first‑ and next‑generation DNA methylation clocks, and I clarify why these procedures can systematically exclude biomarkers of aging and age‑related disease. In contrast to the assumption that regularized linear regression is required to prevent overfitting, I demonstrate that hypothesis‑driven selection of biologically relevant features, combined with ordinary least squares (OLS), yields accurate, well‑calibrated, and generalizable clocks with high interpretability. I further show that disease‑associated shifts in CpG methylation interacting with opposite‑signed model weights can cancel at the aggregate—an incoherence effect that reduces resolution between health and inflammaging. Lastly, I introduce feature rectification, which aligns these shifts to restore coherent signal aggregation and improves separation of DNAm‑age predictions for healthy individuals versus patients with chronic inflammatory diseases.