Fig. 2: Two paradigms of biological age estimation.

In the first paradigm, biomarker data are compressed into a single latent quantity representing either the current state (e.g., first-generation epigenetic clocks) or the rate (e.g., DunedinPACE) of biological aging. Examples include first-generation epigenetic clocks, the Klemera-Doubal model, PCA clocks, and DunedinPACE. The red dashed arrow illustrates the idea that quantities within the first paradigm are not specifically designed for health outcome prediction—yet this is typically expected of them. In the second paradigm, biomarkers are used to train an ensemble of models that directly predict risks of age-related diseases—including all-cause mortality—with these risks then aggregated into a unified biological age estimate. Examples include second-generation epigenetic clocks; additionally, survival ML models, expert-derived risk scores, and Large Health Models (LHMs) fall into this category if paired with a downstream biological age calculation step.