Why clocks trained on health outcomes capture aging rate better
An epigenetic clock estimates biological age from DNA methylation, a chemical mark on DNA that changes with age. The difference between this estimated age and a person's actual age is called epigenetic age acceleration (EAA): a positive value suggests faster biological aging, a negative value slower aging [1][2][3]. A clock trained only to predict chronological age may track the calendar well but miss whether someone is aging faster or slower than normal.
In mice, researchers built three types of clocks: a general clock trained without pre-selecting DNA sites, a developmental clock trained on sites that change during early development, and an interventional clock trained on sites that respond to caloric restriction or growth hormone receptor knockout—interventions known to affect aging [1][2][3]. All three predicted age with similar correlation (average r = 0.89), but the interventional clocks showed larger deviations from chronological age and higher median error [1][2][3]. That larger deviation is not a flaw: it means the interventional clock is picking up biological variation that the general clock smooths over.
How mouse clocks were validated against diet, weight, and lifespan
The mouse clocks were tested in 339 mice from a genetically diverse family (the BXD strains), some fed a high-fat diet and some a control diet [1][2][3]. The interventional clock was the most sensitive to diet: on the control diet, mice averaged a slight age deceleration (−0.04 years), while on the high-fat diet they averaged an age acceleration (+0.11 years) [1][2][3]. The general and developmental liver clocks did not differ between diets, suggesting they are less responsive to this metabolic challenge.
Body weight told a similar story. In mice weighed just one or three days after starting the high-fat diet, higher weight correlated with greater age acceleration on all clocks, but especially on the interventional clocks (r = 0.45 for the pan-tissue clock and r = 0.58 for the liver clock) [1][2][3]. When the researchers checked how well the clocks predicted strain lifespan, only the general pan-tissue clock and the liver interventional clock showed significant inverse correlations—accelerated aging went with shorter strain lifespan—but these correlations were weak, explaining only a small fraction of lifespan variance [1][2][3].
Human clocks: next-generation models and what their validation shows
In humans, the field has moved from first-generation clocks trained solely on chronological age to next-generation clocks trained to associate with health, lifestyle, or age-related outcomes [5]. A review of these models reports that next-generation clocks associate with a greater number of health and disease signals, are often more predictive of age-related outcomes, and appear more responsive to interventions [5]. That review also notes that the choice of clock depends on the research goal, but for health-oriented association and intervention studies, current evidence favors next-generation models [5].
A different approach uses proteins rather than DNA methylation. Organ-specific proteomic aging clocks have been developed and validated in large population cohorts, including the UK Biobank (n = 43,616), where they showed high accuracy and predicted disease and longevity across diverse populations [4]. This suggests that clock validation is not limited to methylation: protein-based clocks are being tested against hard outcomes like disease and death, not just calendar age.
What the validation studies still cannot settle
The mouse studies show that interventional clocks respond more strongly to diet and weight, but the link to lifespan was weak and limited to only two of the clocks tested [1][2][3]. That means a clock can be sensitive to metabolic state without necessarily being a strong predictor of how long an animal will live. In humans, the review of next-generation clocks reports greater predictive power for health outcomes, but it does not detail the validation methodologies used across studies, so it is difficult to compare how rigorously each clock was tested [5]. The proteomic clocks were validated in large cohorts for disease and longevity, but the abstract does not specify how they compare head-to-head with epigenetic clocks [4].
Sources used in this answer
Author response: Genetic loci and metabolic states associated with murine epigenetic aging
In BXD mice, interventional epigenetic clocks showed larger deviations from chronological age and higher predictive error than general or developmental clocks, and were more strongly associated with high-fat diet and body weight, while all clocks predicted age with similar correlation.
Decision letter: Genetic loci and metabolic states associated with murine epigenetic aging
The decision letter highlights that three epigenetic predictors (general, developmental, interventional) were compared for accuracy and associations with metabolic traits and lifespan in genetically diverse mice.
Editor's evaluation: Genetic loci and metabolic states associated with murine epigenetic aging
The editor's evaluation notes that three epigenetic predictors were tested for how they differ between BXD mice, including metabolic traits and lifespan, and that quantitative trait loci were identified for the different predictors.
Organ-specific proteomic aging clocks predict disease and longevity across diverse populations
Organ-specific proteomic aging clocks were developed and validated in the UK Biobank (n=43,616) and other cohorts, showing high accuracy and predicting disease and longevity across diverse populations.
First-generation versus next-generation epigenetic aging clocks: Differences in performance and utility
Next-generation epigenetic clocks trained on health, lifestyle, or age-related outcomes associate with more health and disease signals, are often more predictive of age-related outcomes, and appear more responsive to interventions than first-generation clocks trained only on chronological age.