FraminghamPACE Measures Biological Aging From DNA Methylation
Researchers posted the second version of FraminghamPACE on July 16, a preprint describing a DNA-methylation biomarker trained on the Framingham Heart Study Offspring Cohort. The models used data from 2,069 participants and were tested across independent cohorts and the CALERIE trial, but the work has not been peer reviewed and does not establish a clinical surrogate endpoint.
Researchers have developed FraminghamPACE, a DNA-methylation biomarker designed to estimate how quickly a person is aging rather than how old they are. The second version of the preprint was posted on July 16, 2026. It has not yet been peer reviewed.
How the clock was built
The team adapted a Pace of Aging method previously developed in the Dunedin Study for the mixed ages and uneven measurement schedules in the Framingham Heart Study Offspring Cohort. It modeled change across 17 blood-chemistry and organ-function measures collected over eight examination waves from 1971 through 2008.
Researchers then trained two machine-learning models on DNA-methylation data from 2,069 participants. An elastic-net version selected 1,123 CpG sites, while a ridge-regression version retained 14,465. In held-out Framingham data, the models explained 47% and 46% of the variance in the longitudinal Pace of Aging measure, respectively. Those figures describe agreement with the study's constructed aging phenotype; they are not measures of diagnostic accuracy.
What the validation found
The paper reports high repeat-measure reliability in technical replicates from the LOLIPOP and CARDIA studies, with intraclass correlations of at least 0.95. Across additional observational cohorts, higher FraminghamPACE values were associated with shorter healthspan and lifespan. Repeated CARDIA measurements also rose as participants aged, consistent with an accelerating pace of biological aging across adulthood.
In the CALERIE randomized trial, the FraminghamPACE measures slowed in the calorie-restriction group in parallel with DunedinPACE. The authors present that response as evidence that biomarkers trained on longitudinal physiological change may be more sensitive to some interventions than clocks trained mainly to predict chronological age or survival.
The practical boundary
For data and AI teams, the study is a useful example of target design: the model predicts a longitudinal, multi-system phenotype rather than a convenient calendar-age label. It also shows the trade-off between sparse elastic-net models and broader ridge models without assuming that more features automatically produce a better clinical tool.
The work remains a preprint, and associations with later health outcomes do not make FraminghamPACE a validated surrogate endpoint. Clinical use would require independent replication, calibration across populations and laboratories, and evidence that intervention-driven changes in the score reliably predict meaningful changes in healthspan.
Key Points
- 1FraminghamPACE was trained on DNA-methylation data from 2,069 Framingham Offspring participants against a longitudinal measure built from 17 physiological indicators.
- 2Elastic-net and ridge versions used 1,123 and 14,465 CpG sites and explained 47% and 46% of held-out variance in the constructed Pace of Aging phenotype.
- 3The preprint reports strong technical reliability and intervention sensitivity, but it does not establish FraminghamPACE as a clinically validated surrogate endpoint.
Scoring Rationale
The preprint introduces a reproducible machine-learning biomarker trained on unusually long longitudinal health data and tests it across several cohorts and a randomized trial. Its practical importance is tempered by preprint status and the absence of clinical surrogate validation.
Sources
Primary source and supporting public references used for this report.
Practice with real Health & Insurance data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Health & Insurance problems
