A New Pace of Aging Clock Derived from the Framingham Heart Study Offspring Cohort
Aging clocks of many varieties have been produced in recent years by applying machine learning techniques to a wide range of biological data that changes with age. This approach yields a tool that is disconnected from our understanding of the mechanisms of aging; links between the forms of cell and tissue damage and dysfunction that drive aging and the measures making up the clocks have yet to be determined. This makes it hard to interpret results, and hard to make practical use of a clock to assess the quality of any given approach to slowing or reversing aging. We have no idea in advance as to whether a given clock will perform well for a given intervention, and finding out is a slow process. The primary approach to this challenge taken by the research community is to produce new clocks at a fair pace, and gather as much data as possible on how the clocks behave, in search of patterns of clock behavior.
The outcome most often used to develop aging biomarkers is age itself, i.e. years lived since birth. However, in humans, relying on years lived as an outcome introduces a range of biases, most prominently confounding of aging with survival; humans in their 70s and beyond are, by definition, successful agers, having outlived most of their peers. The results of machine learning analysis differentiating older from younger people could therefore reflect not only aging-related biological damage, but also resilience.
An alternative approach that may overcome this limitation is to apply machine learning to an outcome that represents something closer to what many interventions aim to modify: the current rate of aging-related biological deterioration. We developed such a measure, Pace of Aging, by modeling changes over 20 years of follow-up in a panel of organ-function measurements among participants in the Dunedin Longitudinal Study. Critically, it also proved sensitive to the effects of calorie restriction, the intervention best established to slow aging in a range of laboratory models.
If using Pace of Aging in machine learning analysis to develop aging biomarkers can yield more sensitive endpoints for clinical trials, this would be consequential for the field. However, there are alternative explanations for the calorie restriction trial result. The participants in the trial (CALERIE) were healthy midlife adults, similar to the Dunedin Study members whose data were used to develop DunedinPACE. In contrast, the leading survival-time biomarker, the GrimAge epigenetic clock, was developed using data from older adults, many of whom had prevalent chronic disease. The critical factor could therefore be similarities between the participants whose data were used to develop the biomarker and the participants in the clinical trial.
To adjudicate between these competing hypotheses, biomarker design vs. demographic similarity, we developed a novel Pace of Aging biomarker in the same older-adult cohort used to develop GrimAge and tested its response to intervention in the CALERIE trial. We obtained data from the Framingham Heart Study Offspring Cohort. We adapted our Pace of Aging method for mixed-age cohorts with variable follow-up of organ-function measures and applied it to develop a novel DNA methylation biomarker of Pace of Aging in data from the Framingham Heart Study Offspring Cohort. When applied in independent cohorts, this novel biomarker (1) demonstrated exceptional technical reliability; (2) revealed a pattern of accelerating Pace of Aging with advancing age, replicating a finding first observed for our original Pace of Aging biomarkers developed in the Dunedin Study. In analysis of a randomized controlled trial of calorie restriction in healthy, non-obese humans (CALERIE), our novel Pace of Aging biomarker was slowed by calorie restriction, parallel to our original Pace of Aging biomarker.