A new platform ranks 41 epigenetic aging clocks across 179 blood datasets
TranslAGE harmonizes 179 blood methylation datasets and scores 41 epigenetic clocks on stability, treatment response, associations and risk.
Summary
Epigenetic clocks estimate biological age from DNA methylation, but dozens exist and studies rarely compare them fairly. Researchers at Yale and collaborators built TranslAGE, a free online resource that pulls together 179 human blood methylation datasets covering more than 42,000 samples. They precalculated 41 published clocks for every sample and score each one on four traits: stability against noise, response to treatments, associations with disease and traits, and ability to predict future illness or death. Together these form a composite called the STAR score. The goal is to help scientists and clinical trial designers pick the right clock for their purpose instead of guessing. This is a preprint that has not been peer reviewed. The text provided describes the platform and methods more than final clock rankings.
Detailed Summary
Epigenetic clocks are among the most widely used biomarkers in aging research, but the field has a selection problem. Clocks have grown from first-generation chronological age predictors (Horvath, Hannum) to mortality and pace-of-aging clocks (PhenoAge, GrimAge, DunedinPACE) and system-specific clocks, and thousands of methylation-based algorithms have been reported. Studies usually test only a few clocks, in one dataset, on one property. When results conflict, nobody can tell whether the cause is the clock, the population or the analysis.
TranslAGE, from Yale and collaborators including Albert Higgins-Chen, tackles this directly. It harmonizes 179 human whole-blood methylation datasets from NCBI GEO, EMBL and private collaborators, covering more than 34,000 individuals and more than 42,000 methylation profiles on Illumina 450k and EPIC arrays. Of these datasets, 129 are longitudinal and 51 are cross-sectional. Every dataset was converted to a common metadata schema, and 41 clocks were calculated using an extension of the methylCIPHER R package. Each score was residualized for chronological age and sex.
The platform scores clocks on four domains, which together form the STAR framework. Stability measures reliability using the intraclass correlation coefficient. Five datasets with technical replicates and six with biological replicates (repeat samples within 24 hours, with events such as meals, mild stress or pollen exposure) were pooled by random-effects meta-analysis after Fisher's z-transformation. This matters because noise that exceeds the effect of an intervention inflates the sample size a trial needs.
Treatment response uses longitudinal data. Of 129 longitudinal datasets, 101 passed quality control. Seventy-four involved a putative aging intervention: 25 lifestyle, 11 supplement, 22 pharmacological and 16 other, such as Mediterranean diet, smoking cessation and metformin. Nineteen captured adverse exposures such as radiotherapy, SARS-CoV-2 infection or air pollution, and 8 were untreated controls. Paired t-tests compared baseline and follow-up residuals. Effect sizes were scaled to the population standard deviation of each clock from a large EPICv1 dataset, so negative values mean younger biological age and 1.0 equals one standard deviation.
The Associations domain captures cross-sectional links with age, demographics, disease and other phenotypes. The Risk domain tests prediction of future functional decline, morbidity and mortality. Interactive dashboards at translage.io let users rank clocks by each domain or by the composite STAR score. A researcher planning a lifestyle trial, for example, can find the clocks that are both stable and responsive to similar interventions. The authors describe it as the first standardized, reproducible benchmarking framework of this kind, and plan continual updates.
Several caveats apply. This is a bioRxiv preprint, and the supplied text ends before the detailed Associations and Risk results or the discussion, so specific clock rankings are not summarized here. The Stability analysis rests on only 11 datasets. The platform currently includes only human whole-blood array data.
Key Findings
- TranslAGE integrates 179 harmonized human blood DNA methylation datasets with more than 34,000 individuals and more than 42,000 methylation profiles
- 41 epigenetic biomarker scores are precalculated for every sample, spanning first-generation, mortality, reliability-optimized and system-specific clocks
- Of the datasets, 129 are longitudinal and 51 are cross-sectional, on Illumina 450k and EPIC arrays
- Stability is quantified from 5 technical-replicate and 6 biological-replicate datasets using pooled intraclass correlation coefficients
- 101 of 129 longitudinal datasets passed quality control for treatment response, including 74 intervention datasets (25 lifestyle, 11 supplement, 22 pharmacological, 16 other)
- 19 datasets capture adverse exposures such as radiotherapy, SARS-CoV-2 infection and air pollution, and 8 serve as untreated controls
- The four-domain STAR framework (Stability, Treatment response, Associations, Risk) combines test metrics into one composite score for ranking clocks
Methodology
The authors compiled public (NCBI GEO, EMBL) and private whole-blood human methylation array datasets, standardized their metadata, and calculated 41 clocks with an extended methylCIPHER pipeline. Scores were residualized for age and sex and routed to the Stability, Treatment response, Associations and Risk analyses by dataset type. Stability used intraclass correlation coefficients, Fisher z-transformed and pooled by random-effects meta-analysis. Treatment response used paired t-tests of baseline versus follow-up residuals, with effect sizes scaled to each clock's population standard deviation from a large EPICv1 dataset.
Study Limitations
This is a preprint that has not been peer reviewed, and the supplied text ends before the Associations and Risk results and the discussion, so those findings are not covered. Datasets are limited to human whole-blood array data, and the Stability analysis rests on only 5 technical and 6 biological replicate datasets. Downloadable harmonized data are pending journal publication and limited to originally public datasets. No conflict-of-interest statement appeared in the provided text.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: