Industries / Frontier science / Longevity
Measuring age
This work estimates how old a body is, or how fast it is aging, from death rates, health checks, blood tests and chemical marks on DNA, as distinct from its age in years.
7
Milestones achieved, from 1825 to 2025.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1825
A law of dying
Benjamin Gompertz, a London actuary (a specialist who prices insurance from death rates), read a paper to the Royal Society on 16 June 1825 showing that over much of adult life the rate of dying rises exponentially with age. The slope of his curve became a way to compare how fast populations and species age: George Sacher later put the time it takes the death rate to double at 220 days in house mice and 3,100 days, about eight and a half years, in US white women in 1969. Gompertz also saw that in his life tables the yearly death rate stayed almost flat at 0.25 from age 92, a late-life "plateau" that was later raised as a challenge to his law.
Lessonthe law fits most of adult life but not the oldest ages, an exception Gompertz himself noted in his own tables.
Credit
Benjamin Gompertz; described by Thomas Kirkwood of Newcastle University in 2015
1958
The same people, for decades
Nathan Shock, chief of the Gerontology Branch at the US National Institutes of Health, began the Baltimore Longitudinal Study of Aging (BLSA) in 1958 to study normal aging by testing the same healthy volunteers again and again, which the NIA describes as a break with scientific convention. Women joined in 1978, and by the National Institute on Aging's count more than 3,200 people have taken part, with over 1,300 now returning every one to four years for nearly three days of tests. In 1985 BLSA researchers reported creatinine clearance, a measure of how well the kidneys filter blood, taken 5 to 14 times each in 446 volunteers between 1958 and 1981: in the 254 with no sign of kidney disease and not on diuretics or blood-pressure drugs it fell by 0.75 ml per minute a year on average, yet a third of everyone followed showed no fall at all.
Lessonan average decline with age is not every person's decline, and only repeated measures of the same people can show the difference.
Credit
Nathan W. Shock and the NIA; R. D. Lindeman, J. Tobin and N. W. Shock
2001
Frailty, scored two ways
Frailty is a state in old age that carries a high risk of falls, disability, hospital stays and death, and in 2001 two teams turned it into a score. A. B. Mitnitski, A. J. Mogilner and K. Rockwood defined a frailty index as the share of a list of possible health problems, from symptoms to abnormal test results, that a person has, and found in older Canadians that problems built up at about 3% a year; L. P. Fried of Johns Hopkins and colleagues defined a frailty phenotype as three or more of five signs (unplanned weight loss, exhaustion, weak grip, slow walking and low activity), which marked 6.9% of 5,317 people aged 65 and over in the US Cardiovascular Health Study and predicted each of those four outcomes. In 2016 A. Clegg and colleagues at the University of Leeds built an electronic frailty index from 36 deficits in the family-doctor records of 931,541 UK patients, and by 2025 it was built into UK primary care record systems nationally.
Lessonfrailty can be scored from records doctors already keep, without new tests.
pmc.ncbi.nlm.nih.govebi.ac.ukpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Credit
Mitnitski, Mogilner and Rockwood; Fried and the Cardiovascular Health Study Collaborative Research Group; Clegg and colleagues
2013
Age read from DNA
DNA methylation is a set of chemical tags on DNA whose pattern changes with age. In 2011 a UCLA team had estimated age from two tagged sites in the saliva of 34 pairs of identical twins, to within 5.2 years on average; G. Hannum, T. Ideker, K. Zhang and colleagues at the University of California, San Diego, then built a model from 71 sites in the blood of 656 people aged 19 to 101 that erred by 3.9 years, and by 4.9 years in a second group of 174. In October 2013 S. Horvath of UCLA built a clock from 353 sites, trained on about 8,000 samples covering 51 tissues and cell types, with a median error of 3.6 years in test data; it read close to zero for embryonic stem cells.
Lessonthese clocks were trained to match calendar age, which later work showed is a different task from predicting health.
Credit
S. Bocklandt, E. Vilain and colleagues (2011); Hannum, Ideker, Zhang and colleagues, in Molecular Cell (online November 2012, in print January 2013); Horvath, in Genome Biology
2019
Clocks trained on death
In 2018 M. E. Levine, S. Horvath and colleagues at UCLA built DNAm PhenoAge from 513 methylation sites that track a "phenotypic age" computed from nine blood measures and age, and in 2019 A. T. Lu, Horvath and colleagues built GrimAge from DNA-methylation stand-ins for seven blood proteins and for lifetime smoking, fitted to time to death. In 490 people in the Irish Longitudinal Study on Ageing (TILDA), the age-trained Horvath and Hannum clocks did not predict health, while GrimAge was linked to 8 of 9 outcomes and, after allowing for social and lifestyle factors, still to walking speed, frailty and death within ten years. In 3,581 people in the US Health and Retirement Study, PhenoAge, GrimAge and DunedinPACE predicted cognitive problems at the time, functional limits and chronic illness two years later, and death within four years, but the authors found that demographics, socioeconomic status, mental health and health habits "remain equally, if not more robust, predictors".
Lessontraining on survival made clocks better predictors of health than training on age, but in a national sample a person's circumstances and habits predicted as well or better.
pmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Credit
Levine, Lu, Horvath and colleagues, UCLA; C. McCrory, R. A. Kenny and colleagues, Trinity College Dublin; J. D. Faul, E. M. Crimmins and colleagues, University of Michigan and University of Southern California
2022
Same sample, different ages
In 2020 K. Sugden and colleagues measured 350 blood DNA samples twice and found that many methylation sites read unreliably from one run to the next; in 2022 A. T. Higgins-Chen, M. E. Levine and colleagues at Yale showed that this noise made six well-known clocks differ by up to 9 years between repeat runs of the same sample, and that rebuilding each clock from principal components (summary patterns drawn from many sites at once) brought most repeats within 1.5 years. The same year D. W. Belsky of Columbia University, A. Caspi and T. E. Moffitt of Duke University and colleagues released DunedinPACE, a pace-of-aging score (years of biological aging per calendar year) trained on 20 years of decline in 19 measures of organ health in 1,037 people born in Dunedin, New Zealand, in 1972 and 1973, using only sites that read reliably; its agreement between repeat runs, measured as an intraclass correlation, was 0.96 out of 1. In 2026 the Yale group reanalyzed the data of two published studies, an eight-week diet, exercise and supplement program and two years of folic acid and vitamin B12, that had each reported a younger clock age, and found that none of the other clocks it added confirmed them.
Lessonone clock moving after a treatment can be noise, so the Yale group asks that reliable clocks, clocks trained on death or pace of aging, and several clocks all move the same way.
ebi.ac.ukpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Credit
Sugden and colleagues; Higgins-Chen, Levine and colleagues; Belsky, Caspi, Moffitt and colleagues; D. S. Borrus, A. T. Higgins-Chen and colleagues
2025
One test bench for clocks
Biolearn, an open-source Python library from V. N. Gladyshev's group at Brigham and Women's Hospital and Harvard Medical School and colleagues, computes 39 aging biomarkers the same way on the same public data; across more than 20,000 people, how well a clock predicted calendar age bore almost no relation to how well it predicted death (a correlation of 0.12). The Biomarkers of Aging Consortium used it to score an open competition in 2024 on held-out blood samples from 500 people in the Mass General Brigham Biobank, and 152 teams from 28 countries took part. In the age round the best entries erred by under 2.5 years, against about 4.8 years for Horvath's clock; in the death round, which six teams entered, the winner, built from blood proteins measured in 51,969 UK Biobank participants, reached a concordance index (the chance of correctly ranking which of two people dies first) of 0.844, against 0.812 for GrimAge version 2.
Lessona clock has to be judged against the outcome it is meant for, because the best match for calendar age is not the best predictor of death.
Credit
K. Ying, M. Moqri, V. N. Gladyshev and the Biomarkers of Aging Consortium; US$200,000 in prizes from the Methuselah Foundation and the Volo Foundation
Watch for
Age from a face photo
FaceAge is a deep-learning model, a program trained on many examples, that estimates biological age from a photograph of the face. Its makers trained and checked it on photos of 58,851 people aged 60 or older presumed to be healthy, from two public image sets, then tested it on 6,196 cancer patients in the Netherlands and the US: patients looked on average 4.79 years older than a reference group without cancer, and each decade of extra FaceAge went with a 15% higher risk of death in 4,906 patients with many kinds of cancer, after allowing for other factors. A later paper by the same group extended it to 24,556 cancer patients aged 60 and over.
Achieved whena group outside these institutions tests FaceAge on another set of patients, or on people without cancer, and finds that it predicts survival beyond calendar age.
Claimed by
D. Bontempi, R. H. Mak, H. J. W. L. Aerts and colleagues at Mass General Brigham and Maastricht University, in The Lancet Digital Health, May 2025
Organ ages from blood
H. S.-H. Oh, T. Wyss-Coray and colleagues at Stanford estimated the biological age of 11 organs from 2,916 blood proteins in 44,498 UK Biobank participants. They report that an especially aged brain carried about three times the risk of Alzheimer's disease (hazard ratio 3.1), similar to carrying one copy of APOE4, the strongest genetic risk factor for the common form, and that the risk of death rose with each extra aged organ, to 8.3 times for eight or more. The paper states that Wyss-Coray, Oh and a third author are co-founders of Teal Omics, Inc. and hold equity in it.
Achieved whena group outside Stanford applies the published organ models to a cohort other than UK Biobank and finds similar links to disease and death.
Claimed by
Oh, Wyss-Coray and colleagues, in Nature Medicine, July 2025
Biological age from a wristband
WHOOP, a Boston maker of fitness wristbands, launched Healthspan on 8 May 2025, a feature it says uses nine measures linked to long-term health, among them sleep and heart-rate data, to give each member a "WHOOP Age" and a "Pace of Aging". It says it developed the method with Eric Verdin, chief executive of the Buck Institute for Research on Aging. A Europe PMC search on 20 November 2026 found no published test of WHOOP Age.
Achieved whena study by a group outside WHOOP and the Buck Institute, with its methods published, shows that WHOOP Age or Pace of Aging predicts illness or death, or agrees with an established aging measure, in a defined group of wearers.
Claimed by
WHOOP, May 2025
Open
01
Same clock, same number
A public test suite that feeds the same synthetic methylation data, made by computer with and without missing sites, to every open toolkit that implements a given clock, and publishes how far their answers differ, rerun on each new release.
The briefCredit · unclaimed02
Methylation clocks, appraised
A public appraisal of the most-used DNA methylation clocks with PROBAST+AI, a published checklist for judging bias in prediction models, together with a ledger, for each clock, of every study that tested it against health outcomes, flagged where any of the clock's makers is also an author of the test.
The briefCredit · unclaimed
Open 01 · the brief
Same clock, same number
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test suite that feeds the same synthetic methylation data, made by computer with and without missing sites, to every open toolkit that implements a given clock, and publishes how far their answers differ, rerun on each new release.
Why it is still open
At least six openly available toolkits compute published clocks with their own code, and many clocks appear in more than one. Biolearn says its models "were verified to be correct" against the original software where available, and OmniAge, a 2026 package of 413 aging biomarkers, says its team compared its outputs with the original software or, where none was released, with pyaging, methylCIPHER and methylclock; neither paper reports in its main text how far the toolkits differ from one another on the same input. Missing values matter: in Biolearn's own test, randomly removing 20% of values left Horvath-clock predictions 8.96 years from the full-data predictions on average, and filling them back in cut that to 2.40 years; A. L. Kelly and colleagues found in 2026 that leaving missing sites out of the calculation biased every clock they tested. A search of Europe PMC and GitHub on 20 November 2026 found no shared test that runs one input through several toolkits.
pmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Built with what exists
Biolearn and pyaging (Python); methylCIPHER, its rewrite methylCIPHERv2, and dnaMethyAge (R); methylclock (Bioconductor); OmniAge (R and Python, released July 2026); synthetic data from a fixed random seed; GitHub Actions to rerun the suite on each release.
First step
Generate one synthetic matrix covering the sites of Horvath's 2013 clock, Hannum's clock and PhenoAge, run it through Biolearn, pyaging and methylCIPHER, and compare the three outputs.
Done looks like
For at least 20 clocks found in two or more toolkits, a published table of the largest difference in predicted years between toolkits on the same input, with and without 5% of sites missing; every difference above 0.1 year filed as an issue with the toolkit's maintainers; the suite rerunning automatically for six months.
Open 02 · the brief
Methylation clocks, appraised
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public appraisal of the most-used DNA methylation clocks with PROBAST+AI, a published checklist for judging bias in prediction models, together with a ledger, for each clock, of every study that tested it against health outcomes, flagged where any of the clock's makers is also an author of the test.
Why it is still open
J. Zhang and colleagues appraised 81 clocks built from routine clinical tests with PROBAST+AI and found that 31 (38.3%) had been validated neither within their own data nor on outside data, and that most were at high risk of bias; their review covered clinical-test clocks only, published up to June 2024. Some outside tests of methylation clocks share authors with the clocks: A. T. Lu and S. Horvath, who built GrimAge, are authors of the TILDA study that found GrimAge outperformed other clocks, and M. E. Levine, who built PhenoAge, of the Health and Retirement Study analysis. A Europe PMC search on 20 November 2026 found no PROBAST appraisal of DNA methylation clocks.
Built with what exists
The Europe PMC and OpenAlex APIs (OpenAlex gives each author a stable identifier, so shared authorship can be counted by machine); the PROBAST+AI checklist; a model that scores fixed-answer checklist questions, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"), with two people checking its answers; an open CSV file.
First step
Appraise Horvath's 2013 clock, Hannum's clock, PhenoAge, GrimAge and DunedinPACE by hand with PROBAST+AI, and pull every Europe PMC paper that reports GrimAge against death.
Done looks like
The 30 most-cited methylation clocks appraised by two raters, with at least 85% agreement on each domain rating; for each clock, every outcome study found, with the share run by groups that include none of its makers; all released under an open license.