DunedinPACE vs. GrimAge vs. PhenoAge

DunedinPACE, GrimAge and PhenoAge can appear on the same biological-age report, but they do not measure the same thing. DunedinPACE reports a rate. GrimAge reports an age-like estimate trained around mortality-related signals. PhenoAge reports an age-like estimate built from a clinical-risk profile; DNAm PhenoAge estimates that profile from DNA methylation, the chemical marks on DNA that these tests read. Humanic Blue is comparing the questions behind these three names, not treating them as interchangeable scores.

That makes this less like three thermometers and more like a speedometer, a risk model and a laboratory score sharing a dashboard. Each has human cohort evidence. None is a diagnosis, a lifespan forecast or proof that changing the number will help someone live longer. For living longer and better, the useful first question is not “Which age is mine?” It is “What outcome was this model trained to estimate?”

Why do these three clocks report different kinds of result?

Clock	Main output	What shaped the model	Plain-language question
DunedinPACE	Rate near 1.0	Change in 19 organ-system measures over two decades	How quickly did aging-related physiology change in the reference cohort?
GrimAge / GrimAge2	Age-like value and age acceleration	DNA patterns used to estimate signals related to death risk, including smoking exposure	How does this methylation profile relate to mortality and disease patterns in cohorts?
PhenoAge / DNAm PhenoAge	Age-like value and age acceleration	Chronological age plus nine clinical blood measures	What age does this blood-test pattern suggest in a model built around death risk?

The names make them look like neighboring rulers. Their training targets make them different instruments.

What does DunedinPACE measure?

DunedinPACE was developed from the Dunedin Study, a group of about 1,000 New Zealanders born in 1972–1973. Researchers measured change in 19 cardiovascular, metabolic, kidney, liver, immune, dental and lung indicators at ages 26, 32, 38 and 45, then trained a 173-site blood-methylation algorithm to estimate that pace from one sample.

Its scale is centered near 1.0: roughly one year of physiological change per calendar year in the model’s reference framework. A result of 1.10 is therefore a rate estimate, not an age ten percent older than the person.

The development paper linked faster DunedinPACE with illness, disability and earlier death across several cohorts. It also noted that the founding cohort was modest in size, came from one country and required validation in more diverse populations. The algorithm is commercially licensed to TruDiagnostic, and several authors were named inventors on the licensed intellectual property.

GrimAge: mortality-related signals expressed as age

DNAm GrimAge was developed from about 1,700 Framingham Offspring participants. It combines age and sex with DNA-pattern estimates of smoking exposure and seven proteins in the liquid part of blood, then expresses the result in years. “Age acceleration” is the difference between the modeled value and the value expected for chronological age.

The original model was tested in about 7,000 people across five study groups and was associated with later disease and death. GrimAge2 added DNA-pattern estimates related to C-reactive protein, an inflammation marker, and hemoglobin A1c, a marker of longer-term blood sugar. It was evaluated in blood samples from about 10,000 people across nine study groups.

Smoking information is built into both versions, which helps explain their relationship with respiratory outcomes. That design can be useful for prediction, but it does not show that the clock has discovered a single underlying cause of aging.

PhenoAge: first a clinical score, then a methylation clock

PhenoAge began as a clinical model. Researchers used data from about 10,000 US adults to combine chronological age with nine blood measures: albumin, creatinine, glucose, C-reactive protein, lymphocyte percentage, mean cell volume, red-cell distribution width, alkaline phosphatase and white-cell count.

They then trained DNAm PhenoAge in about 450 people to estimate that combined blood-test score from 513 methylation sites. Both versions report years, but one uses ordinary clinical measurements and the other uses DNA patterns to approximate the blood-based score.

The development study associated DNAm PhenoAge acceleration with mortality and health outcomes in several cohorts. It also reported that the original clinical PhenoAge was more predictive of morbidity and mortality than its methylation counterpart for the comparisons performed. A DNA test is therefore not automatically the more informative version just because it arrived in a more futuristic box.

Do the clocks agree? Only partly

In a comparison of about 1,200 adults, the DunedinPACE developers found that people’s scores on the different clocks agreed only partly. Someone with a faster DunedinPACE result did not necessarily have an equally high age-acceleration result on GrimAge or PhenoAge. The clocks were trained to detect different patterns.

A broader 2025 Nature Communications study followed about 19,000 people in Scotland for ten years and compared 14 clocks across 174 incident diseases and mortality. Second- and third-generation clocks outperformed first-generation birthday clocks, but the authors found no clear best clock for analyses across all diseases. GrimAge2 had the largest number of statistically significant links with later disease among the clocks, while different clocks were informative for different outcomes.

Repeatability is a separate issue. A 2022 Nature Aging study found differences of up to nine years when the same sample was measured again using six clocks. Redesigned versions, using a statistical method called principal components to reduce noise, brought most repeat readings within 1.5 years. The researchers disclosed Elysium licensing, data and employment or advisory relationships.

Population fit matters as well. A 2026 eLife study found that many methylation clocks were less accurate in groups of people with mixed genetic ancestry than in White comparison groups. That finding does not invalidate every clock, but it makes training and validation populations part of any responsible interpretation.

How should a comparison account for clocks trained to answer different questions?

What changed in intervention trials?

In the two-year CALERIE trial analysis, healthy adults without obesity were assigned to a calorie-restriction target or unrestricted intake. DunedinPACE slowed by roughly 2%–3%, while PhenoAge and GrimAge did not change significantly. The clock analysis was additional to the trial’s original work, effects were small and no disease or mortality benefit was tested.

The three-year DO-HEALTH analysis included about 780 generally healthy adults age 70 or older. One gram of omega-3 per day modestly slowed PhenoAge, GrimAge2 and DunedinPACE; the original GrimAge did not show the same reported effect. Vitamin D, omega-3 and home exercise had additive effects on PhenoAge. This was also an additional biomarker analysis, and the authors noted the absence of a generally accepted standard for biological age.

The clocks disagreeing is not necessarily an error. They were built to respond to different patterns. What neither trial established is the clinically important bridge: that making a clock younger causes a person to avoid disease or live longer.

What would a change in one clock, but not another, tell researchers about an intervention?