https://onlinelibrary.wiley.com/doi/epdf/10.1111/acel.70626
A very easily identifiable difficulty of DNA methylation clocks is that they track methylation in most cases in white blood cells and it is the function of other tissue cells that matters. ie they are an expensive way of getting information that is not that useful.
chatGPT(5.5paid):
Overall assessment
This is a useful, tightly focused short communication asking a clinically important question: do DNA-methylation epigenetic clocks improve prediction of future disease beyond inexpensive conventional risk factors? In this Finnish middle-aged cohort, the answer was no. A model using age, sex, smoking, alcohol consumption, BMI and waist-to-hip ratio discriminated future non-communicable disease better than models containing an epigenetic clock.
The paper is persuasive as a caution against marketing epigenetic-age tests as individual disease-risk predictors. It is much less decisive about whether epigenetic clocks are valid measures of biological ageing, useful mechanistic biomarkers, or useful in older and more heterogeneous populations.
1. Summary
Research question
The authors tested whether commonly used epigenetic clocks:
- predict incident age-related disease;
- remain predictive after adjustment for simple conventional risk factors; and
- outperform a model based only on those conventional factors.
The central distinction is between:
- showing that a clock is associated with future disease; and
- showing that measuring the clock adds useful predictive information beyond data already available cheaply.
That distinction is the paper’s main conceptual contribution.
Cohort and outcome
The study used 1,108 participants from the Young Finns Study. Participants were:
- 34–49 years old at baseline;
- free of the diseases included in the outcome at baseline;
- followed for approximately 7–9 years.
During follow-up, 222 participants, or 20%, developed at least one included non-communicable disease or condition. The composite included cardiometabolic diseases, hypertension, cancer and steatotic liver disease, with the detailed definitions relegated to the supporting tables.
Epigenetic measures
The authors evaluated:
- Horvath age acceleration;
- Hannum age acceleration;
- PhenoAge acceleration;
- GrimAge acceleration;
- principal-component versions of several clocks;
- DunedinPACE.
The regression analyses used age-deviation or pace-of-ageing measures rather than simply comparing estimated epigenetic age with disease.
Main results
In models adjusted only for age, sex and methylation-array version, three measures significantly predicted incident disease:
| Clock | Odds ratio | 95% CI | p-value |
|---|---|---|---|
| PCPhenoAge acceleration | 1.18 | 1.01–1.37 | 0.032 |
| GrimAge acceleration | 1.20 | 1.03–1.39 | 0.015 |
| DunedinPACE | 1.19 | 1.02–1.39 | 0.024 |
First-generation clocks such as Horvath and Hannum were not significant. This agrees with the general pattern that second- and third-generation clocks, trained partly on health-related phenotypes or mortality, tend to predict outcomes better than clocks trained primarily to estimate chronological age.
After adjustment for smoking, alcohol use, BMI and waist-to-hip ratio, none of the clocks remained statistically significant. The attenuation appeared to be driven mainly by the anthropometric variables rather than smoking and alcohol alone.
The traditional risk-factor model had an AUC of:
- 0.649 with 95% CI 0.608–0.690.
The three leading epigenetic models had AUCs of approximately:
- PCPhenoAge: 0.602;
- GrimAge: 0.608;
- DunedinPACE: 0.609.
The ROC chart on page 3 shows that all models performed only modestly, although the traditional-factor curve was consistently superior over much of the range. The AUC differences were reported as statistically significant by DeLong tests.
Authors’ interpretation
The authors conclude that epigenetic clocks and conventional disease-risk factors probably capture substantially overlapping information. Because methylation testing is more expensive and technically demanding, clocks need to demonstrate appreciable incremental value before clinical or consumer use for disease prediction can be justified.
They nevertheless acknowledge that epigenetic clocks may remain valuable for:
- investigating biological ageing;
- studying ageing mechanisms;
- assessing interventions;
- populations in which conventional risk factors are less informative.
2. What is novel?
A. It asks about incremental utility, not merely association
The strongest novelty is not the observation that epigenetic age correlates with morbidity. That has been shown repeatedly. The paper instead asks the more demanding translational question:
Does the clock provide enough additional information to justify measuring it?
Many biomarker papers report that a new marker remains statistically significant after adjustment. That does not necessarily mean it materially improves prediction. This paper explicitly contrasts clocks with a simple alternative model.
B. Direct comparison within the same cohort
The clocks and conventional factors were evaluated in the same participants, against the same outcome and follow-up period. This avoids the weak inference involved in comparing AUCs or effect sizes taken from unrelated studies.
C. It includes several generations of clocks
The comparison spans:
- chronological-age-trained clocks;
- phenotype- and mortality-trained clocks;
- principal-component clocks designed to improve technical reliability;
- DunedinPACE, which estimates rate rather than accumulated age.
This helps show that the negative result is not confined to one obsolete clock.
D. Focus on a relatively young, disease-free population
Many clock studies concern older populations with substantial prevalent illness. Here, participants were middle-aged and free of the studied diseases at baseline. That is relevant to consumer testing, because relatively healthy middle-aged people may be among those most likely to buy biological-age tests.
E. Practical emphasis on cost and clinical justification
The paper explicitly connects statistical performance with implementation. A molecular biomarker need not merely work; it should add enough value to compensate for cost, laboratory complexity and uncertainty.
How strong is the novelty?
The novelty is moderate rather than fundamental. Previous studies had already shown attenuation of clock associations after conventional adjustment, and some had found complete loss of significance. The new element is the explicit, compact head-to-head framing and the emphasis on comparative discrimination.
3. Critique
Strengths
1. The question is clinically well formulated
The paper correctly distinguishes:
- biological association;
- independent association;
- discrimination;
- added predictive value.
This is often neglected in ageing-biomarker research.
2. Prospective design
Risk factors and methylation were measured before disease incidence. This is stronger than a cross-sectional comparison between epigenetic age and existing disease.
3. Participants were disease-free at baseline
Excluding prevalent disease reduces the possibility that already diagnosed disease simply altered methylation or body composition.
4. Multiple clocks were compared
The broadly similar result across several measures makes it less likely that the conclusion is an artefact of selecting one poorly performing clock.
5. The claims are generally restrained
The authors do not conclude that clocks are biologically meaningless. They explicitly separate mechanistic ageing research from individual disease prediction.
Important limitations
1. The composite outcome is highly heterogeneous
“Any non-communicable disease” combines conditions with very different biology, risk factors and incidence patterns, including:
- hypertension;
- metabolic and cardiovascular conditions;
- cancer;
- steatotic liver disease.
A biomarker may predict one disease well but another poorly. Combining them can dilute disease-specific effects and make interpretation difficult.
It also means that the traditional model has a built-in advantage: BMI and waist-to-hip ratio are especially direct predictors of hypertension, diabetes and fatty liver, likely important components of the endpoint.
The paper would be more informative with:
- disease-specific analyses;
- cardiometabolic and non-cardiometabolic composites;
- time-to-first-event analyses;
- competing-risk analyses where relevant.
The modest number of events may have prevented these analyses, but that limitation should constrain the generality of the title.
2. The comparison is not fully symmetrical
The traditional model contains six predictors:
- age;
- sex;
- smoking;
- alcohol;
- BMI;
- waist-to-hip ratio.
Each clock model appears to contain:
- age;
- sex;
- array version;
- one clock.
Thus, the study compares a multivariable conventional model with a model containing essentially one molecular summary measure plus basic adjustment. It is unsurprising that several clinically relevant variables collectively outperform one biomarker.
A more clinically relevant comparison would be:
- conventional model alone;
- conventional model plus one clock;
- comparison of performance between those two nested models.
The authors partly address this by fitting fully adjusted clock models, but they do not present a comprehensive incremental-prediction analysis because the clocks were no longer statistically significant. That leaves the key “added value” question only partly answered.
3. Lack of statistical significance does not prove zero incremental value
The authors state that formal nested comparisons were not feasible or justified because none of the clocks remained significant in fully adjusted models. This is debatable.
The relevant clinical question is whether adding a clock improves prediction, not whether its individual coefficient crosses p < 0.05. One can still estimate:
- change in AUC;
- likelihood-ratio improvement;
- Brier score;
- calibration slope and intercept;
- decision-curve net benefit;
- cross-validated log loss;
- net reclassification improvement, used cautiously;
- confidence intervals around incremental performance.
A clock could conceivably produce a small predictive improvement despite an individually non-significant coefficient, especially with limited power. Conversely, a statistically significant coefficient can add almost no useful prediction.
4. The study may be underpowered for small incremental effects
There were 222 events. That is adequate for a compact model, but incremental biomarker effects are often small and difficult to detect.
The confidence intervals around the clock odds ratios are compatible with effects ranging from negligible to moderately important. The null adjusted result therefore should be interpreted as:
no convincing evidence of added value in this cohort,
rather than:
proof that the clocks have no additional value.
5. No external or internal validation is clearly reported
The AUCs appear to be estimated in the same dataset used to fit the models. Without:
- cross-validation;
- bootstrap optimism correction;
- a held-out validation sample;
- external validation,
performance estimates may be optimistic.
This affects both model types, but the amount of optimism may differ depending on model construction and variable handling.
6. Discrimination is weak for all models
An AUC of 0.649 is only modest. The paper’s most robust conclusion is therefore not that conventional risk factors predict disease particularly well, but that they performed less poorly than the tested clocks.
The title “outperform” is statistically accurate, but it could be misread as implying strong clinical prediction. The authors appropriately acknowledge this in the body.
7. Calibration and clinical usefulness are not assessed
AUC measures ranking, not whether predicted absolute risks are correct or useful.
A model can have:
- acceptable AUC but poor calibration;
- a small AUC improvement but worthwhile clinical net benefit;
- statistically different AUCs with no practical consequence.
The paper does not report calibration plots, Brier scores or decision-curve analysis. Therefore, its conclusions about clinical usefulness are incomplete.
8. BMI and waist-to-hip ratio may be mediators, confounders or components of biological ageing
Adjustment has a conceptual ambiguity.
If accelerated biological ageing causes changes in adiposity, insulin resistance or body-fat distribution, then BMI and WHR may lie partly on the pathway:
[
\text{ageing process} \rightarrow \text{adiposity/metabolic dysfunction} \rightarrow \text{disease}
]
Adjusting for them removes this mediated signal. That is appropriate when asking whether clocks add prediction beyond directly observed anthropometry, but not when asking whether the clocks capture biologically meaningful ageing.
Thus, the attenuation does not establish that clocks are invalid. It may show that much of their clinically relevant signal is already expressed in body composition.
9. GrimAge is partly designed around conventional exposure-related biology
GrimAge incorporates methylation surrogates associated with smoking and plasma proteins linked to morbidity and mortality. Its association with disease is therefore not independent of conventional risk-factor biology by design.
The overlap found here is informative, but it should not be interpreted as surprising evidence against methylation biomarkers. Some clocks were explicitly trained to encode risk-related exposures and physiological deterioration.
10. Limited generalisability
The population was:
- Finnish;
- relatively young;
- restricted to a narrow age range;
- initially free of the included diseases.
Results may differ in:
- adults over 65;
- ethnically diverse cohorts;
- populations with wider social and environmental exposures;
- secondary prevention;
- patients with established disease;
- cohorts followed for several decades;
- unusually healthy individuals with little variation in BMI or smoking.
The authors recognise this limitation, but it remains central.
11. Follow-up may be too short for some ageing outcomes
Seven to nine years is meaningful for hypertension and metabolic disease but relatively short for:
- many cancers;
- dementia;
- frailty;
- mortality;
- chronic organ decline.
Epigenetic clocks might be more useful for long-latency outcomes or long-term mortality than for the particular composite observed here.
12. The endpoint is binary rather than time-to-event
Logistic regression treats a participant diagnosed after one year similarly to one diagnosed after nine years and may not fully use variation in follow-up duration.
A Cox or other survival model would generally be more informative if diagnosis dates were available. It would allow:
- censoring;
- varying follow-up;
- hazard estimation;
- time-dependent assessment of discrimination.
13. The paper depends heavily on supporting information
Several crucial matters are not visible in the main short communication:
- exact disease definitions and frequencies;
- fully adjusted coefficients;
- serial adjustment results;
- precise AUC comparisons;
- missing-data handling;
- methylation preprocessing;
- clock calculation;
- scaling of odds ratios.
Because the uploaded document does not include the supplementary tables themselves, the robustness of these details cannot be fully assessed from the five-page paper alone.
14. No comparison with established clinical risk scores
The conventional model is intentionally simple, but comparison with validated scores would be more clinically meaningful, for example a cardiovascular or diabetes-risk algorithm.
It would also be informative to compare clocks against routine biomarkers such as:
- blood pressure;
- HbA1c or glucose;
- lipids;
- liver enzymes;
- kidney function;
- inflammatory markers.
The present result may underestimate how high a bar a clock would need to clear in real clinical practice.
15. No analysis of whether clocks help in apparently low-risk people
The authors suggest clocks might be more informative when conventional risk factors vary little, but do not test this directly.
Useful subgroup analyses could include:
- normal-BMI participants;
- non-smokers;
- participants without hypertension;
- those with favourable conventional risk profiles;
- discordant cases, such as low conventional risk but accelerated methylation age.
That is arguably where an epigenetic test would need to prove its value.
4. Interpretation of the biological meaning
The study supports the view that epigenetic clocks—particularly GrimAge and DunedinPACE—contain information related to future disease, because they predicted incidence in minimally adjusted models.
But their signal appears substantially shared with adiposity and body-fat distribution. Several interpretations are possible:
- Redundancy: clocks indirectly measure exposure and metabolic damage already apparent from BMI and WHR.
- Mediation: epigenetic ageing contributes to disease partly through adiposity or metabolic dysfunction.
- Reverse direction at baseline: adiposity and related metabolic processes alter blood methylation patterns captured by the clocks.
- Common causes: lifestyle, inflammation and social or environmental exposures influence both methylation and body composition.
- Outcome matching: the composite was dominated by diseases for which anthropometric measures are especially predictive.
The study cannot distinguish among these mechanisms.
5. Bottom-line judgement
What the paper establishes reasonably well
In relatively healthy Finnish adults aged 34–49, followed for 7–9 years, a simple combination of conventional demographic, lifestyle and anthropometric variables predicted a broad composite of incident non-communicable disease better than any single tested epigenetic clock model.
It also shows that the apparent predictive effects of GrimAge, PCPhenoAge and DunedinPACE were substantially attenuated by BMI and waist-to-hip ratio.
What it does not establish
It does not show that:
- epigenetic clocks are not measures of biological ageing;
- clocks cannot predict mortality or particular diseases;
- clocks are useless in older adults;
- clocks cannot monitor responses to interventions;
- clocks add exactly zero information;
- methylation biomarkers could not outperform conventional factors if trained specifically for the relevant disease and population.
Overall appraisal
This is a valuable negative or sobering result, with a well-chosen translational question. Its main message is credible:
Association with disease is not enough to justify a costly biomarker when cheap measurements already capture the same information.
However, the evidence is best regarded as cohort- and endpoint-specific. The heterogeneous composite outcome, limited sample size, narrow age range, modest discrimination, absence of detailed incremental-prediction and validation analyses, and lack of disease-specific outcomes prevent a broad verdict against epigenetic clocks.
The paper is strongest as a critique of consumer-facing claims that a single epigenetic-age number provides clinically superior personal risk information. It is weaker as a critique of epigenetic clocks as mechanistic or intervention-response biomarkers.