How epigenetic clocks tick: unpacking the black box by deciphering biological pathways and transcriptomic signatures of accelerated aging (paper 20 July 2026)

#1

https://www.nature.com/articles/s41514-026-00446-x

chatGPT(5.5paid):

Paper

Arpawong et al., “How epigenetic clocks tick: unpacking the black box by deciphering biological pathways and transcriptomic signatures of accelerated aging,” npj Aging, 2026. The uploaded version is an accepted, unedited manuscript, so minor details may change during final production.

Summary

Research question

DNA-methylation clocks predict morbidity and mortality, but different clocks often give different answers and appear to predict different aspects of ageing. The authors ask:

  1. Which gene-expression patterns and biological pathways are associated with each major epigenetic clock?
  2. Are these pathways shared across clocks or largely clock-specific?
  3. Can the associated expression patterns be converted into transcriptomic scores that predict health outcomes?

Study design

The study used blood samples from 3,227 participants in the US Health and Retirement Study, with DNA methylation and RNA sequencing measured at approximately the same time. Mean age was about 70 years, and 58% of participants were women.

Five clocks were examined:

  • Horvath
  • Hannum
  • PhenoAge
  • GrimAge
  • DunedinPACE

For each clock, the authors calculated age acceleration—the component of the clock value not explained by chronological age. They then:

  • split the cohort into an 80% discovery set and a 20% hold-out test set;
  • identified genes whose expression was associated with each clock;
  • used Reactome and Gene Ontology enrichment to infer biological functions;
  • built a Transcriptomic Aging Gene Score, or TAGS, for each clock;
  • tested TAGS against morbidity, physical function, cytokines, telomere length and mortality;
  • examined transferability in three external datasets.

The workflow is clearly depicted in Figure 1 on page 6.

Main findings

1. The clocks have markedly different transcriptomic signatures

The numbers of associated genes differed enormously:

Clock Differentially expressed genes
Horvath 49
Hannum 142
PhenoAge 455
GrimAge 676
DunedinPACE 3,204

There was no gene associated with all five clocks. DunedinPACE was particularly distinctive: 2,419 of its 3,204 associated genes were unique to it.

However, much of the PhenoAge and GrimAge signal was contained within the broader DunedinPACE signature: 86% of PhenoAge genes and 81% of GrimAge genes overlapped with DunedinPACE.

This supports the interpretation that the clocks are not interchangeable measures of one unitary ageing process. Rather, they sample different, partly overlapping molecular states.

2. Reactome pathways were mostly clock-specific

At the more narrowly defined Reactome level, the clocks differed dramatically:

  • Horvath: 2 enriched pathways
  • GrimAge: 6
  • DunedinPACE: 89

No Reactome pathway was enriched across all five clocks, although 14 occurred in at least two.

The broad signatures were:

  • Horvath: metabolism and signal transduction
  • Hannum: homeostasis and vascular-wall interactions
  • GrimAge: interferon and immune signalling
  • PhenoAge: cellular senescence, mitosis and extracellular-matrix activity, with reduced transcriptional and viral-response pathways
  • DunedinPACE: a very broad signature including protein metabolism, immunity, respiration and nervous-system development, together with reduced DNA repair, extracellular-matrix organisation and some neuronal pathways

Several immune pathways—particularly innate immunity and neutrophil degranulation—were shared by Hannum, PhenoAge, GrimAge and DunedinPACE. Figure 3 on page 9 shows how concentrated the Horvath and GrimAge profiles are compared with the much broader DunedinPACE profile.

3. Broader Gene Ontology categories showed much more convergence

The apparent divergence depended heavily on the level of biological description.

Although Reactome pathways were highly clock-specific, Gene Ontology analysis found 1,453 enriched biological-process terms, of which 239 were common to all five clocks. These could be grouped into four broad themes:

  1. metabolism and macromolecular processing;
  2. development and tissue formation;
  3. immune-system activity;
  4. regulatory, inflammatory and signalling processes.

Thus, the clocks appear different at the detailed pathway level but converge on broad ageing-related systems.

4. Direct links between clock CpGs and nearby gene expression were uncommon

The authors also examined whether methylation at clock CpGs correlated with expression of nearby genes. Only 0–7% of proximal CpG–gene pairs showed correlations of at least (|r|>0.25), and even fewer were also differentially expressed with the corresponding clock.

Examples included:

  • CD248 for Hannum;
  • AIM2, MKRN3 and ZNF154 for PhenoAge;
  • AHRR, TAGLN and CD93 for GrimAge;
  • ABCG1 for DunedinPACE.

This suggests that a clock’s transcriptomic associations generally cannot be explained by a simple model in which its constituent CpGs directly regulate the nearest gene.

5. TAGS captured only part of the methylation-clock signal

Correlations between each methylation clock and its corresponding TAGS were moderate:

Clock Correlation with corresponding TAGS
Horvath 0.25
Hannum 0.27
PhenoAge 0.33
GrimAge 0.45
DunedinPACE 0.50

This indicates that TAGS are related to, but not merely RNA-based copies of, the methylation clocks.

The TAGS themselves were often much more highly correlated with one another—up to 0.86—than the original methylation clocks were. This could mean that gene expression represents a more convergent downstream physiological state, while DNA methylation retains more clock-specific information.

6. TAGS often had stronger health associations

In the hold-out sample, TAGS generally showed stronger associations than their parent DNA-methylation clocks with:

  • mortality;
  • frailty;
  • activities-of-daily-living limitations;
  • walking speed;
  • diabetes;
  • heart and lung disease;
  • BMI;
  • telomere length;
  • IL-6 and IL-10.

For example, the mortality hazard ratios per standard-deviation increase were:

  • Horvath DNAm: 1.08 versus Horvath TAGS: 1.67
  • Hannum DNAm: 1.02 versus Hannum TAGS: 1.92
  • PhenoAge DNAm: 1.22 versus PhenoAge TAGS: 1.54

For GrimAge and DunedinPACE, both methylation and expression measures retained some independent predictive information. In joint models, methylation components tended to contribute more to mortality and cognitive prediction, whereas TAGS tended to contribute more to metabolic disease and inflammatory markers.

External datasets produced broadly comparable clock–TAGS correlations, although with substantial variation.

Novelty

1. Simultaneous comparison of five major clocks against contemporaneous RNA sequencing

The strongest novelty is the systematic comparison of five widely used clocks in the same relatively large, population-based cohort with methylation and gene expression measured concurrently. Earlier studies often examined one clock, one tissue or a selected exposure group.

This design makes differences between clocks more interpretable because they are not primarily attributable to different cohorts, laboratory platforms or sampling times.

2. Moving from individual CpGs to downstream whole-transcriptome signatures

Rather than assuming that a clock CpG regulates its nearest gene, the authors associate the complete age-acceleration measure with genome-wide expression. That is valuable because clock values are multivariate constructs, and their biological meaning may emerge from distributed regulatory networks rather than direct local CpG effects.

3. Demonstration of scale-dependent convergence

One particularly interesting conceptual result is that the clocks look:

  • highly distinct at the gene and Reactome-pathway levels;
  • substantially more similar at the broad Gene Ontology level.

This reconciles two apparently conflicting views: clocks can represent distinct molecular mechanisms while still reflecting common high-level features of ageing.

4. Creation of clock-specific TAGS

The authors do not stop at pathway interpretation. They convert the associated genes into testable transcriptomic scores and evaluate them in a hold-out sample and external datasets.

TAGS are therefore both:

  • proposed functional readouts of the methylation clocks; and
  • potentially useful biomarkers in their own right.

5. Evidence that methylation and expression contain complementary information

For GrimAge and DunedinPACE in particular, methylation and TAGS both retained independent associations with some outcomes. This supports a multilayer biomarker model in which methylation captures relatively stable regulatory history while RNA captures the current functional state.

Critique

1. The study identifies correlates, not mechanisms

The title promises to explain “how epigenetic clocks tick,” but the analysis is cross-sectional and observational. It cannot establish whether:

  • methylation causes the expression changes;
  • expression influences methylation;
  • both are consequences of inflammation, disease, smoking or another upstream process.

Indeed, the weak correspondence between clock CpGs and nearby expression argues against a straightforward direct-regulation explanation.

The paper improves biological interpretation, but it does not really open the clock’s causal mechanism. “Transcriptomic correlates of clock acceleration” would be a more precise description.

2. Whole-blood composition is a major interpretive problem

Many of the strongest pathways are immune-related, including neutrophil degranulation and interferon signalling. These may reflect:

  • altered proportions of neutrophils, lymphocytes and monocytes;
  • changes in activation state within a cell type;
  • or both.

The analyses adjust for estimated cell composition, but computational adjustment cannot completely separate composition from intracellular regulation. Residual confounding is especially likely because cell estimates are themselves imperfect and broad leukocyte categories hide many biologically distinct subtypes.

Single-cell RNA sequencing, sorted-cell analyses or methylation and RNA measured within the same purified cell populations would be needed to show that these are genuinely intracellular ageing programs.

3. DunedinPACE’s vast signature may partly reflect statistical structure rather than biological breadth

DunedinPACE was associated with 3,204 genes and 89 Reactome pathways, compared with only 49 genes and two pathways for Horvath.

The paper interprets this partly as biological breadth, but several statistical factors could produce the difference:

  • different reliability and variance of the clocks;
  • different relationships with immune-cell composition;
  • different signal-to-noise ratios;
  • DunedinPACE’s training on multisystem physiological decline;
  • differences in scale or residual distribution.

The number of detected genes is therefore not a clean measure of how biologically comprehensive a clock is. A comparison based on matched numbers of top-ranked genes, effect-size distributions or down-sampling would help distinguish biological breadth from power.

4. The pathway analysis uses inconsistent-looking significance structures

Reactome enrichment used an FDR threshold of 0.25, whereas the initial gene-expression analysis used FDR <0.01. An FDR of 0.25 is conventional in some exploratory GSEA contexts, but it permits a relatively high expected proportion of false discoveries.

This matters because the central claim of clock-specific pathways depends on which pathways cross that threshold. Sensitivity analyses using:

  • FDR <0.05;
  • continuous ranked-list enrichment;
  • matched gene-set sizes;
  • and alternative pathway databases

would indicate how robust the clock differences are.

5. Comparing counts of significant pathways can be misleading

Horvath’s two pathways and DunedinPACE’s 89 pathways are partly a consequence of having 49 versus 3,204 associated genes. The authors acknowledge this, but still give considerable interpretive weight to pathway counts.

A clock with more significant genes has much greater enrichment power. Consequently, “no pathway detected” does not mean “the clock does not represent that pathway.”

The comparison would be stronger if the authors repeatedly sampled equal numbers of genes from each clock or compared full ranked enrichment statistics rather than thresholded lists.

6. TAGS may predict health well because they are direct measures of current illness

RNA expression is dynamic and strongly affected by:

  • active inflammation;
  • diabetes;
  • smoking;
  • medication;
  • infection;
  • recent stress;
  • and existing organ disease.

Therefore, it is unsurprising that TAGS predict contemporaneous morbidity and cytokine levels better than methylation clocks. They may function as broad blood-health or inflammatory-state scores rather than more accurate measures of the underlying rate of ageing.

This is especially relevant because the TAGS genes were selected based on their association with clocks in the same general population and then tested against outcomes closely linked to the pathways discovered.

Longitudinal evidence that TAGS predict future change beyond baseline disease, biomarkers and conventional blood counts would be more compelling.

7. The test sample is independent for model fitting, but not fully external

The 20% hold-out group comes from the same cohort, sample collection system, RNA-sequencing pipeline and population as the training set. This protects against direct overfitting but does not test robustness to:

  • laboratory differences;
  • demographic differences;
  • batch shifts;
  • or different disease distributions.

The external datasets help, but appear heterogeneous, relatively specialised and not equivalent replications of all the outcome analyses. Their clock–TAGS correlations varied considerably, with some as low as 0.06.

8. TAGS construction appears relatively simple

The scores are additive summaries of genes selected by significance and oriented by their direction of association. Such scores are interpretable, but they may be unstable where thousands of correlated genes are included.

Potential issues include:

  • equal or insufficiently differentiated weighting;
  • redundancy among co-expressed genes;
  • sensitivity to RNA-processing and batch effects;
  • dependence on the discovery cohort’s expression distribution.

Penalised regression, nested cross-validation, pathway-level scores or latent-factor methods might produce smaller and more transportable signatures. These should be compared against the additive TAGS rather than assuming the simple construction is optimal.

9. Multiple outcome testing raises false-positive concerns

Five methylation clocks and five TAGS are tested across numerous phenotypes using several model types. The paper presents confidence intervals but does not appear to impose a strong experiment-wide correction across all clock–outcome comparisons.

The overall pattern is probably not due solely to chance, but individual differences—for example, whether a TAGS “outperforms” a parent clock for a particular outcome—should be interpreted cautiously.

Formal tests comparing coefficients, nested model performance and out-of-sample discrimination would be preferable to comparing significance or visual effect-size magnitude.

10. Effect size is not the same as predictive performance

The paper frequently describes TAGS as stronger predictors because their hazard ratios, odds ratios or standardised coefficients are larger.

However, a larger coefficient per standard deviation does not necessarily mean better individual-level prediction. Prediction should primarily be compared using:

  • cross-validated C-statistics;
  • AUC;
  • calibration;
  • explained variance;
  • net reclassification;
  • and decision-curve analysis.

The authors provide some discrimination metrics in supplementary analyses, but report that models containing both layers often performed similarly to TAGS alone. That result is less supportive of strong methylation–RNA complementarity than the coefficient plots initially suggest.

11. Limited age, tissue and population generalisability

The primary sample consists mainly of older US adults and is predominantly White. Results may not apply to:

  • young adults;
  • developmental ageing;
  • non-US populations;
  • substantially different ancestry groups;
  • or tissues other than blood.

This is particularly important for the multi-tissue Horvath clock. A weak blood transcriptomic signature does not imply that Horvath lacks a strong biological interpretation in other tissues.

12. “Unique pathways” should not be interpreted as separate causes of ageing

Failure to achieve significance in one clock and significance in another does not establish that the pathway is uniquely represented by the latter. Statistical significance is not itself evidence that two effect estimates differ.

A direct heterogeneity test comparing pathway enrichment scores across clocks would be needed. Some of the claimed uniqueness may simply reflect differences in power around similar underlying effects.

Overall assessment

This is a strong, useful and methodologically ambitious observational study. Its most convincing conclusion is that epigenetic clocks are not interchangeable: their age-acceleration measures associate with different detailed blood-expression signatures, while converging on broader metabolic, developmental, immune and regulatory processes.

The TAGS work is also promising and supports the idea that methylation and RNA provide different views of biological state. However, the paper goes too far when it frames these associations as revealing the internal mechanisms by which clocks tick. Whole-blood composition, cross-sectional design, unequal statistical power across clocks and the acute disease sensitivity of RNA expression prevent a mechanistic interpretation.

The best interpretation is:

Epigenetic clocks mark partially distinct upstream histories or regulatory states, while transcriptomic scores capture more convergent and more immediate downstream physiological states.

That model explains why TAGS may correlate more strongly with current disease, while GrimAge and DunedinPACE methylation retain additional information about mortality and cognitive outcomes.