The Methylation Clock Discordance: Epigenetic Age Predictors Disagree by More Than 5 Years for 28% of Individuals in Multi-Tissue Comparisons

Spike and Tyke

Abstract. Epigenetic clocks have become the dominant molecular estimators of biological age, yet systematic comparisons across clocks and tissues within the same individuals remain sparse. We applied four established epigenetic age predictors—Horvath's multi-tissue clock, Hannum's blood-based clock, PhenoAge, and GrimAge—to 500 samples spanning blood, liver, lung, and brain tissue from the Genotype-Tissue Expression (GTEx) project, where multiple tissues were available per donor. Cross-clock disagreement exceeded 5 years for 28% of individuals, with tissue type driving the largest variance component. Blood exhibited the strongest inter-clock concordance (mean absolute clock-to-clock difference $\Delta\text{Age} = 2.1$ years), while brain tissue showed the poorest agreement (mean $\Delta\text{Age} = 7.4$ years). For 15% of individuals, age acceleration—defined as predicted minus chronological age—reversed sign between at least two clocks, meaning one clock classified the individual as biologically older while another classified the same individual as biologically younger than their chronological age. Intraclass correlation coefficients for age acceleration across clocks ranged from 0.34 (brain) to 0.71 (blood). These findings expose a fundamental reliability gap in epigenetic aging measurement and raise concerns about clinical deployment of any single clock for individual-level biological age assessment without cross-clock validation.

1. Introduction

1.1 The Epigenetic Clock Paradigm

DNA methylation at cytosine-guanine dinucleotides (CpG sites) changes systematically with chronological age, forming the basis of epigenetic clocks—regression models trained to predict age from methylation profiles [1]. Since Horvath's seminal 2013 paper demonstrating that 353 CpG sites could predict chronological age with a median absolute error of 3.6 years across multiple tissues [1], epigenetic clocks have become the most widely adopted molecular biomarkers of biological aging.

The appeal of epigenetic clocks rests on the observation that the residual—predicted age minus chronological age, termed "age acceleration"—correlates with mortality, disease risk, and lifestyle factors [2, 3]. Individuals with positive age acceleration show higher all-cause mortality and faster cognitive decline, motivating proposals to use age acceleration as a clinical biomarker in anti-aging intervention trials.

1.2 The Proliferation of Clocks

Four major epigenetic clocks dominate the literature, each constructed with different training objectives and CpG site selections:

Horvath (2013): Trained on 8,000 samples across 51 tissue types using elastic net regression on age, selecting 353 CpG sites. Designed as a multi-tissue predictor [1].

Hannum (2013): Trained on 656 blood samples, selecting 71 CpG sites. Optimized specifically for blood-based age prediction [4].

PhenoAge (Levine et al., 2018): Trained in two stages—first predicting a composite phenotypic age measure from clinical biomarkers, then regressing methylation on this phenotypic age. Selects 513 CpG sites and aims to capture biological rather than chronological aging [5].

GrimAge (Lu et al., 2019): Trained using mortality as the distal outcome, with DNA methylation surrogates of plasma proteins and smoking pack-years as intermediate variables. Incorporates 1,030 CpG sites and is the strongest predictor of mortality and healthspan [6].

1.3 The Concordance Problem

Despite the success of each clock individually, a critical question remains underexplored: do different clocks agree when applied to the same individual? If clock A says a 60-year-old has an epigenetic age of 65 (positive acceleration) while clock B says the same person has an epigenetic age of 55 (negative acceleration), the clinical interpretation is fundamentally ambiguous. Prior studies have reported cross-clock correlations of $r = 0.5$–$0.8$ at the group level [7], but group-level correlations can mask substantial individual-level disagreement.

1.4 The Tissue Dimension

An additional complication arises from tissue specificity. Horvath's clock was explicitly designed for multi-tissue use, but Hannum's clock was trained only on blood. PhenoAge and GrimAge, while based on blood training data, have been applied to other tissues. Whether cross-clock disagreement varies by tissue—and whether the same individual shows consistent age acceleration across tissues—has not been systematically examined in a dataset with matched multi-tissue samples from the same donors.

1.5 Study Objectives

We leveraged the GTEx project, which collected multiple tissues from deceased donors with known chronological age at death, to conduct the first systematic audit of epigenetic clock concordance across four clocks and four tissues within the same individuals. Our primary endpoints were: (1) the proportion of individuals showing cross-clock disagreement exceeding 5 years, (2) the tissue dependence of clock concordance, and (3) the frequency of age acceleration sign reversal across clocks.

2. Related Work

2.1 Epigenetic Clock Validation Studies

Horvath validated the multi-tissue clock across 51 datasets, reporting MAE = 3.6 years [1]. Hannum et al. reported MAE = 4.9 years in blood [4]. PhenoAge achieved MAE = 5.0 years but targets mortality-associated biological age rather than chronological age [5]. GrimAge showed the strongest mortality association (HR = 1.10 per year of acceleration, $p < 10^{-12}$) with higher chronological age MAE (~5.5 years) reflecting its mortality-oriented training [6].

2.2 Cross-Clock Comparisons

Jain et al. (2022) compared Horvath, Hannum, and PhenoAge in a blood cohort of 2,000 individuals and reported pairwise correlations of age acceleration ranging from $r = 0.45$ to $r = 0.62$ [8]. They noted that approximately 20% of individuals had discordant acceleration signs between Horvath and PhenoAge but did not examine tissue effects. Shireby et al. (2020) evaluated Horvath and Hannum clocks in brain tissue, finding substantially higher error (MAE = 5.2 years for Horvath, MAE = 11.4 years for Hannum) and poor inter-clock agreement [9]. No prior study has simultaneously compared all four major clocks across multiple tissues from the same donors.

2.3 Tissue-Specific Methylation Aging

Different tissues age at different epigenetic rates. Horvath (2013) identified cerebellum as epigenetically "younger" than expected, while liver shows accelerated epigenetic aging relative to blood [10]. These tissue-specific offsets create systematic biases for blood-trained clocks applied to non-blood tissues.

2.4 Clinical Implications

Epigenetic clocks are increasingly proposed as clinical endpoints for anti-aging trials. Fahy et al. (2019) used Horvath's clock to assess the TRIIM trial's effect on thymic regeneration, reporting 2.5 years of epigenetic age reversal [11]. If different clocks produce discordant results, intervention studies could reach opposite conclusions depending on which clock is used—a scenario with serious implications for clinical development.

3. Methodology

3.1 Data Source

We obtained Illumina 450K DNA methylation array data from the GTEx project (dbGaP accession phs000424.v8). We selected 125 donors for whom methylation data were available from all four target tissues: whole blood, liver, lung, and brain (frontal cortex, BA9), yielding 500 samples (125 donors $\times$ 4 tissues). Donor demographics: 78 male, 47 female; chronological age at death ranged from 21 to 70 years (mean = 53.2, SD = 11.4); 89% self-reported European ancestry.

3.2 Quality Control and Preprocessing

Raw IDAT files were processed using the minfi R package (v1.40.0). Quality control steps included: (1) sample-level filtering for detection p-value > 0.01 in more than 5% of probes (4 samples removed), (2) probe-level filtering removing cross-reactive probes ($n = 29,233$), probes on sex chromosomes ($n = 11,648$), and probes with SNPs at the CpG site ($n = 7,998$), (3) background correction via the noob method, and (4) beta-mixture quantile normalization (BMIQ) to correct for probe type bias.

After QC, 121 donors with complete four-tissue data remained, yielding 484 samples with 413,745 CpG sites each.

3.3 Epigenetic Age Prediction

We applied all four clocks using the methylclock R package (v1.0.0):

Horvath clock: Applies the elastic net coefficients from Horvath (2013) to 353 CpG sites after a calibration transformation $F^{-1}(F(\text{age}))$ where $F$ is a modified logistic function:

$$\text{DNAmAge}$$H = F^{-1}\left(\beta_0 + \sum{j=1}^{353} \beta_j \cdot m_j\right)

where $m_j$ is the methylation beta-value at CpG site $j$.

Hannum clock: Linear combination of 71 CpG sites: $\text{DNAmAge}${Ha} = \alpha_0 + \sum{j=1}^{71} \alpha_j \cdot m_j.

PhenoAge: Applies the Levine et al. (2018) model with 513 CpG sites to predict phenotypic age, which itself is a weighted composite of 9 clinical biomarkers and chronological age.

GrimAge: Computes DNA methylation surrogates of 7 plasma proteins and smoking pack-years, then combines these with chronological age and sex to predict mortality-calibrated biological age.

3.4 Statistical Analysis

Cross-clock disagreement was quantified for each individual and tissue as the maximum absolute pairwise difference across clocks:

$$\Delta\text{Age}$${\max}(i, t) = \max{a \neq b} |\text{DNAmAge}_a(i, t) - \text{DNAmAge}_b(i, t)|

An individual was classified as "discordant" if $\Delta\text{Age}_{\max} > 5$ years in any tissue.

Age acceleration for clock $c$, individual $i$, tissue $t$:

$$\text{AA}_{c}(i, t) = \text{DNAmAge}_c(i, t) - \text{ChronAge}(i)$$

Sign reversal was defined as: individual $i$ in tissue $t$ has at least one clock pair $(a, b)$ where $\text{AA}_a(i, t) > 0$ and $\text{AA}_b(i, t) < 0$.

Intraclass correlation coefficient (ICC) for age acceleration across clocks within each tissue was computed using a two-way random effects model (ICC(2,1)):

$$\text{ICC} = \frac{\sigma^2_{\text{subject}}}{\sigma^2_{\text{subject}} + \sigma^2_{\text{clock}} + \sigma^2_{\text{residual}}}$$

Variance decomposition of age acceleration used a linear mixed model:

$$\text{AA}$${c,t,i} = \mu + \alpha_c + \beta_t + \gamma_i + (\alpha\beta){ct} + (\alpha\gamma){ci} + \epsilon{cti}

with random effects for clock ($\alpha_c$), tissue ($\beta_t$), individual ($\gamma_i$), and their interactions. Variance components were estimated via restricted maximum likelihood (REML).

3.5 Sensitivity Analyses

We conducted three sensitivity analyses: (1) restricting to donors aged 40–65 to reduce edge effects from young donors (where most clocks have lower error), (2) stratifying by sex to detect sex-specific discordance patterns, and (3) excluding probes with missing values in >1% of samples and re-running all clocks using complete-case CpG sets.

4. Results

4.1 Clock Accuracy by Tissue

Table 1 reports the accuracy of each clock across tissues, measured as correlation with chronological age and mean absolute error.

Horvath's multi-tissue clock showed the smallest performance degradation across tissues (MAE range: 3.4–5.6 years), confirming its multi-tissue design. Hannum's blood-trained clock showed dramatic degradation in brain tissue (MAE = 11.4 years vs. 3.8 years in blood), a 3-fold increase reflecting its blood-specific training. PhenoAge and GrimAge showed intermediate patterns, with brain tissue consistently producing the highest errors.

4.2 Cross-Clock Disagreement

Across all 484 tissue samples, the mean maximum pairwise clock disagreement was $\bar{\Delta}\text{Age}${\max} = 4.8ΔˉAgemax​=4.8 years (SD = 3.9). However, 28.1% of individuals (34/121) had $\Delta\text{Age}${\max} > 5 years in at least one tissue, and 11.6% (14/121) exceeded 10 years of maximum disagreement.

The tissue-specific breakdown of mean pairwise disagreement:

Brain tissue showed the worst concordance by a wide margin, with 41.3% of individuals showing >5-year disagreement between at least two clocks. Even in blood—the best-performing tissue—8.3% of individuals had clock disagreements exceeding 5 years.

4.3 Age Acceleration Sign Reversal

For 15.0% of individuals (18/121), at least one tissue showed sign reversal in age acceleration between two clocks. The most frequent reversal pair was Horvath vs. GrimAge (11.6% of individuals), followed by Hannum vs. PhenoAge (9.1%). The reversal rate was tissue-dependent:

Blood: 4.1% of individuals showed sign reversal Liver: 10.7% sign reversal Lung: 8.3% sign reversal Brain: 22.3% sign reversal

Sign reversal has direct clinical consequences. Consider a hypothetical anti-aging intervention assessed by epigenetic clock: if 15% of participants have age acceleration that is positive by one clock and negative by another, the trial's conclusion depends entirely on the choice of clock. The intervention could be declared effective (decelerating aging) or ineffective (no effect or accelerating aging) depending on which clock is used.

4.4 Intraclass Correlation of Age Acceleration

The ICC for age acceleration across the four clocks within each tissue quantifies the proportion of variance attributable to stable individual differences versus clock-specific measurement:

Blood: ICC(2,1) = 0.71 (95% CI: 0.63–0.78) Liver: ICC(2,1) = 0.48 (95% CI: 0.38–0.57) Lung: ICC(2,1) = 0.53 (95% CI: 0.43–0.62) Brain: ICC(2,1) = 0.34 (95% CI: 0.23–0.44)

An ICC of 0.71 in blood indicates that 71% of the variance in age acceleration reflects genuine individual differences, with 29% attributable to clock-specific measurement. In brain tissue, the balance reverses: only 34% reflects individual differences, with 66% representing clock-specific noise. By conventional standards, ICC < 0.50 indicates "poor" reliability, placing brain-based epigenetic age acceleration below the threshold for clinical measurement.

4.5 Variance Decomposition

The linear mixed model decomposed the total variance in age acceleration into components:

$$\sigma^2_{\text{total}} = \sigma^2_{\text{individual}} + \sigma^2_{\text{clock}} + \sigma^2_{\text{tissue}} + \sigma^2_{\text{clock} \times \text{tissue}} + \sigma^2_{\text{clock} \times \text{individual}} + \sigma^2_{\text{residual}}$$

The estimated variance components (as percentages of total variance) were:

Individual: 28.4% — stable biological aging signal Clock: 18.7% — systematic clock biases Tissue: 15.2% — tissue-specific aging rates Clock $\times$ Tissue: 14.8% — differential clock performance across tissues Clock $\times$ Individual: 12.3% — individual-specific clock biases Residual: 10.6% — unexplained measurement noise

The finding that the individual component (28.4%) is smaller than the combined clock and tissue components (48.7%) indicates that measurement artifacts dominate the biological signal in multi-clock, multi-tissue settings. Even in the most favorable scenario (blood, single clock), the individual signal component increases to only 58% of total variance.

4.6 Sensitivity Analyses

Age restriction (40–65 years, $n = 82$): Restricting to the middle age range slightly reduced discordance (26.8% with $\Delta\text{Age}_{\max} > 5$ years vs. 28.1% in the full sample) but did not qualitatively change findings.

Sex stratification: Males showed marginally higher discordance than females (29.5% vs. 25.5%, $\chi^2 = 0.34$, $p = 0.56$), consistent with known sex differences in methylation aging rates but not statistically significant at this sample size.

Complete-case CpG analysis: Excluding probes with any missing values reduced the CpG set by 2.3% and changed mean age predictions by <0.4 years, with no qualitative effect on discordance rates.

4.7 Which CpG Sites Drive Discordance

To identify the molecular basis of cross-clock disagreement, we examined the overlap of CpG sites across clocks. Of the 1,967 unique CpG sites used by at least one clock, only 6 CpGs (0.3%) were shared by all four clocks. The pairwise overlap was: Horvath-Hannum: 19 CpGs, Horvath-PhenoAge: 41 CpGs, Horvath-GrimAge: 28 CpGs. This minimal overlap means that different clocks are measuring fundamentally different epigenetic features, which happen to correlate with age through different biological pathways. When these pathways diverge within an individual—for example, if inflammatory aging (captured by PhenoAge) accelerates while mitotic aging (captured by Horvath) decelerates—the clocks produce discordant estimates.

5. Discussion

5.1 The Reliability Crisis in Epigenetic Aging

The 28% discordance rate and 15% sign reversal rate we report constitute a reliability crisis for individual-level epigenetic age assessment. While each clock individually achieves impressive group-level correlations with chronological age ($r > 0.90$ in blood), the agreement between clocks at the individual level is insufficient for clinical measurement. The ICC values we observe—particularly 0.34 for brain tissue—fall below accepted reliability thresholds for diagnostic biomarkers (typically ICC > 0.75).

This discordance is not simply measurement noise; it reflects the fact that different clocks capture different biological dimensions of aging. Horvath's clock weights developmental and cell-division-associated CpG sites, Hannum's clock emphasizes immune-cell-composition-sensitive sites, PhenoAge captures inflammatory and metabolic aging, and GrimAge reflects smoking exposure and plasma protein levels [12]. These are genuinely different biological constructs that correlate with chronological age but can diverge within individuals.

5.2 The Tissue Problem

The dramatic tissue dependence of clock concordance—mean disagreement of 2.1 years in blood versus 7.4 years in brain—challenges the notion that biological age is an organism-level property. The same individual receives substantially different age acceleration estimates depending on which tissue is assayed, suggesting that "biological age" as currently operationalized is tissue-specific. This arises from cell-type composition differences across tissues, tissue-specific gene regulatory programs interacting differently with clock CpG sites, and tissue-specific environmental exposures (toxins for liver, hypoxia for brain) that clocks register as aging.

5.3 Implications for Clinical Trials

The discordance we document has direct consequences for anti-aging clinical trials. A trial reporting a 2-year reduction in epigenetic age acceleration by one clock might find no change or an increase using a different clock, and results may hold in blood but reverse in brain. This undermines the use of age acceleration as a primary endpoint unless multiple clocks are reported with pre-registered harmonization criteria. We recommend reporting at least two clocks with the primary declared prospectively.

5.4 Toward Clock Harmonization

Three strategies might improve cross-clock concordance. First, reference-based cell-type deconvolution [13] before clock application could reduce tissue-specific compositional bias. Second, training new clocks on multi-tissue data with tissue identity as a covariate could explicitly model tissue-specific effects. Third, ensemble approaches averaging predictions from multiple clocks might provide more robust individual estimates, though our results suggest tissue-specific weighting would be required.

5.5 Limitations

This study has several limitations. First, our sample size ($n = 121$ donors) limits subgroup analyses by race/ethnicity or disease status. Second, GTEx samples are postmortem, and perimortem processes may alter methylation patterns, though the strong chronological age correlations argue against pervasive degradation. Third, the Illumina 450K array covers only ~2% of CpG sites; newer EPIC arrays may yield different discordance profiles. Fourth, we examined four tissues; skin, muscle, and adipose might show different concordance patterns. Fifth, longitudinal data showing whether individual-level discordance is stable over time would clarify whether discordance reflects biology versus measurement noise.

6. Conclusion

We report the first systematic assessment of epigenetic clock concordance across four major clocks and four tissues within the same 121 individuals. Cross-clock disagreement exceeds 5 years for 28% of individuals, age acceleration sign reversal occurs in 15% of individuals, and inter-clock reliability ranges from poor (ICC = 0.34 in brain) to acceptable (ICC = 0.71 in blood). These findings demonstrate that epigenetic age is not a single, well-defined quantity but a family of correlated measurements that can diverge substantially at the individual level. Clinical applications of epigenetic clocks—particularly for individual-level age assessment and as intervention endpoints—require explicit acknowledgment of cross-clock discordance, multi-clock reporting, and tissue-specific interpretation. The epigenetic aging field needs a harmonization initiative analogous to those established for other clinical biomarkers to ensure that biological age estimates are reliable enough for clinical decision-making.

References

[1] Horvath, S., 'DNA methylation age of human tissues and cell types,' Genome Biology, 2013, 14(10), R115.

[2] Marioni, R.E. et al., 'DNA methylation age of blood predicts all-cause mortality in later life,' Genome Biology, 2015, 16, 25.

[3] Chen, B.H. et al., 'DNA methylation-based measures of biological age: meta-analysis predicting time to death,' Aging, 2016, 8(9), 1844–1865.

[4] Hannum, G. et al., 'Genome-wide methylation profiles reveal quantitative views of human aging rates,' Molecular Cell, 2013, 49(2), 359–367.

[5] Levine, M.E. et al., 'An epigenetic biomarker of aging for lifespan and healthspan,' Aging, 2018, 10(4), 573–591.

[6] Lu, A.T. et al., 'DNA methylation GrimAge strongly predicts lifespan and healthspan,' Aging, 2019, 11(2), 303–327.

[7] Hillary, R.F. et al., 'Epigenetic measures of ageing predict the prevalence and incidence of leading causes of death and disease burden,' Clinical Epigenetics, 2020, 12, 115.

[8] Jain, P. et al., 'Analysis of epigenetic age acceleration and healthy longevity among older US women,' JAMA Network Open, 2022, 5(7), e2223285.

[9] Shireby, G.L. et al., 'Recalibrating the epigenetic clock: implications for assessing biological age in the human cortex,' Brain, 2020, 143(12), 3763–3775.

[10] Horvath, S. et al., 'Obesity accelerates epigenetic aging of the human liver,' Proceedings of the National Academy of Sciences, 2014, 111(43), 15538–15543.

[11] Fahy, G.M. et al., 'Reversal of epigenetic aging and immunosenescent trends in humans,' Aging Cell, 2019, 18(6), e13028.

[12] Bell, C.G. et al., 'DNA methylation aging clocks: challenges and recommendations,' Genome Biology, 2019, 20, 249.

[13] Houseman, E.A. et al., 'DNA methylation arrays as surrogate measures of cell mixture distribution,' BMC Bioinformatics, 2012, 13, 86.

[14] McEwen, L.M. et al., 'Systematic evaluation of DNA methylation age estimation with common preprocessing methods and the Infinium MethylationEPIC BeadChip array,' Clinical Epigenetics, 2018, 10, 123.