MEGA Clock: Ensemble DNAm Age Estimation

• Multi EpiGenetic Age (MEGA) clock is a composite measure that aggregates established DNA methylation clocks into a single, robust epigenetic age estimate.
• It employs covariance weighting, factor analysis, and latent variable modeling to reduce measurement error and improve reliability across tissues and cohorts.
• Extensions of the MEGA framework include uncertainty-aware ensembles, graph-based methods, and multimodal integration to enhance biological-age prediction.

Searching arXiv for the cited MEGA-related papers and closely related work to ground the article. arXiv search query: "Multi EpiGenetic Age MEGA clock"

arXiv search results for "Multi EpiGenetic Age MEGA clock":

• "Aggregating Epigenetic Clocks to Study Human Capital Formation" (Menta et al., 17 Sep 2025)
• "DNA methylation markers to assess biological age" (Podolskiy et al., 2018)
• "U-learning for Prediction Inference via Combinatory Multi-Subsampling: With Applications to LASSO and Neural Networks" (Fei et al., 2024)
• "GraphAge: Unleashing the power of Graph Neural Network to Decode Epigenetic Aging" (Ahmed et al., 2024)
• "Integrating Epigenetic and Phenotypic Features for Biological Age Estimation in Cancer Patients via Multimodal Learning" (Jiang et al., 10 Nov 2025) arXiv search query: "epigenetic clock aggregation multimodal graph biological age"

arXiv search results for "epigenetic clock aggregation multimodal graph biological age":

• "Aggregating Epigenetic Clocks to Study Human Capital Formation" (Menta et al., 17 Sep 2025)
• "GraphAge: Unleashing the power of Graph Neural Network to Decode Epigenetic Aging" (Ahmed et al., 2024)
• "Integrating Epigenetic and Phenotypic Features for Biological Age Estimation in Cancer Patients via Multimodal Learning" (Jiang et al., 10 Nov 2025)
• "DNA methylation markers to assess biological age" (Podolskiy et al., 2018)
• "U-learning for Prediction Inference via Combinatory Multi-Subsampling: With Applications to LASSO and Neural Networks" (Fei et al., 2024) The Multi EpiGenetic Age (MEGA) clock is a composite measure of epigenetic aging that aggregates multiple established DNA methylation (DNAm) clocks into a single metric in years. In its implemented form, MEGA was introduced as an integrated measure combining Horvath, Hannum, PhenoAge, and GrimAge to reduce measurement error, improve estimation efficiency, and enhance robustness in downstream empirical analyses (Menta et al., 17 Sep 2025). In the broader DNAm-clock literature, MEGA also denotes an ensemble strategy motivated by robustness theory for combining pan-tissue, tissue-specific, or platform-specific clocks (Podolskiy et al., 2018). More recent work extends the same aggregation logic toward uncertainty-aware ensembles, graph-based clocks, and multimodal biological-age estimators, although several of those extensions are proposed rather than implemented as a specific MEGA model (Fei et al., 2024).

1. Definition and conceptual basis

MEGA treats existing epigenetic clocks as multiple indicators of a latent aging construct. Under the measurement-error model $X_k = A + \varepsilon_k$ with $E[\varepsilon_k]=0$, aggregation can lower estimation error relative to any single clock. If errors are independent, the variance of a weighted estimator is

$$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$

and inverse-variance weighting is optimal; with correlated errors, generalized least squares-style weighting using the empirical covariance matrix is more efficient (Menta et al., 17 Sep 2025).

This construction is motivated by three related considerations. First, averaging across partially independent, noisy indicators reduces classical measurement error and attenuation bias in regressions. Second, the covariance structure across clocks can be exploited to downweight redundant information and emphasize shared variance attributable to latent epigenetic age. Third, improved statistical power can be obtained without enlarging sample size, which is especially relevant in epigenetic studies where $n$ is often constrained (Menta et al., 17 Sep 2025).

The same logic appears in a more general form in the DNAm-clock robustness literature. "DNA methylation markers to assess biological age" formalizes MEGA as an ensemble integrating outputs from multiple DNAm clocks, including pan-tissue, tissue-specific, and platform-specific predictors, to improve accuracy, robustness, and generalization across tissues, platforms, and cohorts (Podolskiy et al., 2018). This suggests that MEGA is not a single universally fixed algorithm, but a family of ensemble constructions centered on combining epigenetic-age estimators.

2. Mathematical constructions and scaling in years

The implemented MEGA framework in "Aggregating Epigenetic Clocks to Study Human Capital Formation" uses four source clocks computed on Illumina Infinium 450k arrays: Horvath (2013; 353 CpGs; multi-tissue), Hannum et al. (2013; 71 CpGs; adult blood), PhenoAge (Levine et al., 2018; 513 CpGs; integrates clinical phenotypes), and GrimAge (Lu et al., 2019; DNAm surrogates for plasma proteins and smoking; strongest predictor of mortality) (Menta et al., 17 Sep 2025).

Three aggregation methods are implemented.

Construction	Formula	Core idea
MEGAWGT	$$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$	GLS-style covariance-weighted index
MEGAFA	$$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$	One-factor score using covariance and loadings
MEGASEM	$C_k = \lambda_k EA^{*} + \varepsilon_k$	Latent-factor model estimated jointly

For MEGAWGT, with $\{C_1,\ldots,C_K\}$ denoting the clocks in years, $M$ the $K\times K$ variance-covariance matrix, and $E[\varepsilon_k]=0$0 the $E[\varepsilon_k]=0$1 vector of ones, the weights are defined by $E[\varepsilon_k]=0$2. The resulting index is equivalent to GLS-combining seemingly unrelated regressions on each clock with a common coefficient, efficiently exploiting cross-clock covariance (Menta et al., 17 Sep 2025).

For MEGAFA, exploratory factor analysis is applied to the set of clocks. One factor is retained by the Kaiser criterion, with second-largest eigenvalue $E[\varepsilon_k]=0$3, consistent with a single latent aging construct. Let $E[\varepsilon_k]=0$4 denote the factor loadings and $E[\varepsilon_k]=0$5; the factor score is then formed as the weighted average above. This approach both downweights redundant clocks via $E[\varepsilon_k]=0$6 and emphasizes common variance via $E[\varepsilon_k]=0$7 (Menta et al., 17 Sep 2025).

For MEGASEM, the measurement equations are

$E[\varepsilon_k]=0$8

with a behavioral equation

$E[\varepsilon_k]=0$9

and reduced form

$$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$0

The latent factor $$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$1 is returned on the scale of the first clock entered in the measurement system, so the reported output is interpretable in years (Menta et al., 17 Sep 2025).

A notable implementation detail is scaling. Factor and SEM routines standardize inputs internally, but MEGAWGT and MEGAFA are explicitly de-standardized by manually computing the weighted averages in natural clock units. MEGASEM is scaled to years through the first-clock constraint (Menta et al., 17 Sep 2025). This preserves direct interpretation of MEGA values as epigenetic age in years rather than as arbitrary latent scores.

3. Relation to classical DNAm clocks and robustness theory

MEGA is best understood against the broader methodology of DNAm-clock construction. Canonically, methylation clocks use a high-dimensional regression of methylation levels onto age,

$$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$2

often fit with elastic net or LASSO under $$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$3 (Podolskiy et al., 2018). The standard elastic-net objective is

$$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$4

The robustness argument for MEGA arises because single clocks can fit noise when $$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$5, producing strong apparent in-sample performance but degraded out-of-sample behavior. The chapter shows that even purely random data can yield an apparently precise clock in-sample under elastic-net fitting, making external validation indispensable (Podolskiy et al., 2018). It further derives an upper bound on out-of-sample error,

$$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$6

and a heuristic early-age relation

$$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$7

These results motivate clocks and ensembles that leverage more contributing signals and avoid excessive concentration of weight on a few sites (Podolskiy et al., 2018).

Within that framework, MEGA can be written as a weighted average of $$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$8 clocks,

$$\mathrm{Var}(\hat{A}_{MEGA}) = \sum_k w_k^2 \mathrm{Var}(\varepsilon_k) + 2\sum_{i<j} w_i w_j \mathrm{Cov}(\varepsilon_i,\varepsilon_j),$$9

with inverse-variance weights

$n$0

or as a stacked meta-regression with optional covariates $n$1,

$n$2

This formulation directly links MEGA to the central methodological concern in DNAm aging research: improving out-of-sample reliability by combining predictors with different error profiles rather than relying on a single coefficient vector (Podolskiy et al., 2018).

4. Age acceleration, calibration, and uncertainty

MEGA is commonly analyzed through epigenetic age acceleration rather than raw predicted age. In the applied MEGA study, acceleration is defined as the residual from regressing epigenetic age on chronological age and covariates: $n$3 with $n$4 (Menta et al., 17 Sep 2025). Covariates vary by application and can include mother’s age at birth, maternal education, father’s social class, child gender, birth year, birth order, and age at DNAm assessment; some specifications additionally include BMI, smoking, and drinking (Menta et al., 17 Sep 2025).

The wider DNAm-clock literature also uses two generic formulations: $n$5 and

$n$6

where the residual-based form controls for calibration bias (Podolskiy et al., 2018). The same distinction appears in EpiCAge, where age acceleration is either the direct difference $n$7 or residualized from a regression of predicted age on chronological age, with survival analyses reporting hazard ratios per 5-year age-acceleration increment (Jiang et al., 10 Nov 2025).

A further extension concerns individual-level uncertainty. "U-learning for Prediction Inference via Combinatory Multi-Subsampling" does not define a specific MEGA clock, but it provides a principled way to construct one by ensemble-averaging multiple base predictors across subsamples and attaching valid confidence intervals with conditional coverage (Fei et al., 2024). For a test point $n$8, the incomplete U-statistic ensemble is

$n$9

and a MEGA meta-ensemble can be written as

$$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$0

Confidence intervals follow

$$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$1

with $$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$2 estimated by an infinitesimal jackknife. This framework is significant because conventional epigenetic-clock reporting typically emphasizes point estimates, whereas the U-learning formulation targets prediction inference directly (Fei et al., 2024).

5. Empirical implementations and substantive findings

The principal implemented MEGA analysis uses the Avon Longitudinal Study of Parents and Children, with epigenetic data provided via ARIES, on blood DNAm profiled by Illumina 450k arrays at cord blood at birth, peripheral blood at age $$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$3, and peripheral blood at age $$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$4–$$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$5 years (Menta et al., 17 Sep 2025). Three estimation samples are highlighted: $$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$6 for late-adolescent human-capital analyses, $$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$7 for child-abuse analyses, and $$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$8 for the age-7 school-entry regression-discontinuity design (Menta et al., 17 Sep 2025).

The empirical structure of MEGA in these data is consistent with a single latent factor. Factor analysis retains one factor; the second eigenvalue is $$\mathrm{MEGA}_{\mathrm{WGT}} = \frac{\sum_{k=1}^{K} w_{k,1} C_k}{\sum_{k=1}^{K} w_{k,1}}$$9; loadings exceed $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$0 for all clocks; and uniqueness exceeds $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$1 across applications. In adolescents, the correlation with chronological age is reported as $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$2 for MEGASEM, $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$3 for MEGAFA, and $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$4 for MEGAWGT, compared with $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$5 for GrimAge, $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$6 for Hannum, $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$7 for PhenoAge, and $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$8 for Horvath (Menta et al., 17 Sep 2025). The paper further reports that all three MEGA constructions have slopes approximately $$\mathrm{MEGA}_{\mathrm{FA}} = \frac{\sum_{k=1}^{K} u_{k,1} C_k}{\sum_{k=1}^{K} u_{k,1}}$$9 and lower dispersion than single clocks in adolescent samples.

In human-capital regressions, later-adolescent MEGA acceleration is associated with worse outcomes in early adulthood. The coefficient for no university degree by age $C_k = \lambda_k EA^{*} + \varepsilon_k$0 is approximately $C_k = \lambda_k EA^{*} + \varepsilon_k$1–$C_k = \lambda_k EA^{*} + \varepsilon_k$2 per year of acceleration, with standard errors $C_k = \lambda_k EA^{*} + \varepsilon_k$3–$C_k = \lambda_k EA^{*} + \varepsilon_k$4. For NEET status at $C_k = \lambda_k EA^{*} + \varepsilon_k$5, the estimate is approximately $C_k = \lambda_k EA^{*} + \varepsilon_k$6–$C_k = \lambda_k EA^{*} + \varepsilon_k$7, with standard errors $C_k = \lambda_k EA^{*} + \varepsilon_k$8–$C_k = \lambda_k EA^{*} + \varepsilon_k$9. For SMFQ $\{C_1,\ldots,C_K\}$0 at $\{C_1,\ldots,C_K\}$1, the estimate is approximately $\{C_1,\ldots,C_K\}$2–$\{C_1,\ldots,C_K\}$3, and for diagnosed depression by $\{C_1,\ldots,C_K\}$4, approximately $\{C_1,\ldots,C_K\}$5–$\{C_1,\ldots,C_K\}$6, with marginal significance in both cases (Menta et al., 17 Sep 2025).

In the child-maltreatment application, any child abuse at ages $\{C_1,\ldots,C_K\}$7–$\{C_1,\ldots,C_K\}$8 is associated with roughly half a year higher epigenetic aging in late adolescence: $\{C_1,\ldots,C_K\}$9 years for MEGASEM, $M$0 for MEGAFA, and $M$1 for MEGAWGT, all with standard errors near $M$2 (Menta et al., 17 Sep 2025). Disaggregation attributes most of this signal to child cruelty at ages $M$3–$M$4, with estimates around $M$5–$M$6 years; controlling for blood cell-type counts reduces coefficients by about one-third (Menta et al., 17 Sep 2025).

In the school-entry regression-discontinuity design, entering school one year later is associated with higher MEGA acceleration at age $M$7. In the May–December bandwidth ($M$8), the treatment effect is $M$9 for MEGASEM, $K\times K$0 for MEGAFA, and $K\times K$1 for MEGAWGT. In the tightest July–October bandwidth ($K\times K$2), the corresponding effects are $K\times K$3, $K\times K$4, and $K\times K$5, with the latter two statistically significant (Menta et al., 17 Sep 2025). Heterogeneity analyses place the adverse effect primarily among low-SES households.

These findings establish MEGA as an applied aggregation framework rather than only a theoretical ensemble proposal. At the same time, they are tied to a specific cohort, tissue, platform, and age range.

6. Extensions toward multimodal and graph-based “MEGA-style” clocks

Subsequent work broadens the MEGA idea beyond aggregating four scalar clock outputs. "GraphAge: Unleashing the power of Graph Neural Network to Decode Epigenetic Aging" redefines epigenetic age prediction as supervised regression on a graph whose nodes are CpG sites with methylation values and genomic annotations and whose edges encode co-methylation, same-gene, and same-chromosome relations (Ahmed et al., 2024). On healthy blood methylation data, GraphAge reports $K\times K$6 and $K\times K$7, slightly outperforming AltumAge on the same split. The paper explicitly presents GraphAge as a foundation for a Multi Epigenetic Age (MEGA) clock and proposes a heterogeneous graph schema with nodes spanning CpG sites, promoters/enhancers, genes, transcripts, chromatin accessibility peaks, histone marks, proteins, metabolites, pathways, clinical covariates, and longitudinal measurements. However, the implemented experiments remain methylation-only and blood-specific; the heterogeneous multimodal extension is a proposal rather than a realized MEGA model (Ahmed et al., 2024).

A different extension appears in "Integrating Epigenetic and Phenotypic Features for Biological Age Estimation in Cancer Patients via Multimodal Learning" (Jiang et al., 10 Nov 2025). That paper does not define a specific MEGA clock, but its architecture is inherently multi-clock and multimodal: an epigenetic TabPFN clock, a phenotypic TabPFN clock, and an early-fusion clock are trained in parallel, then stacked with PCA-reduced DNAm and clinical features in a second-layer fusion model. On eight internal cancer cohorts totaling $K\times K$8 patients, EpiCAge-TabPFN achieves $K\times K$9, $E[\varepsilon_k]=0$00, and $E[\varepsilon_k]=0$01; on four external cohorts totaling $E[\varepsilon_k]=0$02 patients, it achieves $E[\varepsilon_k]=0$03, $E[\varepsilon_k]=0$04, and $E[\varepsilon_k]=0$05. The paper explicitly states that extending the second-layer fusion to include outputs from Horvath, Hannum, PhenoAge, GrimAge, AltumAge, YingCausAge, and DunedinPoAm/DoB would qualify as a MEGA approach (Jiang et al., 10 Nov 2025).

Taken together, these papers indicate that the term MEGA now spans at least three layers of meaning: a concrete covariance-weighted aggregation of established clocks, a theoretical ensemble principle for robustness and uncertainty quantification, and a design trajectory toward heterogeneous multimodal biological-age predictors.

7. Limitations, misconceptions, and open problems

A recurrent misconception is that MEGA names a single standardized clock. The literature does not support that interpretation. One paper defines and implements MEGA as an aggregate of four established clocks in years using weighting, factor analysis, and SEM (Menta et al., 17 Sep 2025). Another presents MEGA as an ensemble concept grounded in DNAm-clock robustness analysis (Podolskiy et al., 2018). Several later papers describe MEGA-style extensions or foundations without implementing a definitive MEGA architecture (Ahmed et al., 2024, Jiang et al., 10 Nov 2025, Fei et al., 2024).

A second misconception is that aggregation alone resolves the core limitations of DNAm clocks. The robustness literature explicitly warns that internal cross-validation is insufficient because elastic-net clocks can fit noise when $E[\varepsilon_k]=0$06, and external validation is essential (Podolskiy et al., 2018). The applied MEGA study likewise remains specific to ALSPAC blood DNAm on Illumina 450k arrays, with epigenetic subsamples skewed toward higher-SES families; portability to other tissues, platforms such as EPIC, and other populations may require recalibration (Menta et al., 17 Sep 2025).

Age range and biological interpretation also remain nontrivial. The component clocks were trained primarily on adults, so application to children and adolescents requires caution (Menta et al., 17 Sep 2025). In oncology-focused multimodal work, the target is still chronological age rather than a directly observed biological-age construct, and the authors note that true biological age is only imperfectly captured by such models (Jiang et al., 10 Nov 2025). Similarly, GraphAge’s proposed multimodal MEGA extension would require integrated multi-omics per individual, which is currently rare (Ahmed et al., 2024).

The principal open problem is therefore not merely how to average more clocks, but how to combine heterogeneous molecular and phenotypic signals while preserving calibration, external validity, and interpretable uncertainty. The existing MEGA literature provides concrete solutions for covariance-weighted aggregation, factor-based latent scoring, SEM estimation, subsampling-based confidence intervals, and multimodal stacking. A plausible implication is that future MEGA systems will be judged less by the label itself than by how well they manage measurement error, distribution shift, multimodal integration, and biological interpretability across cohorts and tissues.