To test whether current methylation clocks are able to predict age accurately in diverse, genetically admixed groups, we evaluated age predictions for the control individuals in the MAGENTA study. We first analyzed the widely used Horvath clock, which was trained on data from several tissues and cell types to accurately predict age across the lifespan using methylation levels at 353 CpG sites.

Age predicted from DNA methylation (DNAm age) in the White cohort using the Horvath clock was strongly correlated with chronological age (Pearson r=0.72) (Figure 2A). While this correlation is lower than reported in the original study (>0.9), it is consistent with previous studies of older individuals (Horvath, 2013; Marioni et al., 2015).

In comparison to the White cohort, the correlation between DNAm age and chronological age was significantly lower for Puerto Ricans (r=0.45, p=0.007, n=74) and African Americans (r=0.51, p=0.016, n=107; Figure 2B). The correlations for Cubans (r=0.68, p=0.385, n=21) and Peruvians (r=0.72, p=0.52, n=41) were similar to the White cohort. These differences in clock accuracy were not driven by systematic differences in the age distributions for the cohorts (Figure 1—figure supplement 1). The methylation clocks were also more accurate in the White cohort than the African American and Puerto Rican cohorts in terms of median absolute error (MAE; Appendix 1—table 1). Combining cases and controls did not qualitatively change the performance relationships; the accuracy remained lower for African Americans and Puerto Ricans relative to the Whites (Figure 2—figure supplement 1).

The two cohorts with lower correlations come from regions where individuals often have substantial amounts of African ancestry. To explore if admixture levels associated with the accuracy of the Horvath clock in predicting age, we estimated the global proportions of African (YRI), European (CEU), and American (PEL) ancestries in each individual from the MAGENTA cohort using reference groups from the 1000 Genomes Project (The 1000 Auton et al., 2015).

Methylation clock accuracy was lowest for the cohorts with substantial African ancestry: African Americans (median 85% African) and Puerto Ricans (median 15% African). In contrast, the clocks performed similarly to the White cohort in groups with the lowest African ancestry: Cubans (6% African) and Peruvians (2% African; Figure 2C). We regressed the Horvath clock error on the proportions of African ancestry in the genomes of the MAGENTA individuals, adjusting for chronological age. The proportion of African ancestry is significantly associated with increased Horvath clock error (p=0.039), with an estimate of 1.46 years more error for 100% African ancestry compared to no African ancestry.

To evaluate if the significantly lower accuracy of the Horvath clock age predictions on individuals with substantial African ancestry holds outside of MAGENTA, we analyzed three independent whole blood methylation datasets. Two focused on African American individuals—the Grady Trauma Project (n=422; Katrinli et al., 2024) and the GENOA study (n=1394; Shang et al., 2023)—and one focused on White Swedish individuals (n=729; Johansson et al., 2013). Each study sampled a wide range of chronological ages, including many older individuals. To enable direct comparison with MAGENTA, we first focused on older individuals (≥55 years old). As observed in MAGENTA, the Horvath clock had lower accuracy for the African American cohorts (Figure 3, bottom row; Grady: r=0.57, GENOA: r=0.67) and higher accuracy for the White Swedish cohort (r=0.80). We then expanded the analysis to all individuals in each cohort. As expected, the overall accuracy of the age predictions improved, but the disparity in performance between cohorts remained (Figure 3, top row; Grady: r=0.88, GENOA: r=0.85, Swedish: r=0.97).

A full summary of the clock accuracy statistics, including MAE, maximum absolute error, correlation, and sample size for each cohort is included in Appendix 1—table 2. Taken together, these results show that the reduced clock accuracy for individuals with substantial African ancestry is not limited to the cohorts in the MAGENTA study, but rather is a general pattern in the portability of methylation clocks.

To investigate the performance of other methylation clocks at predicting chronological age in admixed individuals, we selected several additional publicly available open-source clocks. We considered two other ‘first-generation’ clocks that were trained to predict chronological age: the Hannum clock (Hannum et al., 2013) and a model developed by Zhang et al., 2019 that used large datasets for training and achieved substantially higher performance than previous age predictors. We hereafter refer to this elastic net model as ‘Zhang_EN’.

Both models achieved higher correlations with chronological age than the Horvath clock across the cohorts in the MAGENTA study. For example, Hannum has a correlation of 0.74 in the White cohort, and consistent with previous evaluations, Zhang_EN has a correlation of 0.88. These relative performance trends held across cohorts, but again, the cohorts with substantial African ancestry, African Americans and Puerto Ricans, consistently had the lowest age correlations for each clock (Figure 4).

Next, we evaluated the PhenoAge clock, a ‘second-generation’ clock that is trained on biomarkers of frailty and physiological deterioration (Levine et al., 2018). The correlation between DNAm age and chronological age was lower for this clock in comparison to the other methylation clocks (r=0.53 in the White cohort). This is likely due to the fact that this clock was not trained to predict age directly, but rather markers of aging. This clock did not show as substantial a difference in performance between cohorts as seen for the Horvath clock, but the African American and Puerto Rican individuals again had the lowest correlation of all cohorts.

Overall, these results demonstrate that current methylation clocks vary in the correlation of their predicted DNAm age with chronological age in genetically admixed cohorts. The clocks are also consistently the least accurate in predicting age in cohorts with substantial proportions of African ancestry.

Principal component (PC) versions of many methylation clocks have been developed with the aims of reducing noise and improving replicability and generalizability of age predictions (Higgins-Chen et al., 2022). Thus, we evaluated the accuracy and stability of PC methylation clocks across genetic ancestries. Overall, the PC version of the Horvath clock did not result in consistent improvement in age prediction accuracy or generalization across MAGENTA cohorts (Figure 3—figure supplements 1 and 2). The lower accuracy for age prediction in individuals of substantial African ancestry was present for the PC versions of the clocks in the replication cohorts, just as in the MAGENTA cohorts (Figure 3—figure supplement 3). Taken together, these results indicate that PC-based versions of the existing methylation clocks do not provide generalizable age predictors for individuals of diverse genetic ancestries.

DNAm age has been proposed as a promising biomarker and predictive tool for age-related disease risk, particularly because of associations between accelerated DNAm age (compared to chronological age) and the presence of diseases such as coronary heart disease, Parkinson’s disease, and AD. However, these results have largely been observed in European-ancestry cohorts.

To evaluate the ability of methylation clocks to identify accelerated aging and risk for age-related disease in diverse, genetically admixed individuals, we quantified the association of methylation age acceleration with AD status in cohorts from the MAGENTA study. In addition to the clocks tested in the previous section, we also included a ‘third-generation’ clock, DunedinPACE, that aims to predict the pace of aging as measured by change in biomarkers over time from methylation data, rather than age itself (Belsky et al., 2022).

The cell type composition in blood is known to change with age, which if not accounted for, can confound age acceleration estimates (Jaiswal and Ebert, 2019). Thus, we focused on intrinsic age acceleration estimates computed using established algorithms to correct for cell type composition.

To establish a baseline for this analysis, we tested whether individuals with AD in the White cohort show significantly greater age acceleration than non-demented controls. As expected from previous studies (Levine et al., 2015; Levine et al., 2018), AD cases have modest but significantly greater age acceleration as measured by the Horvath clock than controls (Figure 5—figure supplement 1; median 1.7 vs 1.5 years, p=0.041). For each of the other clocks, AD cases had higher median age acceleration than controls (Figure 5B), though the differences only reached statistical significance for the DunedinPACE clock (median 1.09 vs 1.07, p=0.044).

Having established that previously reported age acceleration in AD was detectable in the MAGENTA White cohort, we evaluated whether the clocks found accelerated aging in the admixed AD cohorts. Focusing first on the Horvath clock, we observed inconsistent relationships between age acceleration and AD status. None of the admixed cohorts showed a significant difference, and controls even had higher median age acceleration in Peruvians and Cubans (Figure 5A). Across the other clocks, none consistently identified greater age acceleration in AD cases across all populations (Figure 5B). Among the first- and second-generation clocks, only PhenoAge demonstrated a significant ability to differentiate AD cases from controls in any of the non-European ancestry groups, specifically in African American individuals (p=0.008), which were included in its training set. Cubans consistently showed greater age acceleration in controls rather than cases, while none of the other admixed cohorts even had consistent directions of effect across methods. While sample size in these cohorts is relatively small, our power calculations suggest that low statistical power is unlikely to explain the lack of signal (see Discussion).

DunedinPACE stood out in the evaluation, as it identified significantly greater aging in AD cases compared to controls in the White (p=0.044), African American (p=0.0019), and Puerto Rican (p=0.0090) cohorts using its ‘pace of aging’ metric. However, no significant differences were found for Cubans (p=0.26) or Peruvians (p=0.81).

Inspired by recent work on the ensembling of PRS to better predict disease risk from genetic variation across populations (Monti et al., 2024), we evaluated whether combining age predictions could lead to greater accuracy in age prediction and AD risk prediction in the admixed cohorts. To investigate this, we averaged the age predictions for each individual in the MAGENTA study across five methylation clocks: Horvath, Hannum, Zhang_EN, Zhang_BLUP, and PhenoAge clocks. (The Zhang_BLUP is a variation of the Zhang_EN clock that does not use strong regularization.)

The ensemble method’s DNAm age prediction is more strongly correlated with chronological age in comparison to the Horvath and PhenoAge clocks, but it did not improve upon the best predictors (Zhang clocks) across populations (Figure 5—figure supplement 2A).

We next evaluated whether the ensemble intrinsic age acceleration estimates would associate more strongly with AD disease status relative to the standalone methylation clocks. Following the same evaluation framework as for the individual clocks, we found only one significant difference in age acceleration. Cuban control individuals had significantly lower age acceleration than AD cases (Figure 5—figure supplement 2B). Thus, a simple ensemble does not lead to stronger performance at either task in admixed cohorts.

Our results so far demonstrate that existing methylation clocks do not perform consistently across genetically admixed individuals. We now shift our attention to investigating possible mechanisms underlying this lack of generalization. We hypothesized the differential methylation levels between genetic ancestries at clock CpGs could contribute to the lower performance of methylation clocks in admixed individuals. To test for this pattern, we analyzed whole blood methylation data from the EWAS Hub for 306 individuals of African (AFR) and 300 of European (EUR) ancestry. We fitted a linear regression model to the whole blood methylation profiles for these two sets of individuals, including covariates for sex and chronological age. We called CpG sites as differentially methylated in AFR vs. EUR individuals controlling the false discovery rate at 5% with the Benjamini-Hochberg correction. We identified 73,819 differentially methylated CpG sites and intersected them with CpGs included in each clock.

All the clocks considered had a substantial proportion of their CpGs differentially methylated in AFR vs EUR individuals (Figure 6A). The PhenoAge and Horvath clocks had the highest fractions with 24.8% (127/513) and 23.8% (84/353) of their CpG sites differentially methylated, respectively. The other clocks also had substantial fractions of their CpGs differentially methylated: Zhang EN clock (12.6%, 65/514), DunedinPACE (9.2%, 16/173), and the Hannum clock (7%, 5/71). This suggests systematic differences in methylation patterns between genetic ancestries at CpG sites considered in methylation clocks are common.

The presence of differentially methylated CpGs in a clock does not necessarily indicate a barrier to accurate prediction across ancestries. For example, the differential methylation could reflect true differences in environmental exposures between the groups that influence aging. To identify CpGs that could drive the lower accuracy for individuals with substantial African ancestry in their genomes, we tested whether methylation levels at individual clock CpGs are associated with increased clock error across all MAGENTA individuals. The number of CpG sites with methylation levels significantly associated (after Benjamini-Hochberg multiple testing correction) with clock error in the MAGENTA individuals varied across the clocks (Figure 6B). The methylation levels at 56 of the 353 Horvath clock CpGs significantly associated with increased clock error, in contrast to 19 of the 71 Hannum clock CpGs. Finally, the methylation levels at 52 of the 514 Zhang EN clock CpGs significantly associated with increased clock error in MAGENTA individuals. Because the PhenoAge and DunedinPACE clocks are second and third generation clocks, respectively, and therefore do not aim to predict age directly but rather a composite biological age, we did not include these two clocks in this analysis.

Given that many CpG sites used in clocks have different methylation levels between genetic ancestries, we sought to investigate potential mechanisms underlying the differences in methylation and decreased performance of some clocks. Environmental differences between populations likely contribute to the decreased performance of the clocks; however, we do not have comprehensive data on differences in environmental exposures between the cohorts. Moreover, several observations suggest that genetic factors also play a substantial role in clock generalizability. First, genetic ancestry explains a substantial fraction of methylation differences between human populations (Galanter et al., 2017). Furthermore, the particularly large decrease in performance for individuals with a substantial fraction of African ancestry in different populations suggests that genetic divergence rather than the presence of environmental differences alone is responsible.

Thus, we now focus on evaluating potential mechanisms by which genetic variation could reduce clock accuracy across human groups. First, we evaluate whether genetic variation in human populations disrupts the potential for methylation at CpG sites. We then explore how often genetic variants that associate with methylation levels (meQTL) influence clock CpGs and differ in frequency between genetic ancestries.

We first quantified how often a genetic variant segregating in human populations disrupted a CpG site included in a clock. This scenario could lead to errors if the variant has different frequencies across cohorts given the loss of potential for methylation and the ability of the site to contribute to the age prediction.

Of the 353 CpG sites considered in the Horvath clock, 245 (69%) have at least one disruptive genetic variant observed in at least one individual in the gnomAD database of variants identified in a cohort of 76,156, including thousands of individuals of non-European ancestry. While this may seem like a substantial concern, these variants are extremely rare both globally and stratifying by ancestry (Figure 6C; Figure 6—figure supplement 1). The average frequency is 0.0001, with the most common case being a variant observed in just one individual. Only one clock CpG disrupting variant had a frequency greater than 1%.

The genetic variant frequency patterns are similar for the other clocks (Figure 6—figure supplement 2). Of the 71 CpG sites considered in the Hannum clock, 54 (76%) have at least one disruptive variant from the gnomAD database, but only one of the variants has a frequency greater than 1%. For the 514 CpG sites that make up the Zhang19_EN clock, 403 (78%) are disrupted by at least one gnomAD variant, and five of the variants in clock CpGs have a frequency greater than 1%. The PhenoAge clock has 381 of its 513 (74%) disrupted by a gnomAD variant, but only two of these variants have an allele frequency greater than 1%. Finally, in the case of DunedinPACE clock and its 173 CpG sites, 158 (91%) are disrupted by at least one gnomAD variant. However, only one of these variants has an allele frequency greater than 1%. Thus, genetic variation in clock CpG sites themselves is unlikely to be the main cause of the lack of generalization of the methylation clocks.

We next assessed the prevalence of meQTLs, genetic variants that associate with clock CpG site methylation, another potential modifier of methylation levels that could lead to spurious DNAm age predictions across individuals. We gathered three sets of meQTLs from Europeans, South Asians, and African Americans (Methods). We intersected the CpGs associated with the meQTLs with clock CpG sites.

Out of the 353 CpGs included in the Horvath clock, 271 (77%) had at least one meQTL. Overall, a total of 29,033 unique variants associated with methylation levels at Horvath clock CpGs. In contrast to CpG disrupting variants, the meQTL had an average allele frequency of 0.19, and 26,500 were common, as determined from gnomAD version 4.1 (global allele frequencies ≥1%). However, only a small proportion of CpG sites in the other first-generation clocks have meQTL (Hannum: 8%, Zhang_EN: 1%). PhenoAge is similar, with 7% of its CpGs affected by at least one meQTL. Finally, DunedinPACE had no meQTLs affecting its 173 clock CpG sites (Figure 6D).

Thus, the clock with the largest decrease in performance in admixed cohorts (in terms of predicting chronological age and identifying age acceleration in AD) has the largest fraction of CpGs with meQTLs. In contrast, DunedinPACE, the best performing clock at identifying AD cases in the MAGENTA study, had no meQTLs. The three other clocks with intermediate performance in the admixed cohorts all have meQTLs for some CpGs, but a much lower fraction than the Horvath clock.

The presence of meQTL influencing clock CpGs does not necessarily indicate a problem for clock generalization. However, differences in the presence or frequency of meQTL that influence clock CpGs between genetic ancestries could lead to decreased DNAm age prediction accuracy (and therefore weaker associations with disease) in admixed cohorts. For example, if a clock is trained on a cohort without an meQTL, the learned weights for the CpG will not have accounted for the effects of the meQTL. To quantify whether differences in meQTL across genetic ancestries could influence methylation clocks, we analyzed the gnomAD allele frequencies for the 29,033 Horvath clock meQTL tag variants in multiple global populations.

The Horvath clock meQTLs are at significantly higher frequencies in African ancestry populations (median 0.068) than in each of the eight other population groups considered (Figure 6E; 0.004–0.046, p <3.85×10⁻²⁵). There were also 2328 meQTL that were only observed in African individuals.

To connect these results to individuals with recent admixture, like many in the MAGENTA study, we also tested whether the Horvath clock meQTL differed in frequency in local ancestry blocks of different origins in genetically admixed individuals from gnomAD. We used pre-computed local ancestry calls for 7612 Latino/Admixed American individuals to compare allele frequencies for each meQTL in three ancestral backgrounds: African, Amerindigenous, and European. The clock CpG-affecting meQTLs were at higher frequencies in African local ancestry backgrounds (Figure 6—figure supplement 4) relative to Amerindigenous and European backgrounds, consistent with our finding that these meQTLs are most frequent in African ancestry individuals at the global population level.

To explore if meQTL influence CpG methylation relevant to clock predictions in the MAGENTA individuals, we tested whether meQTLs described in the previous section are present in MAGENTA. For example, an meQTL (chr3:51639064_A_G) for a Horvath clock CpG with a reported beta of 1.34 and an allele frequency of 2% in gnomAD (version 4.1) is present in MAGENTA individuals (study-wide frequency = 11%; AA frequency = 20%; PR frequency = 7.5%; CUB and PER combined frequency = 6%). In individuals with at least one copy of the variant, the methylation level of the corresponding Horvath clock CpG site was higher, as expected from the reported beta for the meQTL. The median methylation level of the clock CpG site for individuals with at least one copy of the variant was significantly higher than for individuals without the variant (median difference in methylation beta = 0.04; p = 1.64×10⁻⁷). Consistent with our hypothesis that this meQTL could contribute to Horvath clock error, individuals with this meQTL had 0.4 years higher median clock error.

Expanding on this example, we quantified whether differences in meQTL frequency across populations observed in other datasets are present in MAGENTA. The overall frequency of meQTL without substantial differentiation in large databases was similar between the MAGENTA admixed cohorts (Figure 6F, left panel; median variant frequencies: AA = 0.32, PR = 0.31, PER = 0.29, CUB = 0.30). However, as expected, the frequency of AFR-differentiated meQTL was higher in the MAGENTA African American (median = 0.21) and Puerto Rican (median = 0.07) cohorts compared to the Peruvian (median = 0.02) and Cuban (median = 0.04) cohorts (Figure 6F, right panel). Thus, the MAGENTA individuals carry many meQTL at different frequencies in African ancestry that influence clock CpG methylation.

Finally, we tested whether the error-associated CpGs in the clocks were also affected by meQTL. Strikingly, of the 56 Horvath clock CpGs that were significantly associated with increased clock error in MAGENTA individuals, 42 were also affected by an meQTL, and nine were affected by African ancestry-differentiated meQTL. This pattern contrasts with the Hannum and Zhang EN clocks, where only three and one of the error-associated CpGs were affected by an meQTL, respectively (Figure 6G, Figure 6—figure supplement 5).

We applied multiple commonly used first-, second-, and third-generation methylation clocks to all individuals in the MAGENTA study with genome-wide methylation data. We used established implementations of the Horvath, Hannum, Zhang_EN, Zhang_BLUP, and PhenoAge clocks from the methylclock R library (Pelegí-Sisó et al., 2021). We also applied DunedinPACE, a third-generation clock separately, because it was not included in the methylclock library (Belsky et al., 2022). Because this clock does not explicitly predict age, it is not included in the analyses of correlation with biological age. Unless otherwise specified, default options were used for all clocks.

The methylation clocks considered analyze different numbers of CpG sites. For each clock, the sites considered were taken from the methylclock library or the original publication. In the case of missing data, the methylClock library imputes methylation status using the mpute.knn function from the impute R library. The MAGENTA cohorts had low proportions of missing data for clock CpGs. Specifically, there were 3.7% missing for the Horvath clock, 12.7% for the Hannum clock, 4.5% for the Zhang_EN clock, 3.7% for the PhenoAge clock, and 17.9% for the DunedinPACE clock.

We computed the Pearson correlation of estimated methylation age and chronological age for the controls in each cohort. To compare the strength of correlation between cohorts, we computed p-values for the observed differences using Fisher’s z test and the Zou method for computing confidence intervals as implemented in the cocor library (Diedenhofen and Musch, 2015).

In order to quantify methylation age acceleration from blood methylation data, we estimated raw, intrinsic, and extrinsic age acceleration for all clocks, except DunedinPACE, from the methylclock library. Blood cell type composition differs between individuals and over the lifespan; thus, we report results in the main text using intrinsic age acceleration estimates, which capture age acceleration independently of blood cell proportions. We did not have access to empirical estimates of blood cell type counts for the data in MAGENTA, and as such estimated the counts using the functionality for deconvolution implemented in methylclock, specifically with the ‘blood gse35069 complete’ parameter. Given that cell type proportions vary between populations based on genetic ancestry, this is an inherent limitation of our study. Nevertheless, the biology of the aging process should be captured using this panel as a reference on our own datasets.

Using the age acceleration estimates, we compared methylation age association between AD cases and matched controls using a Mann-Whitney U test, as implemented in the stats library in R. We analyzed the study as a whole and stratified by cohort.

To evaluate the reproducibility of clock performance across ancestrally diverse populations, we analyzed three independent datasets obtained from the NCBI Gene Expression Omnibus (GEO) using the GEOquery R package. These cohorts were: (1) a European-ancestry cohort from Sweden (GSE87571, Ntotal=732\begin{document}$N_{\mathrm{total}}=732$\end{document}), (2) an African American cohort from the Genetic Epidemiology Network of Arteriopathy study (GENOA; GSE210255, Ntotal=1,394\begin{document}$N_{\mathrm{total}}=1,394$\end{document}), and (3) an African American cohort from the Grady Trauma Project (GTP; GSE72680, Ntotal=422\begin{document}$N_{\mathrm{total}}=422$\end{document}).

Processed methylation beta-value matrices were loaded using the data.table package. Probes with high missingness (>5%) and detection p-value columns were removed prior to calculation. To ensure demographic alignment with the MAGENTA study cohort, we also analyzed subsets of each dataset including only individuals aged 55 years or older. This resulted in n=280\begin{document}$n=280$\end{document} for the Swedish cohort, n=865\begin{document}$n=865$\end{document} for the GENOA study, and n=62\begin{document}$n=62$\end{document} for the GTP cohort. DNA methylation age (DNAmAge) was calculated using the methylclock package. Clock performance was then quantified using the MAE, mean squared error (MSE), and Pearson correlation between predicted and chronological age.

We tested the PC versions of multiple methylation clocks using the PC Clock R library (Levine Lab, 2022; Higgins-Chen et al., 2022). We then generated the same accuracy metrics as for the ‘regular’ versions of the clocks when applied to the MAGENTA cohorts and all three replicate cohorts.

We tested the performance of a simple ensemble of age prediction methods at both estimating chronological age and distinguishing AD cases and controls. The ensemble was computed as the average of the estimate of each method on each individual. The resulting predictions were evaluated as described for each clock itself.

To identify specific CpG sites whose methylation levels were associated with increased prediction error in the MAGENTA study individuals, we performed a site-wise association analysis across all CpG markers included in the Horvath, Hannum, and Zhang EN clock models. For each clock, we regressed absolute prediction error on the normalized methylation beta values for each individual CpG site within the respective methylation clock. To account for multiple testing, p-values were adjusted for each clock separately using the Benjamini-Hochberg (FDR) correction methods. CpG sites were considered significantly associated with prediction error if they had an FDR-adjusted p-value less than 0.05.

To test whether the CpG sites included in the methylation clocks are variable between individuals, we intersected all the clock CpGs with variants identified in version 3.0 of gnomAD (Karczewski et al., 2020). This database covers 76,156 individuals with whole genome sequencing harmonized from many large-scale sequencing studies. The intersection was performed using bedtools in hg38 coordinates (Quinlan and Hall, 2010).

All code used for these analyses is available in the following GitHub repository: https://github.com/seba2550/methyl-clocks-admixture (copy archived at Cruz-Gonzalez, 2026).