chatGPTpaid(5.6):
Overall assessment
Cruz-Gonzalez et al., “Methylation clocks fail to generalize across genetically admixed individuals” provides convincing evidence that several established blood DNA-methylation clocks—especially the Horvath clock—perform less reliably in cohorts with substantial African genetic ancestry.
The paper is important, but its title is broader than its results. It demonstrates reduced portability of particular clocks in particular populations; it does not show that methylation clocks generally fail in all admixed individuals. The proposed meQTL mechanism is plausible and well supported associationally, but not demonstrated causally.
Summary
The authors ask whether methylation clocks developed largely in European-ancestry datasets accurately estimate age and Alzheimer’s-associated age acceleration in more genetically diverse populations.
They analyse:
- 621 MAGENTA participants: 313 Alzheimer’s disease cases and 308 controls.
- African American, Puerto Rican, Cuban, Peruvian and White cohorts.
- More than 2,500 additional participants from two African American studies and a Swedish White cohort.
- First-generation clocks—Horvath, Hannum and Zhang—plus PhenoAge, DunedinPACE, principal-component clocks and a simple clock ensemble.
For the Horvath clock in MAGENTA controls:
| Cohort | Controls | Age correlation | Median absolute error |
|---|---|---|---|
| White | 65 | 0.72 | 5.10 years |
| African American | 107 | 0.51 | 5.38 years |
| Puerto Rican | 74 | 0.45 | 5.19 years |
| Peruvian | 41 | 0.72 | 4.19 years |
| Cuban | 21 | 0.68 | 5.60 years |
The proportion of African ancestry was associated with greater Horvath-clock error, but the estimated magnitude was modest: approximately 1.46 additional years of error when comparing 100% with 0% African ancestry, with a marginal p-value of 0.039.
The poorer correlations replicated in the external African American cohorts. Similar, though variable, ancestry differences appeared with the Hannum and Zhang clocks. Principal-component transformations and averaging several clocks did not solve the problem.
The clocks were also inconsistent in detecting accelerated ageing in Alzheimer’s disease:
- Horvath age acceleration distinguished Alzheimer’s cases from controls only in the White cohort.
- PhenoAge did so in the African American cohort, which is notable because African American participants were represented in its training data.
- DunedinPACE performed best, detecting a faster pace of ageing in White, African American and Puerto Rican Alzheimer’s cases.
The authors then investigate why performance might differ. They find that:
- About 24% of Horvath-clock CpGs were differentially methylated between African- and European-ancestry samples.
- Variants directly destroying clock CpGs were numerous in principle but almost universally too rare to explain population-level differences.
- In contrast, 271 of 353 Horvath CpGs had at least one previously identified methylation quantitative-trait locus, or meQTL.
- These meQTL variants were more frequent in African genetic ancestry backgrounds.
- Forty-two of the 56 Horvath CpGs associated with prediction error were also affected by known meQTLs.
- No meQTLs affecting DunedinPACE sites were detected in the particular meQTL resources used.
The authors therefore recommend more diverse training datasets and avoiding CpGs strongly influenced by ancestry-differentiated meQTLs.
What is genuinely novel?
The broad observation that methylation clocks can behave differently across ancestry groups is not new. Previous studies had already reported population differences, and Meeks et al. had linked common genetic variation to epigenetic-age estimates in African populations.
The paper’s real novelty is the integration of several elements:
-
Continuous ancestry rather than demographic labels alone.
It relates individual genomic ancestry proportions to clock error within admixed cohorts. -
A wide comparison of clock types.
First-, second- and third-generation clocks, PC clocks and ensembles are tested in a common framework. -
Connection to Alzheimer’s disease.
It shows that portability problems can alter apparent disease-associated age acceleration, not merely chronological-age prediction. -
A specific genetic explanation.
It distinguishes rare variants that destroy clock CpGs from common meQTLs that alter their methylation. -
Evidence about possible solutions.
The relative performance of PhenoAge and DunedinPACE supports diverse training and meQTL-resistant CpG selection, while PC transformation and simple ensembling appear insufficient.
Thus, the novelty is a systematic ancestry–clock–meQTL–disease synthesis, rather than the first discovery that clock performance differs between populations.
Critique
Strengths
The study uses genetic ancestry estimates, rather than treating racial or geographic categories as precise biological variables. Samples within MAGENTA were processed through a common pipeline, several clocks were compared, and the principal finding was tested in independent datasets. The data and analysis code are also publicly available.
The distinction between rare CpG-disrupting variants and common meQTLs is particularly useful. It prevents the paper from settling for a vague claim that “genetics matters” and identifies a testable route for improving clocks.
Principal limitations
-
The title overgeneralises.
Cubans and Peruvians were genetically admixed but showed performance similar to the White cohort. DunedinPACE was comparatively portable. The evidence therefore concerns several clocks in cohorts with substantial African ancestry, not admixed individuals generally. -
Correlation differences look larger than absolute-error differences.
Horvath correlations fell from 0.72 in White controls to 0.51 and 0.45 in African American and Puerto Rican controls, yet median absolute errors differed by only 0.28 and 0.09 years. Correlation is highly sensitive to age range and variance. Calibration slopes, intercepts and age-distribution-matched comparisons would better establish clinically meaningful failure. -
External ancestry and cohort effects are confounded.
The Swedish and African American replication samples came from different studies, environments and ascertainment procedures. This confirms that the performance difference recurs across datasets, but cannot show that genetic ancestry, rather than study-specific technical or environmental factors, caused it. -
The meQTL analysis does not prove mediation.
Clock error is mathematically calculated from the same CpG methylation measurements subsequently tested for association with that error. Some CpG–error association is therefore expected. The authors do not show that adjusting for genotype, removing meQTL-affected CpGs or retraining the clock eliminates the ancestry effect. -
“No meQTL” means none detected, not none exist.
The very different meQTL counts across clocks may reflect discovery-cohort size, linkage disequilibrium, probe coverage and statistical power. DunedinPACE cannot yet be described as genetically uninfluenced merely because the assembled databases contained no detected meQTL for its sites. -
Cell composition and missing CpGs are important technical uncertainties.
Blood-cell proportions were estimated rather than measured, using a reference method not extensively validated across ancestries. In addition, 12.7% of Hannum sites and 17.9% of DunedinPACE sites were missing and imputed. Both issues could create or conceal population differences. -
The Alzheimer’s results are statistically fragile.
Numerous clock-by-cohort comparisons were made without a clearly reported overall multiple-testing correction. The White Horvath result was only a 0.2-year median difference with p=0.041. Several cohorts were small, and reported power of 47–75% is not strong enough to interpret non-significance as evidence that a clock fails. -
There is an internal numerical inconsistency.
For Grady participants aged 55 or older, Figure 3 reports (r=0.49) and MAE 4.88, whereas the text and appendix report approximately (r=0.57) and MAE 5.49. This requires clarification.
Bottom line
The paper supports a strong warning: absolute epigenetic-age estimates and cross-population comparisons should not be interpreted without knowing a clock’s training population, calibration and sensitivity to genotype.
It does not establish that DNA-methylation clocks are useless, that admixture itself is the cause, or that meQTLs fully explain the observed differences. A decisive follow-up would retrain or genotype-adjust the clocks in a single large, uniformly processed multi-ancestry cohort and test whether portability and health-outcome prediction actually improve.
For your citrate → acetyl-CoA → histone-acetylation hypothesis, the paper is only indirectly relevant. It concerns the statistical portability of DNA methylation biomarkers, not histone acetylation, transcription or splicing, and therefore neither supports nor contradicts that mechanistic axis.