Methylation Clocks Do Not Predict Age or Alzheimer's Disease Risk Across Genetically Admixed Individuals

This study demonstrates that current DNA methylation clocks lack portability and accuracy in predicting age and Alzheimer's disease risk across genetically admixed populations, particularly those with African ancestry, due to widespread differences in methylation patterns and higher frequencies of methylation-associated genetic variants (meQTLs) in these groups.

The Gist

Imagine you have a very sophisticated, high-tech watch that doesn't tell you the time of day, but rather tells you how "old" your body feels on the inside. Scientists call this a Methylation Clock.

These clocks look at tiny chemical tags (like little sticky notes) attached to your DNA. As we get older, these sticky notes change in a predictable pattern. By counting them, the clock tries to guess your biological age. If the clock says you are 60, but you were born 50 years ago, it suggests your body is aging faster than it should. This is a big deal because it could help doctors predict who might get diseases like Alzheimer's before symptoms even show up.

But here's the problem: This paper is like a reality check for those watches. The researchers found that these clocks work great for some people, but they often break or give the wrong time for others.

Here is the story of what they found, explained simply:

1. The "One-Size-Fits-All" Watch Doesn't Fit Everyone

The scientists tested these clocks on a huge group of people with and without Alzheimer's. They looked at people of different backgrounds: White, African American, Puerto Rican, Cuban, and Peruvian.

• The Result: The clocks worked reasonably well for the White participants. But for people with significant African ancestry (like African Americans and many Puerto Ricans), the clocks were like a broken compass. They couldn't accurately guess the person's age.
• The Analogy: Imagine you trained a dog to fetch a specific type of ball. If you throw that exact ball, the dog catches it perfectly. But if you throw a slightly different ball (one that looks similar but has a different texture), the dog gets confused and misses. The clocks were "trained" mostly on data from White people. When they tried to "fetch" the age of people with different genetic backgrounds, they missed the ball.

2. Why Did the Clocks Fail?

The researchers asked: Why do these clocks fail for people with African ancestry? They looked for two main culprits:

• Culprit A: Broken DNA Sites (Rare): Sometimes, a person's DNA has a tiny mutation that physically destroys the spot where the clock looks for its "sticky note." The researchers found these mutations, but they were so rare (like finding a specific grain of sand on a beach) that they couldn't be the main reason the clocks failed.
• Culprit B: The "Volume Knob" (Common): This was the real smoking gun. They found thousands of genetic variations that act like volume knobs for those sticky notes. These "knobs" turn the chemical tags up or down.
  • The Twist: These volume knobs are set to very different levels in people of African ancestry compared to people of European ancestry.
  • The Analogy: Imagine the clock is a radio station. The White training data was recorded at a volume of "5." But for people with African ancestry, their genetic "volume knob" is naturally set to "8." The clock hears the signal, thinks, "Wow, that's loud! That must mean the person is super old!" and gives a wrong answer. The clock doesn't know the volume knob is different; it just assumes everyone is listening at the same volume.

3. The Alzheimer's Connection

The researchers also tried to use these clocks to see if they could spot Alzheimer's patients by seeing if their bodies were "aging too fast."

• The Result: In the White group, the clocks could spot the difference (Alzheimer's patients looked biologically older). But in the mixed-race groups, the clocks were too confused to tell the difference. They couldn't distinguish between a healthy person and someone with Alzheimer's because the "genetic volume knob" noise was drowning out the disease signal.

4. What Does This Mean for the Future?

The paper concludes that we cannot just take a tool built on one group of people and assume it works for everyone.

• The Warning: If we use these clocks on diverse populations today, we might tell a healthy person they are "aging fast" (causing unnecessary panic) or tell a sick person they are "fine" (missing a chance for treatment).
• The Solution: We need to build "Universal Clocks." This means training the AI on data from all kinds of people, not just one group. We also need to build clocks that ignore those "volume knobs" (genetic variations) so they measure the actual aging process, not just the genetic background.

The Bottom Line

Think of these methylation clocks as maps. Right now, we have a very detailed map of the United States, but if you try to use that same map to navigate through Africa or South America, you'll get lost. The landmarks (the DNA patterns) look similar, but the terrain is different.

This paper is a friendly reminder to scientists: "Don't assume your map works everywhere. We need to draw new maps for every part of the world to make sure everyone gets the right directions."

Technical Summary

1. Problem Statement

DNA methylation clocks are statistical models trained to predict chronological age or biological aging rates based on methylation patterns at specific CpG sites. While these clocks have shown promise as biomarkers for age-related diseases (e.g., Alzheimer's Disease, AD) and environmental stressors, they have been predominantly trained and validated on populations of European ancestry.

• The Gap: There is a critical lack of validation in genetically diverse and admixed populations. Given that methylation patterns are influenced by both environment and genetics (specifically methylation quantitative trait loci, or meQTLs), and that genetic ancestry explains a significant portion of methylation variance, the authors hypothesize that current clocks may fail to generalize to non-European populations, potentially leading to biased risk assessments in precision medicine.

2. Methodology

The study employed a multi-cohort approach to evaluate the portability of methylation clocks and investigate the genetic mechanisms behind performance disparities.

• Cohorts:
  • Primary Cohort (MAGENTA): 621 individuals (AD patients and controls) from the Americas, including White, African American, Puerto Rican, Cuban, and Peruvian participants.
  • Replication Cohorts: >2,500 individuals from three independent datasets: Swedish Whites (NSPHS), African Americans (GENOA study), and African Americans (Grady Trauma Project).
• Clocks Evaluated: A comprehensive set of first-, second-, and third-generation clocks:
  • First-Gen (Chronological Age): Horvath, Hannum, Zhang EN (Elastic Net), Zhang BLUP.
  • Second-Gen (Mortality/Frailty): PhenoAge.
  • Third-Gen (Pace of Aging): DunedinPACE.
  • Variants: Principal Component (PC) versions and ensemble averages of the clocks were also tested.
• Analytical Workflow:
  1. Accuracy Assessment: Calculated Pearson correlations between predicted DNAm age and chronological age, and Mean Absolute Error (MAE), stratified by ancestry.
  2. Disease Association: Tested for "age acceleration" (residuals of predicted vs. chronological age) in AD cases vs. controls across cohorts.
  3. Mechanistic Investigation:
     • Differential Methylation: Compared methylation levels of clock CpGs between African (AFR) and European (EUR) ancestry individuals using EWAS Hub data.
     • Genetic Disruption: Checked if CpG sites were disrupted by genetic variants (SNPs) using gnomAD data.
     • meQTL Analysis: Intersected clock CpGs with known meQTLs from multiple populations (European, South Asian, African American) to determine if genetic variants influencing methylation differ in frequency across ancestries.

3. Key Results

A. Reduced Accuracy in Admixed and African Ancestry Populations

• Correlation Drop: The Horvath clock showed strong correlation with chronological age in White cohorts ($r \approx 0.72$) but significantly lower correlations in African American ($r \approx 0.51$) and Puerto Rican ($r \approx 0.45$) cohorts.
• Ancestry Gradient: Accuracy decreased as the proportion of African genetic ancestry increased. Groups with low African ancestry (Cubans, Peruvians) performed similarly to Whites.
• Replication: This trend was replicated in independent cohorts: Swedish Whites ($r=0.97$) outperformed African American cohorts (GENOA: $r=0.85$; Grady: $r=0.88$).
• Generational Consistency: This performance gap persisted across first-, second-, and third-generation clocks, as well as PC-based versions and ensemble methods.

B. Failure to Predict Alzheimer's Disease Risk

• Lack of Signal: While the Horvath clock detected modest age acceleration in White AD cases compared to controls, it failed to consistently identify accelerated aging in admixed AD cohorts.
• Inconsistency: In many admixed groups, controls showed higher age acceleration than cases, or no significant difference was found.
• Exception: The DunedinPACE clock (third-generation) showed some ability to differentiate AD cases in White, African American, and Puerto Rican cohorts, but failed in Cuban and Peruvian groups.

C. Mechanistic Drivers of Failure

• Differential Methylation: A substantial fraction of clock CpGs are differentially methylated between African and European ancestry individuals (e.g., 23.8% of Horvath CpGs).
• Genetic Disruption is Rare: While many clock CpGs are disrupted by genetic variants (SNPs), these variants are extremely rare (average frequency 0.0001) and unlikely to drive population-level performance differences.
• meQTLs are the Primary Culprit:
  • The Horvath clock has a high fraction of CpGs influenced by meQTLs (77%).
  • Crucially, these meQTLs have significantly higher allele frequencies in African ancestry populations (median 0.068) compared to other groups (0.004–0.046).
  • Many CpGs associated with increased clock error in the MAGENTA cohort were also affected by these African-differentiated meQTLs.
  • In contrast, DunedinPACE (which performed better) contains no CpGs affected by meQTLs.

4. Key Contributions

1. Empirical Evidence of Bias: The study provides robust, multi-cohort evidence that current methylation clocks are not portable across genetic ancestries, specifically underperforming in individuals with African genetic heritage.
2. Identification of meQTLs as a Barrier: The authors pinpointed population-specific meQTLs as a major genetic mechanism causing clock failure, distinguishing this from simple environmental differences or rare variant disruptions.
3. Clinical Implications for AD: The research demonstrates that using these clocks to assess AD risk or biological age in diverse populations may yield false negatives (missing at-risk individuals) or false positives (mislabeling healthy individuals as "accelerated"), potentially exacerbating health disparities.
4. Evaluation of Mitigation Strategies: The study tested and largely dismissed simple fixes like ensemble averaging or Principal Component (PC) versions, suggesting that fundamental changes to clock architecture are required.

5. Significance and Future Directions

• Precision Medicine Warning: The findings serve as a critical warning against the uncritical application of epigenetic biomarkers in global health and social epidemiology. Using Eurocentric clocks on diverse populations could lead to spurious conclusions about aging and disease risk.
• Path Forward: The authors propose two main strategies to improve portability:
  1. Diverse Training Data: Including individuals from multiple genetic ancestries in the training sets of new clocks.
  2. meQTL-Aware Design: Developing clocks that utilize CpG sites not influenced by population-differentiated meQTLs. The success of DunedinPACE (which lacks meQTL-affected CpGs) supports this hypothesis.
• Broader Context: The limitations identified mirror those of Polygenic Risk Scores (PRS), reinforcing the need for "genetics-aware" biomarkers to ensure equitable healthcare outcomes.

In conclusion, the paper establishes that methylation clocks are currently unreliable for predicting age and Alzheimer's risk in genetically admixed populations, primarily due to the influence of population-specific genetic variants (meQTLs) on methylation patterns. Addressing this requires a paradigm shift toward inclusive training data and biologically robust model design.