Common DNA sequence variation influences epigenetic aging in African populations

Abstract

Aging is associated with genome-wide changes in DNA methylation in humans, facilitating the development of epigenetic age prediction models. However, these models have been trained primarily on European-ancestry individuals and none account for the impact of methylation quantitative trait loci (meQTL). To address these gaps, we analyze the relationships between age, genotype, and CpG methylation in 3 understudied populations: central African Baka (n = 35), southern African ‡Khomani San (n = 52), and southern African Himba (n = 51). We show that published prediction methods yield higher mean errors in these cohorts compared to European-ancestry individuals and find that unaccounted-for DNA sequence variation may be a significant factor underlying this loss of accuracy. We leverage information about the associations between DNA genotype and CpG methylation to develop an age predictor that is minimally influenced by meQTL and show that this model remains accurate across a broad range of genetic backgrounds. Intriguingly, we also find that the older individuals and those with lower epigenetic age acceleration carry more genetic variants linked to reduced epigenetic age. These findings support the hypothesis that multiple heritable factors collectively influence healthspan and longevity in human populations.

Fig. 1. Distributions of age-adjusted prediction error across diverse cohorts.

Beeswarm plots show differences in age prediction error for samples predicted to be saliva by dnamage.clockfoundation.org, adjusted for individual age, among Himba (n = 49), ‡Khomani San (n = 46), Baka (n = 35), Hispanic/Latino (n = 69), and European (n = 130) samples across 8 published epigenetic clocks. We tested for significant differences in age-adjusted prediction error among all populations by ANOVA, followed by a Tukey test to identify significant pairwise differences. * indicates an adjusted p-value of <0.05, ** <0.01, and *** <0.001.

Fig. 2. Strong correlation of epigenome-wide association effect sizes across populations.

(A–C) show the correlation of estimated effect sizes of DNA methylation on age for all pairwise comparisons among the Himba (n = 51), ‡Khomani San (n = 52), and Baka (n = 35). The black points indicate effect sizes for CpG sites that were significantly associated with age in at least one of the 3 populations. The red kernel density shows the distribution of effect sizes for CpG sites that were not significantly associated with age in any population. The Pearson correlation between effect sizes for the significant CpGs (black points) and the significance of the correlation is indicated at the bottom right of each panel. (D) is a Manhattan plot depicting associations between DNA methylation and age along the entire genome from a meta-analysis of the individual epigenome-wide association studies run in the three populations. A total of 3211 CpG sites exceed the threshold for significance, a p-value of 0.05 corrected for the number of CpG sites tested (red dotted line). Only non-significant effect sizes with an absolute value less than 1000 were included in the kernel density to restrict axes ranges for visualization purposes.

Fig. 3. Shared cis-genetic architecture of CpG methylation among populations.

(A–C) show the Pearson correlations of estimated effect sizes of SNP genotype on DNA methylation level from baseline cis-meQTL scans of the Himba (n = 51), ‡Khomani San (n = 52), and Baka (n = 35) for cases where the same SNP-CpG relationship was identified in both populations. (D–F) show the Pearson correlations in cis-heritability measures for significantly heritable (p-value < .05) CpG sites across all pairwise combinations of populations. (G–I) show the Pearson correlations of cis-SNP weights on DNA methylation levels estimated from the FUSION regression models for the instances where the same SNP was estimated to have a non-zero weight across different populations, but the selected model was allowed to vary between populations. Weights were scaled within model type. In each panel, the dashed line represents the line of equality. The significance of the correlations is noted beneath the Pearson R values.

Fig. 4. Accounting for meQTL genotype improves power to detect age associations.

(A–D) show p-values for the association between CpG methylation and chronological age from the unadjusted epigenome-wide association study (x-axes) versus p-values from the meQTL-adjusted epigenome-wide association study (y-axes) in the Himba (n = 51), ‡Khomani San (n = 52), Baka (n = 35), and meta-analysis of the 3 populations. Red points show the results of 100 permutations where a random SNP’s genotype was regressed out rather than the true meQTL, whereas black points show the results from regressing out the true meQTL. (E–G) highlight the diamond points from (A–C), respectively, illustrating the influence of genotype on DNA methylation at CpG sites showing particularly large p-value improvements in the adjusted EWAS relative to the unadjusted EWAS.

Fig. 5. Differentiated meQTL influence CpG predictors in published epigenetic clocks.

(A–C) show the allele frequencies of meQTL identified in each of the three African populations, Himba (n = 51), ‡Khomani San (n = 52), and Baka (n = 35), relative to their frequency in 1000 Genomes Phase 3 Europeans. The color of the points corresponds to the density of neighboring points, i.e, yellow points are in high-density regions relative to dark blue points. Red points are meQTL influencing CpGs in published age prediction models. (D–F) show the influence of genotype on baseline methylation level for the meQTL highlighted with a diamond from the top row, examples of meQTL that are invariant in Europeans but segregate in the African population.

Fig. 6. Performance of epigenetic clocks trained on heritable CpG sites versus non-heritable CpG sites.

(A) and (B) are each based on a single model that exhibited the closest to the mean accuracy amongst the heritable models (A) or amongst the non-heritable models (B) in the African test samples (n = 44). (C) depicts the distributions of the mean absolute error from all 100 heritable and non-heritable models as applied to the African test subset (n = 44), Hispanic/Latino (n = 69), European (n = 130), African-American (n = 64), and Japanese (n = 19) cohorts. Boxplots depict the medians of the distributions and the 1.5× interquartile ranges. Models based on CpG predictors that are not significantly impacted by cis-genetic variation exhibit lower absolute error and less bias when applied to our test samples and to out-of-cohort samples than models based on CpG sites that are significantly heritable, except for when applied to the African-American samples. Table 1 shows the mean absolute values and standard deviations of each of the distributions in panel C and exact p-values of the two-sided T-tests. * indicates a p-value of <0.05, ** <0.01, and *** <0.001.

Fig. 7. Relationship between epigenetic aging score (EAS) and aging metrics.

Scatterplots show the relationship between the genotype-based epigenetic aging score (EAS) and various metrics of epigenetic age acceleration for the Himba, ≠Khomani San, and Baka. Each EAS model was built from the respective population’s epigenome-wide association study and baseline meQTL scan results. Individuals’ EAS values were plotted against A chronological age itself, B ‘Intrinsic Epigenetic Age Acceleration’ based on the Horvath multi-tissue age predictor, C ‘Extrinsic Epigenetic Age Acceleration’ based on the Hannum predictor, D epigenetic age acceleration based on PhenoAge, E epigenetic age acceleration based on GrimAge, adjusted for predicted age, and F DNA methylation-based telomere length, adjusted for true age. Only samples predicted to be saliva or blood from the online platform were used: Himba (n = 49), ‡Khomani San (n = 46), Baka (n = 35). Pearson correlations and unadjusted p-values for the significance of the correlation are shown in the bottom right of each panel.