Development of an epigenetic clock resistant to changes in immune cell composition

Abstract

Epigenetic clocks are age predictors that use machine-learning models trained on DNA CpG methylation values to predict chronological or biological age. Increases in predicted epigenetic age relative to chronological age (epigenetic age acceleration) are connected to aging-associated pathologies, and changes in epigenetic age are linked to canonical aging hallmarks. However, epigenetic clocks rely on training data from bulk tissues whose cellular composition changes with age. Here, we found that human naive CD8⁺ T cells, which decrease in frequency during aging, exhibit an epigenetic age 15-20 years younger than effector memory CD8⁺ T cells from the same individual. Importantly, homogenous naive T cells isolated from individuals of different ages show a progressive increase in epigenetic age, indicating that current epigenetic clocks measure two independent variables, aging and immune cell composition. To isolate the age-associated cell intrinsic changes, we created an epigenetic clock, the IntrinClock, that did not change among 10 immune cell types tested. IntrinClock shows a robust predicted epigenetic age increase in a model of replicative senescence in vitro and age reversal during OSKM-mediated reprogramming.

Fig. 1. CpG site changes during T-cell differentiation.

A Experimental design for determining impact of CD8⁺ differentiation on epigenetic clock age prediction. B UMAP dimensionality reduction of CD8⁺ DNA methylation profiles. C Differences between predicted epigenetic age as a function of donor age and CD8⁺ T-cell subset. D Comparison of shared CpG site changes between age in CD8⁺ T cells and CD8⁺ cell subset. E Percent of sites in four epigenetic clocks that are correlated with CD8⁺ T-cell differentiation. Comparison of the F Hannum (p = 1.1 * 10⁻⁷), G Horvath (p = 0.001), H Horvath skin and blood (p = 2.8 * 10⁻⁶), and I PhenoAge (p = 4.8 * 10⁻⁸) epigenetic age acceleration predictions for four CD8⁺ T-cell subsets. *** ANOVA p-value less than or equal to 0.001. Samples derived from N = 7 individuals. Boxplots are centered at median and bound one quartile on each side.

Fig. 2. IntrinClock design strategy and performance.

A Filtering strategy for CpG sites. B Filtering strategy for samples. C Visualization of the filtering process for differentiation-independent age-related CpGs. Blue CpGs (those correlated with age but not with being a naive cell) were included in the feature set, whereas gray CpGs were not. Dashed line indicates linear least-squared regression line of relationship between CpG age correlation and CpG CD8⁺ naive cell correlation. D Correlation between age and IntrinClock predicted age in a variety of tissues from the test set. E–H Individual correlation plots for specific tissues in the test set. I Epigenetic age vs. chronological age correlation plot for semen samples.

Fig. 3. Epigenetic age accelerations measured by different clocks.

A Differences in epigenetic age accelerations in different CD8⁺ subsets generated in this study. Horvath clock predictions overlaid in light gray. B Epigenetic ages of CD8⁺ naive cells and effector memory cells, based on data from GSE66564 and GSE83156. C Epigenetic ages of CD4⁺ naive cells and central memory cells, based on data from GSE121192 and GSE71825. D Epigenetic ages of PBMCs, CD8⁺ naive, CD8⁺ central memory, CD8⁺ combined effector and TEMRA, CD4⁺ naive, CD4⁺ central memory, B-cell naive, B-cell switched memory, CD16⁺CD56_dim NK, and classical monocyte cells. E–I Association of percentage of e, effector memory CD8⁺ cells, f, central memory CD4⁺ cells, g, class-switched B cells, h, CD16⁺ CD56_dim NK cells, and i, classical monocytes with epigenetic age acceleration. Samples derived from N = 9 individuals.

Fig. 4. Distributions of CpG positions.

A Distributions of CpG positions relative to genes in IntrinClock sites that are hyper-methylated with age relative to background. B Distributions of CpG positions relative to genes in IntrinClock sites that are hypo-methylated with age relative to background. C Genomic distribution of IntrinClock CpG positions. D HOMER analysis of the top 12 motifs enriched within 19 bp on either side (5′ or 3′) of IntrinClock sites (40 bp total). *** one-sample proportion t-test p-value < 0.001; * < 0.05.

Fig. 5. Impact on disease and in vitro interventions on the IntrinClock.

A IntrinClock epigenetic age in HIV⁺ and HIV^- individuals, DNA methylation data from GSE67751. Samples derived from N = 92 individuals. B Correlation plot of HIV status, clock residuals, and predicted immune cell type proportions. C IntrinClock epigenetic age in COVID positive and COVID negative individuals, DNA methylation data from GSE167202. Samples derived from N = 525 individuals. D Correlation plot of COVID status, clock residuals, and predicted immune cell type proportions. E Epigenetic reprogramming affects fibroblast predicted IntrinClock age. DNA methylation data from GSE54848. N = 3 independent biological samples. F Induced replicative senescence in fibroblasts leads to an increase in IntrinClock predicted age. DNA methylation data from GSE91069. N = 3 independent biological samples. T-test p-values # < 0.10; * < 0.05; *** < 0.001. Boxplots are centered at median and bound one quartile on each side.