Universal DNA methylation age across mammalian tissues

In separate articles, we described the application of the mammalian methylation array to individual mammalian species^10–19. These studies already demonstrate that one can build dual-species epigenetic age estimators (for example, human–naked mole rat clocks)^10–17, in contrast to first- and second-generation clocks that measure human age^4,20,21 and mortality risk^22,23, respectively. However, it is not yet known whether one can develop a mathematical formula to estimate age in all mammalian species. Here we present three such pan-mammalian age estimators.

The first, basic clock (clock 1), regresses log-transformed chronological age on DNA methylation levels of all available mammals. Although such a clock can directly estimate the age of any mammal, its usefulness could be further increased if its output were adjusted for differences in the maximum lifespan of each species as well, as this would allow biologically meaningful comparisons to be made between species with very different lifespans. To this end, we developed a second universal clock that defines individual age relative to the maximum lifespan of its species; generating relative age estimates between 0 and 1. Because the accuracy of this universal relative age clock (clock 2) could be compromised in species for which knowledge of maximum lifespan is inaccurate, we developed a third universal clock, using age at sexual maturity (ASM) and gestation time instead of maximum lifespan, as these traits are better established and explain over 69% of maximum lifespan variation on the log scale (Supplementary Data 2). This third clock is referred to as the universal log–linear age clock (clock 3). The non-linear mathematical function underlying the age transformation of clock 3 reflects the fact that epigenetic clocks tick faster during development, an observation that led to the establishment of the first pan-tissue clock for humans⁴ (Extended Data Fig. 1a,b,d,e).

Using LOSO cross-validation, the clocks displayed age correlations as high as r = 0.941 (Supplementary Table 1), suggesting their applicability to species not included in the training set. However, for certain species, such as bowhead whales, the basic clock’s predicted epigenetic age poorly aligns with chronological age (Extended Data Fig. 2a).

For the basic clock 1, the mean discrepancy between LOSO DNAmAge and chronological age (Delta.Age) is negatively correlated with species maximum lifespan (r = −0.84, P = 1.0 × 10⁻¹⁹) and ASM (r = −0.75, P = 7.9 × 10⁻¹⁴; Extended Data Fig. 2c,d). Here, the strengths of clocks 2 and 3 come to fore as they adjust for these species characteristics during their construction (Extended Data Fig. 1).

Universal clocks 2 and 3, arguably more biologically meaningful than clock 1, achieve a correlation of r ≥ 0.95 between DNAm transformed age and observed transformed age, respectively (Fig. 1d,f). We will focus on them in the following text. They are pan-tissue clocks offering comparable accuracy in LOFO estimates across numerous tissue types (Fig. 1 and Supplementary Data 4.2). For instance, clock 2 yielded high age correlations in humans (LOFO estimate of r = 0.959 across 20 tissue types), mice (r = 0.948, 26 tissues) and bottlenose dolphins (r = 0.945, two tissues). Fig. 2 displays circle plots for the age correlation estimates in different species sorted by maximum lifespan.

Visual inspection indicates no relationship between age correlation from clocks 2 and 3 and maximum lifespan (dashed line, Fig. 2, circle). While accurately predicting age for the humpback whale and other mammals, the clocks sometimes underestimated bowhead whale reported age (mammalian species index 4.11.1 in Fig. 1a,c), possibly due to overestimation of older whales’ ages by aspartic acid racemization.

Clocks 2 and 3 provide similarly accurate LOSO age estimates between evolutionarily distant species (Supplementary Data 5.2), including dogs (n = 742, 93 breeds, r = 0.94, MAE < 2.28 years), African elephants (r = 0.96, MAE < 4.0 years) and flying foxes (r = 0.97, MAE < 2.3 years) (Fig. 1j–l). Such accuracy demonstrates these clocks’ broad relevance, tapping into conserved age-related mechanisms across mammals, including species not in the training data (Supplementary Data 5.1–5.2).

The three universal clocks performed well for 114 species with fewer than 15 samples each (r ≈ 0.90, MAE ≈ 1.2 years for clocks 1–3; Extended Data Fig. 3a–c), exhibiting strong correlation for relative age (r = 0.91 for clock 2; Extended Data Fig. 3d).

The significantly distinct epigenomic landscape across tissue types^24,25 prompted an assessment of these clocks’ performance in different tissues. We assessed the tissue-specific accuracy of clock 2 for estimating relative age (r = 0.95, Fig. 1d) across 33 distinct tissue types, observing a median correlation of 0.91 and a median MAE for relative age of 0.027 (Supplementary Data 4.3). High age correlation was consistently observed in brain regions: whole brain (r = 0.991), cerebellum (r = 0.963), cortex (r = 0.957), hippocampus (r = 0.954) and striatum (r = 0.935; Extended Data Fig. 5a,d,f,g,i and Supplementary Data 4.3) as well as in organs: spleen (r = 0.982), liver (r = 0.963) and kidney (r = 0.963; Extended Data Fig. 5b,c,e). Blood and skin also showed high estimates of relative age correlations across different species: blood (r = 0.952, MAE = 0.022, 124 species) and skin (r = 0.942, MAE = 0.027, 92 species; Extended Data Fig. 5h,k).

The universal pan-mammalian clocks, derived from multiple tissue types, are essentially pan-tissue clocks. We also constructed analogous clocks solely based on blood (Universal BloodClock 2 and Universal BloodClock 3) and skin (Universal SkinClock 2 and Universal SkinClock 3), the tissues most readily accessible across all species. These tissue-specific clocks tend to demonstrate slightly higher accuracy than the pan-tissue clocks when analyzing their respective tissues. Both the blood and skin clocks exhibit robust age correlations (r ≈ 0.983–0.987 for blood and r ≈ 0.951–0.968 for skin; Extended Data Fig. 4c,g).

We evaluated the cross-sectional associations of lifestyle factors and clinical biomarkers with clocks 2 and 3 in the same cohorts. Robust correlation analysis (biweight midcorrelation (bicor)²⁹) revealed associations of both clocks with inflammation (C-reactive protein, bicor = 0.12, P = 9.9 × 10⁻¹⁶) and dyslipidemia (triglyceride levels, P = 3.2 × 10⁻⁷; Supplementary Table 2). Less significant associations were for fasting glucose levels (P = 0.0093), body mass index (P = 0.011), smoking status (P = 0.027) or physical exercise (P = 0.0064). While these are nominally significant, they are far weaker than those observed with custom clocks for human mortality risk^23,28.

To investigate whether genetic control within a species influences the epigenetic aging rates measured by pan-mammalian clocks, we used human pedigree data from the FHS. Pedigree-based polygenic models of epigenetic age, adjusted for age and sex, yielded significant narrow-sense heritability estimates for clock 2 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}^{2}$$\end{document}h2 = 0.44, P = 3.4 × 10⁻⁸) and clock 3 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}^{2}$$\end{document}h2 = 0.41, P = 4.0 × 10⁻⁷). These heritability estimates for pan-mammalian clocks are on par with that of Horvath’s human pan-tissue clock (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}^{2}$$\end{document}h2 = 0.39, P = 4.0 × 10⁻⁷)⁴.

Epigenetic clocks, such as the human pan-tissue clock, suggest that cellular reprogramming based on the Yamanaka factors (collectively termed as OSKM: OCT4, SOX2, KLF4, and c-MYC) induces age reversal^4,30. To examine whether the universal clocks show a similar age-reversal pattern during reprogramming, we applied clock 2 and clock 3 to a previously published reprogramming dataset in human dermal fibroblasts³¹. We imputed the mammalian methylation array data on the basis of the existing human Infinium array data. Both clocks suggest age reversal after OSKM transduction (Fig. 3c,d). Notably, universal clock 2 showed a decrease in epigenetic age in partially reprogrammed cells after 11 d (Fig. 3c), mirroring observations with human epigenetic clocks^4,30,32.

Growth hormone, generated by somatotropic cells, stimulates body tissue growth, including bone. The somatotropic axis (growth hormone and insulin-like growth factor 1 (IGF-1) levels and their cognate receptors) is central to aging and longevity studies³³. Decreased growth hormone–IGF-1 signaling extends longevity in various species, including mice³⁴. A full-body growth hormone receptor-knockout (KO) (GHRKO) mouse holds the official record for being the longest-lived representative of Mus musculus, living 1 week shy of 5 years³³.

We examined whether reduced growth hormone–IGF-1 pathway activity slows universal pan-mammalian clocks, using three mouse models: (1) Snell dwarf mice, lacking growth hormone production and hence living longer^35,36, (2) full-body GHRKO mice with increased lifespan³⁷ and (3) liver-specific GHRKO mice, showing lowered serum IGF-1 levels but not lifespan increase.

Clock 2 and 3 analyses revealed that Snell dwarf mice exhibit a significantly lower epigenetic age across all considered tissues than wild-type mice (cerebral cortex, Student’s t-test, P = 2.0 × 10⁻⁸; kidney, P = 6.0 × 10⁻¹⁰; liver, P = 1.0 × 10⁻⁷; tail, P = 1.0 × 10⁻⁶; blood, P = 2.0 × 10⁻³; spleen, P = 0.03; Fig. 3e,f). Similarly, full-body GHRKO mice showed lower epigenetic age in several tissues (liver, P = 3.0 × 10⁻⁵; kidney, P = 2.0 × 10⁻⁵; cerebral cortex, P = 0.02; Fig. 3e,f).

Growth hormone receptor signaling stimulates IGF-1 liver synthesis, suggesting that dwarf mice’s epigenetic age reversal may be due to lower circulating IGF-1 levels. This hypothesis, however, is not supported by our epigenetic age measurements of liver-specific GHRKO mice, which exhibit a non-significant difference from the wild-type controls (Fig. 3e). Both clocks 2 and 3 show that the liver-specific GHRKO mice are not epigenetically younger than wild-type mice (Fig. 3e). Unlike full-body GHRKO mice, liver-specific GHRKO mice do not possess a longevity advantage^38,39.

Caloric restriction (CR), which also slows the somatotrophic axis (growth hormone–IGF-1), is associated with prolonged lifespan in several mouse strains^40,41. Previous studies using mouse clocks have shown that CR reduces the rate of epigenetic aging in liver samples^5–7. Using existing methylation data from a murine study of CR⁴², we find that clocks 2 and 3 yield a reduced epigenetic age for mouse liver samples (P = 6.0 × 10⁻¹² for clock 2, P = 7.0 × 10⁻¹⁵ for clock 3; Fig. 3e,f). These results for pan-mammalian clocks align with those obtained with mouse-specific clocks^5,43,44.

TET enzymes are instrumental in active DNA demethylation. Because hydroxymethylation mediated by TET enzymes is prevalent in brain tissue, we applied the universal clocks to brain tissue samples from Tet1-, Tet2- and Tet3-KO mice. Analysis with our universal clocks revealed that Tet3-KO mice exhibit a reduced rate of epigenetic aging (cerebral cortex, P = 3.0 × 10⁻⁹ and striatum, P = 2.0 × 10⁻¹²; Fig. 3e,f). By contrast, significant epigenetic age-reversal effects in brain tissue were relatively weak for Tet1 (cerebral cortex, P = 6.0 × 10⁻³ and striatum, P = 2.0 × 10⁻⁴; Fig. 3e) and could not be observed for Tet2-KO mice (P > 0.6; Fig. 3e).

The differential effect of Tet3 KO versus Tet1 or Tet2 KO in neurons echoes the results of an epigenetic reprogramming study in mouse retinal ganglion cells (Oct4, Sox2 and Klf4 (ref. ⁴⁵)).

Universal clocks, founded on penalized regression models, consist solely of CpG sites that are most predictive of age. Consequently, most other age-related CpG sites are not included in the final regression models.

Imposing a stringent unadjusted significance threshold of α = 10⁻²⁰⁰ limited our analysis to fewer than 1,000 CpG sites across all eutherian species and tissues (Fig. 4a and Supplementary Data 6.1). Of the 832 resulting age-related CpG sites, those most significantly associate with age were cg12841266 (P = 1.4 × 10^−1,001) and cg11084334 (P = 2.6 × 10⁻⁸⁹¹), both located in exon 2 of LHFPL4 (hg38). Notably, cg12841266 exhibited a correlation ≥0.8 in 28 species (Supplementary Data 7; three examples are shown in Fig. 4b–d). Another CpG, cg09710440, resides in exon 1 of LHFPL3 (P = 5.0 × 10⁻⁷⁸⁷), a paralog of LHFPL4 (Fig. 4a, Extended Data Fig. 6 and Supplementary Data 6.1–6.7). As LHFPL4 and LHFPL3 are in human chromosomes 3 and 7, respectively, their consistent age-related gain of methylation is not due to physical proximity.

Beyond LHFPL4 and LHFPL3, other significant gene pairs among the top 30 age-related CpG sites include ZIC1 (chromosome 3) and ZIC2 (chromosome 13), PAX2 (chromosome 10) and PAX5 (chromosome 9) and CELF6 (chromosome 15) and CELF4 (chromosome 18; Supplementary Data 6.1). Located on separate chromosomes, their shared age-related methylation changes cannot be due to physical proximity, indicating a likely functional role in aging. Intriguingly, each gene pair encodes proteins with activities in development.

We extended the age-related EWAS analysis to marsupials and monotremes. The top age-related CpG sites for marsupials were found near genes involved in development, including GRIK2 (P = 8.8 × 10⁻²¹; Supplementary Data 6.8), encoding a neurotransmitter-associated glutamate receptor, and ZIC4 (P = 2.7 × 10⁻¹⁹), encoding a zinc finger protein. The age-related EWAS in monotremes implicated cg22777952 in FOXB1 (P = 8.1 × 10⁻¹⁰; Supplementary Data 6.9), encoding a forkhead box protein. Moderate positive correlation with eutherian age-related methylation changes was observed (r = 0.295 in marsupials, Fig. 4e; r = 0.227 in monotremes, Fig. 4f), in part due to the lower sample numbers in these groups. However, the age effect on methylation of cg11084334 (not cg12841266) in LHFPL4 is preserved in marsupials (P = 4.8 × 10⁻⁷; Fig. 4e) and monotremes (P = 2.4 × 10⁻⁵; Fig. 4f), despite these limitations.

To understand age-related CpG sites across species and tissues, we focused on six tissues with many available species: brain (whole and cortex), blood, liver, muscle and skin. We performed an EWAS meta-analysis on 935 whole brains (18 species–brain tissue categories, eight species), 391 cortices (six species), 4,513 blood samples (56 species), 1,063 livers (ten species), 354 muscle samples (five species) and 2,363 skin samples (65 species; Supplementary Data–). 1.6 1.11

Consistently across all tissues, CpG sites with positive age correlations outnumbered those with negative correlations (Extended Data Fig. 6). While many age-related cytosines were either specific to individual organs (Supplementary Data 6.2–6.7) or shared between several organs, 51 CpG sites (48 positively and three negatively age related) were common to all five organs (Fig. 4g and Supplementary Table 3). In total, 35 genes were proximal to the 48 positive CpG sites, and three genes were proximal to the three negative CpG sites. Interestingly, 20 of these 35 genes encode transcription factors (TFs), including 11 homeobox proteins, seven zinc finger TFs and two paired box proteins, involved in developmental processes including embryonic development (Supplementary Table 3). The relevance of this becomes evident below, where the chromatin state, function and tissue-specific accessibility associated with the location of age-related CpG sites are described.

We observed that 57% of the top 1,000 positively age-related CpG sites were situated in a CpG island (human genome), while only 2% of the top 1,000 negatively age-related CpG sites resided there (EWAS of age across all tissues; Supplementary Data). 6.1

To understand the epigenetic context of age-related CpG sites, we accessed a detailed universal chromatin state annotation of the human genome. This resource, derived from 1,032 experiments mapping 32 chromatin marks across 100+ human cell and tissue types⁴⁶ (Fig. 4h, Extended Data Fig. 7 and Supplementary Data 8.2–8.9), allowed us to overlay the positions of the top 1,000 age-related CpG sites. We found that positively age-related CpG sites were significantly enriched in states associated with polycomb repressive complex 2 (PRC2)-binding sites (states BivProm1, BivProm2, ReprPC1). These CpG sites localized to PRC2-binding sites, as defined by embryonic ectoderm development (EED), enhancer of zeste 2 PRC2 subunit (EZH2) and PRC2 subunit (SUZ12) binding (the first row of Fig. 4h). This PRC2 enrichment could be observed for all tissue types collectively (odds ratio (OR) = 22.8, hypergeometric P = 1.9 × 10⁻⁴⁴⁹) and when analyzed individually: blood (OR = 29.8, P = 2.9 × 10⁻⁵¹⁰), liver (OR = 14.3, P = 7.3 × 10⁻³³⁸), skin (OR = 14.3, P = 9.9 × 10⁻³³⁷), cortex (OR = 6.5, P = 3.7 × 10⁻¹⁶³) and brain (OR = 3.2, P = 9.7 × 10⁻⁵⁷). Indeed, the majority of the top 1,000 positively age-related CpG sites were significantly enriched in PRC2-binding sites: 80.8% (808 CpG sites) in blood, 67.5% in liver and 67.2% in skin (Supplementary Data 8.1).

PRC2, a transcriptional repressor complex, is a key contributor to H3K27 methylation, a chromatin modification linked to transcriptional repression⁴⁷. Importantly, PRC2-mediated histone 3 lysine 27 (H3K27) methylation is crucial for establishing bivalent promoters, which house histones with both H3K27 trimethylation (H3K27me3) and histone 3 lysine 4 trimethylation (H3K4me3). As such, it is consistent that positively age-related CpG sites are also found to be enriched in bivalent promoter states (rows 3 and 4 of Fig. 4h). They show even greater presence in a bivalent state associated with more H3K27me3 than H3K4me3 (BivProm2) than in BivProm1, associated with more balanced levels of these histone modifications⁴⁶. The top EWAS hit, LHFPL4 cg12841266, in a bivalent state (BivProm2) and PRC2-binding region (EED-, EZH2-, SUZ12-binding sites), exemplifies this (Supplementary Data 8.1). These mammalian results echo those from human studies^48,49, in which tissue-independent age-related gain of methylation is characterized by cytosines that are located in PRC2-binding sites and bivalent chromatin domains.

We found that ORs for the overlap between positively age-related CpG sites and PRC2-binding sites were markedly higher in proliferative tissues (blood, skin, liver) than in non-proliferative tissues (skeletal muscle, brain, cerebral cortex; Fig. 4h). The distinction between proliferative and non-proliferative tissues also manifested when considering negatively age-related CpG sites (those that lose methylation levels with age). In highly proliferative tissues (blood, skin), age-related loss of methylation was seen in CpG sites located in select heterochromatin (HET1, HET7), which are marked by histone 3 lysine 9 trimethylation, or inactive chromatin states (Quies1, Quies2), as listed in Supplementary Data 8.2 and Vu & Ernst⁴⁶. Conversely, in non-proliferative tissues, age-related methylation loss could be seen in the exon- and high-expression-associated transcription state TxEx4 (OR = 12.9, P = 1.6 × 10⁻⁵² in the cerebral cortex and OR = 6.7, P = 3.7 × 10⁻²² in skeletal muscle). TxEx4 is far less enriched with age-related cytosines that lose methylation in proliferative tissues such as blood (OR = 2.6, P = 1.7 × 10⁻⁴) or skin (OR = 0.7, P = 0.25).

Our chromatin state analysis of age-related loss of methylation demonstrated that it is important to distinguish proliferating tissues (blood, skin) from non-proliferative tissues (brain, muscle). Consequently, we examined the correlation between DNA replication and methylation. Late-replicating genome domains, prone to partial methylation, show pronounced methylation loss in solo WCGW cytosines (CpG sites flanked by A or T on either side⁵⁰). We overlaid the top 1,000 age-related CpG sites (positive or negative) on the reported late-replicating domains, which are enriched with partially methylated domains (PMDs)⁵⁰. As previously reported for human tissues⁵⁰, we observed age-related loss of methylation in PMDs and solo WCGW sites in mammalian tissues that proliferate, such as blood and skin (Extended Data Fig. 8 and Supplementary Data 9). Notably, the top 1,000 negatively age-related CpG sites overlap significantly with CpG sites that are both common PMDs and solo WCGW sites (hg19): skin (OR = 7.9, P = 1.6 × 10⁻⁹⁰), blood (OR = 5.3, P = 1.5 × 10⁻⁵⁰) and all tissues (OR = 7.3, P = 4.4 × 10⁻⁸¹; Extended Data Fig. 8). Contrastingly, non-proliferative tissues, such as the brain, show a different pattern: CpG sites losing methylation with age are enriched in highly methylated domains (HMDs, OR = 3.3, P = 1.9 × 10⁻⁷⁴) over PMDs (OR = 0.2, P = 4.9 × 10⁻⁶⁴). CpG sites gaining methylation with age show weaker overlap with both PMDs and highly methylated domains. Similar findings were observed in late-replicating mouse genome domains (mm10; Extended Data Fig. 8). In summary, pan-mammalian CpG sites losing methylation with age are enriched in late-replicating regions of highly proliferative tissues.

Analysis of CpG sites positively correlated across all tissues revealed ‘nervous system development’ as a highly significant gene ontology (GO) term (P = 1.3 × 10⁻²⁰³). This term was consistent across blood (P = 1.9 × 10⁻²²⁴), liver (P = 2.6 × 10⁻¹³⁷), muscle (P = 3.4 × 10⁻¹⁴), skin (P = 1.7 × 10⁻¹⁴⁵), brain (P = 6.4 × 10⁻³⁵) and cortex (P = 1.0 × 10⁻⁷⁸). Other significant GO terms included ‘developmental process’, ‘regulation of RNA metabolic process‘, ‘nucleic acid-binding TF activity‘, ‘pattern specification’ and ‘anatomical structure development’ (Fig. 7). The GREAT analysis also indicated that a significant proportion of the top 1,000 positively age-related CpG sites are located in PRC2 target sites (P = 8.3 × 10⁻²¹²), which was also true for individual core PRC2 subunits (SUZ12, EED or EZH2; Fig. 7). It follows that these CpG sites were also enriched in promoters with the H3K27me3 modification in embryonic stem cells for all tissues (P = 1.9 × 10⁻²⁶⁹), blood (P = 2.8 × 10⁻²⁸⁵), liver (P = 3.4 × 10⁻¹⁸²), muscle (P = 9.0 × 10⁻¹⁷), skin (P = 7.8 × 10⁻²⁰²), brain (P = 1.4 × 10⁻⁵⁴) and cortex (P = 2.7 × 10⁻¹¹⁵; Fig. 7). As PRC2 plays a critical role in development, these results reinforce the epigenetic link between development and aging. This connection is supported by observations that developmentally compromised mice, due to growth hormone receptor (GHRKO) ablation or anterior pituitary gland removal (Snell mice), show reduced rates of epigenetic aging in multiple tissues, as measured by universal epigenetic age clocks (Fig. 3e).

While positively age-related CpG sites (across all tissues) were enriched in 2,961 GO or Molecular Signatures Database terms at a false discovery rate of 0.05 (Supplementary Data 10.1), negatively age-related CpG sites were enriched in only three. Negatively age-related CpG sites in brain and muscle were enriched in genes associated with circadian rhythm (brain, P = 3.3 × 10⁻¹⁵; cerebral cortex, P = 4.0 × 10⁻¹⁹; muscle, P = 2.3 × 10⁻⁸; Fig. 7) and Alzheimer’s disease-related gene sets (for example, P = 1.8 × 10⁻²⁹ in brain and P = 2.4 × 10⁻²² in the cerebral cortex in Fig. 7). These CpG sites also overlapped with gene sets related to mitochondrial function in brain, cerebral cortex and muscle (for example, P = 3.6 × 10⁻⁷; Supplementary Data 10.2).

The GREAT analysis showed enrichment of both positively and negatively age-related CpG sites in mortality or aging gene sets, cancer (Fig. 7) and targets of three Yamanaka factors: SOX2, MYC and OCT4 (Supplementary Data 10.3). Of the 341 genes proximal to positively age-related CpG sites, 162 were implicated in mortality or aging (P = 6.3 × 10⁻¹³⁸; Fig. 7). Similar enrichments were seen in specific tissues: blood (P = 3.8 × 10⁻¹⁸⁴), liver (P = 2.7 × 10⁻¹¹²), muscle (P = 6.2 × 10⁻⁵), skin (P = 9.1 × 10⁻⁸⁴), combined brain tissues (P = 1.2 × 10⁻²¹) and the cerebral cortex (P = 5.0 × 10⁻⁵⁰).

As inflammation increases with aging, we assessed the overlap with inflammation-related gene sets (Supplementary Data 10.4). Positively age-related CpG sites are enriched in the gene set associated with inflammation in the murine pancreas (all tissues, P = 8.4 × 10⁻²¹ and skin, P = 9.4 × 10⁻²⁰). Negatively age-related CpG sites are enriched in Toll-like signaling (GO:0034121↗) genes (muscle, P = 9.2 × 10⁻⁸).

Both positively and negatively age-related CpG sites are enriched in immunologic signature gene sets associated with interleukin (IL; for example, IL-6, IL-23) and transforming growth factor (TGF)-β1 exposure in type 17 helper T cells (Supplementary Data 10.4) for notably brain (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{negative}}}$$\end{document}Pnegative = 6.1 × 10⁻¹¹) and cerebral cortex (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{negative}}}$$\end{document}Pnegative = 9.1 × 10⁻⁸) and, to a lesser extent, skin (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{positive}}}$$\end{document}Ppositive = 4.0 × 10⁻⁴) tissues.

Concerns that these highly significant enrichments may be a result of potential biases in the mammalian methylation array platform could be discounted after sensitivity analysis, as reported in Supplementary Note. 3

We used the CellBase⁵² and ENCODE databases⁵³ to annotate CpG sites with binding sites for 68 TFs identified through chromatin immunoprecipitation followed by sequencing (ChIP–seq) in 17 cell types. If a CpG site overlapped with the binding site of a TF (hg19) in at least one cell type, it was assigned to that TF. Analysis of the most significant age-related CpG sites across mammals showed that the REST TF was the most significant TF for the top 1,000 positively age-related CpG sites across all tissues (OR = 8.4, P = 3.1 × 10⁻⁵⁴), especially in proliferative tissues such as blood (OR = 5.8, P = 2.7 × 10⁻³²), skin (OR = 8.7, P = 6.8 × 10⁻⁵⁹) and liver (OR = 5.4, P = 1.5 × 10⁻²⁸). REST TF enrichment was less significant in non-proliferative tissues such as muscle (OR = 1.8, P = 2.2 × 10⁻³), cerebral cortex (OR = 1.6, P = 0.01) and brain (OR = 1.4, P = 0.09; Extended Data Fig. 9 and Supplementary Data 11).

REST TF ChIP–seq analysis was performed on five cell lines, including a human embryonic stem cell line (Supplementary Data 11.1). REST is known for repressing neuronal genes in non-neuronal tissues, which could explain the weak enrichments in brain regions. Notably, CpG cg12841266 near LHFPL4 is within the REST-binding region.

Substantial binding enrichments were observed for transcription factor 12 (TCF12) and histone deacetylase 2 (HDAC2). TCF12 is part of the basic helix–loop–helix (bHLH) E-protein family, associated with neuronal differentiation, and top positively age-related CpG sites are proximal to another bHLH gene, NEUROD1 (Supplementary Table 3 and Supplementary Data 11). Lower enrichments were noted for CCCTC-binding factor (CTCF) and Nanog homeobox (NANOG). For the top 1,000 negatively age-related CpG sites, fewer significant TF binding enrichments emerged, with JUN (c-Jun) in blood (OR = 2.8, P = 2.6 × 10⁻⁹) and brain (OR = 1.5, P = 0.024; Extended Data Fig. 9) being exceptions.

We studied whether the top 1,000 positively and negatively age-related CpG sites neighbor genes with age-correlated mRNA levels. Using GenAge⁵⁴ and Enrichr^55,56 databases, we scrutinized age-specific transcriptome-wide association studies (TWAS) in four mammalian species. The EWAS–TWAS overlap analysis (Fig. 7, Extended Data Fig. 10a and Supplementary Data 12) indicates significant overlaps between age-related CpG sites and transcriptomic age changes in several species, including Genotype–Tissue Expression (GTEx) human tibial nerve samples, normal monkey hippocampal samples (P = 9 × 10⁻¹⁵) and various rat and mouse tissues. However, the age-related EWAS and TWAS overlap is generally weak and tissue specific.

We compared proximal genes of the top 1,000 positively and negatively age-related CpG sites with the top 2.5% of genes implicated in various human genome-wide association studies (GWAS). Notable enrichments were seen in genes associated with waist-to-hip ratio for positively age-related CpG sites in livers (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{positive}}}$$\end{document}Ppositive = 1.0 × 10⁻¹⁶), and with human length at birth for positively age-related CpG sites in the cortex (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{positive}}}$$\end{document}Ppositive = 1.0 × 10⁻¹²) and liver (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{positive}}}$$\end{document}Ppositive = 2.0 × 10⁻¹⁰; Fig. 7). Significant enrichments (defined here as nominal P < 5.0 × 10⁻⁴) were also seen with genes linked to mother’s longevity (mother attained age; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{positive}}}$$\end{document}Ppositive = 2.0 × 10⁻⁴; Fig. 7, Extended Data Fig. 10b and Supplementary Data 13.1–13.7), human longevity for negatively age-related CpG sites in muscle (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{negative}}}$$\end{document}Pnegative = 8.0 × 10⁻⁶), epigenetic age acceleration on the mortality clock (GrimAge \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{positive}}}$$\end{document}Ppositive = 7.0 × 10⁻⁷ in muscle), age-related macular degeneration (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{positive}}}$$\end{document}Ppositive = 2.0 × 10⁻⁸ in all tissues), Alzheimer’s disease (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{negative}}}$$\end{document}Pnegative = 1.0 × 10⁻⁴ in brain), leukocyte telomere length (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{negative}}}$$\end{document}Pnegative = 3.0 × 10⁻¹³ in muscle and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{negative}}}$$\end{document}Pnegative = 2 × 10⁻¹¹ in brain) and age at menarche (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{positive}}}$$\end{document}Ppositive = 4.0 × 10⁻⁵ in all tissues). Overall, our GWAS overlap analysis indicates that pan-mammalian age-related CpG sites are proximal to genes influencing human development (birth length, menarche), obesity and longevity.

We calculated the percentage of cells per individual with the respective peak. A strong, statistically significant negative correlation (Fig. 8b) was found between age and the number of cells with the ATAC peak overlapping cg12841266 in LHFPL4. This shows that, with age (as methylation increases), open chromatin cell number decreases. Of 17 gene regions, 16 correlated negatively with age, with seven being statistically significant (P < 0.05; Fig. 8a). The hypermethylated sites were highly enriched for this age-associated accessibility loss (P < 0.001; Fig. 8b). The significant genes (LHFPL4, TLX3, ZIC2, PAX2, NR2E1, NEUROD1, DLX6-AS1) are related to developmental processes (Supplementary Table 3). ZIC5, another Zic family gene, also showed a nearly significant negative age correlation (r = −0.54, P = 0.07; Supplementary Data 14.1). No scATAC-seq signal was detected in the cg09710440 region of LHFPL3, possibly due to proximity to a bivalent gene’s TSS (232 bp).

We examined whether the seven significant ATAC peaks identified a particular cell type subset. Due to the sparsity of scATAC-seq data, we determined the fraction of each cell group containing at least one of these regions. We found that stem cell–progenitor populations had a higher proportion of open chromatin at these sites than differentiated cells (mean of 14.9% versus mean of 2.9%; Fig. 8c). This suggests that the observed age-related reduction of open chromatin states could be due to the loss (for example, death or differentiation) of progenitor cells in the tissue.

We studied three cell groups: hematopoietic stem cells (HSCs), progenitor cells and differentiated cells. Age showed a negative correlation with the percentage of HSCs (r = −0.69, P = 0.01) but no significant correlation with progenitor or differentiated cells (Fig. 8g–i). Next, we analyzed the correlation between age and the proportion of cells containing an ATAC peak in at least one of the seven significant CpG regions (Fig. 8d–f). Differentiated cells demonstrated a significant loss of ATAC signal in these regions with age (r = −0.68, P = 0.01; Fig. 8f), whereas no change was seen in HSCs or progenitor cells (Fig. 8d,e). This suggests that these regions, gaining methylation and losing accessibility with age, belong to a differentiated cell population. Lastly, analyzing increasing lists of positively age-related CpG sites, we noted that the percentage of cells with an ATAC peak at these locations decreasing with age in human BMNCs (median correlation < −0.2 across the top 500 or 1,000 positively age-related CpG sites).

We tested whether our human HSC findings extended to murine HSCs by analyzing another public scATAC-seq dataset from murine HSCs with four replicates each in young (10-week) and old (20-month) mice⁵⁹. This dataset provided access to our age-related CpG sites in 4,492 young and 3,300 old HSCs. Of the top 35 positively age-related CpG sites, 33 overlapped with ATAC peaks (Supplementary Data 14.2). We then calculated the proportion of HSCs in each age group with the respective peak. The proportion of old HSCs with a peak near Lhfpl4 was not significantly different from that of young HSCs (OR = 0.94, P = 0.7), implying no observable age-related chromatin compactification in murine HSCs. This was also true for the other 32 CpG sites and their associated peaks. Contrarily, the proportion of old HSCs with an ATAC peak was significantly higher than that of young HSCs for five CpG sites (near Bdnf, Isl1, Twist1, Nr2e1, Sall1; Fisher exact P value < 0.05; Supplementary Data 14.2), indicating age-related chromatin opening (Fig. 8j), aligning with Itokawa et al.’s report⁵⁹.

All local ethical guidelines were followed, and necessary approvals from respective human ethical review boards and animal ethical committees were duly obtained. Details can be found in Supplementary Notes,and. 1 2 4

Data collection and analysis were not performed blind to the conditions of the experiments. In the ensuing sections, we delineate the quality-control measures for our samples and the statistical methods employed in each analysis, with additional details provided in Supplementary Notesand. 1 5

We used a subset of the data from the Mammalian Methylation Consortium for which age information was available⁹. The tissue samples are described in Supplementary Data 1.1–1.4, and related citations are listed in Supplementary Notes 1 and 2. We used the SeSAMe normalization method⁸⁵.

Our epigenetic clocks were trained and evaluated on samples with highly confident age assessments (less than 10% error). We focused on typical aging patterns, hence excluding tissues from preclinical anti-aging or pro-aging intervention studies.

Species characteristics such as maximum lifespan (maximum observed age) and ASM were obtained from an updated version of AnAge⁸⁶ (https://genomics.senescence.info/species/index.html↗). To facilitate reproducibility, we have posted this modified and updated version of AnAge in Supplementary Data 1.13.

We applied elastic net regression models to establish three universal mammalian clocks for estimating chronological age across all tissues (n = 11,754 from 185 species) in eutherians (n = 11,439 from 174 species), marsupials (n = 210 from nine species) and monotremes (n = 15 from two species). The three elastic net regression models, implemented using the glmnet 4.1-7 package in R, corresponded to different outcome measures described in the following:log-transformed chronological age: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\log ({\rm{Age}}+2)$$\end{document}log(Age+2), where an offset of 2 years was added to avoid negative numbers in case of prenatal samples,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-{\rm{log}}(-{\rm{log}}({\rm{RelativeAge}}))$$\end{document}−log(−log(RelativeAge)) andlog–linear transformed age.

DNAmAge estimates of each clock were computed via the respective inverse transformation. Age transformations used for building universal clocks 2 and 3 incorporated a selection of three species characteristics: gestational time, age at sexual maturity () and maximum lifespan. All of these species variables surrounding time are measured in units of years. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({\rm{GestationT}})$$\end{document} ( ) GestationT \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{ASM}}$$\end{document} ASM \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({\rm{MaxLifespan}})$$\end{document} ( ) MaxLifespan

To adjust for age, we defined epigenetic age acceleration (AgeAccel) as the raw residual resulting from regressing DNAmAge (from universal clocks 2 and 3) on chronological age. By definition, the resulting AgeAccel measure is not correlated with chronological age.

We applied our universal clocks 2 and 3 to 4,651 individuals from (1) the FHS Offspring cohort (n = 2,544 Caucasians, 54% women)⁸⁷ and (2) the WHI cohort^88,89 (n = 2,107, 100% women; Supplementary Note 4). Methylation levels were profiled in blood samples using the Illumina 450k arrays. The FHS cohort had a mean (s.d.) age of 66.3 (8.9) years at blood draw, with 330 deaths during an average follow-up of 7.8 years. The WHI cohort, which enrolled postmenopausal women 50–79 years in age, consisted of three ethnic groups: 47% of European ancestry, 32% African Americans and 20% of Hispanic ancestry. These groups exhibited similar age distributions, with a mean (s.d.) age of 65.4 (7.1) years, and a mean (s.d.) follow-up time of 16.9 (4.6) years. During the follow-up, 765 women died.

None of the samples from the murine anti-aging studies were used in the training set of the universal clocks, that is, these are truly independent test data. Clocks 2 and 3 were evaluated in five mouse experiments (independent test data): (1) Snell dwarf mice (n = 95), (2) GHRKO experiment 1 (GHRKO, n = 71 samples), (3) GHRKO experiment 2 (n = 96 samples), (4) three Tet experiments: Tet1 KO (n = 64), Tet2 KO (n = 65) and Tet3 KO (n = 63) and (5) CR (n = 95). Details can be found in Supplementary Note 6.

In our primary EWAS of age, we focused on samples from eutherians (n = 65 species) for which each species has at least 15 samples from the same tissue type. In secondary analyses, we also studied aging effects in marsupials (n = 4 marsupial species that had at least ten same-tissue type samples) and monotremes (only n = 2 species). Data distribution was assumed to be normal, but this was not formally tested.

Our meta-analysis for EWAS of age in eutherian species combined Pearson correlation test statistics across species–tissue strata that contained at least 15 samples each. The minimum sample size requirement resulted in 143 species–tissue strata from 65 eutherian species (Supplementary Data). To counter the dependency patterns resulting from multiple tissues from the same species, the meta-analysis was carried out in two steps. First, we meta-analyzed the EWAS of different tissues for each species separately. These tissue-specific summary statistics were combined within the same species to represent the EWAS results at species level. Second, we meta-analyzed the resulting 65-species EWAS results across species to arrive at the final meta-EWAS of age. In each meta-analysis step, we used the unweighted Stouffer’s method as implemented in R. In more detail, we gathered 68 blood samples from 27 distinct lemur species and 23 skin samples from 23 distinct lemur species, each species–tissue stratum with at most three samples. We therefore combined those 68 blood samples to perform blood EWAS in lemurs. Similarly, we combined the 23 skin samples for skin EWAS in lemurs. As listed in Supplementary Data, the combined species in lemurs was denoted by Strepsirrhine in the column ‘Species Latin Name’. 1.5 1.5

EWAS of age in marsupials was based on a two-step meta-analysis in which we relaxed the threshold of sample size in the species–tissue category to n ≥ 10 (Supplementary Data 1.12). Due to a small sample size in monotremes (n = 15), we combined all monotreme samples into a single dataset.

Polycomb repressive complex annotations were defined based on the binding of at least two transcriptional factor members of polycomb repressor complex 1 (PRC1 with subgroups RING1, RNF2, BMI1) or PRC2 (with subgroups EED, SUZ12 and EZH2) in 49 available ChIP–seq datasets from ENCODE⁵³.

We identified 640 and 5,287 CpG sites in the array that were located in regions bound by PRC1 and PRC2, respectively. We performed a one-sided hypergeometric analysis to study both enrichment (OR > 1) and depletion (OR < 1) patterns for our age-related markers based on the top 1,000 CpG sites increased with age and the top 1,000 CpG sites decreased with age from the EWAS of age.

To annotate our age-related CpG sites based on chromatin states, we assigned a state for all our mammalian CpG sites based on a recently published universal ChromHMM chromatin state annotation of the human genome⁴⁶. The underlying hidden Markov model was trained with over 1,000 datasets of 32 chromatin marks in more than 100 cell and tissue types. This model then produced a single chromatin state annotation per genomic position that is applicable across cell and tissue types, as opposed to producing an annotation that is specific to one cell or tissue type. A total of 100 distinct states were generated and categorized into 16 major groups according to the parameters of the model and external genome annotations⁴⁶ (described in Supplementary Data 8.2).

We performed a one-sided hypergeometric analysis to study both enrichment (OR > 1) and depletion (OR < 1) patterns for our age-related markers based on the top 1,000 CpG sites with a positive correlation with age and the top 1,000 CpG sites with a negative correlation with age across different eutherian species.

The annotation of late-replicating domains (hg19 and mm10) was obtained from Zhou et al.⁵⁰, as described in Supplementary Note 5.

We applied the GREAT analysis software tool⁵¹ to the top 1,000 positively age-related and the top 1,000 negatively age-related CpG sites from the EWAS of age. GREAT implemented foreground–background hypergeometric tests over genomic regions where we input all CpG sites of the mammalian array as background and the genomic regions of the 1,000 CpG sites as foreground. This approach yielded hypergeometric P values that were not confounded by the number of CpG sites within a gene (Supplementary Note 5).

Our EWAS–TWAS-based overlap analysis related the gene sets found by our EWAS of age with the gene sets from our in-house TWAS database. The TWAS database, along with our analytical approaches, is described in Supplementary Note. 5

Our EWAS–GWAS overlap analysis linked the gene sets discovered in our EWAS of age with those identified in published large-scale GWAS studies of various phenotypes (Supplementary Note). 5

We used the CellBase database⁵², incorporating ENCODE⁵³ TF binding sites for our analysis (Supplementary Note 5).

Recent advances have enabled the sequencing of ATAC profiles within single cells, enabling assessment of the proportion of cells containing an open chromatin region⁵⁸. We cross-referenced the top 35 CpG sites with positive age correlation across mammalian tissues with publicly available scATAC-seq data (Supplementary Table 3). We downloaded 10x Multiome count data in AnnData format as H5AD from the Gene Expression Omnibus (accession number GSE194122↗). The ATAC array data were managed using the Python package anndata⁹². hg38 ATAC peak locations were extracted from the metadata ‘.var’ section using anndata. Peak locations were overlapped with probe locations using GenomicRanges⁹³ for the top 35 CpG sites. The overlapping peaks were then used to extract the processed counts for each cell. The proportion of cells containing an ATAC peak for each individual sample was calculated. A correlation was calculated by comparing this value against the age of each individual sample. The cell type for each barcode was extracted from the observable object. We subsequently computed the proportion of each cell type containing an ATAC peak in one of the seven significantly correlated regions (LHFPL4, TLX3, ZIC2, PAX2, NR2E1, NEUROD1 and DLX6-AS1). Progenitor cells were grouped as MK/E progenitors, G/M progenitors, lymph progenitors and proerythroblasts, and differentiated cells were grouped as CD14⁺ monocytes, CD16⁺ monocytes, CD8⁺ T naive, CD8⁺ T, CD4⁺ T naive, CD4⁺ T activated, naive CD20⁺ B, B1 B, transitional B and NK. The percentage of each of the three populations (HSC, progenitor and differentiated cells) was calculated, and the proportion of cells containing an ATAC peak in one of the seven significantly correlated regions was calculated. To confirm enrichment for the hypermethylated sites showing decrease in chromatin accessibility with age, we randomly selected 1,000 sets of 17 ATAC peaks and compared the mean correlation with age of the selected regions to the 1,000 sampled sets of regions.

The following URLs are available: AnAge (https://genomics.senescence.info/species/index.html↗), GREAT (http://great.stanford.edu/public/html/↗), late-replicating domains (https://zwdzwd.github.io/pmd↗), UCSC Genome Browser (http://genome.ucsc.edu/index.html↗).

Further information on research design is available in thelinked to this article. Nature Portfolio Reporting Summary