Somatic mutation as an explanation for epigenetic aging

Abstract

DNA methylation marks have recently been used to build models known as epigenetic clocks, which predict calendar age. As methylation of cytosine promotes C-to-T mutations, we hypothesized that the methylation changes observed with age should reflect the accrual of somatic mutations, and the two should yield analogous aging estimates. In an analysis of multimodal data from 9,331 human individuals, we found that CpG mutations indeed coincide with changes in methylation, not only at the mutated site but with pervasive remodeling of the methylome out to ±10 kilobases. This one-to-many mapping allows mutation-based predictions of age that agree with epigenetic clocks, including which individuals are aging more rapidly or slowly than expected. Moreover, genomic loci where mutations accumulate with age also tend to have methylation patterns that are especially predictive of age. These results suggest a close coupling between the accumulation of sporadic somatic mutations and the widespread changes in methylation observed over the course of life.

Practically since the elucidation of the DNA double helix, it has been postulated that progressive damage to this fundamental structure is the cause of aging^1,2. The primary support for this theory relates to somatic mutations, which accumulate in the genomes of most tissues and species throughout life^2–4. Such accumulation has been associated with multiple characteristics of old age, including immune dysfunction⁵, neurodegeneration^6,7 and cancer⁸.

Aging has also been associated with other major types of molecular changes beyond DNA mutations, leading to a debate on which of these aging ‘hallmarks’ are fundamental causes^9,10. In particular, much recent attention has been given to associations of age with DNA methylation, a dynamic epigenetic mark found primarily at CG dinucleotides (CpG sites) throughout the genome¹¹. CpG methylation has diverse functional consequences, including X chromosome inactivation¹², chromatin and transcriptional regulation¹³, cell-type specification and maintenance of pluripotency^14–16. DNA methylation patterns have been found to change regularly over the course of life, prompting the creation of statistical models termed ‘epigenetic clocks’, which attempt to predict an individual’s age using their DNA methylation profile^17,18. Subsequent research has shown that epigenetic clock predictions correlate with a host of age-related biological attributes, including frailty, Alzheimer’s disease, all-cause mortality and time to death^19–22. Such observations have bolstered epigenetic theories of aging, which propose that progressive remodeling of the epigenome leads to aging phenotypes through the dysregulation of gene expression, cellular function and senescence^23–25. However, the degree to which epigenetic changes are direct causes of aging remains unclear.

Despite the separate interest in DNA mutations and DNA methylation as theories of aging, the relationship between the two processes is not well understood. One recent study reported that somatic mutations in DNA-binding sites for ten–eleven translocation methylcytosine dioxygenase 1 (TET1), the primary enzyme involved in the removal of methylation marks²⁶, are associated with local hypermethylation²⁷. Another demonstrated an association among somatic mutations, subsequent clonal expansion of blood cells and accelerated epigenetic aging²⁸. Most other research linking DNA sequence and methylation has focused on inherited germline variants rather than acquired somatic mutations, such as efforts to identify methyl quantitative trait loci linking common polymorphisms to methylation levels of specific CpG sites^29,30.

An intrinsic biochemical connection between DNA mutation and methylation occurs at 5-methylcytosine residues^3,31,32, which spontaneously deaminate over time to yield thymine³³. A prerequisite for this mutational event is cytosine methylation, relating somatic mutation to prior epigenetic modification of DNA. Conversely, a prerequisite for DNA methylation is the presence of a CpG site, which may be eliminated by prior somatic mutation. Given this interdependence, we considered that the separate links that have been established between DNA mutation and aging and between DNA methylation and aging might each reflect a common underlying process whereby methylation potentiates mutation and/or mutation potentiates changes in methylation.

To explore this hypothesis, we set out to comprehensively examine the relationship between somatic DNA mutations and DNA methylation in large collections of human tissue samples characterized for both layers of molecular information. In what follows, we identify several types of interaction between somatic mutation and DNA methylation, both one-to-one and one-to-many (Extended Data Fig. 1a). Based on these findings, we use somatic mutations as a surrogate for epigenetic marks in measures of aging, indicating the degree to which epigenetic aging is explained by somatic mutations (Extended Data Fig. 1b). These results are initially demonstrated in cancer cohorts, which have by far the greatest numbers of individuals with matched DNA mutation and methylation profiles. We then show that our findings extrapolate to normal tissues drawn from the same patients.

Results

Genome-wide hypomethylation of mutated CpG sites

To study the connections between somatic mutations and DNA methylation marks, we first analyzed data from human patients cataloged in The Cancer Genome Atlas (TCGA)^34–36 and the Pan-Cancer Analysis of Whole Genomes (PCAWG)³⁷. Tumor biopsies had been drawn from a diversity of tissue types and characterized by whole-exome sequencing (TCGA, 8,680 exomes across 33 tissues) or whole-genome sequencing (PCAWG, 651 genomes across 3 tissues). In each case, DNA from the tumor sample was compared to a second DNA sample drawn from the same individual, with differences used to define somatic mutations (typically comparing the tumor DNA sequence to that in whole blood; Methods). These data had been complemented by methylation profiling of the same tissues using the Illumina Infinium HumanMethylation450 BeadChip, which provides methylation fraction readouts (the fraction of DNA reads that are methylated) for approximately 450,000 CpG sites across the entire genome³⁸. To facilitate the validation of our findings in a nontumor context, we additionally obtained somatic mutation (n = 111) and DNA methylation (n = 187) profiles from normal, noncancerous tissues from a subset of these same individuals (Methods).

From these data, we considered all single base-pair substitution mutations (n = 3,457,875 mutation events) and CpG sites for which all individuals had a reliably measured methylation value (n = 326,751 CpG sites; Methods and Extended Data Fig. 2a–d). Consistent with previous reports^3,31, CpG sites were the most frequently mutated dinucleotide, accounting for 13.5% of all somatic mutations across the genome (Fig. 1a). The vast majority of these were C > T transitions (82.3%; Extended Data Fig. 2e) occurring at CpG sites that tended to be more heavily methylated than average (Extended Data Fig. 2f).

Fig. 1 |. Frequency and methylation status of CpG mutation events.

a, Percentage of genome-wide somatic mutations classified as CpG (n = 467,079 mutations) or non-CpG (n = 2,990,796 mutations). Expected percentages were calculated supposing mutation probability to be uniform across the genome (Methods). b, Diagram showing two categories of CpG sites: those where no individual is mutated (nonmutated CpG site, gray) and those where a mutation has occurred in at least one individual (mutated CpG site, red; bottom) and the remaining individuals are nonmutated (blue; top). c, Distribution of CpG methylation values for the categories of CpG sites from b. The methylation fractions of mutated individuals (red) and nonmutated individuals (blue) are shown for the 1,000 CpG sites with the highest MAF (corresponding to MAF > 0.53; Methods). d, Methylation change between mutated and nonmutated individuals at n = 8,037 mutated CpG sites. Methylation change is the difference between the median methylation fraction in mutated individuals and the median methylation fraction in nonmutated individuals of matched age and tissue. CpG sites are binned into five groups based on MAF, with violin plots summarizing the distribution of methylation changes within each group. Vertical bars inside each violin represent the interquartile range. The two-sided P value was calculated based on the exact distribution of Pearson’s r modeled as a beta function. MAF, mutant allele fraction.

We next asked whether individuals harboring a mutated CpG site exhibit lower detected levels of methylation at that site compared to nonmutated individuals (Fig. 1b). We reasoned that once mutated, the site would no longer constitute a CG dinucleotide, reducing its likelihood of methylation. Indeed, we found a significant decrease in detected methylation in individuals with a mutation at a CpG site compared to nonmutated individuals at the same site (Mann–Whitney P = 3.90 × 10⁻⁹; Fig. 1c and Methods), with the decrease of detected methylation related to the frequency of reads with the mutant allele (Pearson’s r = −0.17, P = 2.08 × 10⁻⁵³; Fig. 1d and Extended Data Fig. 2g). These results support a model in which CpG mutations occur primarily at hypermethylated sites (due to the spontaneous deamination of methylcytosine) and can become fixed in the genome of daughter cells, causing a decrease in observed methylation corresponding to the mutated clonal population.

Mutated sites show extensive remodeling of the surrounding methylome

During this exploration, we noted numerous cases in which somatic mutations coincided not only with hypomethylation at the mutated CpG site but also with atypical methylation of numerous CpGs in the surrounding genome. An illustrative example was the C > T mutation at base pair 56,642,556 of chromosome 16 in the individual TCGA-GV-A3QI (Fig. 2a). CpG sites adjacent to this somatic mutation were strikingly hypermethylated in this individual, with such hypermethylation extending over a contiguous region more than 30 kilobases (kb) downstream. This effect encompassed the metallothionein 2A gene (MT2A) as well as additional metallothionein family members (MT1E and MT1M), for which methylation-linked repression has been associated with metastasis in multiple cancer types^39–42.

Fig. 2 |. Association of mutations with regional methylation patterns.

a, Example mutated site where the individual TCGA-GV-A3QI has a C > T mutation at chr16:56,642,556 of the hg19 human genome. Top, ideogram of chromosome 16, with a red bar indicating the location of the mutated site. The first underlying track shows hg19 base-pair coordinates, the second the documented genes in the region (encoding five metallothionein factors) and the third the locations of CpG sites measured on the Illumina 450k methylation array (vertical bars). Bottom, heatmap of CpG methylation fractions. Rows are samples (1 mutated, 28 matched), and columns are the measured CpGs within a ±50-kb window proximal to the mutation (n = 62 CpG sites). Color corresponds to the methylation fraction of each CpG. The mutated sample row and mutated site column are labeled in red, with the mutation event indicated by a lightning bolt. b, Calculation of the change in methylation fraction $\left(\text{ΔMF}\right)$ with reference to a specific mutated site. Left, heatmap of methylation fractions of the mutated site and CpGs in the surrounding window, replicated from a. Right, heatmap of the corresponding differences in methylation between each sample (row) and all other samples in the matrix (median of other rows), computed separately for each site in the window (columns). The final $\text{ΔMF}$ value was calculated as the overall methylation change of the mutated sample, taking the median across all sites in the window (Methods). Matched background samples were defined as those without any somatic mutations in the window and that were of the same tissue type and approximate age (±5 years) as the mutated sample. UCSC, University of California, Santa Cruz.

To move beyond anecdotal observations, we devised a general test for whether somatic mutations are associated with remodeled methylation at surrounding CpGs. We computed a quantity we called $\text{ΔM}{\text{F}}_w$, the median change in methylation fraction observed for CpGs within a window of $\pm w$ kb near a mutated site, comparing the mutated individual to matched nonmutated individuals (Fig. 2b and Methods). Examination of $\text{ΔM}{\text{F}}_{1\text{kb}}$ values at increasing distances from mutated sites revealed significantly greater methylation change at these sites compared to randomly chosen nonmutated sites, with this effect extending for up to ±10 kb (Fig. 3a and Methods). Calculating the methylation change across this entire ±10-kb range, we observed that $\text{ΔM}{\text{F}}_{10\text{kb}}$ tended toward substantially more extreme values than expected at random (n = 2,600 mutated sites with sufficient nearby CpGs, P < 10⁻¹²⁴; Fig. 3b), with mutated loci more than four times as likely to have an extreme decrease in nearby methylation ($\text{ΔM}{\text{F}}_{10\text{kb}}$ of <−0.3; Fig. 3c). Mutated loci were also enriched for nearby methylation increases (similar to the example in Fig. 2a), albeit more weakly (Fig. 3c). This phenomenon was observed in all tissue types for which we had data (Extended Data Fig. 3a). Overall, 15.5% of mutated sites had a more extremely positive or negative $\text{ΔM}{\text{F}}_{10\text{kb}}$ than expected (against a reference of random control sites, $\mid z\text{score}\mid >1.96$; Methods).

Fig. 3 |. Magnitude and extent of methylation changes near somatic mutations.

a, Absolute $\text{ΔM}{\text{F}}_{1\text{kb}}$ as the window center is moved away from the mutated site (n = 2,600, red). This quantity is also shown for nonmutated random control sites (n = 260,000, blue) (Methods). Points indicate the mean value, and error bars denote the 95% confidence interval. A significant difference in the distribution of absolute $\text{ΔM}{\text{F}}_{1\text{kb}}$ values (two-sided t test) is marked (***P ≤ 0.001, **P ≤ 0.01). Other comparisons are nonsignificant (NS, P > 0.05). b, Probability distribution of $\text{ΔM}{\text{F}}_{10\text{kb}}$ values calculated in a ±10-kb window surrounding mutated (red) versus random control (blue) sites. Mutated sites include n = 2,600 mutated sites with MAF ≥ 0.8, ≥15 matched individuals (individuals of the same tissue type within ±5 years of age) and one or more measured CpGs within the window. Random control sites include n = 260,000 nonmutated sites (Methods). The same was found when controlling for the initial methylation state of mutated and random loci (Extended Data Fig. 3b,c). The P value is shown for a two-sided Mann–Whitney test for a difference in median absolute deviation (MAD) of $\text{ΔM}{\text{F}}_{10\text{kb}}$ between the mutated and nonmutated random control loci. c, Line plot depicting the fold enrichment for mutated over nonmutated sites as a function of $\text{ΔM}{\text{F}}_{10\text{kb}}$. Fold enrichment is the ratio of the probability of observing a given $\text{ΔM}{\text{F}}_{10\text{kb}}$ for mutated sites versus the probability of that $\text{ΔM}{\text{F}}_{10\text{kb}}$ for nonmutated control sites. $\text{ΔM}{\text{F}}_{10\text{kb}}$ is divided into equally spaced bins from −0.4 to 0.4. d, Enrichment of extreme $\text{ΔM}{\text{F}}_{10\text{kb}}$ values at CpG sites and CpG islands. Top versus bottom bar charts show the 25% of mutations with the most positive versus most negative $\text{ΔM}{\text{F}}_{10\text{kb}}$ values in a (n = 650 mutations each). The enrichment of these mutations (bars, y axis) was considered for different types of sites, depending on whether the site is a CpG and/or falls within a CpG island (x-axis categories). Enrichment was compared to the genomic baseline (Methods), with significance determined by a one-sided binomial test. Significant enrichment (P ≤ 0.001) is marked with asterisks (***) and nonsignificant (P > 0.01) with ‘NS’. CpG islands are defined as genomic regions with ≥200 bp, ≥50% GC content and a high CpG occurrence. e, Boxplot of the absolute $\text{ΔM}{\text{F}}_{10\text{kb}}$ value as a function of the MAF. The plot includes all mutated sites with ≥15 matched samples and one or more measured CpGs within ±10 kb (n = 3,880 mutated loci). The two-sided P value was calculated based on the exact distribution of Pearson’s r modeled as a beta function. The central line represents the median; the edges of the box represent the interquartile range; the whiskers indicate 1.5 times the interquartile range; and the points represent all $\text{ΔM}{\text{F}}_{10\text{kb}}$ values outside these ranges.

Deeper explorations showed that the methylation increases and decreases (1) have a direction of change that depends on local CpG density (that is, whether they are inside CpG islands¹³; Fig. 3d and Extended Data Fig. 3d), (2) are specific to the genomic context, enriched for being mutations to CpG sites (Fig. 3d) in intergenic regions (Extended Data Fig. 3e,f), and (3) increase with the fraction of DNA in the sample harboring the mutation (Fig. 3e and Extended Data Fig. 3g). In noncancerous tissues, methylation disturbances were found surrounding 8.0% of somatic mutations ($\mid z\text{score}\mid >1.96$ compared to random control sites; Methods and Extended Data Fig. 4a), each extending for, on average, ±1 kb from the site of mutation (Extended Data Fig. 4b). This decreased frequency and shorter range of effect may be attributed to the lower mutated allele frequency (MAF) of somatic mutations in normal tissues compared to tumor tissues (mean ± s.d. MAF: normal = 0.10 ± 0.013, tumor = 0.83 ± 0.24). These results indicate that our earlier observations were not anecdotal, but CpG mutations are generally associated with an atypical methylation pattern in the surrounding DNA.

Somatic mutations mirror epigenetic predictions of age

While mutation of any particular CpG site is exceedingly rare in the human population, and thus a poor predictor of age, its corresponding CpG methylation fraction varies regularly in a manner often associated with age¹⁷. However, we considered that the one-to-many relationship revealed by our previous analysis (Fig. 3), by which a single CpG mutation maps to a broad profile of methylation changes in the surrounding DNA, might bridge this apparent gap between sporadic mutation accumulation and consistent methylation change. Accordingly, we compared two procedures for predicting human chronological age: the first using an individual’s profile of CpG methylation values, as in previous epigenetic aging models (methylation clock), and the second using their profile of somatic mutations, including the counts of somatic mutations within 10 kb of each of these same CpGs (mutation clock; Methods and Fig. 4a). Evaluating these models using a nested cross-validation procedure (Methods), we found that the methylation clock predicted age with an accuracy of r = 0.83 (Pearson’s correlation), whereas the mutation clock had an accuracy of r = 0.67 (Fig. 4b,c). When predictions were examined within each tissue, both the mutation and methylation models were most accurate at predicting age in brain samples and least accurate in kidney and bone samples (Extended Data Fig. 5a).

Fig. 4 |. Association among mutation age, methylation age and chronological age.

a, Methylation clock: the methylation fractions of CpGs are used in a gradient boosted tree model to predict chronological age. Mutation clock: the count of mutations around the same CpGs are used in an identical model to predict chronological age. Both models incorporate similar covariates and whole-genome features (Methods). b, Scatter plot of human individuals, showing age predictions from the mutation model versus their chronological age. The plot includes n = 1,601 individuals with samples from five tissues (Methods). c, Similar to b but showing age predictions from the methylation rather than mutation model for the same individuals. d, Violin plots of the methylation age residual versus the mutation age residual (Methods). The plot includes the same individuals as in b and c. Pearson’s r refers to the correlation between the methylation age residual and the mutation age residual (that is, partial correlation, P = 1.48 × 10⁻⁸², two-sided P value calculated based on the exact distribution of Pearson’s r modeled as a beta function). The central line of the inner boxplot represents the median; the edges of the box represent the interquartile range; and the whiskers represent 1.5 times the interquartile range. e, Distribution of methylation age residuals for the same individuals as in b and c, computed according to each of four previous methylation clocks. ‘This study’ refers to the methylation clock shown in c (Methods). For each clock, the 20% (n = 320) of individuals with the youngest mutation age for their chronological age are shown in a lighter color (low mutation age residual), and the 20% (n = 320) of individuals with the oldest mutation age for their chronological age are shown in a darker color (high mutation age residual). The asterisks (***) indicate a significant (P ≤ 10⁻⁵¹) difference in distribution between the low and high mutation residual age groups, based on a two-sided Mann–Whitney U test. f, Bar plot depicting the ratio of observed versus expected overlap between sets of age-associated CpG sites. For the same individuals and CpG sites as in c, the CpGs with maximal (top 1%, 5% and 10%) Pearson’s correlation between local mutation burden (±10 kb) and age and between methylation fraction and age were chosen. The intersection (overlap) between these sets was compared to the expected intersection assuming random selection (Methods). Significant enrichment based on a two-sided binomial test (P ≤ 10⁻⁵, Bonferroni corrected) is marked with asterisks (***). g, Mean mutation burden (left y axis) or mean methylation fraction (right y axis) plotted versus chronological age (x axis) for CpG site cg19236454. Data were from brain (LGG, low grade-glioma) samples, considering individuals with a nonzero mutation burden (±10 kb) at this site (n = 67). Pearson’s correlation with chronological age: mutation burden = 0.18, methylation = −0.18. Error bars denote the standard error. h, Diagram summarizing the relationships among three measures of age: mutation, methylation and chronological time. The variance explained was calculated as the squared Pearson’s correlation between each pair of measures for the same individuals as in b and c. MAE, mean absolute error.

Beyond their accuracies of age prediction, we found that the two clocks agreed significantly in several other key aspects. First, the predictions from both models were highly correlated across individuals (Pearson’s r = 0.74), and this relationship persisted even after controlling for calendar age (partial correlation = 0.45, P = 1.48 × 10⁻⁸²; Fig. 4d and Methods). For example, for individuals predicted by mutations to be 1 year older than their calendar age, the methylation clock yielded a corresponding overprediction of 0.75 ± 0.53 years (mean ± s.d.). This same agreement in over-/underprediction of each individual (similarity in model residuals) was observed when comparing the mutation clock to previously published methylation clocks (Fig. 4e and Extended Data Fig. 6a,b)^17,18,21. Second, CpG sites for which the surrounding mutation burden was most associated with age also tended to have the most age-associated methylation values, as highlighted by the 80-fold greater-than-expected overlap among the top 1% of age-associated sites of each kind (Fig. 4f and Methods). One example was CpG site cg19236454 (chr19:42,799,926), for which the local mutation burden progressively increased with age (±10 kb, r = 0.18), whereas the methylation of this site was progressively lost (r = −0.18; Fig. 4g). This overlap demonstrates a local coupling of age-related mutation accumulation and methylation change, independent of any clock or model.

To confirm that the synchronization between the mutation and methylation clocks is not unique to tumor tissues, we applied the same modeling procedure to normal tissues (n = 111 somatic mutation profiles, n = 187 methylation profiles). We found that models trained to predict chronological age from the mutation and methylation profiles of these normal tissues (1) were predictive of age (Pearson’s r: mutation clock = 0.74, methylation clock = 0.92; Extended Data Fig. 5b), (2) agreed in the age prediction of the same individuals as clocks trained on tumor data (Pearson’s r of normal versus tumor predicted age: mutation clocks = 0.79, methylation clocks = 0.89; Extended Data Fig. 5c,d) and (3) continued to exhibit an association between mutation and methylation age (partial correlation of normal clocks = 0.48, P = 0.0018; Methods and Extended Data Fig. 5e–g). Thus, mutation and methylation profiles were synchronized with respect to predictions of age (Fig. 4h), both globally (throughout the genome) and locally surrounding individual CpG sites, in normal and tumor tissues.

Discussion

In this study, we observed notable associations between CpG mutation and methylation at multiple scales. At the scale of single nucleotides, CpG sites altered by somatic mutation (the most frequent mutation type across the genome; Fig. 1a and Extended Data Fig. 2e,f) exhibit a decrease in detected methylation at that site (Fig. 1c,d). At a larger scale, such mutated sites coincide with sweeping methylation changes across numerous CpGs within the surrounding genomic region (Figs. 2 and 3a–c). Plausibly as a result of this larger-scale relationship, individuals whose mutations indicate increased genomic age also tend to have older methylomes (Fig. 4d,h and Extended Data Fig. 5g).

A fundamental tension addressed in this study is that two individuals rarely share a somatic mutation at the same CpG site; thus, mutations would initially seem too sparse to explain the numerous CpG sites at which methylation reliably changes with age. However, our findings show that single mutations can correspond to appreciable shifts in the methylome, with a graded relationship that depends on the frequency of the mutated allele (that is, clonality of the mutant cell population; Fig. 3e). Consistent with these findings, we see that, within individuals of the same calendar age, mutation and methylation clocks agree on which individuals are aging faster or slower (Fig. 4d,e), and somatic mutations explain more than 50% of the variation in methylation age across individuals (Fig. 4d,h). Notably, this relationship can be observed directly, without relying on predictive modeling, in the colocalization of age-associated mutation and methylation sites (Fig. 4f,g).

The mechanisms by which a CpG mutation affects its methylation state, or conversely by which CpG methylation potentiates its own mutation, are already established. The prior methylation of a CpG makes a subsequent somatic mutation more likely due to methylcytosine deamination³³. In turn, when either nucleotide of a CpG site is mutated, the site is no longer a CpG, substantially decreasing the likelihood of future methylation by a DNA methyltransferase⁴³. For mutations exhibiting larger-scale gains or losses of methylation in the surrounding kilobases, it is conceivable that either methylation or mutation, or neither, could be the primary causal agent. The observed association between somatic mutation and local hypermethylation (Fig. 3b) may occur if hypermethylation creates an environment prone to methylcytosine deamination events, giving rise to rare somatic mutations embedded within hypermethylated regions. However, this model does not explain the frequently observed co-occurrences of mutations with neighboring hypomethylation throughout the genome (Fig. 3b), as hypomethylation should decrease, not increase, the local probability of mutation.

An alternative possibility is that mutations are the primary causes of subsequent changes in methylation. Mutations within the DNA-binding site of a methylase or demethylase enzyme could plausibly affect enzyme activity, dysregulating the methylation state of the surrounding genome^44–46. Such a relationship has been reported explicitly for somatic mutations in the DNA-binding sites of TET1, a demethylase, leading to local gains of methylation²⁷. More broadly, it is well known that germline DNA sequence variants govern the methylation patterns of many CpG sites, affecting as many as 40% of CpGs throughout the genome (i.e. the effect underlying methyl quantitative trait loci)⁴⁷. Somatic mutations in these sequences may yield effects on methylation analogous to those observed for inherited variants.

A third possibility is that mutation and methylation events are not causal of each other but both are downstream of some earlier event. One such event might relate to the repair of DNA double-strand breaks, which have been demonstrated to result in both somatic mutations and methylation changes near the site of repair^48–50. Here, the mutation and methylation changes would be indicative of an earlier repair of double-strand breaks, an activity recently suggested to cause epigenetic aging²³. Similarly, the formation of 8-hydroxyguanine lesions has been associated with both somatic mutation and loss of methylation in the surrounding DNA⁵¹.

Regardless, understanding the causality among mutations, methylation and aging has important implications for how we seek to prevent or reverse aging. In particular, if mutations are the fundamental driver of aging phenotypes, and epigenetic changes simply track this process, then strategies aimed at epigenetic reversal^23,52–54 may be treating a symptom rather than a cause.

Some limitations of this study are as follows. First, most of the samples we analyzed came from tumor biopsies, as the current large human datasets measuring both somatic mutations and methylation pertain to cancer patients. While analyses of normal tissues from a subset of these same individuals supported the findings in tumor tissues (Extended Data Figs. 5 and 7), the relevance of these findings to normative aging should be further examined in larger datasets of normal individuals and tissues by, for example, applying mutational clocks to independent datasets. This limitation notwithstanding, we note that mutation burdens in cancerous and normal tissues are similar^55,56, and most mutations found in tumors are believed to represent normal mutational processes unrelated to cancer^55,57. Second, most of the somatic mutations we analyzed were derived from exome sequencing, limiting our conclusions regarding the association of mutations with age and methylation to those regions. Third, our analysis was cross-sectional rather than longitudinal, with each individual measured at a single time point only. In the future, a longitudinal study design could greatly inform the actual order of events. Fourth, methylation values were assessed using Illumina 450k technology, which scores the methylation fractions of approximately 450,000 defined CpG sites through bisulfite conversion of unmethylated cytosines (to thymines) and subsequent oligonucleotide hybridization⁵⁸. For CpGs at which the cytosine has been mutated (for example, to TpG), the technology will read out this mutation event as a loss of methylation, whether or not a concomitant loss of methylation has occurred. Finally, there exist other factors associated with epigenetic aging that do not explicitly implicate somatic mutations. Some epigenetic changes clearly reflect alterations in tissue composition with age^59,60, and other changes are associated with the expression of developmental genes^61,62, such as in the binding sites of the polycomb repressive complex^63,64. Some of these factors may nonetheless relate to DNA mutations; for instance, somatic mutations can drive alterations in tissue composition^65,66.

Methods

Data access and preprocessing

We obtained paired DNA methylation (Illumina 450k array) and somatic mutation data from two public consortia: TCGA^34–36 and PCAWG³⁷. Relevant to TCGA, we used the Pan-Can cohort (http://xena.ucsc.edu/), which includes 8,680 samples from 33 cancer types with both Illumina 450k methylation data and somatic mutation calls. In addition to these tumor samples, we generated somatic mutation profiles from normal, noncancerous tissues of 111 TCGA individuals (see ‘Somatic mutation identification’ in the Methods) and obtained DNA methylation profiles from normal tissues of 187 TCGA individuals (40 individuals had both normal somatic mutation and normal DNA methylation profiles). Data from the PCAWG consortium (https://xenabrowser.net/datapages/?hub=https://pcawg.xenahubs.net:443) include 651 samples from three cancer types with both Illumina 450k methylation array data and whole-genome somatic mutation calls. Methylation data from both cohorts were further processed as follows. First, we removed CpG sites for which any sample had a missing value, leaving 273,202 CpG sites for TCGA and 326,749 CpG sites for PCAWG (Extended Data Fig. 2a–d). Second, we removed samples for which the mean methylation fraction (over all remaining CpGs) was more than 3 s.d. outside its expected (mean) value over all samples. Third, each sample was quantile normalized.

Somatic mutation identification

In addition to the somatic mutations called by the TCGA consortium in tumor tissues³⁵ (n = 8,860 individuals), we identified somatic mutations in the normal tissues of individuals from TCGA who had DNA sequencing carried out in three locations: tumor, tumor-adjacent tissue and blood (n = 111). We called somatic mutations using Mutect2 (ref. 67) by comparing the tumor-adjacent DNA sequence to the blood DNA sequence in each individual. Mutect2 was run with default parameters following the tumor versus normal somatic mutation calling pipeline described by the Genome Analysis Toolkit (GATK) but treating the tumor-adjacent DNA sequence as the ‘tumor’ sample and the blood DNA sequence as the ‘normal’ sample (that is, reference) (https://gatk.broadinstitute.org/hc/en-us/articles/360035531132--How-to-Call-somatic-mutations-using-GATK4-Mutect2). In so doing, we identified variants present in the tumor-adjacent tissue that were not present in the blood of the same individual, indicating their somatic origin. The germline resources provided by the GATK were used to filter out common germline mutations⁶⁷. Mutations at loci with coverage of fewer than 50 reads in either the tumor-adjacent or blood sample were discarded.

Characterizing CpG mutation frequency

Based on the University of California, Santa Cruz, hg19 human genome annotations⁶⁸, the number of nucleotides that comprise CpG residues equals 2 bp × 28,299,634 CpG sites, within a total genome length of 3,137,144,693 bp. Therefore, 1.8% of randomly distributed mutations are expected to be CpG mutations, and the remaining 98.2% of mutations are not (Fig. 1a). As CpG sites are palindromic, CG on one DNA strand is equivalent to GC on the complementary strand; thus, for simplicity, we refer to all CpG mutations on either strand as alterations to the C residue in the first position. This convention was used to record the frequency of each dinucleotide sequence resulting from a CpG mutation (Extended Data Fig. 2e). For these cumulative analyses relating to the overall frequency of CpG mutation (Fig. 1a and Extended Data Fig. 2e,f), the PCAWG samples were used exclusively as they have whole-genome sequences, encompassing all CpG sites, rather than exome sequences only.

Characterizing methylation at mutated CpG sites

The methylation status of two categories of CpG sites was compared: ‘nonmutated sites’, where no mutation was observed in any individual, and ‘mutated sites’, where at least one individual had a mutation (Fig. 1b). For CpG sites of the first category (265,399 nonmutated sites), the distribution of methylation fractions was plotted (Fig. 1c). For CpG sites of the second category (8,037 mutated sites), some individuals harbor a mutation at that particular CpG and some do not. In this case, the distributions of CpG methylation fractions were plotted separately for the mutated versus nonmutated individuals (Fig. 1c). For analyses of methylation associated with mutated CpG sites (Fig. 1c,d), tumor TCGA samples were used exclusively as there were many more occurrences of CpG mutations in this dataset due to its much larger sample size. We note that the effect of a CpG mutation on the overall methylation fraction of a tissue can be masked by nonmutated cells of that tissue having higher-than-expected methylation levels—this is a possible explanation for the rare increases in methylation observed at mutated CpG sites and the modest association between MAF and methylation change (Fig. 1d).

Calculating mutation-associated methylation change

Somatic mutation events were defined as site–sample pairs for which the nucleotide present at that site in that particular sample assumed a different (A, C, G, T) value than in the matched normal tissue (typically whole blood from the same sample). For each mutation event, the genomic ‘locus’ was defined as a $\pm w$-kb window (upstream and downstream) proximal to the mutated site. Matched background samples were defined as those without any somatic mutations in this window and that were of the same tissue type and approximate age (±5 years) as the mutated sample (Figs. 2 and 3). Using these definitions, we calculated $\text{ΔM}{\text{F}}_w$, the median normalized change in methylation fraction $m_{k,j}$ at each measured CpG site $k$ in the $\pm w$-kb window surrounding mutated site $i$ in mutated sample $j$ relative to matched background samples $b\in$ background:

$$\begin{gathered} \text{ΔM}{\text{F}}_{i,j}=\text{media}{\text{n}}_{k\in \text{window}(i,w)}\left(m_{k,j}-{\text{median}}_{b\in \text{background}}\left(m_{k,b}\right)\right) \\ \text{Window}(i,w)=\{\text{CpG sites within}\pm w\text{from site}i\} \end{gathered}$$

We calculated $\text{ΔMF}$ in ±10-kb windows for each mutation event in the PCAWG data, where the mutation was a single base-pair substitution; there were at least 15 viable matched background samples; there was no other mutation within ±10 kb of the mutation in the mutated or matched background samples; the MAF was ≥0.8; and there was at least one CpG site within 10 kb of the mutated base (Fig. 3b,c). For the normal tissue TCGA data, we calculated $\text{ΔMF}$ in ±1-kb windows at each mutation event, where the mutation was a single base-pair substitution; there were at least 10 viable matched background samples; there was no other mutation within ±10 kb of the mutation in the mutated or matched background samples; and there was at least one CpG site within 1 kb of the mutated base (Extended Data Fig. 4a–c). These criteria left 2,600 and 463 mutation events for the tumor PCAWG and normal TCGA datasets, respectively. Random control events were chosen to create a background distribution of $\text{ΔMF}$ values at genomic loci lacking somatic mutations in any sample. For each true mutation event, we randomly chose 100 nonmutated nucleotides from the corresponding mutated sample and calculated $\text{ΔMF}$ at these loci. To perform this calculation, we treated the randomly chosen nucleotide as if it were a mutation and calculated the $\text{ΔMF}$ of CpG sites within ±1 or ±10 kb (for normal TCGA or tumor PCAWG, respectively). To control for the initial methylation state, we repeated this analysis by selecting random control sites with a methylation profile matched (in terms of the median methylation fraction of comparison sites) to that of the matched samples of each true mutation event (Extended Data Fig. 3b,c). To quantify the frequency at which mutated sites had $\text{ΔMF}$ values that were more extreme than expected, we calculated the z score of the $\text{ΔMF}$ values at each mutated site relative to the distribution of $\text{ΔMF}$ at all random control sites.

Extent of mutation-associated methylation remodeling

To investigate the extent of methylation remodeling associated with a somatic mutation, we computed $\text{ΔMF}$ in a 1-kb window at increasing distances (up to 1 Mb) from each mutated and random control site. Then, we aligned these $\text{ΔM}{\text{F}}_{1\text{kb}}$ values by their linear genomic distance from their respective mutated site to observe the marginal methylation change at each distance (Fig. 3a and Extended Data Fig. 4c).

CpG and CGI enrichment

The genomic background rate of CpG islands (CGIs) and CpGs was calculated based on the hg19 annotation^69,70 (https://hgdownload.soe.ucsc.edu/goldenPath/hg19/database/cpgIslandExt.txt.gz). The frequency of CpGs in CGIs was based on previously published statistics³¹. To understand whether mutation type (CpG or non-CpG) and location (CGI or non-CGI) were related to the degree of mutation-associated methylation change, we divided the frequency of each mutation type by the expected genome-wide rate (‘fold enrichment’, focusing on the 25% of mutation events with the most positive or negative $\text{ΔM}{\text{F}}_{10\text{kb}}$). A one-sided binomial statistic was used to test for an increase in mutation frequency above the genomic background rate of each mutation type (Fig. 3d).

Clock datasets and features

The tumor tissue samples used for all clock-related analyses were from the LGG (low-grade glioma; brain), GBM (glioblastoma multiforme; brain 2), SARC (sarcoma; bone), KIRP (kidney renal papillary cell carcinoma; kidney) or THCA (thyroid carcinoma; thyroid) cancer types in TCGA (n = 1,601 individuals). Samples from normal tissues were chosen from these same tissue types as the tumor samples when available (n: thyroid = 23 individuals and kidney = 27 individuals), with the addition of two other tissue types chosen owing to their available sample size (n: colon = 34 individuals and liver = 27 individuals). Somatic mutations in documented cancer genes⁷¹, as well as mutations within the 25 genes having the highest mutation frequency for that tissue type, were removed to mitigate the influence of driving cancer genes on the analysis. Samples with a mutation burden greater than 3 s.d. from the mean mutation burden of that tissue were discarded. We created a shared feature set for training all clock models, selecting the 100,000 CpG sites with the greatest average somatic mutation burden across samples within ±10 kb of the CpG site.

Mutation clock

The mutation clocks were based on a gradient boosted tree model, an XGBoost Regressor⁷² with default parameters, which we trained to predict chronological age using features derived from somatic mutations at the 100,000 CpG sites described above. A separate model was trained for the tumor (n = 1,601) and normal (n = 111) samples because of the different rates of mutation accumulation and mutational patterns, as described in previous literature^3,73. The features used to describe an individual sample were (1) the counts of all somatic mutations within ±10 kb of each CpG site (‘mutation burden’; Fig. 4a), (2) the one-hot-encoded tissue type, (3) the genome-wide mutation burden, summing the MAFs across all 100,000 CpG sites, and (4) the mutation burden associated with each possible trinucleotide context (for example, CCT > CTT or TCT > TTT). To assess the influence of the whole-genome features (3 and 4), we also constructed models with only local mutation counts and tissue type (Extended Data Fig. 7). When the genome-wide features were removed from the models, the performance of the methylation clocks was unchanged (tumor: Pearson’s r = 0.83; normal: Pearson’s r = 0.92; Extended Data Fig. 7a,b); the mutation clocks remained significantly predictive of age (tumor: Pearson’s r = 0.53, P = 1.54 × 10⁻¹¹³; normal: Pearson’s r = 0.73, P = 9.8 × 10⁻⁰⁸; Extended Data Fig. 7a,b); and the two measures continued to agree in the over-/underprediction of age in the same individuals (tumor: partial correlation = 0.51, P = 6.66 × 10⁻¹⁰⁵; normal: partial correlation = 0.52, P = 5.32 × 10⁻⁰⁴; Extended Data Fig. 7c,d). For the tumor tissues, the accuracy of age prediction was assessed using nested cross-validation, in which 64% of samples were used for model training, 16% for hyperparameter tuning and 20% for testing, with this entire procedure repeated over five folds (Fig. 4b and Extended Data Figs. 5 and 7). As relatively few normal tissue samples were available (n = 111), we assessed the performance of the normal tissue mutation clock only on the normal samples for which we had both mutation and methylation profiles (n = 40) to maximize the number of samples used in training while not reducing the number with calculable mutation and methylation ages. Here, we conducted a similar nested cross-validation procedure in which, in each fold, we held out 20% of the normal samples with both data types for testing (n = 8 samples each fold) and performed training and hyperparameter tuning on the remaining samples (n = 71 samples with only mutation data and n = 32 samples with both data types). Following hyperparameter tuning, the number of features selected for use in the trained tumor mutation clock model ranged from 1,189 to 1,293 with a mean of 1,237.2 across folds and from 307 to 344 with a mean of 322 across folds for the normal tissue mutation clock.

Methylation clock

The methylation clock was also based on an XGBoost Regressor model, with identical parameters⁷² as for the mutation clock described above. A separate model was trained for the tumor (n = 1,601) and normal (n = 187) samples. The definition of features was also closely matched, focusing on the same 100,000 CpG sites but using methylation rather than mutation data. In particular, the features used to describe an individual sample were (1) the methylation fraction of each CpG, (2) the one-hot-encoded tissue type and (3) the overall degree of genome-wide methylation, computed as the sum of methylation fraction values across all 100,000 CpG sites (Fig. 4a,c and Extended Data Fig. 5). Models without the global feature were also constructed for comparison (Extended Data Fig. 7). Nested cross-validation was performed as for the mutation clock. Following hyperparameter tuning within each fold of cross-validation, the number of features selected for use in the trained tumor methylation clock ranged from 5,131 to 5,221 with a mean of 5,171.8 across cross-validation folds and from 1,987 to 2,139 with a mean of 2,051 for the normal methylation clock.

Application of existing clocks

The Hannum, Horvath and PhenoAge clocks were refitted to the TCGA pan-cancer dataset. This step was necessary as some CpG sites included in these clocks did not have methylation values passing quality controls (see ‘Data access and preprocessing’ in the Methods). The remaining 66%, 63% and 61% of features were used in the refitted Hannum, Horvath and PhenoAge clocks, respectively. Briefly, the CpGs used in these clocks were obtained from the respective publications^17,18,21, and an elastic-net regression model⁷⁴ with default parameters was trained on these features to predict the chronological age of TCGA samples. Nested fivefold cross-validation was used to assess the performance of each of the previously published clocks before refitting (Extended Data Fig. 6a) and all clocks (including the clock trained in this study) after being refitted to the TCGA data (Extended Data Fig. 6b). For each sample, the residual of each methylation clock was compared to the residual of the mutation clock (Fig. 4e).

Local association of methylation, mutation burden and age

Across the 1,601 individuals and 100,000 CpG loci used in the tumor mutation and methylation clocks, we compared the association between the methylation fraction of each CpG site, its mutation burden in the surrounding 20 kb and the chronological age of the samples (Fig. 4f). First, we calculated the Pearson’s correlation coefficient between chronological age and the quantile-normalized methylation fraction and mutation burden (±10 kb) at each CpG locus. Second, we selected the 1%, 5% and 10% of CpG sites with the largest methylation–age correlation and mutation burden–age correlation and counted the number of CpGs shared between these groups. Third, we compared this overlap to the expected rate of overlap assuming random selection from the 100,000 original CpGs. A two-sided binomial test was applied to assess statistical significance.

Statistics and reproducibility

All analyses were performed in the Python 3.10 and R 3.6.1 environments. Data analysis was conducted using Pandas 1.5.3, SciPy 1.10.0, Pingouin 0.5.3 and Statsmodels 0.13.5. Data were visualized with Seaborn 0.12.1 and Matplotlib 3.7.1. No statistical method was used to predetermine sample sizes. The experiments were not randomized, and the investigators were not blinded to allocation during experiments and outcome assessment. Samples with a mean methylation fraction (across all CpG sites) greater than 3 s.d. from the mean methylation fraction of all samples together were excluded from the analyses. Specific statistical approaches used are noted in the respective Methods sections and figure captions.

Extended Data

Extended Data Fig. 1 |. Links among CpG mutations, methylome remodeling, and aging.

a) Various mutational processes affect the genome. Here, we show that some of these mutations associate with an aberrant DNA methylation pattern at both the mutated site and at numerous neighboring CpGs. b) An individual’s DNA mutation profile and DNA methylation profile make similar predictions of their calendar age and rate of aging. Panel a created with BioRender.com.

Extended Data Fig. 2 |. Supplemental characterization of CpG mutations.

a) The distribution of methylation fraction values of each CpG site in the TCGA and PCAWG datasets separately (TCGA = 273,202 and PCAWG = 326,749 CpG sites) in each sample (TCGA = 8,680 and PCAWG = 651 samples). b) The CpG density (number of CpGs per base pair) in the 50 and 125 base pairs surrounding each of the CpG sites in (a). The central line of the inner boxplot represents the median, the edges of the box the interquartile range (IQR), and the whiskers 1.5-times the IQR. c) Violin plots of the distribution of mean methylation fraction of non-mutated individuals at the same mutated CpG sites as in Fig. 1d (n = 8,037 sites), stratified by CpG mutation type. d) As in (c), but the distribution of CpG density in the 125 bp surrounding each CpG site. e) Pie chart showing the proportion of CpG mutations (n = 467,079 mutations) that result in specific mutated nucleotides. Note that 5’-CpG-3’ sites are palindromic, corresponding to a 3’-GpC-5’ sequence on the opposite strand; thus, mutation of the C residue is equivalent to mutation of the complementary G residue. For simplicity, we refer to all CpG mutations by the status of the C residue. f) Violin plot showing the mean methylation fraction across all PCAWG samples, considering CpG sites where a mutation has occurred in at least one sample (left, n = 1,137 CpG sites), CpG sites where no mutation has occurred in any sample (middle, n = 325,614 CpG sites), and all measured CpG sites (right, n = 326,751). Significant difference of distribution (p ≤ 3.03 × 10^–50) is marked with (***) and non-significant (p > 0.05) with (n.s.), based on a two-sided Mann-Whitney test. g) Methylation fraction at the same mutated CpG sites as Fig. 1d (n = 8,037 sites). CpG sites are binned into five groups based on MAF, with violin plots summarizing the distribution of methylation fraction within each group. Vertical bars inside each violin represent the interquartile range. Two-sided p value calculated based on the exact distribution of Pearson’s r modeled as a beta function.

Extended Data Fig. 3 |. Magnitude of methylation change near somatic mutations by tissue and genomic context.

a) Boxplots of the distribution of $\text{ΔM}{\text{F}}_{10\text{kb}}$ values for mutated (red) versus random control (n = 260,000, blue) sites for each tissue type separately (n = 813, 144, and 1,643 mutated sites from Pancreas, Brain, and Ovary tissues, respectively). P value shown for a two-sided Mann-Whitney test for a difference in median methylation fraction between the mutated and non-mutated random control loci. P value shown for a two-sided Mann-Whitney test for a difference in median absolute deviation (MAD) of $\text{ΔM}{\text{F}}_{10\text{kb}}$ between the mutated and non-mutated random control loci. The central line represents the median, the edges of the box the interquartile range (IQR), and the whiskers 1.5-times the IQR. b) A histogram of the median methylation fraction across comparison sites within ±10 kb of mutated (n = 2,600, red) and random control sites (n = 260,000, blue). Mutated sites are the same as Fig. 3b. Random control sites have been selected as before, with the additional criteria of having a methylation profile matched to that of the matched samples at mutated sites (as measured by the median methylation fraction of comparison sites, Methods). P value shown for a two-sided Mann-Whitney test for a difference in median methylation fraction between the mutated and random control loci. c) Probability distribution of $\text{ΔM}{\text{F}}_{10\text{kb}}$ values for mutated (red) versus random control (blue) sites. Mutated and random sites are the same as (b). P value calculated as in (a). d) Line plot depicting the fold enrichment for mutated over non-mutated random control sites as a function of $\text{ΔM}{\text{F}}_{10\text{kb}}$, for the same sites as Fig. 3b. Sites are stratified depending on whether the site is a CpG and/or falls within a CpG island (n = 419 CpG-non-CGI, 21 CpG-CGI, 2,120 non-CpG-nonCGI, and 39 non-CpG-CGI sites). Fold enrichment is the ratio of the probability of observing a given $\text{ΔM}{\text{F}}_{10\text{kb}}$ for mutated sites versus non-mutated random control sites. $\text{ΔM}{\text{F}}_{10\text{kb}}$ is divided into equally spaced bins from –0.4 to 0.4. e) Barchart showing the fold-enrichment of mutated sites with the most extreme methylation changes (absolute $\text{ΔM}{\text{F}}_{10\text{kb}}\mid \text{Z}-\text{score}\mid >1.96$, n = 401 mutated sites) in various genomic regions, compared to all other mutated sites (n = 2,199 mutated sites). P values were calculated using a two-sided Fisher exact test. The categories ‘Upstream gene’ and ‘Downstream gene’ refer to variants located within 1 kb of the 5’ transcription start site and the 3’ transcription stop site, respectively, but outside the gene itself. f) As in (e), but comparing the mutated sites with the most extreme gains of methylation (Z-score of $\text{ΔM}{\text{F}}_{10\text{kb}}>1$) to those with the most extreme losses of methylation (Z-score of $\text{ΔM}{\text{F}}_{10\text{kb}}<-1$). g) Boxplot of the $\text{ΔM}{\text{F}}_{10\text{kb}}$ value as a function of the mutated allele frequency (MAF). Same sites and samples as Fig. 3e (n = 3,880 mutated loci. The Pearson correlation is shown for the association of MAF with $\text{ΔM}{\text{F}}_{10\text{kb}}$ and the absolute value of $\text{ΔM}{\text{F}}_{10\text{kb}}$. Two-sided p values were calculated based on the exact distribution of Pearson’s r modeled as a beta function. The central line represents the median, the edges of the box the interquartile range (IQR), the whiskers 1.5-times the IQR, and the points all $\text{ΔM}{\text{F}}_{10\text{kb}}$ value outside of these ranges.

Extended Data Fig. 4 |. Mutation-associated methylation change in normal tissues.

a) Probability distribution of $\text{ΔM}{\text{F}}_{1\text{kb}}$ values for mutated (red) versus random control (blue) sites. Includes n = 463 mutated sites (n = 146 samples) with MAF ≤ 0.15, ≥10 matched individuals (individuals of same tissue type within ± 10 years of age), and ≥1 measured CpG within the window. Random control sites include n = 46,300 non-mutated sites (n = 146 samples, Methods). P value shown for a two-sided Mann-Whitney test for a difference in median absolute deviation (MAD) of $\text{ΔM}{\text{F}}_{1\text{kb}}$ between the mutated and non-mutated random control loci. b) Line plot depicting the fold enrichment for mutated over non-mutated sites as a function of $\text{ΔM}{\text{F}}_{1\text{kb}}$. Fold enrichment is the ratio of the probability of observing a given $\text{ΔM}{\text{F}}_{1\text{kb}}$ for mutated sites versus the probability of that $\text{ΔM}{\text{F}}_{1\text{kb}}$ for nonmutated control sites. $\text{ΔM}{\text{F}}_{1\text{kb}}$ is divided into equally spaced bins from –0.45 to 0.45. c) Absolute $\text{ΔM}{\text{F}}_{1\text{kb}}$ as the window center is moved away from the mutated site (n = 463, red). This quantity is also shown for non-mutated random control sites (n = 46,300, blue) (Methods). Points indicate the mean value and error bars denote the 95% confidence interval. A significant difference in distribution of absolute $\text{ΔM}{\text{F}}_{1\text{kb}}$ values (two-sided t-test) is marked (**, p ≤ .01), (*, p ≤ .05). Other comparisons are non-significant (n.s., p > 0.05).

Extended Data Fig. 5 |. Supplemental age prediction accuracy.

a) Bar plot indicating the correlation of chronological age with the age predictions of mutation clocks (left) or methylation clocks (right). Correlations are shown across all tumor tissues (n = 1,601) and in each of five TCGA tumor tissues individually: LGG (Brain), GBM (Brain-2), SARC (Bone), KIRP (Kidney), and THCA (Thyroid). b) As in (a) but for age predictions using samples from normal (that is non-cancerous) tissues (n = 40 individuals). c) Heatmap indicating the pairwise consistencies (Pearson correlation) among the mutation age in normal tissue, mutation age in tumor tissue, and chronological age. Data shown for n = 22 individuals with mutations measured in both normal and tumor tissues (the same individuals as from panel b with the exception of 11 colon samples and 7 liver samples as these were not available in the tumor samples). d) As in (c), but comparing predictions from methylation clocks. e) Scatter plot of human individuals, showing age predictions from the mutation model versus their chronological age. Shared area denotes the 95% confidence interval of the line of best fit. Includes 40 individuals from four normal tissues (Methods). A two-sided p value was calculated based on the exact distribution of Pearson’s r modeled as a beta function. f) Similar to panel (b) but showing age predictions from the methylation rather than mutation model. g) Violin plots of the methylation age residual versus mutation age residual (Methods). Plots include the same individuals as in panels (b,c). Pearson r refers to the correlation between methylation age residual and mutation age residual, controlling for chronological age (that is, partial correlation, p = 1.76 × 10^–3). The central line of the inner boxplot represents the median, the edges of the box the interquartile range (IQR), the whiskers 1.5-times the IQR, and the points all the methylation age residual values. Statistics calculated as in (e).

Extended Data Fig. 6 |. Performance comparison to previous epigenetic clocks.

a) Pearson r between predicted and chronological age for Hannum, Horvath, and PhenoAge clocks across the same samples as Fig. 4b (n = 1,601). Predictions were done using the subset of features from each clock that existed in our methylation data after quality control (66%, 63%, and 61% of CpG sites from the Hannum, Horvath, and PhenoAge clocks, respectively). The performance of this study’s methylation clock is not shown as it is inherently fit to the TCGA dataset in 5-fold CV. b) Pearson r between predicted and chronological age for Hannum, Horvath, and PhenoAge clocks after re-fitting (Methods). Same samples as (a). The performance of the methylation clock trained in this study (‘This study’) is shown for reference.

Extended Data Fig. 7 |. Mutation age prediction without whole-genome features.

a) Correlation of chronological versus predicted age, shown for mutation or methylation clocks built without whole-genome features (n = 1,601 individuals). Correlations are shown across all tissues and in each of five TCGA tissues individually: LGG (Brain), GBM (Brain-2), SARC (Bone), KIRP (Kidney), and THCA (Thyroid). b) As in (a) but for age predictions using samples from normal (that is non-cancerous) tissues (n = 40). c) The methylation age residual is plotted versus the mutation age residual, using clocks without whole-genome features (Methods). Violin plots summarize the same samples as in panel (a). Pearson r refers to the correlation between methylation age residual and mutation age residual, controlling for chronological age (that is, partial correlation, p = 6.66 × 10^–105). The central line of the inner boxplot represents the median, the edges of the box the interquartile range (IQR), and the whiskers 1.5-times the IQR. A two-sided p value was calculated based on the exact distribution of Pearson’s r modeled as a beta function. d) Similar to (c), but for the samples in (b). The central line of the inner boxplot represents the median, the edges of the box the interquartile range (IQR), the whiskers 1.5-times the IQR, and the points all the methylation age residual values. Statistics calculated as in (c).

Supplementary Material

The online version contains supplementary material available at https://doi.org/10.1038/s43587-024-00794-x.

Data availability

All data analyzed were from The Cancer Genome Atlas Pan-Can cohort^34–36 (http://xena.ucsc.edu/) and the Pan-Cancer Analysis of Whole Genomes⁴⁸ (https://xenabrowser.net/datapages/?hub=https://pcawg.xenahubs.net:443). Data can be accessed from the provided links and are described further in the respective publications (https://doi.org/10.1038/ng.2764, https://doi.org/10.1038/s41586-020-1969-6)^35,37. Data to replicate the figures in this manuscript can be found on figshare (‘Somatic mutation as an explanation for epigenetic aging (Koch et al. 2024)’, https://figshare.com/projects/Somatic_mutation_as_an_explanation_for_epigenetic_aging_Koch_et_al_2024_/224232)⁷⁵. The panel of normal and gnomAD resources used for filtering the somatic mutation calls can be accessed by downloading Mutect2 (https://gatk.broadinstitute.org/hc/en-us/articles/360037593851-Mutect2). A file containing Illumina 450k array CpG locations and characteristics can be accessed on the Illumina website (https://webdata.illumina.com/downloads/productfiles/humanmethylation450/humanmethylation450_15017482_v1–2.csv). The hg19 genome annotation can be accessed through the University of California, Santa Cruz, website (https://hgdownload.soe.ucsc.edu/goldenPath/hg19/database/cpgIslandExt.txt.gz).