Main

Organisms display enormous variation as the result of evolution, spanning many orders of magnitude in characteristics such as size, energy requirements and lifespan. Despite this remarkable diversity, it has been observed that biological traits often share underlying mechanisms and constraints1. These fundamental connections between organisms can be reflected in scaling laws, which mathematically describe an association between two physical quantities over several orders of magnitude.

A notable example of a scaling law in the field of biology is Max Kleiber’s observation that an animal’s metabolic rate is proportional to its mass to the power of three-quarters2. This observation was later shown to hold across not just whole organisms but also cells and mitochondria, spanning a total of 27 orders of magnitude in mass3. It has been proposed that this relationship arises from the transport of materials through branching fractal-like networks and that evolution tends to minimize the energy required to supply these materials4. Such an explanation demonstrates the power of scaling laws to reveal fundamental processes that govern biological systems.

DNA methylation is an epigenetic modification in which a methyl group is added to a cytosine base followed by a guanine (CpG). Methylation status at a given CpG site can vary between cells, meaning a methylation proportion can be calculated for each CpG site across a population of cells. Methylation proportions of some CpG sites change in a predictable way with age. This observation led to the development of the first ‘epigenetic clocks’ in the early 2010s (refs. 5,6,7), which used methylation proportions of selected CpG sites to predict chronological age in humans. Since then, epigenetic clocks have been extended to numerous other organisms, including the development of clocks that measure age across mammalian species8.

Recently, in mammals, DNA methylation rates have been shown to generally correlate with a species’ maximum lifespan, although no scaling has been observed and the biological mechanisms behind the correlation remain unclear. Lowe et al.9 looked at age-related CpG sites across six mammals and found a negative trend between methylation rates and maximum lifespan. Similarly, Wilkinson et al.10 looked at age-related CpG sites in 26 bat species and again found a negative correlation between methylation rate and longevity. More generally, methylation dynamics have recently been used to develop epigenetic predictors of life history traits11 and to attempt to identify specific CpG sites and associated genes involved in both aging and longevity12. Findings such as these have led to the prediction that a scaling relationship might exist between methylation rates and maximum lifespan13.

We compared the methylation rates of conserved age-related CpG sites in blood and skin in a total of 42 mammalian species, representing nine taxonomic orders and covering almost the entire range of mammalian lifespans (Supplementary Tables 1 and 2). In contrast to previous studies, we removed the impact of potential statistical artifacts, which arise when comparing linear rates in a bounded space of methylation values across species of different lifespans, by developing a statistically robust framework and analyzing the effect of CpG selection (described below). We found that methylation rates in both tissues scaled tightly with maximum lifespan. The emergence of explicit scaling suggests that epigenetic mechanisms are, or are linked to, an underlying evolutionary constraint on lifespan that is shared across species.

We aimed to compare methylation rates, defined as the slope from linear regressions of methylation proportion versus age, in conserved age-related sites across mammals. An overview of our workflow is depicted in Fig. 1a. We initially curated our data for each tissue by removing outliers using density-based clustering14 on principal components (PCs; step 1 in Fig. 1a and Extended Data Fig. 1). Additionally, we removed samples below the age of sexual maturity known to exhibit non-linear dynamics7 (step 2 in Fig. 1a).

Fig. 1: DNA methylation rates scale with maximum lifespan.
figure 1

a, Workflow overview. M, methylation proportion. b, Methylation rate (*ratio compared to baseline species) versus maximum lifespan in blood samples. The y coordinate of each point is the cumulative product of the median rate ratio (Methods). The regression line is plotted from the transformed log-linear association shown in c. The shaded region represents the 95% CI. c, Same data as in b but with axes log transformed. The regression line is from a simple linear regression of the form y ~ x. d,e, Equivalent analysis as b,c but in skin samples. Unlabeled points are various bat species (see Supplementary Table 2 for details). Created with BioRender.com.

Full size image

Next, we aimed to avoid any biases arising from the calculation of rates across different lifespans. First, as methylation levels are constrained between 0 and 1, they are more likely to reach these boundaries in age-related sites and start to stabilize in longer-lived mammals. Consequently, simply fitting linear regression lines to these data will result in slower methylation rates for longer-lived animals (Extended Data Fig. 2a). Second, R2-based thresholds to select age-related sites may bias the selection of CpG sites toward those with slower rates in longer-lived animals. This is because shorter-lived species might not be sampled long enough for small trends to become statistically apparent. Working with cohort data, these concerns arise when comparing mammalian species across different ranges of sampled ages rather than different lifespans. In a simulation based on the lifespans observed in our data, we show that not accounting for these differences in sample age ranges results in an artificial negative association with maximum lifespan (Extended Data Fig. 2b).

To develop a statistically robust framework, we therefore ordered the datasets by maximum observed age and compared each mammal with its neighbors in a sequential pairwise manner (step 3 in Fig. 1a and the Methods). We started from the mammal with the shortest observed age, which we refer to as the baseline species. For each comparison, we restricted the datasets of both mammals to be as close as possible to each other (step 4 in Fig. 1a and the Methods).

For each pairwise comparison, we selected the common set of age-related CpG sites passing an R2 threshold and sharing the same directionality. For each CpG in this set, we calculated the methylation rate for each species (step 5 in Fig. 1a). We then calculated the methylation rate ratio of the longer-sampled species compared to the shorter-sampled species and extracted the median ratio across all selected CpG sites. Next, we computed the cumulative product of median ratios to compare all species together (step 6 of Fig. 1a and the Methods). This cumulative product can be thought of simply as the methylation rate of each species (indirectly) compared to that of the baseline species. In this way, we were able to compare the methylation rates of the shortest-sampled animals with those of the longest-sampled animals while at all steps comparing rates over the same timespans.

We explored incremental R2 thresholds from 0 to 0.2 to define an age-related site and show the emergence of a stable scaling law for each tissue (Extended Data Fig. 3).

To validate that our methodology removes any biases resulting in artificial scaling, we applied it to simulated data that use the ages observed in our dataset but with rates randomly drawn for each site from all the observed rates across all mammals. The absence of any scaling law observed in this simulation emphasizes the robustness of our approach in contrast to previous methods (Extended Data Fig. 4)9,10,15.

To explore the existence of a scaling law between methylation rate and lifespan, we plotted the methylation rate for each species (as explained above) against maximum lifespan in two tissues for which we had sufficient data: blood and skin (Fig. 1b–e). For each tissue, this revealed a relationship in which methylation rates decay to an asymptote as lifespan increases (Fig. 1b,d). Taking the logarithm of the x and y axes (see the Methods for details) resulted in strong linear associations with slopes equal to −0.95 in blood (95% confidence interval (CI), −1.14 to −0.76) and −0.60 in skin (95% CI, −0.74 to −0.47) (Fig. 1c,e). This implies power law relationships for each tissue in which methylation rates are proportional to lifespan to the power of −0.95 (blood) and −0.60 (skin). The relationships were strong and consistent in both tissues, with relatively little variation in methylation rates unexplained by differences in maximum lifespan (R2 = 0.81 in blood, R2 = 0.80 in skin). Similar associations were seen when CpG sites were stratified into hypermethylating and hypomethylating sites (Extended Data Fig. 5) and in a sensitivity analysis in which we omitted the initial trimming of ages up to the age of sexual maturity (Extended Data Fig. 6). Notably, the three largest outliers in blood samples were livestock species (pigs, sheep and cattle).

Overall, our analysis of DNA methylation data in mammals reveals scaling between maximum lifespan and DNA methylation rate over approximately two orders of magnitude and in two distinct tissue types. For blood, for example, this relationship means that the methylation rate of humans is about half as fast as that of chimpanzees, given that our lifespan is about twice as long. An interesting application of such scaling relationships is that it allows estimations of maximum lifespan for newly discovered species through longitudinal sampling, in which only the time interval between samples is needed instead of any knowledge of chronological ages.

Many physiological characteristics exhibit scaling with lifespan because they are indirectly associated through body mass4,16. However, the relationship we observed is largely independent of body mass, with no clear trend seen when regressing against mass instead of lifespan (Extended Data Fig. 7). For example, the naked mole rat, which is an outlier in body mass relationships, scaled appropriately in our data17.

The fact that specific and quantitative relationships exist between methylation rate and maximum lifespan suggests that there is an evolutionary constraint acting across diverse mammalian lineages. This means that, when an organism’s lifespan evolves, its methylation rates also change. A scaling law emerges when these changes occur in a predictable way.

Methylation changes over a lifespan can be most simply described by the occurrence of epimutations in stem cells and their inheritance through cell divisions. As such, methylation rate in a mammal, ML, of lifespan L can be thought of as the product of two underlying quantities: R, the rate of stem cell division, and p, the probability that a cell division results in a change in methylation state18,19.

$${M}_{L}\propto {pR}\propto {L}^{a},$$
(1)

where a denotes the scaling law. Under this model, the quantity pR must scale with lifespan. As for which of these factors may be responsible for the scaling we observe, we discuss two non-mutually exclusive scenarios below.

First, the probability p of methylation changes with each stem cell division may scale with maximum lifespan. In this scenario, aberrant methylation levels themselves are an evolutionary constraint on maximum lifespan. In other words, epimutation burden is deleterious, and so mechanisms to reduce it are selected for in longer-lived organisms (that is, p decreases as L increases in equation (1)). This scenario would support an instructive role of DNA methylation in the associations observed between epigenetic changes and physiological outcomes in aging19,20,21,22.

Second, it is possible that stem cell replication rates scale with maximum lifespan (that is, R decreases as L increases in equation (1)). In the hematopoietic system, it has previously been suggested that the rate of stem cell divisions in mammals decreases with lifespan such that the total number of divisions per stem cell is approximately constant, regardless of lifespan23. Additionally, it has recently been observed by Cagan et al. that somatic mutation rates also exhibit negative scaling with maximum lifespan24. Given that cell division plays a role in both processes, one possibility is that the scaling of both methylation and mutation rates is driven by stem cell replication rates.

While our study unambiguously shows the existence of scaling between methylation rates and maximum lifespan in mammals, the precise value of the scaling is subject to some uncertainty. First, the data are one source of error. Specifically, although the Mammalian Methylation Consortium dataset provided an unprecedented resource for the community, the number of observations per mammal was sometimes limited. Additionally, the sampling distribution of ages in some mammals was uneven or sparse or covered only a small proportion of their maximum lifespan. Combined, these factors added uncertainty to the calculation of methylation rates. Similarly, the age and maximum lifespan estimates in some mammalian species are imprecise, adding further uncertainty to the calculations. Second, while it is well established that methylation dynamics can be approximated by linear functions, they do not provide a comprehensive model. Some age-associated CpG sites exhibit non-linear dynamics later in life as they approach a stable value and begin to plateau. This phenomenon would result in an underestimation of the true methylation rate in faster methylating mammals and likely bias our results toward the null (Extended Data Fig. 8), although we partially mitigated this phenomenon by excluding CpG sites with methylation values concentrated near the boundaries. Given all these limitations, there is some uncertainty around the exact value of the scaling laws in blood and skin, and it is unclear whether they are distinct or converge on a common value. However, regardless of the precise values, it is striking that such strong scaling relationships exist even with all these potential sources of error. Future studies could evaluate whether these scaling relationships hold for other classes and tissues and whether non-linear models may shed more light on the precise values of the scaling laws.

Methods

Considerations in the calculation of methylation rates

In contrast to previous studies9,10, we restricted our analysis to CpG sites that were related to age in each mammal being compared. We did this because even a conserved CpG site may behave markedly differently between species. For example, the ELOVL2 CpG site (cg16867657) is very strongly associated with age in humans and other primates but shows no association with age in most other species (Extended Data Fig. 9). As such, using this CpG site to compare methylation rates between a primate and non-primate species may not be appropriate.

Additionally, we compared species across the same timespans because various statistical issues may arise when comparing methylation rates of species across different age ranges. In our study, there were two main considerations. First, use of an R2 threshold (or equivalent) to select age-related sites may bias results toward slower rates in longer-lived animals. This is because slowly methylating sites may not be detected in shorter-lived animals due to smaller timespans and the fact that methylation data are often noisy. In other words, shorter-lived animals might not be sampled long enough for small trends to become statistically apparent. Second, methylation proportions are bounded at 0 and 1. This is important, as it means that, if any given site is related to age, then methylation levels may be more likely to have approached these boundaries and begun to plateau in longer-lived mammals. If this is the case, then simply fitting linear regression lines to these data would result in slower methylation rates for longer-lived animals even with the same underlying dynamics (Extended Data Fig. 2).

Our initial approach was to find age-related CpG sites common to all species and compare the average slope. However, very few CpG sites satisfied this criterion, resulting in unstable estimates. This method would also extend poorly to additional animals, as the number of common CpG sites would decrease with each addition. Furthermore, we were not able to compare animals over the same timespan given the vastly different sampling ranges between the shortest- and longest-lived animals.

Because of all the above reasons, we decided to compare each species in a pairwise manner. This maximized the number of common age-related sites we could use in each comparison. Additionally, if we first ordered our dataset by maximum observed age, we could compare neighboring species across the same timespan by appropriately trimming the datasets in each comparison. This would ensure a fair comparison while maximizing the amount of data retained in each comparison. Finally, as we can sequentially move across the dataset, at each point calculating how the rates of the next mammal compare to those of the one before it, we can (indirectly) compare the methylation rates of the shortest-sampled animals with those of the longest-sampled animals while at all points comparing rates over the same timespans.

Primary analysis

We aimed to compare methylation rates, defined as the slope from linear regressions of methylation proportion versus age, in conserved age-related sites across mammals. An overview of our workflow is depicted in Fig. 1a. We initially curated our data for each tissue by conducting PC analysis on all species combined. We then projected each tissue onto the PC1 and PC2 components to detect and remove outlier samples using density-based spatial clustering (DBSCAN) for each species separately (step 1 in Fig. 1a and Extended Data Fig. 1; see code for further details). Additionally, we removed samples below the age of sexual maturity known to exhibit non-linear dynamics7 (step 2 in Fig. 1a). The age of sexual maturity, as reported in the AnAge database25, was then subtracted from all ages so that 0 represented the age of sexual maturity.

To develop a statistically robust framework, we ordered the datasets by maximum observed age and compared each mammal with its neighbors in a sequential pairwise manner (step 3 in Fig. 1a and the Methods). We started from the mammal with the shortest observed age, which we refer to as the baseline species. For each comparison, we restricted the datasets of both mammals to be as close as possible to each other (step 4 in Fig. 1a and the Methods). Specifically, we calculated the maximum sample age of the shorter-observed species and then found the sample with closest age from the comparison species (either above or below the maximum sample age of the shorter-observed species). If the differences in samples was greater than 2% of the lifespan of the shortest mammal or 1 year (whichever was smallest), we used the next-oldest sample in the shorter-observed species and repeated the process. Once two ages were found that satisfied these criteria, we then restricted the observations in each of the compared species accordingly. Mammals that had fewer than 15 samples after this restriction (or 20 initially) were excluded, as were species for which the maximum sampled age was below 25% of the reported maximum lifespan.

For each pairwise comparison, we selected the common set of age-related CpG sites passing an R2 threshold and sharing the same directionality. A CpG site was considered associated with age if the R2 value from a simple linear regression passed a certain threshold (see Methods). CpG sites with a mean methylation proportion above 0.9 or below 0.1 (calculated after any data trimming) were removed, as these sites tend to display non-linear dynamics due to being near the maximum or minimum methylation values.

For each CpG site satisfying these criteria in both mammals in each pairwise comparison, we calculated the methylation rate for each species (step 5 in Fig. 1a). We then calculated the methylation rate ratio of the longer-sampled species compared to the shorter-sampled species. That is, for each CpG site, methylation rate ratio = (rate of longer-sampled species)/(rate of shorter-sampled species). For example, a ratio of 0.5 would mean that the methylation rate of the longer-sampled species is half as fast as the methylation rate of the shorter-sampled species in a particular CpG site. We calculated these ratios across all CpG sites included for each comparison and extracted the median ratio. Use of the median was chosen over the mean, as the mean was severely affected by large outliers (resulting from the division of very small numbers in some ratio calculations).

Next, we computed the cumulative product of median ratios to compare all species together (step 6 of Fig. 1a and the Methods). This cumulative product can be thought of simply as the methylation rate of each species (indirectly) compared to that of the baseline species. For example, the first mammal in the list (for example, the rat) is given a rate ratio of 1. Suppose that the next mammal in the list (mouse) is compared to the rat, yielding a rate ratio of 0.94 (that is, its median methylation rate is 0.94 as fast as that of a mouse in common conserved age-related CpG sites). Suppose the next mammal in the list (sheep) is then compared to the mouse, yielding a rate ratio of 0.53. We can indirectly compare the rates of the sheep to those of the rat by the cumulative rate ratio of 0.94 × 0.53 = 0.49 and so on. In this way, we can compare the methylation rates of the shortest-sampled animals with those of the longest-sampled animals while at all steps comparing rates over the same timespans.

We include a more detailed mathematical description of this process below.

Mathematical description of scaling and power laws

Mathematically, scaling between two variables x and y can generally be described as a power law relationship of the form

$${y}\,{=}\,{a}{{x}}^{s},$$

where s is the scaling law and a is a constant. Alternatively, taking the logarithm on both sides allows for linear inference of both s and a,

$$\log (y)=\,\log (a)+s\,\log (x).$$

In our case, we are interested in inferring the scaling law relating the lifespan lm of a mammal, m, with the slope sc,m of the methylation values in a CpG site c. This translates to the following scaling relation:

$$\log ({l}_{m})=\,\log ({b}_{c})+s\,\log ({s}_{c,m}),$$

where bc denotes the baseline slope in site c or the slope predicted for a mammal with a lifespan of 1 year.

Our pairwise comparison algorithm exploits the following relation for any two mammals m0 and m1:

$$\log\left(\frac{{s}_{c,{m}_{1}}}{{s}_{c,{m}_{0}}}\right)=s\,\log\left({l}_{{m}_{1}}\right)-s\,\log\left({l}_{{m}_{0}}\right).$$

In other words, the ratio between the slopes in two mammals replaces the intercept in the linear relation with one that is relative to the lifespan of the baseline species m0.

Finally, given an arbitrarily ordered set of species m0, m1, … mi, the cumulative product of slopes results in an estimator of the desired scaling law s:

$$\log\left(\frac{{s}_{c,{m}_{1}}}{{s}_{c,{m}_{0}}}\frac{{s}_{c,{m}_{2}}}{{s}_{c,{m}_{1}}}\ldots \frac{{s}_{c,{m}_{i}}}{{s}_{c,{m}_{i-1}}}\right)=s\,\log\left({l}_{{m}_{i}}\right)-s\,\log\left({l}_{{m}_{0}}\right).$$

Stability of the scaling law

We calculated the impact of varying the minimum R2 threshold to define an age-related CpG site (Extended Data Fig. 3). A grid of R2 values between 0 and 0.2 was explored for each tissue (Extended Data Fig. 3a,c). We then assessed the stability of our results using the kernel density estimate of all reported values and selected the optimal scaling as the point of maximum density (Extended Data Fig. 3b,d). R2 values that resulted in less than ten CpG sites in any one comparison were not considered, even if below the threshold of 0.2.

Biases and statistical robustness

We conducted various analyses to explore the robustness of our results.

First, we conducted a random null simulation, showing that not accounting for differences in sample age ranges results in an artificial negative association with maximum lifespan (Extended Data Fig. 2b). Specifically, we created synthetic data representing a scenario that was as similar as possible to the real data, except with no scaling of methylation rates. We used the observed species and associated lifespans in our datasets but with synthetic methylation data. For each species and site, we uniformly sampled ages within its lifespan and randomly drew slopes and initial methylation values from normal distributions to create linear data with random noise and constrained their values to within 0 and 1. In this analysis, we simply calculated the methylation rates of each mammal across their entire sampled age ranges.

Second, we again used synthetic data but instead conducted our primary analysis method (Fig. 1a) on it. In this case, we used the observed species and used exact sampled ages in our datasets but randomly drew slopes and intercepts from all those observed across all mammals in this study to create synthetic linear data.

Reporting summary

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