Longer-lived mammals exhibit greater evolutionary conservatism and epigenetic modification at cancer-associated loci

Abstract

Peto’s paradox underscores the puzzling disconnect between organismal lifespan, body size, and cancer incidence. Here, we investigate this relationship by examining the evolutionary rates of human cancer-associated sites across 58 mammalian species. While overall evolutionary rates at cancer-associated loci do not differ significantly between long-lived and short-lived species, cancer-associated sites show greater evolutionary conservation relative to non-cancer sites within long-lived mammals but not within short-lived mammals. Furthermore, long-lived species exhibit a pronounced enrichment of m⁶A at cancer-associated loci (based on m⁶A modification sites derived exclusively from male samples) despite lower global m⁶A levels. Notably, tumor suppressor genes in long-lived mammals harbor proportionally more m⁶A-modified transcripts, suggesting a potential association between mRNA methylation and post-transcriptional regulation at cancer-relevant loci. Together, these results suggest that enhanced sequence conservation at cancer loci, coupled with targeted mRNA modification, may be part of a multilayered molecular strategy that reduces cancer risk in long‑lived mammals.

Peto’s paradox establishes how cancer incidence mostly does not scale predictably with body size or lifespan among different species. Here, the authors explore this relationship by examining the evolutionary rates of human cancer-associated sites across 58 mammalian species, finding greater conservation of cancer-associated sites relative to cancer-unrelated sites within long-lived mammals, as well as an association with mRNA methylation.

Introduction

Cancer is a disease primarily caused by accumulation of random mutations in crucial genes influenced by various physical, chemical, and biological factors^1–4. Mutations escaping cell cycle regulation can trigger the malignant transformation of normal cells, gradually leading to cancer progression⁵ and also contributing to aging^6,7, as posited by the mutation accumulation hypothesis⁸. This observation leads to the hypothesis that animals with greater body size and longer lifespans are more susceptible to cancer, as they experience increased exposure to carcinogenic factors and undergo a larger number of cell divisions throughout their lives. Intraspecific studies support this hypothesis, demonstrating a positive correlation between body size, lifespan, and cancer risk within individual species^9–13. However, comparative analyses across species do not support this relationship; instead, cancer incidence mostly does not scale predictably with body size or lifespan among different species¹⁴. This apparent contradiction is commonly referred to as Peto’s paradox^15–17.

Recently, accumulated evidence has shown that longer-lived species, such as naked mole rats, elephants, and some bat species display enhanced resistance to cancer^18–20. Previous studies have framed mutations at cancer-associated loci as outcomes of evolutionary processes^21–23, providing a basis for investigating how these loci evolve across taxa with divergent life histories.

In this study, we identify genomic sites in 58 mammalian species that are homologous to human cancer-associated regions and compare their evolutionary dynamics to those of cancer-unrelated sites. By contrasting mutation rates in longer-lived versus shorter-lived species, we evaluate whether evolutionary constraint at these loci correlates with lifespan or body size. Because epigenetic modifications are intimately linked to oncogenesis^24–26, and synonymous changes can affect tumorigenesis by perturbing m⁶A deposition²⁷, we also examine the distribution of 5mC and m⁶A marks at cancer-associated versus unrelated sites across the two lifespan groups. These comparative analyses aim to clarify relationships among sequence evolution, epigenetic regulation, and longevity, and to explore molecular mechanisms that contribute to reduced cancer incidence in long-lived species.

Results

The conservation of cancer-associated sites is independent of lifespan and body size

Whole-genome alignments for 100 vertebrates were obtained from the UCSC Genome Browser (Multiz100way track), which included coding-sequence alignments for 62 mammals. After removing four phylogenetically distant taxa, 58 mammalian species were retained for downstream analyses (Fig. 1A). Maximum lifespan record for these species were retrieved from the AnAge database²⁸. For nine species lacking maximum-lifespan data, lifespan estimates were obtained using PhyloPars²⁹, a phylogenetically informed restricted maximum likelihood method³⁰, applied to life-history data from 816 species (Table S1). We applied two complementary approaches to classify longer-lived species. The first employed an absolute lifespan threshold of 26 years, based on our previous estimate³¹, which yielded 23 longer-lived and 35 shorter-lived species (Table S1). The second used the longevity quotient (LQ) to identify species with relatively extended or reduced lifespans, resulting in seven species classified as relatively longer-lived and seven as relatively shorter-lived. Using these two classification schemes, we carried out comparative analyses of evolutionary rates at human-homologous, cancer-associated sites (Figs. 1B and S1).

Fig. 1. The conservation of cancer-associated sites is independent of lifespan and body size.

A Circular diagram integrating lifespan classification and phylogenetic relationships across 58 mammalian species spanning 19 taxonomic orders. Teal and yellow bar charts denote maximum lifespans of absolutely longer-lived (typeABS-long) and shorter-lived (typeABS-short) species; dark red and dark blue arcs indicate relatively long-lived (typeRELA-longR) and short-lived (typeRELA-shortR) species after adjusting for body size allometry. B Scatter plot of adult body weight [log₁₀(grams)] versus maximum lifespan [log₁₀(years)] across 62 mammalian species. Two-sided linear regression: P = 7.50 × 10⁻¹¹, R² = 0.51. LQ values computed from 987 species in the AnAge database (mean LQ = 1.17, SD = 0.8); species with LQ > LQmean ± 1 SD were classified as relatively long- or short-lived (seven species per group). Species abbreviations in Supplementary Table 1. C Age distribution of cancer occurrence across 12 tissue types from COSMIC; n indicates sample size per tissue. D Transitions-to-transversions ratio (Ts/Tv) for somatic mutations across cancer types. Dashed line: human genomic background (1.9). Ratios below 1.9 suggest mutation saturation; 10 tissues show Ts/Tv <1. (E–F) Ratio of mutation rates at cancer-associated (E) or cancer-unrelated (F) sites in longer- versus shorter-lived species, across genic segments from 5’UTR to 3’UTR. Mutation rates (G–I) and root-to-tip distances (J–L) of cancer-associated sites compared between longer- vs. shorter-lived species (G, J), phylogenetically matched subgroups of 20 species each (H, K), and larger- vs. smaller-bodied species (top/bottom 17 by body mass) (I, L). Two-sided Wilcoxon rank-sum tests; no multiple comparison adjustment. Numbers on box plots indicate medians. Analyses used matched cancer-unrelated sites within 50 bp and 1000-iteration bootstrapping for validation. Corresponding mutation rates (M–O) and root-to-tip distances (P–R) for cancer-unrelated sites across the same comparisons as G–L. Box plots show median (centre), IQR (box), and 1.5 × IQR (whiskers). Significance: *P < 0.05, **P < 0.01, ***P < 0.001. The panel A and B were created with BioRender.com. Created in BioRender. Du, J. (2026) https://BioRender.com/po7aeaj. Source data are provided as a Source Data file.

All human cancer-associated sites were retrieved from the COSMIC database³². In the human reference genome, these sites mapped to 7125 genes and totaled 3,145,051 cancer-associated positions; 85% of the genes (6086 genes) contained fewer than 60 such sites, while the maximum observed number of cancer-associated sites per gene was 723 (Fig. S2). We designate homologous positions in other mammalian genomes that correspond to these human cancer-associated loci as “cancer-associated sites.” For comparison, we randomly sampled nearby, cancer-unrelated positions in the human genome and matched their count to the cancer-associated sites within each gene; these controls are referred to as “cancer-unrelated sites.” Cancer-associated sites were further categorized into 13 groups based on available histological and pathological annotations; the largest proportions originated from stomach, prostate, and pituitary gland tissues (Figs. S2–S4). Sample metadata indicated that the majority of bone (90.1%) and soft-tissue tumors (72.6%) were obtained from individuals younger than 25 years (Fig. 1C), consistent with the known epidemiology of primary bone sarcomas. This pattern aligns with established observations that osteosarcoma and Ewing sarcoma principally affect children and young adults, reflecting an age-dependent incidence peak in the second decade of life³³.

It is believed that transitions are generally more likely to produce synonymous changes than transversions, a bias reflected in the human genome-wide transitions-to-transversions ratio (R_Ts/Tv) of approximately 1.9³⁴. When R_Ts/Tv below ~2, sequences are often considered to approach mutational saturation, a condition that can increase the likelihood that accumulated mutations will have functional (including carcinogenic) consequences^34,35. Across most tissues, cancer-associated sites exhibited lower R_Ts/Tv than cancer-unrelated sites, consistent with an increased relative contribution of transversions and a higher propensity for nonsynonymous (potentially deleterious) changes at these loci (Fig. 1D and Table S2). An exception was skin (melanoma), which showed an elevated R_Ts/Tv of ~2.1; this pattern is plausibly attributable to ultraviolet (UV) mutagenesis, which preferentially produces C → T transitions and thus raises the transition proportion in exposed tissues^36,37. These observations are compatible with a model in which many cancer-associated sites lie within a background of mutational saturation and reduced evolutionary constraint, making them more susceptible to nonsynonymous substitutions that can contribute to oncogenesis.

To evaluate differences in evolutionary conservation between cancer-associated and cancer-unrelated sites, we first computed sitewise conservation scores using Rate4Site³⁸, which estimate relative evolutionary rates across aligned sequences. Across gene segments, mean mutation-rate estimates indicate that cancer-associated positions at both 5’ and 3’ gene termini tend to be less conserved than interior positions in both longer- and shorter-lived species; cancer-unrelated controls show a broadly similar positional trend (Fig. S5a). Comparing conservation between life-history groups, we observed a greater number of conserved sites in shorter-lived species than in longer-lived species, with the 3’ ends showing particularly reduced conservation (Figs. 1E, F and S5b). To account for phylogenetic depth and divergence, we also estimated root-to-tip branch distances for cancer-associated and control sites using RAxML^39,40. These root-to-tip measures, which provide an alternative summary of accumulated substitutions, revealed no significant difference for cancer-associated sites between longer- and shorter-lived groups (Figs. 1G, J and  S6). To control for sampling imbalance, we repeated analyses on matched subsets (20 longer-lived versus 20 shorter-lived species) chosen to equalize total tree branch length; results remained non-significant for mutation-rate differences at cancer-associated sites (Figs. 1H, K and  S7). Similarly, stratification by body mass (17 larger-bodied versus 17 smaller-bodied species) did not reveal significant differences in mutation rates at cancer-associated loci (Fig. 1I, L).

By contrast, Rate4Site scores indicated that cancer-unrelated sites were significantly more conserved in shorter-lived species than in longer-lived species (Fig. 1M, N). This pattern did not persist when conservation was assessed using root-to-tip phylogenetic distances (Fig. 1P, Q), suggesting that the difference may be sensitive to the conservation metric or to properties of the underlying alignments and trees. Interestingly, the relationship reversed when species were stratified by body size: cancer-unrelated sites in larger-bodied species showed higher conservation than those in smaller-bodied species (Fig. 1O, R). The results indicate that the evolutionary constraints on cancer-associated sites are largely independent of organismal lifespan and body size, whereas conservation of cancer-unrelated sites appears more context-dependent and may reflect life-history traits related selective pressures, or methodological differences between rate-estimation approaches.

Cancer-associated sites are more conserved than cancer-unrelated sites in longer-lived species

We then compared evolutionary conservation between cancer-associated and cancer-unrelated sites within longer- and shorter-lived species. Using Rate4Site, cancer-associated sites in longer-lived species exhibited significantly lower evolutionary rates (i.e., higher conservation) than matched cancer-unrelated sites (Figs. 2A and  S8). This pattern was corroborated by root-to-tip estimates from RAxML³⁹ (Fig. 2B) but was not observed in shorter-lived species (Figs. 2C, D and  S9). To reduce potential biases arising from unequal sample sizes and local sequence context, we applied two complementary controls. First, we performed a matched analysis by selecting an equal number of cancer-unrelated sites located within 50 bp of each cancer-associated site. Second, we carried out 1000-fold bootstrap resampling to assess the stability of mean differences. Although these approaches cannot fully exclude residual biases related to gene expression levels, broader mutation-rate heterogeneity, or cross-species variation, both consistently supported a relative conservation advantage of cancer-associated sites in long-lived species. Importantly, stratifying species by relative longevity (LQ) yielded results concordant with those obtained using an absolute lifespan threshold (Fig. S1), indicating that the observed pattern is robust to alternative definitions of lifespan.

Fig. 2. Cancer-associated sites are more conserved than cancer-unrelated sites in longer-lived species.

Numbers on boxplots represent median values for each group. Statistical methods are consistent with those used in Fig. 1G, H. A Mutation rate for cancer-associated sites, cancer-unrelated sites, adjacent numerically matched cancer-unrelated sites, and randomly sampled numerically matched cancer-unrelated sites in longer-lived species. Rates were calculated using Rate4Site. Each point represents an individual site; total numbers of sites are indicated below the boxplots. B The root to tip distance for cancer-associated sites, cancer-unrelated sites, adjacent numerically matched cancer-unrelated sites, and randomly sampled numerically matched cancer-unrelated sites in longer-lived species. Distances were calculated using RAxML. Each point represents a gene; total numbers of genes are indicated below the boxplots. C Mutation rate for cancer-associated sites, cancer-unrelated sites, adjacent numerically matched cancer-unrelated sites, and randomly sampled numerically matched cancer-unrelated sites in shorter-lived species. D The root to tip distance for cancer-associated sites, cancer-unrelated sites, adjacent numerically matched cancer-unrelated sites, and randomly sampled numerically matched cancer-unrelated sites in shorter-lived species. E Mutation rate for cancer-associated sites, cancer-unrelated sites, adjacent numerically matched cancer-unrelated sites, and randomly sampled numerically matched cancer-unrelated sites in larger-bodied species. F The root to tip distance for cancer-associated sites, cancer-unrelated sites, adjacent numerically matched cancer-unrelated sites, and randomly sampled numerically matched cancer-unrelated sites in larger-bodied species. G Mutation rate for cancer-associated sites, cancer-unrelated sites, adjacent numerically matched cancer-unrelated sites, and randomly sampled numerically matched cancer-unrelated sites in smaller-bodied species. H The root to tip distance for cancer-associated sites, cancer-unrelated sites, adjacent numerically matched cancer-unrelated sites, and randomly sampled numerically matched cancer-unrelated sites in smaller-bodied species. Box plots show the median as the centre line, the interquartile range as box bounds, and whiskers extending to 1.5× the interquartile range. Significance: *P < 0.05, **P < 0.01, ***P < 0.001. Source data are provided as a Source Data file.

We also implemented a model-free conservation metric that quantifies deviation from random site-pattern expectations directly from alignments; this independent measure reproduced our principal result (see Figs. S10 and S11). Because tumor age distributions differed between tissue types, we examined genes associated specifically with bone and soft-tissue cancers. Although the number of annotated pathogenic sites per gene was small (bone: mean 3.31, median 2; soft tissue: mean 3.59, median 2), the same trend—stronger constraint on loci linked to early-onset tumors in longer-lived species—was apparent, suggesting targeted protection of key developmental loci in addition to more general cancer-defense mechanisms.

To test robustness to phylogenetic composition, we performed leave-one-out sensitivity analyses that sequentially excluded groups notable for exceptional longevity, body size, or cancer resistance—specifically cetaceans (e.g., Orcinus orca, Tursiops truncatus), bats (e.g., Myotis davidii, Myotis lucifugus, Eptesicus fuscus, Pteropus alecto, Pteropus vampyrus), the naked mole-rat (Heterocephalus glaber), and the African elephant (Loxodonta africana). For each exclusion we rebuilt alignments and phylogenies and repeated the comparisons; the primary results remained unchanged (Fig. S12), indicating that findings are not driven by a few outlier clades.

Finally, we validated our observations using two independent cancer genomics resources (COSMIC and cBioPortal). Across databases and under rigorous statistical controls, cancer-associated sites—especially within tumor suppressor genes—consistently showed stronger evolutionary constraint in longer-lived species (Fig. S13). Together, these convergent lines of evidence support a robust association between organismal lifespan and selective pressure on cancer-relevant loci.

Subsequently, we explored the functional profiles of genes harboring highly conserved cancer-associated sites in long-lived and short-lived species, respectively (Figs. S14-S15). Gene Ontology (GO) enrichment analysis revealed that, in longer-lived mammals, these sites were preferentially associated with pathways implicated in long-term cellular homeostasis and genome stability, including p53 signaling, autophagosome assembly, and TOR signaling (Fig. S16). Conversely, highly conserved cancer-associated genes in shorter-lived species were primarily enriched in core intracellular processing pathways, such as RNA splicing and proteasome-mediated protein degradation (Fig. S16). The observed divergence in enrichment profiles suggests that, in long-lived mammals, conserved cancer-associated sites are more frequently located within genes involved in pathways supporting sustained cellular maintenance. Although the functional consequences of this overrepresentation remain to be established, these patterns imply that genes—or specific sites—linked to genome stability may experience stronger purifying selection in long-lived species.

Interestingly, contrary to the expectation that longer-lived and larger-bodied species should show similar patterns of evolutionary constraint at cancer-associated loci, our comparative analysis revealed an unexpected pattern in large-bodied species: cancer-unrelated sites were significantly more conserved than cancer-associated sites (Fig. 2E). This result was robust to matched-site sampling (equalizing the number of cancer-associated and cancer-unrelated sites) and to 1000-fold bootstrap resampling, and it was concordant across independent summaries of sequence change, including per-site mutation-rate estimates and root-to-tip phylogenetic distances (Fig. 2F). By contrast, smaller-bodied species exhibited no appreciable difference in conservation between cancer-associated and cancer-unrelated sites, a finding that parallels the lack of differential conservation observed in shorter-lived species (Fig. 2G, H). Although body size and lifespan are positively correlated across taxa, our results suggest that extended lifespan—entailing prolonged cell turnover and cumulative exposure to endogenous mutagenic processes—may be the primary selective pressure favoring stronger purifying selection on cancer-relevant genes, rather than body size per se. These patterns are consistent with the possibility that different life-history strategies mitigate cancer risk via distinct genomic routes: longer-lived species may prioritize heightened conservation at cancer-associated loci, whereas larger-bodied species may instead exhibit broader conservation of cancer-unrelated genomic regions or deploy alternative somatic protections (e.g., altered replication fidelity or enhanced DNA repair capacity), thereby reducing overall mutation accumulation. Further functional and comparative studies are warranted to disentangle these mechanisms and to test whether the observed conservation patterns reflect targeted selection on cancer-related functions or more general constraints associated with species-specific physiology and life history.

Cancer-associated sites within TSGs appear to experience stronger selective constraint in longer-lived species

To investigate the selective forces underlying the conservation patterns, we applied codon-based likelihood tests using PAML⁴¹. Longer-lived lineages were defined as the foreground branches, and for each set of sites (cancer-associated and cancer-unrelated) we compared the fit of an alternative model allowing different selection parameters on the foreground branches to a null model that did not. Across all cancer-associated sites taken together, this analysis did not detect a significant difference in selection pressure between longer- and shorter-lived species (Fig. 3A–C). In other words, the null model (H0: no lifespan-associated shift in selection) had a higher log-likelihood than the alternative (H1), and the change in likelihood (2ΔL) was smaller for cancer-associated sites than for cancer-unrelated sites (Fig. 3C). However, stratifying by gene class revealed a clear distinction. In tumor suppressor genes, cancer-associated sites showed a more pronounced difference in selection between longer- and shorter-lived species than did cancer-unrelated sites (Fig. 3G–I), and selection at sites within tumor suppressor genes (TSGs) was positively associated with lifespan. By contrast, oncogenes exhibited no consistent lifespan-related differences in selection for either cancer-associated or cancer-unrelated sites (Fig. 3D–F).

Fig. 3. Cancer-associated sites within TSGs of longer-lived species may experience stronger selective pressure than cancer-unrelated sites.

Each point represents a gene; total numbers of genes per group are indicated below the boxplots. A The log-likelihood comparison of cancer-associated sites between longer-lived and shorter-lived species, analyzed using the PAML branch model. The same set of genes was used to obtain log-likelihood values for both shorter-lived (background branches) and longer-lived (foreground branches) species; therefore, the sample sizes for the two groups are identical. B The log-likelihood comparison of cancer-unrelated sites between longer-lived and shorter-lived species. Cancer-unrelated sites are homologous sites adjacent to cancer-associated sites, with an equal number of sites in both categories. C Differential log-likelihood values between cancer-associated and cancer-unrelated sites in long-lived versus shorter-lived species. D The log-likelihood comparison of cancer-associated sites between longer-lived and shorter-lived species in oncogenes, analyzed using the PAML branch model. E The log-likelihood comparison of cancer-unrelated sites between longer-lived and shorter-lived species in oncogenes. F Differential log-likelihood values between cancer-associated and cancer-unrelated sites in longer-lived versus shorter-lived species in oncogenes. G The log-likelihood comparison of cancer-associated sites between longer-lived and shorter-lived species in TSGs, analyzed using the PAML branch model. H The log-likelihood comparison of cancer-unrelated sites between longer-lived and shorter-lived species in TSGs. I Differential log-likelihood values between cancer-associated and cancer-unrelated sites in longer-lived versus shorter-lived species in TSGs. J Pathways enriched with polysel in genes exhibiting higher conservatism in longer-lived species. See the Gene enrichment analysis section in Methods for details. Circle colour represents the pathway score [dimensionless], with darker blue indicating a higher score. Circle size represents the gene set size [genes], defined as the number of analysed genes assigned to each pathway. Box plots show the median as the centre line, the interquartile range as box bounds, and whiskers extending to 1.5× the interquartile range. Significance: *P < 0.05, **P < 0.01, ***P < 0.001. Source data are provided as a Source Data file.

The likelihood metric quantifies the magnitude of selection-pressure differences between longer- and shorter-lived lineages but does not directly convey relative site-wise mutation or conservation rates across those groups. To address this limitation, we defined Q_ratio as the per-site ratio of conservation scores in longer-lived versus shorter-lived species (see Methods). Q_ratio > 1 denotes greater conservation in longer-lived species at that site, whereas Q_ratio < 1 indicates greater conservation in shorter-lived species. To identify biological processes enriched for sites with elevated conservation in longer-lived lineages, we applied the polysel pathway-enrichment framework^42,43, which is designed to detect pathways with pleiotropic signals across many genes. Genes with Q_ratio > 1 were significantly enriched for pathways involved in gene expression regulation, cell-cycle control, oncogenesis, inflammation, metabolism, and stress-response processes (Fig. 3J). This multidimensional or pleiotropic enrichment is consistent with established mechanisms of longevity⁴⁴, and suggests that enhanced conservation at cancer-associated sites in longer-lived species may simultaneously contribute to tumor suppression and to broader systemic resilience against physiological stress. Taken together, these results indicate that the intensified purifying selection observed in longer-lived species is concentrated at cancer-associated positions within TSGs. This pattern suggests that tighter evolutionary constraint on TSG loci may constitute a central component of the genomic strategies by which longer-lived organisms reduce cancer susceptibility over evolutionary time.

Elevated m⁶A modification at cancer-associated sites in long-lived mammals

Given epigenetic modifications, including 5-methylcytosine (5mC) and N6-methyladenosine (m⁶A), are implicated in genome stability, DNA repair, and cancer-relevant processes^45–49, we further explore the enrichemnet of 5mC and m⁶A modifications at cancer-associated sites versus cancer-unrelated sites. Across the full dataset, cancer-associated sites showed a modest but highly significant excess of 5mC relative to cancer-unrelated sites (P < 2.2×10⁻¹⁶, Chi-squared test). Matched random sampling of nearby control sites confirmed this difference (5mC: 12.92% in cancer-associated sites vs. 12.72% in controls; Fig. 4A, Table S3). However, the 5mC signal varied with longevity: in longer-lived species the proportion of 5mC at cancer-associated sites (14.68%) was essentially indistinguishable from controls (14.61%; P = 0.16), whereas in shorter-lived species cancer-associated sites were significantly more methylated (12.43% vs. 12.19%; Chi-square P = 4.8 × 10⁻⁸) (Fig. 4B, C; Table S4). These results indicate that 5mC enrichment at cancer-associated loci is context-dependent and may be more pronounced in shorter-lived lineages.

Fig. 4. Epigenetic modifications correlate with cancer-associated site conservation and lifespan in TSGs.

Association between 5mC modification and cancer status in all species (A), longer-lived (B), and shorter-lived species (C). Cancer-unrelated sites were randomly sampled to match cancer-associated site counts. Two-sided Chi-squared test: A, P < 2.2 × 10⁻¹⁶; B, P = 0.16; C, P = 4.8 × 10⁻⁸. Association between m⁶A modification and cancer status in all species (D), longer-lived (E), and shorter-lived species (F), with the same sampling approach (two-sided Chi-squared test: D–F, all P < 2.2 × 10⁻¹⁶). Median 5mC levels (G) and m⁶A site proportions (H) in longer-lived (n = 7) vs. shorter-lived (n = 14) species (two-sided Wilcoxon rank-sum test: G, P = 0.68; H, P = 0.012). Species classification in Supplementary Table S5. I Average m⁶A-modified cancer-associated sites per transcript in longer-lived vs. shorter-lived species (*P < 0.05, two-sided Wilcoxon rank-sum test). J GO enrichment analysis of biological processes associated with m⁶A and cancer sites. Mutation rates of TSGs vs. proto-oncogenes in longer-lived (K) and shorter-lived (L) species (TSGs sites n = 108,954; proto-oncogenes sites n = 93,943; two-sided Wilcoxon rank-sum test: K, P = 1.81 × 10⁻³; L, P < 2.2 × 10⁻¹⁶). Proportions of TSGs and proto-oncogenes with cancer-associated sites (M) or m⁶A sites (N). In M: 1.8% of TSGs and 3.1% of proto-oncogenes harbor cancer-associated sites. m⁶A sites per transcript in TSGs vs. proto-oncogenes in longer-lived (O) and shorter-lived (P) species (two-sided Wilcoxon rank-sum test: O, P = 0.0037; P, P = 0.58). Expression levels of proto-oncogenes (Q) and TSGs (R) in longer-lived (n = 21) vs. shorter-lived (n = 42) species (CPM, ONT sequencing; Q, P = 0.039; R, P = 0.051); and TSGs vs. proto-oncogenes in longer-lived (S) and shorter-lived (T) species (S, P = 0.002; T, P = 0.03; all two-sided Wilcoxon rank-sum tests). Box plots: median (centre), IQR (bounds), 1.5× IQR (whiskers). Significance: *P < 0.05, **P < 0.01, ***P < 0.001. Source data provided as a Source Data file.

By contrast, m⁶A modification showed a robust and consistent enrichment at cancer-associated sites across comparisons. Cancer-associated sites exhibited significantly higher m⁶A incidence than controls genome-wide (Chi-square P < 2.2 × 10⁻¹⁶), and matched-site resampling yielded an elevated m⁶A proportion in cancer-associated positions (14.25%; Fig. 4D; Tables S5– S6). Importantly, the enrichment of m⁶A persisted within both longevity groups: in longer-lived species, cancer-associated sites contained 12.66% m⁶A compared to 0.34% in controls (Fig. 4E), and in shorter-lived species the respective figures were 15.68% versus 0.46% (Fig. 4F). These marked disparities suggest that m⁶A is a distinguishing epigenetic feature of cancer-associated loci and may contribute functionally to their evolutionary conservation.

Building on the observed site-specific epigenetic patterns, we compared genome-wide modification levels between longer- and shorter-lived species. Median global 5mC levels did not differ significantly between the two groups (Fig. 4G), indicating broadly similar DNA methylation loads across species of differing lifespan. By contrast, global m⁶A abundance was significantly lower in longer-lived species (median 0.20%) than in shorter-lived species (median 0.25%) (Fig. 4H). Strikingly, despite this reduced global m⁶A burden, transcripts from longer-lived species carried more m⁶A- modified cancer-associated sites per transcript (mean = 3.71) than those from shorter-lived species (mean = 2.57) (Fig. 4I). Functional enrichment of genes that concurrently harbor m⁶A marks and cancer-associated sites identified pathways linked to methylation biology, cancer-relevant processes (including negative regulation of the mitotic cell cycle and apoptotic regulation by tumor suppressors), and immune surveillance (Fig. 4J). This targeted enrichment is consistent with the possibility that, in longer-lived species, m⁶A marks are non-randomly associated with cancer-associated sites and genes involved in tumor-related processes. However, whether this pattern reflects an evolved epitranscriptomic strategy or contributes functionally to transcript stability, splicing, translation, surveillance, or stress responses remains to be experimentally tested.

Targeted m⁶A enrichment and enhanced conservation of TSGs in long-lived mammals

Since tumorigenesis reflects the balance between oncogenic drivers and tumor suppressors, we next analyzed evolutionary rates and epigenetic marking separately for these two gene classes. TSGs were significantly more conserved in longer-lived species, with lower inferred mutation rates than proto-oncogenes (Figs. 4K and S17a). In shorter-lived species the pattern was reversed (Figs. 4L and  S17b), consistent with the idea that extended lifespan selects for greater genomic stability specifically at loci that prevent malignant transformation. At the epigenetic level, a larger absolute number of TSGs carried m⁶A marks (831 genes; 68%) than did proto-oncogenes (455 genes; 77% of that class), reflecting the greater size of the TSG set (Figs. 4N and  S18). Crucially, in longer-lived lineages m⁶A deposition was preferentially concentrated at cancer-associated sites within TSGs (Fig. 4E, F, N), whereas in proto-oncogenes m⁶A marks were dispersed more broadly across non-critical positions.

Consistent with this model, the density of m⁶A on TSG transcripts was higher than on proto-oncogene transcripts in longer-lived species (Fig. 4O), whereas no significant difference was observed in shorter-lived species (Fig. 4P). These epitranscriptomic patterns corresponded to expression differences: proto-oncogene expression was higher in shorter-lived species than in longer-lived species (Fig. 4Q), while TSG expression was higher in longer-lived species (Fig. 4R). Within longer-lived species, TSG expression exceeded proto-oncogene expression (Fig. 4S; P < 0.01), whereas that advantage was attenuated in shorter-lived species (Fig. 4T; P < 0.05). Stratifying by m⁶A status further clarified these relationships. TSGs carrying m⁶A exhibited higher expression than m⁶A-marked proto-oncogenes, an effect that was most pronounced in longer-lived species (Fig. S19A, B). When m⁶A was absent, expression differences between TSGs and proto-oncogenes were not significant (Fig. S19C, D). Moreover, within both gene classes, m⁶A-positive genes showed higher expression than their m⁶A-negative counterparts, with the TSG effect stronger in longer-lived lineages (Fig. S19E-H).

Our data indicate that higher m⁶A density on tumor suppressor genes in longer‑lived species is associated with increased—not decreased—expression of those genes, consistent with a protective, pro‑survival role for targeted m⁶A marking. We hypothesize that long‑lived lineages economize a reduced global m⁶A pool by concentrating modifications at functionally critical cancer‑associated sites in tumor suppressors, thereby enhancing transcript stability or translation where it most effectively reduces cancer risk while limiting systemic costs. Although direct in vivo functional validation is lacking, extensive literature supports the core regulatory function of m⁶A modification in suppressing cancer and aging. Actually, m⁶A modifications can serve as a double-edged sword in cancer biology, with both tumor-promoting and tumor-suppressing roles documented across various malignancies. However, mounting evidence suggests that context-dependent elevation of m⁶A levels can exert significant anti-tumor effects. Although causal validation is still needed, a growing body of work supports context‑dependent anti‑tumor and pro‑homeostatic roles for m⁶A: for example, m⁶A writers and erasers can respectively promote tumor suppression or tumorigenesis depending on substrate and cellular context, and modulation of m⁶A has been shown to affect immune responses and therapeutic sensitivity^50–55.

Conversely, loss or misregulation of m⁶A frequently promotes malignant phenotypes, underscoring the importance of finely tuned epitranscriptomic regulation. Emerging evidence also links m⁶A to aging processes—through effects on stem cell function, proteostasis, and DNA repair—providing a plausible mechanistic bridge between lifespan, cancer resistance, and the selective allocation of m⁶A in long-lived species^{25,27,56–60}. Overall, our findings motivate targeted experimental work to test whether strategic m⁶A placement on tumor suppressors causally contributes to longevity‑associated cancer resistance, thereby providing important insights into the cancer resistance mechanisms of longer-lived species.

Discussion

Our comparative analysis shows that cancer-associated sites in longer-lived mammals are more evolutionarily conserved than cancer-unrelated sites and are disproportionately marked by m⁶A. This enrichment is not global across all cancer genes but is concentrated at functionally important sites within tumor suppressor genes. Longer-lived species display greater conservation of tumor suppressors compared with proto-oncogenes and a higher number of m⁶A-marked TSGs, consistent with a model in which m⁶A marks are preferentially associated with key cancer-relevant loci in longer-lived species. However, whether this association reflects strategic allocation of a limited m⁶A pool or directly enhances transcript stability or translation remains to be determined. Such targeted epitranscriptomic placement may represent an economical, lineage-specific mechanism that complements other cancer-resistance strategies in long-lived species. Concurrently, though robust correlations among conservation, m⁶A enrichment, and expression were established, causal molecular mechanisms remain unproven. Experimental validation is needed to determine how m⁶A at these conserved sites affects transcript fate, protein function, and cancer susceptibility across species. Future work combining functional genomics, perturbation of m⁶A machinery, and cross-species models will be crucial to test causality and to elucidate whether strategic m⁶A targeting contributes directly to longevity-associated cancer resistance.

Methods

Species and cancer-associated sites

We employed longevity quotient (LQ)⁶¹ to assess whether different mammalian species are longer-lived on an evolutionary scale, as LQ considers the mutual influence of body weight and maximum lifespan based on the allometric equation of Austad and Fisher for nonflying eutherian mammals. The average LQ for all mammalian species was calculated to be 1.17, with a standard deviation (SD) of 0.8. According to the general standard, if the LQ of a species is greater than LQ_mean ± SD, it is considered a longer-lived species. Consequently, seven relatively longer-lived species and seven relatively shorter-lived species were identified. Using these two classification methods, mutation rates of homologous cancer-associated sites were calculated separately for longer-lived and shorter-lived species. We obtained multiz100way alignments from UCSC and filtered out 8,504 1-1 homologous genes present in at least two species for further analysis. All human cancer-associated sites were downloaded from the COSMIC database (Cosmic_GenomeScreensMutant_v99_GRCh38.tsv.gz dataset, accessed March 5, 2024).

To mitigate potential biases arising from alignment gaps and low-quality regions in cross-species comparisons, we implemented stringent filtering criteria for all multiple sequence alignments prior to conducting conservation analyses. Human sequences served as the reference anchor, and orthologous sequences from other species were aligned accordingly using OrthoFinder⁶². Alignment positions were excluded if the human residue was present but gaps were present in more than 70% of the non-human species, or conversely, if the human sequence contained a gap while a majority of other species retained aligned residues. Furthermore, to ensure sufficient information content for reliable site-specific evolutionary inference, alignments shorter than 10 amino acids after filtering were excluded from downstream analyses.

“Cancer‑associated sites” were defined as human genomic positions that recurrently harbor pathogenic or driver somatic mutations across tumors, as identified by intersecting entries from the Catalogue Of Somatic Mutations In Cancer (COSMIC) database version 95 and the cBioPortal for Cancer Genomics to maximize the reliability and stringency of our selection. Specifically, a site was deemed cancer-associated only if it was annotated as “pathogenic” or “driver” in both resources. “Cancer-unrelated sites” were defined as homologous positions lacking documented pathogenic somatic variants in either COSMIC v95 or cBioPortal, indicating a lack of evidence for a direct role in tumorigenesis based on current knowledge. To control for local sequence and functional context, we implemented a multivariable matched-pair design within protein-coding sequences. For each cancer-associated site, we selected a control site from the same gene within a 50-bp flanking window, prioritizing matches based on synonymous substitution context (to control for local mutation rate and codon usage) and amino-acid polarity (to account for biochemical environment and potential structural constraints). The matching was performed using a custom Python script. We only retained matched pairs where the difference in the proportion of polar amino acids was less than 5%. This pairing strategy minimizes confounding effects arising from local mutation rate, codon usage biases, and differences in residue chemistry.

Statistical inference combined three complementary resampling procedures to ensure the robustness and validity of our findings: (1) 1000 independent iterations of the matching algorithm were performed to obtain a stable median difference and to account for matching variability; (2) 10,000 bootstrap replicates were generated to compute bias-corrected and accelerated (BCa) confidence intervals for effect estimates; and (3) 10,000 permutation tests were conducted to derive nonparametric p-values under the null hypothesis of no difference between cancer-associated and control sites. All P values were corrected for multiple testing using the Benjamini–Hochberg false discovery rate (FDR) procedure with an alpha level of 0.05.

Site-specific conservation was then quantified using both phylogeny-aware methods (Rate4Site and RAxML) and model-independent conservation scores. The phylogeny-aware methods, specifically Rate4Site and RAxML explicitly account for phylogenetic tree topology and branch lengths, thereby normalizing for differences in background evolutionary rates among species.

Identifying conserved amino acids using Rate4site

Enter the alignment of the species and a phylogenetic tree with a time scale into Rate4site³⁸, and the conservation scores which estimated of the rates considers the topology and branch lengths of the phylogenetic tree, as well as the underlying stochastic process, are printed based on the amino-acids in this sequence. The position of each cancer-associated site in the genome was corresponding to the conserved amino acid fraction, and the evolutionary conserved differences between cancer-associated sites and cancer-unrelated sites, as well as the evolutionary conserved differences between longer-lived species and shorter-lived species were compared. The conservative score of each site is evaluated by Rate4site, and these sites are corresponding to the positions of cancer-associated sites, and then statistical analysis is performed. Rate4site will generate two result files. The r4s.res file standardizes the conservative scores of the r4sOrig.res file so that the average value is zero and the standard deviation is 1. Since we occasionally see the species evolutionary tree, the evolutionary units of all species are the same, and there is no need to standardize the conservative scores of each amino acid position within each gene, so the statistical statistics r4sOrig.res file. Compare the distribution of the conservative scores of longer-lived species and shorter-lived species, as well as the average values of the conservative scores of longer-lived and shorter-lived species in different segments of genes, to explore whether it is position-specific.

We defined the Q_ratio, where the rate of cancer-unrelated in longer-lived species divided by the rate in shorter-lived species is the numerator, and the rate of cancer mutations in longer-lived species divided by the rate in shorter-lived species is the denominator. n represents a cancer-associated sites on a gene, i represents a cancer-unrelated sites near the cancer-associated sites on a gene, i and n have the same number of sites. L_N is the mutation rate of cancer-unrelated sites in longer-lived species, S_N is the mutation rate of cancer-unrelated sites in shorter-lived species, L_M is the mutation rate of cancer-associated sites in longer-lived species, and S_M is the mutation rate of cancer-associated sites in shorter-lived species.

$$Q_{\mathrm{ratio}}=\frac{{\sum }_i{L}_N/S_N}{{\sum }_n{L}_M/S_M}$$ 1

Comparison of longer-lived and shorter-lived species ratios between different organs. Calculate the average distribution of ratio in each segment of the gene to judge the conservation of longer-lived species relative to shorter-lived species in different regions.

Calculating root-to-tip distance using RAxML

The root-to-tip distances of longer-lived and shorter-lived species were calculated by RAxML software³⁹. Each alignment distinguishes between the alignment of cancer-associated sites, the alignment of cancer-unrelated sites, and the alignment of all species. Longer-lived and shorter-lived species were compared respectively, and root-to-tip distance of cancer-associated sites and cancer-unrelated sites were compared. In addition, the rational distance of longer-lived species’ distance / shorter-lived species’ distance is calculated, and the distribution difference between root-to-tip distance results and Rate4site results is compared.

Model free conservative score calculation

We also estimated sitewise conservation using a simple⁶³, model-free score that quantifies deviation of the observed local sequence composition from a random expectation, using a formula below:

$$M_k\left(i\right)={\sum }_x{\left(p_i\left(x\right)-p_e\right)}^2/p_e$$ 2

This formula calculates a conservation score for a specific position (i) in a multiple sequence alignment by comparing the deviation between the observed frequency (P_i(x)) of k-mers and their expected frequency (p_e) under a random model. A higher score indicates that the sequence at that site deviates more from random expectation, i.e., is more conserved; this can more directly reflect the constraints acting on the patterns of raw mutational accumulation. We implemented a script to compute this score for every single site in our genome-wide alignment.

Calculating ω using the branch model

The branch model of PAML software⁴¹ was used to calculate the selection of longer-lived and shorter-lived species. Genes containing cancer pathogenic sites with significant differences in selection pressure between longer-lived and shorter-lived species were counted, and genes containing cancer pathogenic sites under positive selection in longer-lived species were identified. Statistical significance was assessed based on bias and standard error estimates. The 95% confidence interval was calculated by adding or subtracting 1.96 times the standard error from the point estimate. When this interval excluded zero, the null hypothesis was rejected, indicating significant differences between datasets.

To assess the evolutionary conservation and likely functional importance of human cancer-associated sites across mammals, we mapped COSMIC somatic mutation positions onto homologous columns of a multi-species whole-genome alignment. Homologous sequences were retrieved from the alignment and sites with low alignment confidence or excessive gaps were excluded to reduce mapping errors and spurious rate estimates. We then estimated site-specific evolutionary rates using two complementary, widely used approaches: LEISR (a likelihood-based site-rate estimator implemented in HyPhy)⁶⁴ and site models implemented in PAML⁴¹. These analyses tested whether mapped cancer‑associated sites evolve under stronger purifying selection than putatively neutral background positions (see Figure S4).

To ensure robustness, we compared rate estimates between methods and performed sensitivity checks that varied alignment depth and gap thresholds. Statistical significance of rate differences was assessed against matched control sets, and multiple-test correction was applied where appropriate.

As an independent validation of biological relevance, we cross-referenced mapped positions with recurrent somatic mutation data from cBioPortal. We observed substantial concordance between datasets, with the majority of mapped sites overlapping across resources (86.8% overlap with cBioPortal and 69.5% retained from COSMIC after alignment filtering; Fig. S5), supporting the reliability of our mapping and the conserved nature of many cancer-associated loci.

Gene enrichment analysis

To elucidate the functional implications of the observed evolutionary constraints, we conducted a systematic functional characterization of cancer-associated sites exhibiting high levels of conservation. For both longer-lived and shorter-lived mammalian groups, we ranked all cancer-associated sites based on their evolutionary conservation scores as calculated using Rate4Site³⁸. To identify a robust and biologically meaningful set of conserved loci, we defined a group-specific conservation score threshold based on the mode of the conservation score density distribution. This approach aimed to capture a set of sites exhibiting a distinctly higher level of conservation compared to the background distribution. For longer-lived species, a threshold of 0.09 was applied, identifying 22,567 highly conserved sites across 377 genes. For shorter-lived species, a threshold of 0.02 was used, corresponding to 11,016 sites across 239 genes.

The resulting sets of genes harboring these highly conserved sites were then subjected to Gene Ontology (GO) enrichment analysis to identify overrepresented biological processes. We utilized the enrichGO function⁶⁵ using the following parameters: the background gene set was defined as all human protein-coding genes, the organism was set to Homo sapiens, and the P-value and q-value cutoffs were set at 0.05 with Benjamini-Hochberg correction for multiple testing. Redundant GO terms were then simplified using the simplify function. Pathways related to cellular stress resistance, metabolic regulation, and macromolecular quality control were specifically examined for their relevance to longevity-associated cancer resistance, based on prior literature and established biological knowledge.

We use “Polysel”⁶⁶ for gene set enrichment analysis which is possible to detect pathways containing pleiotropic signals. In addition, other variables (such as gene length and genetic distance) can also be used to adjust statistical variables. For the species-specific expression, we used the sum of conservative score calculated by Rate4site software as the gene score (SUMSTAT) to detect pathways enriched with more conserved cancer-associated sites in longer-lived mammals. For mutation rate variation, we used the coefficients in the PGLS regression as SUMSTAT to enrich genes that are positively related to longevity (SUMSTAT of negatively related genes is set to 0) and negatively related genes (SUMSTAT of positively related genes is set to 0 and converted to Absolute value) to detect longevity-related pathways with genetic minor effects. We used the function “RescaleBins” to adjust SUMSTAT. We used “ks.test” in R to check whether the gene score (SUMSTAT) is normally distributed or not. If not, 400,000 random data set was generated to construct an empirical distribution.

The intersection of 5mC, m⁶A and cancer-associated sites

The m5C modification sites of DNA and the m⁶A modification sites of RNA were obtained from a prior study conducted by our group; the majority of samples in that study were derived from males⁶⁷. The homologous positions of human cancer-associated sites in 21 species were obtained using Clustal Omega⁶⁸. Then, the homologous positions of cancer-associated sites in each species were matched with the modification sites.

Reporting summary

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

Author contributions

Xuming Zhou and Xiaoxiao Zhang conceived the study and designed the project; Xiaoxiao Zhang performed the data collection and curation, Meng Li, Xuming Zhou, Gaoming Liu provided assistance; Xiaoxiao Zhang conducted bioinformatics analysis, visualization and wrote the manuscript; Xuming Zhou and Xiaoxiao Zhang designed figures; Xuming Zhou acquired funding and supervised the study; Xuming Zhou, Xiaoxiao Zhang, and Gaoming Liu provided suggestions for data analysis; Xiaoxiao Zhang, Gaoming Liu, Xin Tong, Weixiao Chen, Pingfen Zhu, Juan Du, Xiao Wang and Yue Sun contributed to data interpretation and revised this manuscript.

Peer review

Peer review information

Nature Communications thanks Yang Li, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This project was granted by the National Natural Science Foundation of China (No. 32270437), Investigation of Nationally Protected Wildlife Species in Tongren Region, Guizhou Province (GZKPC-2025-01), the Institute of Zoology, Chinese Academy of Sciences (2023IOZ0104), the Innovative Talents International Cooperative Training Project of China Scholarship Council (No. CXXM20230124) and the Prevention and Control of Emerging and Major Infectious Diseases-National Science and Technology Major Project (2026ZD01911000).

Data availability

Source data are provided with this paper as a Source Data file. All other data supporting the findings of this study are available within the Article and its Supplementary Information. The processed m⁶A and 5mC site-level datasets, together with the associated analysis code and other processed data, have been deposited in Zenodo and are publicly available at https://doi.org/10.5281/zenodo.20244947. The raw Oxford Nanopore direct RNA sequencing (ONT-DRS) and next-generation sequencing (NGS) data used in this study are publicly available in the Genome Sequence Archive at the National Genomics Data Center, China National Center for Bioinformation/Beijing Institute of Genomics, Chinese Academy of Sciences, under accession number CRA015510. The 62-species mammalian multiple-sequence alignment based on the human GRCh38/hg38 reference genome was obtained from the UCSC Genome Browser hg38 100-way alignment resource (http://hgdownload.soe.ucsc.edu/goldenPath/hg38/multiz100way/). Cancer mutation data aligned to the GRCh38 human reference genome were obtained from COSMIC v99. COSMIC data are available to registered non-commercial users following registration and login, whereas commercial access is subject to the applicable COSMIC licensing terms (https://cancer.sanger.ac.uk/cosmic/download). Additional cancer mutation data were retrieved from the publicly accessible cBioPortal for Cancer Genomics (https://www.cbioportal.org/) through its API. The custom R script used to retrieve the mutation data and extract their genomic coordinates is publicly available at https://github.com/XiaoxiaoZhang57/cancer_m6A_program/blob/main/scripts/get_mutation_from_cbioportal.R. The processed cBioPortal-derived cancer mutation data are included in the Zenodo repository associated with this study (https://doi.org/10.5281/zenodo.20244947). Except for COSMIC data, which require registration and are subject to the applicable licensing terms, all datasets listed above are publicly accessible without restriction. Source data are provided with this paper.

Code availability

All custom code required to reproduce the analyses reported in this study is available on GitHub at https://github.com/XiaoxiaoZhang57/cancer_m6A_program and has been permanently archived in Zenodo under https://doi.org/10.5281/zenodo.20244947⁶⁹. The code is released under the MIT License.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76479-3.