Age distinguishes selection from causation in cancer genomes Authors: David Cheek, Martin Blohmer, Martin A Nowak, Tibor Antal, Kamila Naxerova Published: 2026 ABSTRACT AbstractCancer-causing mutations have been identified primarily from positive selection signals in cancer genomes. However, positive selection is also a ubiquitous feature of normal tissue aging. Here we develop a statistical framework to disentangle selection in normal tissue and causation of carcinogenesis. By comparing cancer and normal tissue genomes, we estimate the effects of mutations on cancer risk in the blood, esophagus and colon. We determine that stronger cancer-causing mutations are enriched at younger patient ages. This enables cancer-causing mutations to be identified from patient age distributions, even without normal tissue data. Moreover, we show for acute myeloid leukemia that the age-dependence of purported causal mutations can be explained largely by normal blood evolution, challenging the long-standing notion that childhood cancers require distinct mutations. Broadly, our framework delineates carcinogenesis from normal tissue aging, improving the assessment of cancer risk conferred by mutations.Subject terms: Cancer, Genetics Main The question of which mutations cause cancer, after decades of research, remains central to understanding carcinogenesis and developing treatments. Most answers have come from cancer genomes viewed through a Darwinian lens. Mutations found in cancer genomes more commonly than expected relative to neutral mutations are inferred to be under positive selection and are classified as cancer drivers1–9. However, positively selected mutations are widespread in aging tissues even in the absence of cancer10–16, and positive selection is not equivalent to cancer causation13,16. For example, NOTCH1 shows strong signals of positive selection in esophageal squamous cell carcinoma (ESSC), but NOTCH1 is even more frequently mutated in the normal esophageal epithelium17,18, suggesting that NOTCH1 mutations are positively selected in the normal esophagus while inhibiting carcinogenesis. Murine experiments support this interpretation, contradicting the prior classification of NOTCH1 as a cancer driver gene19. This example exposes a broader uncertainty surrounding the identities and effect sizes of cancer-causing mutations. Here, we develop a statistical inference framework to estimate the contributions of mutations to cancer risk from cancer and normal tissue genomes. Given that high-resolution normal tissue data remain limited for many tissues, we further investigate whether causal information may be extracted from cancer genomes without normal tissue data. We examine patient age data alongside cancer genomes to distinguish mutations that cause cancer from mutations that merely accumulate in aging normal tissues. Results Defining the carcinogenic effect of mutations We defined the carcinogenic effect of a mutation by its influence on the probability of a cell initiating a cancer that is eventually diagnosed. Consider two hypothetical cells, statistically identical except that one cell has mutation z while the other does not. The cell with mutation z initiates cancer at some rate (probability per unit time), r1, while the other cell initiates cancer at rate r0. The per-cell hazard ratio for cancer initiation given the mutation, w = r1 / r0, defines the carcinogenic effect of the mutation (Fig. 1a,b). Mutation frequencies in normal tissue and cancer genomes estimate carcinogenic effects. a, Carcinogenic effect definition. Cell with mutation z initiates cancer at rate r1. Cell without mutation z, otherwise statistically identical, initiates cancer at rate r0. The carcinogenic effect of mutation z is r1 / r0. b, Relationship between the carcinogenic effect of the mutation (x axis), fraction of cells with mutation in tissue (color) and cancer rate in tissue given the presence of the mutation, normalized by cancer rate without the mutation (y axis). c, Relationship between the carcinogenic effect of the mutation (light green, carcinogenic; gray, protective), fraction of normal cells with mutation among many people (x axis) and fraction of cancer founder cells with the mutation (y axis). d,e, Carcinogenic effects estimated per gene for ESSC and AML. d, Fraction of 63 ESSC samples with non-synonymous mutations or small indels in each gene (y axis; with jitter for visualization) against estimated fraction of cells mutated in patient-matched normal esophageal epithelium samples (x axis). Data from Yokoyama et al.18; genes exhibiting positive selection in ESSC are plotted. Genes mutated in fewer than five samples are light colored. e, Fraction of 2,340 AML samples23–25 with non-synonymous mutations or small indels in each gene (y axis) against estimated fraction of cells mutated in 2,410 age-matched normal blood samples21,22. AML driver genes from Bailey et al.26 are plotted. Horizontal and vertical error bars represent 95% bootstrap and binomial confidence intervals for mutation frequencies, respectively. Dotted lines indicate carcinogenic effect estimates. The carcinogenic effect can also be conceptualized as an imaginary randomized controlled trial in which normal tissue cells from many people are randomly assigned or disallowed mutation z, and some of the cells subsequently transform to cancer. If fraction x of the normal tissue cells is assigned the mutation and fraction y of cancer founders carry the mutation, then the odds ratio of mutation frequencies in cancer relative to normal tissue estimates the carcinogenic effect of the mutation (Fig. 1c and Methods). Cancer and normal tissue genomes estimate carcinogenic effects Mutations can never be artificially induced in humans to test causality; therefore, the carcinogenic effect, by our definition, can never be perfectly measured. Nevertheless, the hypothetical experiment that defines carcinogenicity has a measurable natural analog provided by the genetic heterogeneity of human tissues, whereby some cells carry mutation z while others do not. The frequency, x, of mutation z among normal tissue cells can be estimated from deep sequencing of normal tissue samples, while the frequency, y, of mutation z among cancer founder cells can be estimated by the fraction of cancers in which mutation z is clonal, enabling calculation of the odds ratio (equation 1). It is important to emphasize that somatic evolution does not necessarily distribute the mutation among cells independently of other variables that might influence carcinogenesis. If the presence of a mutation in a normal tissue cell is associated with another variable that causes cancer (such as another mutation), then the odds ratio of that mutation in cancer relative to normal tissue (Equation 1) could overestimate or underestimate the carcinogenic effect, depending on whether the association is positive or negative. See Supplementary Notes 1 and 2 for mathematical discussions of confoundment. With this caveat in mind, we used the odds ratio as a rough, practical estimate for the carcinogenic effect. We turned first to the esophagus, a tissue whose somatic evolution is relatively well documented. Yokoyama et al.18 performed targeted sequencing of purported driver genes among 68 ESSC samples and 68 physiologically normal esophagus samples from the same 68 patients. For a fair comparison of malignant and non-malignant cells, we excluded five patients with shared mutations between cancer and ‘normal’ samples. Then, focusing on non-synonymous mutations and insertions and deletions (indels), we compared mutation frequencies between ESSC and the normal esophageal epithelium to estimate carcinogenic effects per gene (Fig. 1d). A limitation of our estimates is that we do not distinguish the carcinogenic effects of different mutations within a gene, which would be possible with larger datasets. The resulting panorama showcases a spectrum of cancer-causing potency among genes (Fig. 1d). TP53 and NFE2L2 have carcinogenic effect estimates standing in the hundreds, while NOTCH1 mutations inhibit carcinogenesis. Several other purported cancer driver genes (for example, EP300, FAT1, PIK3CA) have carcinogenic effect estimates in the statistical vicinity of one, despite exhibiting strong signals of positive selection in ESSCs18. The estimate for PIK3CA is compatible with a recent mouse study showing only a modest cancer-causing effect20. Overall, our estimates support that cancer causation and positive selection in the esophagus are not equivalent. Next, we turned to the blood. We compared mutation frequencies in 2,410 normal blood samples21,22 to 2,340 acute myeloid leukemia (AML) samples23–25 with the same age range (36–103 years; see Methods). We focused on 21 genes that were sequenced in all normal blood and AML studies and that are part of a standard list of AML driver genes26. As in the esophagus, we estimated a wide span of carcinogenic effects in the blood (Fig. 1e and Extended Data Fig. 1a): approximately ten for the most common clonal hematopoiesis driver genes ASXL1, DNMT3A, TET2 and SF3B1, and approximately 100 for TP53; meanwhile, mutations in FLT3, CEBPA, IDH2 and WT1 increase the transformation rate of a cell by about 800-fold to 10,000-fold. The most frequently mutated gene in AML, yet unmutated in the normal blood data, is NPM1, indicating an exceptionally powerful carcinogenic effect. Our estimates of the effects of mutations on leukemogenesis broadly agree with associations between normal blood mutations and subsequent AML development21,27,28 (Extended Data Fig. 1b–e), supporting the consistency of our framework against a broader repertoire of data. Carcinogenic effect estimates for AML and colorectal cancer. (a) Carcinogenic effects for adult AML: original estimates from Fig. 1e (x axis) vs. alternative estimates from regression model that accounts for age as covariate (y axis). See Methods for details of regression. Estimates for NPM1 are infinite but presented at finite values for visualization. Dashed line is y = x. Spearman’s correlation rho=0.98, p = 4.8×10−6. Wilcoxon paired test p = 0.49. (b–d) Carcinogenic effect estimates for adult AML based on odds ratio of mutation frequencies in AML relative to normal blood (Fig. 1e; x axis) vs. estimated effects of normal blood mutations on AML risk from prior literature (y axis): (b) Hazard ratio for AML development given mutation observed among 453,584 individuals in UK Biobank, estimated by Gu et al.28; Spearman’s correlation rho=0.96, p = 0.0028. (c) Odds ratio for AML development given 10% increase in VAF of mutation in normal blood among 808 individuals, estimated by Abelson et al.21; Spearman’s correlation rho = −0.098, p = 0.69. (d) Odds ratio for AML development given mutation at VAF > 1% in normal blood among 360 individuals, estimated by Desai et al.27; Spearman’s correlation rho=0.67, p = 0.083. Mutations in SF3B1, SRSF2 and U2AF1 are denoted ‘spliceosome’. Mutations in IDH1 and IDH2 are denoted ‘IDH’. (e) Carcinogenic effects for adult AML from Fig. 1e (x axis) vs. carcinogenic effects for adult AML estimated from relationship between normal blood mutations and subsequent AML incidence, using data from Abelson et al.21 and theory from Fig. 1b. The inference does not work for genes with few mutations in normal blood, because no association with AML incidence can be measured. Genes significantly associated (q < 0.05) with AML incidence are plotted. Estimate for FLT3 is infinite but presented at finite value for visualization. Spearman’s correlation rho=0.77, p = 0.008. Wilcoxon paired test p = 0.9. (a–e) All statistical tests are two-sided. All horizontal error bars are 95% bootstrap intervals. Vertical error bars in (a) and (e) are 95% Wilks intervals. Vertical error bars in (b–d) are 95% intervals as previously published. (f) Non-synonymous SNV burdens in 1224 AML samples from COSMIC and 3592 single normal hematopoietic cell-derived colonies from Mitchell et al.58. (g) Exome-wide SNV burdens in 1509 ESSC samples from COSMIC and 3 single normal esophageal epithelial cell-derived colonies from drinkers and smokers in Yokoyama et al.18. (h) Exome-wide SNV+indel burden distribution in 2411 COADREAD samples from COSMIC and 1387 normal colonic crypts, from donors with and without inflammatory bowel conditions, in Lee-Six et al.33 and Olafsson et al.14. Mutational types differ between (f), (g), and (h) due to data accessibility. (i) Mutation frequencies per COADREAD driver gene in COADREAD samples and normal crypts. Error bars are 95% binomial confidence intervals. (j) Carcinogenic effect estimates for COADREAD driver genes with 95% Wilks confidence intervals, based on logistic regression model that accounts for patient age and mutation burden (Methods). (i,j) Using samples as in (h) but excluding COADREAD samples whose exonic mutation burden is greater than 251 (maximum of normal crypt mutation burdens). We note that carcinogenic effect estimates are higher for the blood than for the esophagus. One conceivable explanation for this difference is that copy number alterations, rather than single nucleotide variants and small indels (SNVs/small indels), are more important for esophageal carcinogenesis. It could also be that a greater number of mutations is required for—and therefore each mutation contributes less to—ESSC than AML. There is also a technical difference between the cancer types: the normal esophagus samples are patient-matched to the cancer samples, whereas the normal blood samples are from people with no history of blood malignancies. The blood analysis is therefore more exposed to confounding variables. People who develop AML may statistically differ from those who do not with respect to multiple variables, such as smoking, diet or germline genetics, which may influence mutation frequencies in normal blood29–31 and the propensity of an individual to develop AML, potentially biasing our estimates. For AML and ESSC, cancer samples have mutation burdens similar to normal tissue (Extended Data Fig. 1f,g), suggesting that mutagenesis is not a major source of bias in our estimates. However, other cancers such as colorectal cancer may have a far greater mutation burden than normal tissue32 (Extended Data Fig. 1h). We focused on 1,724 colorectal cancer samples (from COSMIC33) for which the mutation burden is in the same range as 1,387 normal colonic crypts14,34, and we used a logistic regression model with mutation burden and age as covariates to estimate carcinogenic effects (Methods and Extended Data Fig. 1i,j). Again, we measured a large range: TP53, APC, KRAS and PIK3CA mutations have carcinogenic effects in the hundreds, FBXW7 and PTEN appear one order of magnitude less important, while ARID1A has a carcinogenic effect that does not significantly differ from one. For the future, normal tissue studies with larger cohorts and error-corrected sequencing35–39 will enable increasingly well-controlled and high-resolution quantifications of the effects of mutations, both on clonal expansion in normal tissue and on the risk of cancerous transformation in a cell. However, the need to determine cancer causation is immediate. For now, cancer genetic data are copious while normal tissue data remain limited. We therefore endeavored to disentangle causation and selection from cancer genomes without a normal tissue control. An evolutionary model predicts age-dependent cancer genetics In the remainder of the paper, we explore whether the age of patients with cancer can be used to distinguish selection and causation. We hypothesized that (1) mutations that are positively selected but not carcinogenic appear preferentially in older patients with cancer, because these mutations accumulate in our aging normal tissues and are passively inherited to cancer, and (2) potent cancer-causing mutations appear preferentially in the cancer genomes of younger patients, because these mutations accelerate the onset of carcinogenesis (Fig. 2a). Selection and carcinogenicity differentially shape the age distributions of mutations in cancers. a, Conceptual cartoon of how patient age might distinguish carcinogenic mutations from non-carcinogenic mutations. Top, non-carcinogenic mutation that is positively selected in normal tissue occupies an increasingly large proportion of normal tissue towards older age. Cancer originates from a cell carrying a mutation with a probability equal to the frequency of that mutation in the tissue. Bottom, a strongly carcinogenic mutation accelerates the onset of cancer. b, Branching process model of a somatic evolution in normal tissue that illustrates stochastic cell division, loss (through death or differentiation), acquisition of mutation z and cancer initiation. The difference between the growth rate of cells with and without mutation z is the selective effect in normal tissue. The ratio of rates at which cells with and without mutation z initiate cancer is the carcinogenic effect. c, Simulations of the number of cells with mutation z in normal tissue, according to the branching process model shown in b. Each line represents a clone that initiates at an age uniformly distributed between 0 and 80 years. d,e, Theoretical mutation frequency among normal tissue cells (d) and cancer-initiating cells (e) as a function of the selective effect of the mutation in normal tissue, carcinogenic effect and tissue age. f, Selection in normal tissue and cancer causation lead to similar mutation frequencies in cancer genomes. Colors represent the frequency of cancer with mutation z divided by the frequency of cancer with a neutral mutation of the same acquisition rate. g, Selection in normal tissue and cancer causation lead to differing age distributions for mutations in cancer genomes. Colors represent the mean patient age of cancers carrying mutation z. In c–g, parameter values: lifespan, 80 years; stem cell pool size, n = 105; mutation rate per year, v = 10−8; division and loss rate, a0 = b0 = 1 per year without mutation and a1 = 1 + s, b1 = 1 with mutation; baseline carcinogenesis rate, r0=10ut3, where u is uniformly distributed on (−16,−13), caricaturing that cancer risk rises with age and varies between people. To express the hypothesis quantitatively, we formulated a branching process model of somatic evolution (Fig. 2b,c). The model illustrates a tissue that stochastically acquires a mutation, and that mutation may edit the rate of cell division, death, differentiation or transformation to cancer. This model minus the carcinogenesis component has been applied to somatic blood evolution15,29,40, with predictions largely consistent with thousands of blood samples15,22,40. For other tissues, the model can be prudently considered a cartoon. We calculated the expected frequency of mutation z among normal tissue cells (Fig. 2d) and among cancer-initiating cells (Fig. 2e and Extended Data Fig. 2a), observing increasing mutation frequencies towards older ages. Then, collating ages, we calculated the odds ratio for the presence of mutation z in cancer genomes relative to a neutral mutation with the same mutation rate, mimicking dN/dS in cancer genomic analyses. This statistic has a positive relationship with both the selective effect in normal tissue and with the carcinogenic effect (Fig. 2f and Extended Data Fig. 2a), illustrating that dN/dS alone cannot distinguish selection from causation. Parameter space exploration for branching process calculations. (a,b) Calculations for model of Fig. 2. (a) Frequency of cancers with mutation z divided by frequency of cancers with neutral mutation of same acquisition rate. (b) Mean patient age for cancers with mutation z minus mean patient age for cancers with neutral mutation of same acquisition rate. (a,b) Model and parameter notation as in Fig. 2. Next, to assess the preference of mutation z towards younger or older cancer genomes, we calculated the mean age of patients whose cancers carry mutation z. We found that for a carcinogenic effect of 1 (no cancer causation), the selective effect in normal tissue and mean age have a positive relationship, whereas for any specific selective effect, the carcinogenic effect and mean age have a negative relationship (Fig. 2g and Supplementary Note 3). Although the precise numerical details of these relationships are sensitive to other somatic evolutionary parameters, which may vary greatly between tissues, the directionality of the relationships is robust (Extended Data Fig. 2b). Our simple model predicts that the selective and carcinogenic effects of a mutation influence the age distribution of patients with the mutation, and, thus, differences between mutations in their effects should lead to age-dependent mutational profiles of cancers. However, age-dependent cancer genetics could also have other explanations. We note a complication to the model: it could be that the carcinogenic effects of the mutation are themselves age-dependent. Age-dependent effects might arise, for example, from mutational interactions with the aging epigenetic state of stem cells, or with other somatic mutations that accumulate with age or as a result of age-dependent cell-extrinsic factors. Relatedly, young patients with cancer might be more likely to have cancer-predisposing germline mutations, potentially interacting with specific somatic mutations. Overall, there are several possible reasons to expect that driver mutation frequency spectra differ between older and younger patients with cancer41,42. Careful empirical investigation is therefore required to assess whether age distributions may help to distinguish selection and causation in real-world cancer data. Hence, we asked whether (1) patient age distributions can differentiate mutations and (2) age differences between mutations can be explained by age-dependent carcinogenic effects. Ideally, the answers to (1) and (2) would be positive and negative, respectively, inviting the next two questions: whether (3) powerful carcinogenic mutations are found preferentially in younger patients with cancer, and (4) mutations that are positively selected in normal tissue are found preferentially in older patients with cancer. We begin with question (1). Cancer types differ in age divergence between driver genes That cancer genome alterations depend on age has already been established for many cancer types41–44. To quantify the degree of age divergence between purported driver genes per cancer type in COSMIC33, we calculated the variance between purported driver genes (using an established list of driver genes26) of patient ages at which non-synonymous SNVs/small indels appear, normalized by the overall variance of the patient age distribution. Comparing cancer types, we observed varying age divergence between driver genes, with esophageal and colorectal cancers exhibiting relatively little and AML exhibiting the most (Fig. 3a). Age distributions of AML mutations provide information on selection and carcinogenicity. a, Among cancer types, AML exhibits the greatest age divergence between driver genes. The age divergence statistic is defined by the variance between driver genes of non-synonymous SNVs/small indels for patient age, normalized by the overall variance of patient age per cancer type (Methods). Cancer type-specific driver genes are listed in Bailey et al.26. P values (color) from the Kruskal–Wallis test of patient age versus driver gene. Data are from COSMIC (Methods). See Supplementary Table 3 for unabbreviated cancer types. b, Mean age of AML patients with non-synonymous SNVs/small indels per driver gene (Methods). Error bars are 95% confidence intervals. Four AML datasets were collated (Extended Data Fig. 3a). c, Carcinogenic effect estimates for older AML (Fig. 1e; mean age, 61 years, s.d., 12 years) versus relative carcinogenic effect estimates for younger AML (normalized by median carcinogenic effects in older AML; Extended Data Fig. 3; mean age, 11 years, s.d., 7 years). NPM1 has an infinite carcinogenic effect estimate for older AML, represented at finite value for visualization. The line represents regression excluding NPM1. Spearman’s correlation, rho = 0.72, P = 0.00038. d, Fraction of older-onset AML versus fraction of younger-onset AML samples with non-synonymous SNVs/small indels per gene. rho = 0.12, P = 0.62. e, Mean AML patient age for gene mutations versus number of mutations per gene per kilobase per normal blood cell (rho = 0.6, P = 0.0045). Mutation densities in normal blood were estimated from 2,410 samples21,22; donor age range, 36–103 years. Genes under positive selection in normal blood according to dN/dS > 1 with q < 0.05 (ref. 22) are denoted by a star shape; these genes have older mean AML patient ages (two-sided Wilcoxon test, P = 0.016). f, Mutation frequencies in normal blood (normalized and unnormalized by gene length; data as in e) and carcinogenic effect estimates (younger and older AML, and combined score; see Extended Data Fig. 3e) have distinct relationships with AML patient age, according to Spearman’s correlation among driver genes. Mutation frequencies in AML (normalized and unnormalized by gene length) do not have a significant correlation with AML patient age. In c–f, P values are from two-sided tests, unadjusted for multiple hypothesis testing. Indeed, it is well known that the genetic makeup of childhood and adult AMLs differs starkly45. For a higher-powered view of AML, we collated four datasets (including data in Fig. 1e and a pediatric AML dataset45) totaling 4,736 patients with AML ranging in age from 0 to 100 years (Extended Data Fig. 3a; see Methods). We calculated the mean patient age for non-synonymous SNVs/small indels per driver gene, observing wide separation of genes (Fig. 3b). For example, the mean age for KIT was 36 years, while the mean age for TET2 was 63 years. Given the age differences observed for these driver genes, and that normal somatic evolution has been documented far more comprehensively in the blood than in other organs, we decided that AML would be the ideal primary exploration ground for questions (2), (3) and (4). Supplementary analyses on AML. (a) AML data overview. Numbers of samples and ages of individuals among normal blood and AML datasets used in our analyses. (b) Mutation frequencies per gene among 1481 patients in young-onset AML group vs. gene lengths. Contour lines denote estimates of relative carcinogenic effects (carcinogenic effects normalized by median carcinogenic effect in older-onset AML, see Methods). Vertical lines are 95% binomial confidence intervals for mutation frequencies. (c) Robustness analysis for comparison of older- and young-onset AML in Fig. 3c,d. Spearman’s correlation between age groups in terms of mutation frequencies (light line) and carcinogenic effect estimates (dark line) among AML driver genes. X-axis specifies upper age limit for patient to be included in younger group while subplot title specifies lower age limit required for patient and normal blood donor to be included in older group. (d) Gene-specific comparison of carcinogenic effect estimates between age groups. Genes are ranked by their carcinogenic effect estimates separately in adult and childhood AML, then difference between rankings is calculated. Error bars represent 95% bootstrap intervals for difference based on 1000 bootstrap resamples of individuals. (e) Mean patient age negatively associates with carcinogenic effect estimate. X-axis shows genes ranked by carcinogenic effect estimates separately for older- and younger-AML (triangle, square), and average of older- and younger-AML rank (circle). Mean patient age correlation with carcinogenic effect estimate is rho = −0.44, p = 0.048 for older-AML; rho = −0.50, p = 0.023 for younger-AML; rho = −0.51, p = 0.018 for average rank. Carcinogenic effects align between young-onset and older-onset AML For question (2), we wanted to compare our carcinogenic effect estimates for adult-onset AML (Fig. 1e) against those for young-onset AML. Our adult-onset AML cohort comprises 2,973 patients with a mean age of 61 years (s.d., 12 years); for a young-onset AML cohort, we selected patients younger than 25 years old from the collated AML datasets (Extended Data Fig. 3a), totaling 1,481 patients with a mean age of 11 years (s.d., 7 years). The new difficulty for inference is that mutations in normal childhood blood are rarely observed, with variant allele frequencies (VAFs) too low for reliable detection46. We approximated mutation frequencies in normal blood as proportional to gene length, thereby estimating carcinogenic effects for young-onset AML up to an unknown normalizing constant (Methods). We then normalized by the median of the adult carcinogenic effects (Extended Data Fig. 3b). We found that carcinogenic effect estimates strongly correlate between the old and young age groups (rho = 0.72, P = 0.00034; Fig. 3c). By contrast, driver mutation frequencies do not significantly correlate between the age groups (rho = 0.12, P = 0.62; Fig. 3d). In other words, the frequencies of driver mutations are radically revised with the progression of age, whereas their carcinogenic effects are comparatively stable in ranking. This result is robust to changes in the age boundaries that define the patient groups (Extended Data Fig. 3c). In conclusion, although some degree of age-dependence in carcinogenic effects is plausible (Extended Data Fig. 3d and Supplementary Note 4), this by itself cannot explain the age-dependence of mutation frequencies in AML. Carcinogenic effects in AML negatively associate with patient age We next wanted to test whether mutations found in the AML genomes of younger patients tend to be more powerful cancer-causing mutations (question (3); Fig. 2g). For a summary statistic of age, we again used the mean patient age of mutations per driver gene (Fig. 3b). As expected, mean age negatively correlates with carcinogenic effect (rho = −0.44, P = 0.048 and rho = −0.5, P = 0.023, respectively for adult-onset and young-onset estimates; Extended Data Fig. 3e), consistent with the concept that causation corresponds to a quickening of carcinogenesis. According to our framework, the other explanatory factor that may underlie the age distributions of mutations in AML is selection in normal blood. Selection in normal blood positively associates with AML patient age To compare genes in terms of selection in normal blood (question (4)), we calculated for each gene the number of non-synonymous mutations per kilobase per normal blood cell (Methods). We observed that mutation density in normal blood positively correlates with mean age in AML (rho = 0.61, P = 0.0038; Fig. 3e). For another view of selection, we turned to dNdScv8 measurements from Fabre et al.22: seven out of 21 genes are under significant positive selection in normal blood (dN/dS > 1, q < 0.05), and these genes have older mean ages in AML (Wilcoxon test, P = 0.016; Fig. 3e). In summary, positive selection in normal blood manifests preferentially in AML genomes of older patients, which stands contrary to the inverse relationship between carcinogenic effect estimates and mean ages in AML (Fig. 3f). Therefore, for the best-documented tissue of the body, causation of cancer and selection in normal tissue leave distinct signals in patient age distributions. Young-age biases of cancer mutations indicate carcinogenicity We extended our investigation to cancer types lacking normal tissue data, seeking to extract causal information from age distributions. More humbly than parameter inference, we looked to test the null hypothesis that a mutation has no influence on the rate of malignant transformation. We defined the null hypothesis by a broad class of models of somatic evolution—including our original branching process model (Figs. 2b and 4a) and models in which clonal expansions follow sub-exponential growth curves (Extended Data Fig. 4a)—with the unifying constraint that the carcinogenic effect of the mutation is one (Supplementary Note 5). A general prediction emerged: for a mutation that is positively selected in normal tissue but does not influence cancer, the mutation is found in cancer genomes of older patients relative to neutral mutations (Figs. 2g and 4b, Extended Data Fig. 4b and Supplementary Note 5). Crucially, the prediction is falsifiable. If a mutation both exhibits positive selection in cancer genomes (elevated frequency relative to neutral mutations) and is enriched towards younger patient ages relative to neutral mutations, then the null hypothesis would be rejected in favor of the alternative hypothesis that the mutation is carcinogenic (Fig. 4c). Young-age biases in cancer genomes indicate carcinogenicity. Mutation frequency and age statistics according to the branching process model. a, Expected growth of positively selected mutant clones in normal tissue. b, Positively selected non-carcinogenic mutations are found preferentially in cancers in older patients relative to neutral mutations. The y axis shows the mean patient age for cancers with mutation z minus the mean patient age for neutral mutations. The x axis shows the expected frequency of mutation z in cancers divided by the expected frequency of neutral mutation with the same acquisition rate. Color in a and b denotes the selective effect of mutation z in normal tissue. Generalization is provided in Extended Data Fig. 4a,b and Supplementary Note 5. c, Null hypothesis test schematic. Mutations with elevated frequency and young-age bias in cancer genomes relative to neutral mutations are incompatible with the null model of no carcinogenicity. d,e, Null hypothesis test for driver genes in breast invasive carcinoma(d) and glioblastoma (e). Total of 1,562 breast invasive carcinoma samples and 775 glioblastoma samples from COSMIC. The y axis shows the mean patient age for non-synonymous mutations and small indels per driver gene minus the mean patient age for synonymous mutations in the exome (Methods; for short, age difference). Error bars are 95% Wald confidence intervals. Color denotes the quantile of the age difference statistic with respect to the null distribution, generated by permuting patient ages (Methods). The x axis is the number of mutations per gene divided by the number of mutations in the exome across the cohort, normalized by coding sequence length. Genes are from Bailey et al.26; the top 12 genes with the greatest number of mutations per kilobase are plotted. f,g, Mean patient age per driver gene negatively associates with carcinogenic effect estimate in esophageal (rho = −0.66, P = 0.012) and colorectal cancers (rho = −0.65, P = 0.0023). P values are from two-sided tests. The y axis shows mean age per driver gene relative to synonymous mutations, as in d and e. Patient ages from 908 ESSC samples and 1,724 colorectal cancer samples in COSMIC. Colorectal cancer samples with low mutation burden, as in Extended Data Fig. 1j. Carcinogenic effect estimates from Fig. 1d and Extended Data Fig. 1j. Genes with infinite carcinogenic effect estimates are denoted by ‘>’. Genes whose carcinogenic effect estimates are based on fewer than five samples are colored gray and excluded from correlation statistics and regression lines. Age distributions of non-synonymous relative to synonymous mutations. (a) Expected growth of positively selected mutant clones in normal tissue, according to exponential, Gompertz, logistic, or polynomial models. (b) Positively selected non-carcinogenic mutations are found preferentially in older patients’ cancers relative to neutral mutations. Y-axis shows the mean patient age for cancers with mutation z minus the mean patient age for neutral mutations. X-axis shows expected frequency of mutation z in cancers divided by expected frequency of neutral mutation with the same acquisition rate. Dot colors correspond to clonal expansion growth curves in normal tissue from (a). (a,b) Full parametrization in Supplementary Note 5. (c) Null hypothesis test for driver genes in AML. 939 AML samples from COSMIC. Y-axis shows mean patient age for non-synonymous mutations and small indels per driver gene minus the mean patient age for synonymous mutations in exome. Error bars are 95% Wald confidence intervals. Color denotes quantile of age difference statistic with respect to null distribution, generated by permuting patient ages (Methods). X-axis is number of mutations per gene divided by number of mutations in exome across the cohort, normalized by coding sequence length. Genes as in Figs. 1e and 3. (d,e) Genes with young-age bias in AML (q < 0.05 in c) are less mutated in normal blood (d; Wilcoxon test p = 0.21) and have greater carcinogenic effect estimates (e; Wilcoxon test p = 0.035). (d) Y-axis is number of mutations in gene per kilobase per normal blood cell, from Fig. 3e. (e) Y-axis is rank carcinogenic effect estimate (mean of ranking for young- and older-onset AML). (f) Fraction of driver genes per cancer type with p < 0.05, where p is quantile of age difference statistic with respect to null distribution (Methods). Vertical lines denote null expectation 0.05. Cancer types with at least 10 putative driver genes listed in Bailey et al.26 and at least 300 samples in COSMIC are plotted. (g) Null hypothesis test for driver genes in breast cancer, restricted to 403 estrogen and progesterone receptor-positive samples from TCGA database. Axes, color, and error bars as in (c). Genes as in Fig. 4d, excluding AKT1 due to absence of mutations. We applied the null hypothesis test to cancer types in COSMIC (Supplementary Table 1 and Extended Data Fig. 4f). Among 939 AML samples, six out of the 21 AML-associated genes have a significant young-age bias relative to exome-wide synonymous mutations, and these genes tend to have greater carcinogenic effects (Extended Data Fig. 4c–e). Among 1,562 breast invasive carcinoma samples, GATA3 and TP53 mutations are significantly young-age biased, whereas PIK3CA is not (Fig. 4d). Restricting to estrogen and progesterone receptor-positive breast cancers in The Cancer Genome Atlas (TCGA) yielded the same results (Extended Data Fig. 4g), reinforcing long-established evidence47–50 for TP53 contribution to breast cancer risk. Among 775 glioblastoma cases, TP53, IDH1, PIK3CA, BRAF and ATRX show young-age bias, but PTEN and EGFR do not (Fig. 4e). We emphasize that failure to reject the null hypothesis does not imply lack of carcinogenicity, but rather that further evidence is required. We also emphasize that causality is not binary. We observed a strong inverse relationship between carcinogenic effects and mean age among 908 esophageal carcinoma samples (rho = −0.66, P = 0.012; Fig. 4f) and among 1,724 colorectal cancer samples in COSMIC (rho = −0.65, p = 0.0023; Fig. 4g). Together with AML (Fig. 3f and Extended Data Fig. 4e), these results support that stronger cancer-causing mutations are found preferentially towards younger patient ages (Fig. 2g). Age biases of AML chromosomal alterations reflect carcinogenicity Having focused on SNVs/small indels so far, we next turned to somatic copy number alterations (SCNAs). SCNAs are a hallmark of cancer genomes51 and are thought to exert significant carcinogenic effects, which we would like to estimate. However, SCNA acquisition can be non-clocklike and explosive, with a flood of alterations late during carcinogenesis potentially obscuring information on selection and causation. We therefore decided to investigate relatively simple cancer genomes with few SCNAs. For a case study cancer type with few chromosomal abnormalities and uniquely voluminous normal tissue data, we looked again to AML. We examined gains and losses of autosomal chromosome arms and whole chromosomes among 2,113 AML samples from Tazi et al.25 and 482,789 normal blood samples from the UK Biobank52. As expected, SCNAs are far more numerous in AML (Extended Data Fig. 5a). To ameliorate confoundment by genome instability, we restricted attention to the 1,809 AML samples and the 482,043 normal blood samples with one or zero SCNAs. We observed that chromosome 8, for example, is amplified in 5% of AMLs and 0.02% of normal blood cells, leading to a carcinogenic effect estimate of ~1,800 (Fig. 5a). Selection and carcinogenicity shape age distributions of somatic copy number alterations. a, SCNA frequencies in AML and normal blood estimate carcinogenic effects. Fraction of 1,809 AML samples with each SCNA from Tazi et al.25 plotted against the mean cell fraction estimate among 482,043 normal blood samples from the UK Biobank52 (normal blood donors aged 40–75 years). Samples with more than one SCNA were excluded. Square-root scale. Vertical and horizontal error bars, respectively, represent 95% binomial and bootstrap intervals, shown for common SCNAs. b, Mean age of AML patients with SCNA positively associates with the mean cell fraction estimate of SCNA in normal blood (rho = 0.49, P = 0.028). The top 20 most frequent alterations over the normal blood and AML cohorts are plotted (alterations ranked by frequency in each cohort independently, and we plot alterations with the highest mean of the two rankings). c, Mean age of AML patients with SCNA negatively associates with carcinogenic effect estimate (rho = −0.54, P = 0.015). d, Pan-cancer comparison of age biases between SCNAs and SNVs/small indels. Left, for each cancer type and tumor suppressor gene, age bias of non-synonymous SNVs/small indels (rank biserial correlation between ages of patients with vs. without mutation in gene) among samples without SCNA is compared to the age bias of deletions among samples without non-synonymous SNVs/small indels in that gene. rho = 0.62, P = 2.6 × 10−6. Right, for each cancer type and oncogene, the age bias of non-synonymous SNVs/small indels among samples without SCNA is compared to the age bias of amplifications among samples without non-synonymous SNVs/small indels. rho = −0.13, P = 0.45. See Supplementary Table 3 for unabbreviated cancer types. Driver genes are from Bailey et al.26. Genes with an SCNA and an SNV/small indel in at least 5% of samples per cancer type are plotted. For b–d, P values are from two-sided tests, unadjusted for multiple hypothesis testing. We then calculated the mean age of patients with AML with each amplification and deletion. We observed a positive correlation between mean age in AML versus frequency in normal blood (rho = 0.49, P = 0.028; Fig. 5b and Extended Data Fig. 5b), and a negative correlation between mean age in AML versus carcinogenic effect estimate (rho = −0.54, P = 0.015; Fig. 5c and Extended Data Fig. 5b), agreeing with our theory. These correlations no longer hold if we include the minority of samples with many SCNAs (Extended Data Fig. 5b), consistent with our expectation that genome instability weakens signals of selection and carcinogenicity. Supplementary analyses for somatic copy number alterations. (a) Fraction of AML and normal blood samples with x autosomal SCNAs, at the level of whole chromosomes or chromosome arms. (b) Robustness analysis for Fig. 5b,c. X-axis shows the upper bound filter for the number of SCNAs per sample. Color represents Spearman’s correlation between mean patient age of SCNA vs. mean cell fraction estimate in normal blood (left), or mean patient age vs. carcinogenic effect estimate (right). Correlations are calculated among the top n most frequent SCNAs in the AML and normal blood cohorts, n represented on the y-axis. (c) Robustness analysis for Fig. 5d. Subplots show upper bound filter for number of exonic SNVs/small indels per cancer sample. X-axis shows lower bound filter for fraction of samples per cancer type with gene altered; after removing cooccurrences of SCNAs and SNVs/small indels, SCNAs and SNVs/small indels must both exceed lower bound frequency for gene-cancer type pair to be included. Values plotted only if the number of gene-cancer type pairs is at least 20. Y-axis shows Spearman’s correlation among gene-cancer type pairs between age bias of SCNAs vs. non-synonymous SNVs/small indels. (d) Linear regression among gene-cancer type pairs. Age bias of deletions is dependent variable. Age bias of non-synonymous SNV/small indels and cancer type are independent variables. Point estimates and 95% confidence intervals are plotted for the effect of each covariate. (e) Same as (d) but for oncogenes and amplifications. (d,e) Same data as in Fig. 5d. Distinct disruptions of tumor suppressor genes agree on age biases For cancer types other than AML, interpreting SCNA age distributions becomes more challenging. The difficulty is firstly that genome instability is more pervasive and secondly that normal tissue data are more limited. For a pan-cancer assessment of whether our framework applies to SCNA age distributions, in the absence of normal tissue controls, we used SNVs/small indels in cancer genomes as a benchmark. We sought to compare age distributions of SCNAs against age distributions of SNVs/small indels that might be expected to exert similar phenotypic effects. The prime candidates for alterations with plausibly similar phenotypic effects are deletions of tumor suppressor genes (TSGs) versus loss-of-function mutations in TSGs, as both alteration types ultimately abrogate gene function. Other candidates are amplifications of oncogenes versus gain-of-function variants in oncogenes4. We reasoned that agreement between SNVs/small indels versus SCNAs on their age distributions would support that our age distribution theory applies to SCNAs as well as supporting the hypothesis that the distinct mutational types exert comparable carcinogenic and selective effects. To put alterations (SCNAs or gene mutations) of different cancer types on the same scale, we measured the age bias of each alteration per cancer type by the rank biserial correlation between patient age and alteration status. We applied this metric to the pan-cancer TCGA database, focusing on samples without whole-genome doubling. We calculated the age bias for non-synonymous SNVs/small indels per driver gene among samples without any SCNA in that gene. We additionally calculated the age bias for each deletion and amplification per gene among samples without any non-synonymous SNVs/small indels in that gene (Supplementary Table 2). Thus, we could compare age biases of distinct alterations of a gene without the complication of cooccurrences. We observed that age biases of deletions versus SNVs/small indels of TSGs have a strong positive correlation (rho = 0.62, P = 2.6 × 10−6; Fig. 5c). By contrast, the age biases of amplifications versus SNVs/small indels of oncogenes are not significantly correlated (rho = −0.13, P = 0.45; Fig. 5d). These results are robust to filtering parameters (Extended Data Fig. 5c), and they are recapitulated by linear regressions that account for cancer type as a covariate (Extended Data Fig. 5d,e). These age comparisons are helpful on several counts. First, that the distinct mutational disruptions of TSGs agree on age distributions supports the notion that age distributions of SCNAs provide information on selection and cancer causation. Second, the comparisons support that deletions and mutations of TSGs exert broadly similar effects, whereas amplifications and mutations of oncogenes overall have more divergent effects. Indeed, some mutations within oncogenes idiosyncratically alter protein structure and function in a way that amplifications cannot. Third, young-age biases provide orthogonal lines of evidence supporting the carcinogenicity of disrupting some TSGs (Supplementary Table 2). Discussion In the search for cancer-causing mutations, hundreds of thousands of cancer genomes have been sequenced, and signals of positive selection have been identified. The interpretation of these signals needs revisiting, however, in light of recent findings that positive selection is ubiquitous in our healthy tissues. Cancer causation needs to be conceptually and inferentially distinguished from positive selection. Towards one possible solution, we have defined the carcinogenic effect of a mutation in terms of the per-cell rate of cancerous transformation. Our metric is simplistic. The multi-stage evolutionary process of carcinogenesis and its combinatorial genetic interactions are boiled down, by our metric, to a single number per alteration. This one-dimensional probabilistic assessment of cancer-causing power has the disadvantage of missing innumerable biological details; on the other hand, its minimalism brings the benefits of broad application and unified interpretation across cancer types and mutations. Assessing the carcinogenicity of mutations has several purposes. In addition to the basic question of the origins of cancer, there may be implications for the modulation of cancer risk. For a mutation with a known carcinogenic effect, perturbing the frequency of the mutation has a predictable consequence on the risk of cancer in a tissue (Fig. 1b). Perhaps one day soon, preventative treatments will reduce the frequency or inhibit the effects of carcinogenic mutations in our tissues before cancer develops. Conversely, it is thought that exogenous carcinogens increase cancer risk largely by elevating frequencies of carcinogenic mutations in our tissues, not only through mutagenesis but also by exerting selective pressures, for example, chemotherapies cause expansion of TP53 mutations in the blood53, while air pollution is associated with expansion of EGFR mutations in the lungs54. Broadly, therefore, for mutations across tissues, estimating carcinogenic effects could contribute to our understanding and management of cancer risk in the population. Carcinogenic effect estimates may also help to establish priority mutational targets for cancer treatments. Mutations in cancers, regardless of their carcinogenic effects, can, in principle, be therapeutically targeted to identify and kill cancer cells. Yet treatment resistance is a pervasive problem. We reason that resistance is more problematic for treatments that target mutations with lower carcinogenic effects because, conceivably, cancers can more easily tolerate the absence of these mutation pathways while keeping the malignant phenotype. Perhaps this explains the generally disappointing performance of anti-PI3K therapies55. Indeed, carcinogenic effects are limited for PIK3CA mutations in the esophagus (Fig. 1d) and possibly breast (Fig. 4d). Further supporting limited carcinogenic effects, there is no clear association between PIK3CA-related overgrowth syndrome and adult cancers56,57. We have proposed to estimate the carcinogenic effects of mutations by comparing cancer and normal tissue genomes, leveraging the genetic diversity of human tissues as a natural experiment for testing mutational causality. Moreover, we have proposed that patient age data can contribute valuable, previously underappreciated causal information. Although the age of cancer onset is shaped by many factors—such as germline genetics, carcinogenic exposures and more—we have determined that the carcinogenic effects of the mutations nevertheless leave a visible mark on patient age distributions. Yet age has been undervalued in cancer genomics, with nearly 30% of COSMIC samples missing age annotations. Looking ahead, cancer genomes and age will be further integrated with the expanding data on normal and pre-cancerous tissues, towards improved causal inferences. This synthesis will clarify how mutations and exposures shape somatic evolution across the lifespan from zygote to malignancy, ultimately informing cancer prevention. Methods IRB approval was not required for the study. Data sets This project only uses publicly available data. No data points were excluded from the analyses except where explicitly stated. For each cancer type analyzed, we obtained some or all the mutation data from the pan-cancer whole-exome sequencing database COSMIC (v.98)23. This database is a collation of many cancer genomic studies, and the classification of some cancer types requires clarification. COSMIC contains TCGA, a dataset that has a standardized classification scheme. We used the TCGA classification of cancer types, which we extended to the whole COSMIC dataset as follows: for each cancer sample in COSMIC but not in TCGA, if the anatomical site and histology of the sample matched those of samples in TCGA with a specific cancer type classification, then we assigned that classification to the sample; otherwise, we disregarded the sample. For the cancer types in our cancer type-specific analyses, namely AML, glioblastoma, ESSC, COADREAD and breast invasive carcinoma, we performed a manual review of the studies that contributed to the COSMIC database to ensure correct classifications. We excluded a small subset of studies in which the cancer type classification was ambiguous or erroneous. We also excluded samples without a patient age annotation. For our pan-cancer COSMIC analyses (Fig. 3a), we focused on cancer types with at least 300 samples. Our filtered version of the COSMIC data is available at 10.5281/zenodo.18343689. For mutations in normal blood, we used two datasets. The first dataset (supplementary Table 2 of Abelson et al.21) comprises 817 normal blood samples from 676 donors who did not develop any hematological malignancies within ~10 years of sampling. A total of 141 of the donors had multiple blood samples available, spanning a median of 10.5 years. The donors are aged between 36 and 90 years. Targeted deep sequencing of AML-associated genes was performed, with a minimum reported VAF of 0.001. The second dataset (supplementary Table 2 in Fabre et al.22) comprises 1,593 blood DNA samples from 385 donors aged between 54 and 103 years at the time of sampling. The participants had no history of hematological malignancy. Their blood was sampled up to five times each over a period of 3–16 years. Targeted deep sequencing of AML-associated and clonal hematopoiesis-associated genes was performed with a mean coverage of ~1000× and a minimum reported VAF of 0.0002. For AML mutations, we leveraged four datasets. One database (supplementary Table 3 of Tazi et al.25) presents sequencing data for 2,113 patients with AML aged between 15 and 100 years. The second database (http://www.cbioportal.org/study/summary?id=aml_target_2018_pub) is part of a pediatric AML trial of the Children’s Oncology Group, National Cancer Institute45, and comprises data on somatic mutations in 886 patients with AML aged between 0 and 30 years. The third database (http://www.cbioportal.org/study/summary?id=aml_ohsu_2022) is part of the Oregon Health & Science University Beat AML cohort24, comprising 742 patients with AML aged from 0 to 80 years. The fourth database contains AML mutation data from COSMIC, comprising 995 patients with AML ranging from 0 to 88 years old. Together, these four AML and two normal blood datasets include a variety of ranges of the genome sequenced, including the whole genome, the whole exome and specific genes. We focused on 21 genes that were covered by all the AML and normal blood studies and were listed in Bailey et al.26 as AML driver genes. For our comparison of driver gene mutations between ESSC and normal esophageal epithelium, we used data from Yokoyama et al.18 (their supplementary Table 8). The cohort comprises 68 patients with ESSC aged between 47 and 79 years. Each patient donated one ESSC sample and one ‘physiologically normal’ epithelium sample. The biopsy area is ~8 mm2. See their supplementary Table 2 for donor and sample information. A total of 24 genes were sequenced. The sequencing depth was not specified, although the minimum VAF reported is 0.05. For a fair comparison of cancer and normal tissue, we wanted to avoid a scenario in which the normal tissue samples were part of a clone that gave rise to the cancer, so we removed five patients who had mutations in common between their cancer and normal tissue samples. We then focused on the subset of genes with signals of positive selection in ESSCs, as shown in supplementary Table 5 of Yokoyama et al.18. For mutations in the normal colon, we turned to data published in Lee-Six et al.34 and Olafsson et al.14, the former including 42 individuals without diagnosed inflammatory bowel disease and the latter comprising 28 patients with ulcerative colitis and 18 patients with Crohn’s disease, with ages ranging from 11 to 80 years. Both studies performed whole-genome sequencing on single colonic crypts; coding mutations were collated in Olafsson et al.14 (their supplementary Table 3), excluding one patient from Lee-Six et al.34 whose colon was heavily mutagenized by prior chemotherapy. In total, the data comprise 1,395 crypts among 91 donors. We obtained data on cytogenetic abnormalities among 2,113 AML samples from Tazi et al.25 (their supplementary Table 4), representing the same patients whose driver gene mutations we analyzed. The chromosomal variants declared include amplifications and deletions of whole chromosomes and chromosome arms, as well as inversions and translocations. Copy number neutral loss of heterozygosity variants are not included in this AML dataset. We obtained data on amplifications and deletions at the level of whole chromosomes and chromosome arms among 482,789 normal blood samples in the UK Biobank from https://github.com/blundelllab/mCA-mutation-rates-fitness-consequencesref. 40; these copy number alterations were originally called by Loh et al.52, and they include estimates of the fraction of blood cells per sample with each alteration. The donor ages range from 40 to 75 years. We obtained copy number annotations for the TCGA pan-cancer dataset from https://github.com/VanLoo-lab/ascat. To map their copy number annotations to the list of putative oncogenes and tumor suppressor genes from Bailey et al.26, we used genomic coordinates for genes from https://useast.ensembl.org/index.html. No declared copy number alteration has a genomic boundary that lands within a putative driver gene; that is, each copy number alteration unambiguously covers a specific gene or does not cover that gene. We removed from the analysis samples annotated as whole genome doubled. We further removed samples of uterine corpus endometrial carcinoma and lung adenocarcinoma because these cancer types show significant negative correlations between patient age and number of synonymous mutations (r = −0.15, P = 5 × 10−4 and r = −0.13, P = 0.3 × 10−3, respectively), in violation of our model’s prediction that neutral mutations accumulate with age. Estimating mutant cell fractions in bulk normal tissue We focused on non-synonymous SNVs and small indels. For each observed mutation in a normal esophagus or blood sample, we multiplied the VAF by two to estimate the fraction of cells in the sample with the mutation. This estimate is based on the simplifying assumption of diploidy and heterozygosity. We adjusted the estimates in two special scenarios. Firstly, for some mutations in some samples, the VAF multiplied by two exceeded a value of one, which we interpreted as possible loss of heterozygosity. Therefore, we set the mutant cell fraction estimate equal to the VAF without multiplying by two. Secondly, for males and the X chromosome, we again set the cell fraction estimate of each mutation equal to the VAF. Then we extended our mutant cell fraction estimates to the level of genes. For this procedure, calculation elaboration is needed for situations in which there are multiple mutations per gene per sample. To estimate the fraction of cells in a sample with a specific gene mutated, we either summed the cell fraction estimates of mutations in that gene, or, if this sum exceeded one, we took the maximum cell fraction estimates of mutations in the gene. The former case represents the assumption that the mutations in the gene land in distinct cells within a sample, while the latter case represents the assumption that the mutations land in the same cells. These estimates should be regarded as ballpark figures, given that mutational cooccurrences per cell, copy number status and sample purity are unknown, and mutations at low VAF (<0.05 for the esophagus and <0.001 for blood) are unreported. Next, for each gene, we took the mean across samples of the per-sample mutant cell fraction estimates, thereby obtaining estimates for the fraction of cells mutated per gene in the cohort. Here, to assess sampling noise (as opposed to uncertainty introduced by measurement or processing errors), we calculated 95% bootstrap confidence intervals for these cell fraction estimates. For each of 1,000 bootstrap resamples of the tissue samples, we calculated mean mutant cell fractions per gene, thereby generating a gene-specific bootstrap distribution for the mean mutant cell fraction; the 2.5 and 97.5 percentiles of the distribution define the confidence interval. We note the theoretical caveat that bootstrap intervals in general struggle to quantify sampling noise for small cohort sizes. Carcinogenic effect estimates The ESSC carcinogenic effect estimates are based on the patient-matched ESSC and normal esophageal epithelium data from Yokoyama et al.18. The adult AML carcinogenic effect estimates are based on the normal blood data from Abelson et al.21 and Fabre et al.22 and the AML data in COSMIC, Bottomly et al.24 and Tazi et al.25, with a focus on patients with AML who were older than 36 years (given that the youngest normal blood donor is 36 years old). For each driver gene, we compared the fraction, y, of cancer samples with the gene mutated (non-synonymous SNVs and indels) against the estimated fraction, x, of normal tissue cells across the cohort with that gene mutated: the odds ratio y/(1−y)x/(1−x) is the carcinogenic effect estimate. See Supplementary Notes 1, 2 and 6 for critical discussions of the validity of this estimate. We performed 1,000 bootstrap resamples of the donors in each of the cohorts to obtain 95% bootstrap confidence intervals for the carcinogenic effect estimates. To more clearly account for age in the comparison of adult AML and normal blood mutations, we additionally estimated carcinogenic estimates through a regression model (Extended Data Fig. 1a). The model has four real-valued gene-specific parameters: a, b, ϕ and w. The first component is that for an AML sample at patient age t, a gene is mutated with probability ww+ea+bt. Here, w > 0 is the carcinogenic effect, and 11+ea+bt is the expected fraction of normal blood cells with the gene mutated. The second component of the model is that for a normal blood sample at age t, the fraction of cells with the gene mutated follows a beta distribution with shape parameters eϕ1+ea+bt and eϕ+a+bt1+ea+bt; so the mean of the distribution is 11+ea+bt while the spread of the distribution is controlled by ϕ. To address the invisibility of mutations at low cell fractions, we then specified that the observed mutant cell fraction in a normal blood sample is distributed as X×I(X≥x0), where X~beta(eϕ1+ea+bt,eϕ+a+bt1+ea+bt) is the actual mutant cell fraction, I is the indicator function and x0 denotes the minimum cell fraction estimate among observed mutations. As such, we obtained a likelihood function for mutation frequencies in the AML and normal blood data, with which we performed maximum likelihood estimation for each gene independently. The carcinogenic effect estimates from this regression model strongly agree with the carcinogenic effect estimates from the unadulterated odds ratio of mutation frequencies between adult AML and normal blood (rho = 0.98, P = 4.8 × 10−6). To estimate carcinogenic effects from the data of Abelson et al.21 on the relationship between normal blood mutations and subsequent AML development (Extended Data Fig. 1e), we used the theoretical relationship shown in Fig. 1b (H = 1 – x + xw) between the carcinogenic effect of a mutation (w), the frequency of that mutation in normal tissue (x) and the hazard ratio for developing cancer in that tissue given the mutation (H). Further notation is required to explain the calculation. We denote the set of normal blood donors by J. For donor j∈J, fraction xj of the cells have a mutation in gene z (averaged over samples in the case of multiple samples per donor). At time tj after the first sample was taken, either the donor develops AML or the donor has not developed AML and they drop out of the study. For the first scenario, aj = 1; for the second scenario, aj = 0. Without loss of generality, denote the set of donors by J=1,2,…,n and suppose that t1,t2,…,tn is an increasing sequence. Then, for each gene, in the spirit of standard inference for Cox proportional hazards models, we perform maximum likelihood estimation for the carcinogenic effect w based on the partial likelihood To estimate carcinogenic effects for the younger AML group (<25 years) in the absence of age-matched normal blood data, we used the same data as for the adult AML analysis, supplemented with the childhood AML data from Bolouri et al.45. We began with the simplifying approximation that a normal haematopoietic stem cell carries a mutation within a specific gene with a probability that is proportional to the gene’s length: that is, for a gene with length L, the gene is, on average, mutated in fraction cL of a child’s haematopoietic stem cell, for some unknown constant c. Then, supposing that a mutation in this gene confers a w-fold elevation in the rate of cancerous transformation, the gene is expected to be mutated in fraction f=wcL/(1−cL+wcL) of AMLs. Approximating further that cL is small relative to one, wc=L−1f/(1−f). By the functional invariance of maximum likelihood estimation, the maximum likelihood estimate for the relative carcinogenic effect wc is given by L−1fˆ/(1−fˆ), where fˆ is the observed fraction of samples mutated. We obtained 95% binomial confidence intervals for f, which we transformed by the equation wc=L−1f/(1−f) to obtain 95% confidence intervals for the relative carcinogenic effect. We note an important methodological distinction between the childhood and adult AML carcinogenic effect estimation: for childhood AML, the normal blood benchmark is formed by gene lengths that approximately represent the total number of possible mutations in a gene, whereas for adult AML, the normal blood benchmark is formed by bulk DNA samples that exhibit only those mutations at VAF high enough to be distinguished from sequencing errors. Thus, the adult carcinogenic effect estimates refer primarily to those mutations in a gene that drive expansion in normal blood, whereas the childhood carcinogenic effect estimates refer to the aggregate of all mutations per gene. More broadly, we expect the age-group-specific modeling assumptions to drive divergence between childhood and adult AML carcinogenic effect estimates, suggesting that Fig. 3c understates the real biological agreement between the age groups. To estimate carcinogenic effects for colorectal cancer, we used cancer data from COSMIC and normal colon data from Lee-Six et al.34 and Olaffson et al.14 We used a logistic regression model: for each colorectal cancer driver gene listed in Bailey et al.26, we modeled that the likelihood of the gene being mutated in a sample takes the form wχwχ+ea+bt+cm, where t is the age of the individual, m is the exonic mutation burden of the sample, χ takes the value 1 or 0 to represent that the sample is a colorectal cancer or a normal colonic crypt, respectively and a, b, c and w are the parameters of the model. We calculated maximum likelihood estimates for the carcinogenic effect w along with 95% profile confidence intervals. Statistics for the age distribution of cancer mutations For each cancer type and driver gene, we calculated the mean age of patients in the cohort weighted by the number of non-synonymous substitutions and small indels in the gene, along with 95% Wald confidence intervals for this mean age statistic. To assess the degree of age divergence between driver genes, we then calculated the variance between driver genes of the mean age statistic, weighted by the number of mutations per gene, which we divided by the variance of the patient age distribution per cancer type. To assess whether the age divergence between driver genes is significantly greater than zero for each cancer type, we calculated a P value from a Kruskal–Wallis test of the null hypothesis that mutations among driver genes have patient ages that are independent of the identity of the driver genes in which they reside. We also calculated the mean age of non-synonymous substitutions and small indels per driver gene minus the mean age of exome-wide synonymous mutations for each cancer type. To generate a null distribution, we recalculated this statistic for each of 1,000 permutations of patient ages among samples of the cancer type in question. For these comparisons of driver gene mutation age distributions against synonymous mutations, we focused on samples in COSMIC whose mutation burden is no greater than ten times the median mutation burden per cancer type, to avoid the statistical signal being diminished by a small number of samples with extreme mutation burden. To compare the age distributions of SCNAs versus SNVs/small indels, we introduced the age bias statistic, defined by the rank biserial correlation between the ages of patients with versus without a particular alteration. Specifically, for each driver gene and cancer type, among cancer samples with no SCNA in that gene, we calculated the rank biserial correlation between the ages of patients with versus without a non-synonymous SNV/small indel in the gene. Separately, among cancer samples without a non-synonymous SNV/small indel in that gene, we calculated the rank biserial correlation between the ages of patients with an amplification of the gene versus no SCNA in that gene, and we calculated the same statistic for deletions. We used the R package ‘effectsize’ for point and interval estimates of the rank biserial correlation (Supplementary Table 2). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Online content Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41588-026-02593-z. Supplementary information Supplementary Notes 1–6 and Supplementary Figs. 1 and 2. Reporting Summary Supplementary Table 1. Statistics for driver gene mutations in COSMIC. Supplementary Table 2. Statistics for SCNAs and driver gene mutations in TCGA. Supplementary Table 3. Cancer type acronyms. Acknowledgements We thank I.-H. Lee and B. Laurent for helpful discussions and data curation support. This work was supported by the National Institutes of Health (R37CA225655, R01CA279054, R01CA269281 and P01HL142494), the Glenn Foundation for Medical Research and an Emerging Leader Award from the Mark Foundation for Cancer Research. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript. Extended data Author contributions D.C. and K.N. conceived of the project. D.C. developed the conceptual framework and performed analyses. M.B. helped to process data. M.B., M.N. and T.A. provided key advice. K.N. supervised the project. D.C. and K.N. wrote the manuscript. Peer review Peer review information Nature Genetics thanks Jacob Scott and the other, anonymous, reviewer(s) for their contribution to the peer review of this work Data availability No new data were generated during the study. We obtained pan-cancer mutation data from COSMIC (v.98)23 (https://cancer.sanger.ac.uk/cosmic/download/cosmic), further mutation data for acute myeloid leukemia from http://www.cbioportal.org/study/summary?id=aml_target_2018_pub (ref. 45), http://www.cbioportal.org/study/summary?id=aml_ohsu_2022 (ref. 24) and supplementary Table 3 of 10.1038/s41467-022-32103-8 (ref. 25), and normal blood mutation data from supplementary Table 2 of 10.1038/s41586-018-0317-6 (ref. 21) and supplementary Table 2 of 10.1038/s41586-022-04785-z (ref. 22). Data for mutations in the normal and cancerous esophageal epithelium were obtained from supplementary Table 8 of 10.1038/s41586-018-0811-x (ref. 18), and data for mutations in the normal colon from supplementary Table 3 of 10.1016/j.cell.2020.06.036 (ref. 14). We obtained copy number annotations for the TCGA pan-cancer dataset from https://github.com/VanLoo-lab/ascat, copy number annotations for blood samples in the UK Biobank from https://github.com/blundelllab/mCA-mutation-rates-fitness-consequences and copy numbers for acute myeloid leukemia from supplementary Table 4 of 10.1038/s41467-022-32103-8 (ref. 25). Summary data to reproduce all figures of this study are provided at 10.5281/zenodo.18343689 (ref. 58). Code availability R (v.4.4.0) code generated during this study is available at 10.5281/zenodo.18343689 (ref. 58). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information David Cheek, Email: david_cheek@hms.harvard.edu. Kamila Naxerova, Email: kamila_naxerova@hms.harvard.edu. Extended data is available for this paper at 10.1038/s41588-026-02593-z. Supplementary information The online version contains supplementary material available at 10.1038/s41588-026-02593-z. References Associated Data Supplementary Materials Supplementary Notes 1–6 and Supplementary Figs. 1 and 2. Reporting Summary Supplementary Table 1. Statistics for driver gene mutations in COSMIC. Supplementary Table 2. Statistics for SCNAs and driver gene mutations in TCGA. Supplementary Table 3. Cancer type acronyms. Data Availability Statement No new data were generated during the study. We obtained pan-cancer mutation data from COSMIC (v.98)23 (https://cancer.sanger.ac.uk/cosmic/download/cosmic), further mutation data for acute myeloid leukemia from http://www.cbioportal.org/study/summary?id=aml_target_2018_pub (ref. 45), http://www.cbioportal.org/study/summary?id=aml_ohsu_2022 (ref. 24) and supplementary Table 3 of 10.1038/s41467-022-32103-8 (ref. 25), and normal blood mutation data from supplementary Table 2 of 10.1038/s41586-018-0317-6 (ref. 21) and supplementary Table 2 of 10.1038/s41586-022-04785-z (ref. 22). Data for mutations in the normal and cancerous esophageal epithelium were obtained from supplementary Table 8 of 10.1038/s41586-018-0811-x (ref. 18), and data for mutations in the normal colon from supplementary Table 3 of 10.1016/j.cell.2020.06.036 (ref. 14). We obtained copy number annotations for the TCGA pan-cancer dataset from https://github.com/VanLoo-lab/ascat, copy number annotations for blood samples in the UK Biobank from https://github.com/blundelllab/mCA-mutation-rates-fitness-consequences and copy numbers for acute myeloid leukemia from supplementary Table 4 of 10.1038/s41467-022-32103-8 (ref. 25). Summary data to reproduce all figures of this study are provided at 10.5281/zenodo.18343689 (ref. 58). R (v.4.4.0) code generated during this study is available at 10.5281/zenodo.18343689 (ref. 58).