Do species that accumulate mutations more slowly live longer?

The inverse association in colorectal crypts is robust within these analyses, but it does not establish that mutation rate determines lifespan. Division, repair, tissue renewal and life history may evolve together. An intervention can also affect competing processes: the genetic DREAM model associated with fewer brain mutations originated in a study reporting systemic amyloidosis and shorter survival.

Correction, September 20, 2026: Our original phylogenetic covariance reversed relatedness. A large diagonal repair and incomplete likelihood calculation then produced the erroneous finding that the association vanished at λ=1. We obtained the authors' original tree and recalculated the models independently in Python and R. We withdraw the deep-clade attribution and claimed conflict with the original study, correct aggregation, missingness, partial-correlation degrees of freedom and species naming, and add relevant division and DREAM-model evidence.

01What constitutes a data point?

A colorectal crypt can be occupied by a stem-cell lineage, allowing accumulated somatic mutations to be detected in its cells. Cagan and colleagues studied 16 mammalian species, with 15 contributing to lifespan-rate analysis. Harbour porpoise lacked age information; our Chinese edition incorrectly called it a seal.s1

The public crypt table contains 208 rows and 53 distinct “Individual” IDs. Two porpoise rows lack age. Another two ring-tailed lemur rows have an age but no reported substitution rate. There are therefore 204 available rates from 52 table IDs in 15 species. The four exclusions are not all missing-age cases. Cross-species lifespan information primarily comes from species, not 204 independent lifespan comparisons.s2s4

The source code averages within individuals and then within species. Our original article averaged crypts directly, giving more weight to individuals with more sampled crypts. Changing the weighting reproduces 14 species' published means essentially exactly. The human discrepancy also has an arithmetic explanation: equal weighting of six table IDs gives 46.395/year; splitting PD36813 into the ac, ad and x sample-prefix groups yields exactly the published 47.125/year. Those groups also use distinct matched reference samples.s10s11s14

Arithmetic agreement does not establish three independent people. The species summary lists eight humans and two lemurs, while the sample table identifies six and one, respectively. Their biological correspondence remains incompletely documented. We use the published species means to check the source model and retain equal weighting of table IDs as a separate sensitivity analysis.s12

02The association survives a valid phylogenetic model

For the 15 published species means, regression of log₁₀ annual mutation rate on log₁₀ lifespan gives a slope of −0.862, 95% CI −1.078 to −0.645, and R²=0.851. The interval includes −1, making inverse proportionality compatible with these data without proving that exact law.

Related species share evolutionary history. Our previous code instead made more distant species more correlated: it used time before the present to the common ancestor as covariance, making human–mouse similarity exceed lion–tiger similarity. It then repaired the invalid matrix with a diagonal addition far larger than rounding error. The result no longer represented the intended phylogenetic model.

We retrieved the actual TimeTree tree from the authors' Zenodo archive and computed covariance from shared branch length between the root and common ancestor, without arbitrary repair.s13

λ describes phylogenetic correlation in residuals; it is not the percentage of the association explained by evolution. Estimating it also requires estimating residual variance, which our old likelihood omitted. The reported intervals remain conditional on the tree and λ, without combining all tree, parameter and lifespan uncertainty.

A second distinction matters: the authors' published PGLS regresses annual rate on inverse lifespan with no intercept, rather than fitting the free-intercept log–log model above. Using their formula and tree, the coefficient k changes from 2439 under ordinary regression to 2000 under PGLS; omitting mouse and rat gives 3326 and 3078. These are of similar order, but a positive k does not establish inverse proportionality: the inverse form and zero intercept are imposed by the model.s10

Our Python tree parser and GLS calculations agree with independent R implementations using ape/caper across 225 covariance entries and 90 parameter checks. The previous disappearance of the association was our implementation error, not evidence against the published study.

03Robustness still has conditions

Five analyzable species have only one table Individual ID. Resampling IDs within species yields a narrow 95% bootstrap interval of approximately −0.907 to −0.824. This describes the observed within-species variation conditional on those species. Repeatedly resampling one ID creates no new individual variation. It does not include all species sampling, phylogenetic or biological-identity uncertainty.

Lifespan is also a defined measurement. Lifespan80 is the age by which 80% of individuals surviving to adulthood have died. Animal data predominantly come from captivity; human data come from 1900 birth cohorts in three countries, and dog records pool dogs and wolves. This is neither mean lifespan nor a uniformly measured trait across species.s1s3

The association persists with median survival age, the 90% mortality age and AnAge maximum longevity. Substituting lifespan confidence limits or the 90% measure drops different species with missing values, however. The six analyses do not all contain the same 15 species. Restricting every definition to the same 11 complete species gives slopes from −0.906 to −0.816. This is definition sensitivity, not statistical elimination of measurement error.

Partial correlation controlling log body mass is approximately −0.907, with p≈7.8×10⁻⁶ after correctly deducting one covariate from the degrees of freedom. A single mass adjustment does not remove the association; it neither excludes other coevolving factors nor decomposes a causal effect of body mass.

04What does a similar lifetime burden mean?

Multiplying published species rates by Lifespan80 gives projected lifetime burdens of 1828–5379 substitutions/genome, a 2.94-fold range.

This convergence merits explanation but is not independent validation: a rate approximately proportional to inverse lifespan naturally produces a more similar product. Donors were not all sampled at their species' Lifespan80. Damage contributing to lifespan limitation, longer life histories selecting for maintenance and slower turnover, or both together can produce such a pattern. A universal lethal mutation threshold is not identified.

05Cell division remains a competing explanation

The 2022 study used limited division-time estimates from different methods to argue that division alone could not explain every difference. Our original claim that it could not explain the principal contribution was stronger than the evidence.

Montazid and colleagues directly examined mouse and naked-mole-rat intestinal stem-cell dynamics in 2023 and integrated human data. Crypt-base division rates also varied with lifespan. Using their colonic estimates of 181 divisions/year in mice and 50 in humans, the mouse–human annual mutation-rate ratio of 16.90 becomes a division-normalized ratio of 4.67. Division therefore deserves serious consideration. The cross-species comparison involved only three species and cannot independently establish a general law.s16

This is not a causal mediation percentage. If annual rate r=d×μ+ν, where d is division frequency, μ a division-related contribution and ν other time-dependent accumulation, then r/d=μ+ν/d. Dividing mutation counts by division frequency does not remove non-division sources.

Tissue specificity is also supported directly. Kapadia2025 estimated a mouse haematopoietic stem/progenitor age slope of 45.3 SBS/year (42.2–48.4), versus approximately 14–17 in related human studies. Matched blood and intestinal-crypt measurements in the same aged mice supported tissue differences. The age slope of 45.3 and Cagan's cumulative-burden/age summary are not identical estimands, so their ratio is not a universal “18-fold tissue difference.” Colorectal findings cannot define a single body-wide mutation clock.s6s7 The correction changes the duplicated upper5×10⁻⁵ tick in Figure4e to5×10⁻⁴; it does not change the annual slope cited here.s8

06DREAM perturbation changes more than mutations

Koch2026 connects expression-derived DREAM-associated activity with cross-species lifespan, cellular mutations and human neuropathology. Some expression and mutation analyses use the same RNA data. Human disease analyses relate postmortem brain expression to an earlier diagnosis; these do not establish prospective protection or causal direction.s9

The genetic experiment is more direct. DREAM assembly was disrupted at eight weeks, and archived brains from four loss-of-function mice and five controls were ultimately analyzed. Covariate-adjusted SBS and indel burdens were reported to be 4.2% and 19.6% lower. Both groups initially contained five mice, with one outlier excluded. Tissue was collected at death, and models included age, sequencing coverage and read families. Small samples, model selection and terminal sampling matter; this is not a measured longitudinal slowing at matched ages.

Crucially, those tissues came from the Perampalam2021 model. Its loss-of-function cohort contained 30 mice and its control cohort 37. The authors reported approximately 16% shorter survival (log-rank p=0.0236), amyloid deposition in heart, liver, spleen and kidney, and organ damage. Survival included euthanasia at humane endpoints. The 2026 paper describes natural-death brain collection, but we lack individual linkage confirming identical sampling definitions between reports.s15

This does not establish that reducing mutations is harmful, or that selective, temporary DREAM modulation would reproduce broad genetic disruption. It shows why a lower tissue mutation burden is insufficient evidence of net health benefit when an intervention affects multiple processes. Testing the proposed chain requires mutation, function, safety and survival outcomes under a clearly specified intervention and time window.

07What the evidence supports

The inverse relationship between colorectal-crypt annual mutation rate and species lifespan remains after correcting phylogenetic implementation, aggregation and lifespan-definition checks. Repair, division, selection and life history can contribute jointly, and correlation does not determine causal direction. Convergent lifetime burden is a derived description of that relationship. DREAM demonstrates a route to influencing mutation accumulation, while the source model's disease and survival outcomes prevent treating it as lifespan-extension evidence.

We did not call variants again from sequencing reads or fully refit the original hierarchical mixed and Bayesian models. This audit covers public tables, species-level models aligned to the authors' code, and additional sensitivities. Source-ID correspondence, cross-study estimands and tree/λ uncertainty remain explicit. Evidence was checked through September20,2026. Codex performed revision, self-review and editorial sign-off; this is not human professional review or intervention advice.

The reproducibility package includes pinned source tables, the original tree, code, independent R validation and outputs. Third-party data retain CC BY4.0 attribution; original paper figures are not redistributed.