PREreview of Tentative evidence that aging is caused by a small number of interacting processes

by sean lim

Published: August 25, 2025
DOI: 10.5281/zenodo.16938777
License: CC BY 4.0

This review provides author-actionable feedback on a preprint investigating whether telomerase activation slows actuarial aging (Gompertz slope) and what that implies about the number and coupling of aging processes.

Summary

Question. Does telomerase activation in mice reduce the Gompertz slope (actuarial aging rate) rather than merely shifting baseline mortality, and what does that imply about the number/strength of interacting aging processes?

Design. Secondary analysis of three mouse telomerase-activation studies (AAV-TERT at 12 mo; CMV-TERT at 18 mo; constitutive Tert knock-in across five generations); Gompertz fits to survival (two datasets digitized); simple ODE model with symmetric cross-talk; extrapolation using 2022 Swedish survivorship; Mathematica NonlinearModelFit with CIs.

Key results.

AAV-TERT (1-yr C57BL/6): median +24%, max +13%; G: 0.008→0.0074 (−7.5%); r²≈0.996–0.997; ΔG not significant; n=43 ctrl, 21 TERT; Fig 1a.
CMV-TERT (18-mo females C57BL/6J): median/max from birth +40%; G: 0.010→0.0081 (−22.9%); p<0.05; n=16/16; Fig 1b.
Tert knock-in (EF1α): max +25%; G reduced 19% (p<0.05); n=22 ctrl, 98 KI; Fig 1c.
ODE cross-talk model (n=3 processes): required cross-talk ratio for ΔG=20% ≈ 33%; for ΔG=10% ≈ 12.5% (analytic for symmetric coupling); theoretical upper bound for ΔG by removing one process ≤1/n; Figs 2–3.
Human extrapolation (Sweden 2022): imposing ΔG=−20% shifts modeled median from 85→104 y; Fig 5.

Implication. Mouse data are consistent with (but do not yet prove) a slope reduction; if true, simple linear-algebra places strong constraints on the number/strength of interacting, feedback-amplified aging processes, and naïve human extrapolations are highly model-dependent.

List of major concerns and feedback

The current survival analysis is not appropriate: fitting Gompertz survivorship curves with least-squares and reporting r² values ignores censoring and left-truncation, which can bias slope estimates (ΔG) and p-values, especially in the 18-month gene therapy cohort (where animals had to survive to treatment) and in pooled knock-in generations. A more robust approach is to reanalyze using maximum-likelihood parametric survival models with age as the time scale and proper handling of truncation/censoring, for example Gompertz, Gompertz–Makeham, or logistic/Kannisto fits. In R this can be done with flexsurvreg(Surv(start, stop, event) ~ group, dist="gompertz") (and analogous models for the others). Results should be reported as hazard ratios or time-ratios with 95% confidence intervals, with model comparisons via ΔAICc and likelihood-ratio tests, and presented in a table summarizing slope and Makeham parameters by group.
Digitization error has not been quantified and could easily explain modest slope shifts (e.g., the 7.5% ΔG), meaning current p-values may be overstated. A simple way to address this is to run a Monte Carlo sensitivity check: have two independent annotators re-digitize both published survival curves, add random jitter of ±0.5–1 week to event times, and repeat the model fitting 1,000 times. Report the resulting distribution of ΔG, hazard ratios, and p-values across resamples, and include a Bland–Altman plot comparing annotators to show consistency. This will make clear whether the observed ΔG is robust to digitization noise.
The current analysis may be confounded by slope–intercept identifiability and model mis-specification. In Gompertz fits, an apparent change in slope (ΔG) can actually reflect an unmodeled Makeham term (baseline hazard) or hidden heterogeneity (frailty), both of which can mimic slowed aging. To address this, re-fit the data with more flexible models: (i) Gompertz–Makeham, which adds an age-independent hazard, and (ii) gamma-Gompertz frailty, which allows for cage- or cohort-level heterogeneity. Report whether the ΔG effect remains once these extensions are included. To check fit quality, provide calibration plots comparing observed vs. predicted survival in deciles of age, and Brier scores as a quantitative measure. This will clarify whether the slope reduction is real or an artifact of model choice.

List of minor concerns and feedback

The manuscript currently reports “p<0.05” without showing confidence bounds, which weakens the claims. Please report exact 95% confidence intervals for all effect sizes: the change in Gompertz slope (ΔG), hazard ratios or time-ratios from survival models, and changes in median and maximum lifespan. For datasets obtained via digitization, add bootstrap-based confidence intervals (e.g. 1,000 resamples of event times with jitter to reflect digitization error). Present these CIs alongside p-values so readers can see both the precision and the uncertainty in the estimates.
The survival comparisons rely entirely on ΔG, but readers will expect standard survival tests as a cross-check. Please add a log-rank test (with age as the time scale) to provide a non-parametric p-value for each comparison. In addition, report an accelerated failure time (AFT) model using Gompertz or Gompertz–Makeham distributions, which yields a time-ratio (e.g. “treated mice lived 1.22× longer, 95% CI …”). Present these results alongside ΔG so the slope-based inference is complemented by conventional survival metrics. This makes the statistical evidence clearer and easier to compare across studies.
The text asserts that telomerase activation did not increase cancer incidence, but no numbers are shown, so the claim isn’t verifiable. Please provide a simple table of gross pathology and histopathology tumor incidence by study arm (treated vs. control) for each dataset. For each comparison, report a risk ratio with a 95% confidence interval (Clopper–Pearson exact if counts are small). If the sample size is too low to draw reliable conclusions (e.g., n<30 per arm or <5 tumors per group), state this explicitly as an underpowered analysis. This ensures that the “no cancer increase” claim is evidence-based and transparently qualified.

Conclusions and limitations discussed

The mouse studies are consistent with, but do not yet prove, a reduction in actuarial slope under telomerase activation. The strongest evidence comes from the CMV-TERT and knock-in datasets, but the effect size and significance remain uncertain because of methodological issues: survival curves were fit by least-squares rather than maximum likelihood, censoring and left-truncation were ignored, Makeham and frailty terms were omitted, and digitization error was not propagated. These factors could easily shift or inflate ΔG estimates. The most serious vulnerability is model mis-specification (particularly unmodeled Makeham/frailty structure) compounded by digitization noise. What would move the paper from “plausible” to “compelling” is a reanalysis that (i) uses MLE-based parametric survival models with truncation and censoring handled correctly, (ii) tests Gompertz–Makeham and frailty alternatives, (iii) stratifies by cohort/sex rather than pooling, and (iv) quantifies digitization uncertainty. If ΔG remains significant with coherent confidence intervals across these models, the core claim would be much stronger.

Concluding remarks

This is a bold and promising idea, but the inference needs tightening. The quickest way forward is: (1) re-fit survival with MLE (Gompertz–Makeham, left-truncation, frailty) and report hazard ratios or ΔG with 95% CIs; (2) test robustness with a two-annotator Monte Carlo/bootstrapped digitization sensitivity; (3) provide a short analytic derivation of the ΔG=α/(1+(n−1)α) relation, correcting the 10%→12.5% cross-talk example and specifying the mortality link h(t)=∑AᵢPᵢ(t). If implemented, these steps will likely yield slightly smaller but still positive ΔG effects, and position the paper as a more rigorous foundation for the claim that aging can be slowed via single-process interventions.