Does the Body Suddenly Age at 44 and 60?

A large number of significant molecular tests within an age window requires further justification before it can be interpreted as faster biological change at that age. Sample sizes, actual age differences between groups, smoothing-induced dependence, multiplicity correction and the rule for selecting peaks all affect the curve. These analyses do not identify individual transition ages or measure disease, functional or survival outcomes.

Correction, 20 September 2026: Our previous reanalysis omitted age sorting, did not reselect smoothing spans within permutations, and mishandled missing predictions outside bootstrap samples' age support. We correct these problems and withdraw claims of identified genuine nonlinear regions and peak-age confidence intervals. Selecting one maximum in each predefined interval is no longer described as automatically discovering two peaks. Both editions, eight figures and the reproduction package are updated.

01The original study and subsequent methodological review

Shen and colleagues' 2024 study included 108 participants aged 25–75, with median follow-up of 1.7 years and a maximum of 6.8 years. The count analysis first summarized repeated samples within individuals and then compared people along the age axis. A longitudinal source cohort does not make that particular curve an individual's continuous aging trajectory. Original study

Our recovered transcriptomic matrix contains 97 people and 8,556 genes. The combined-omics grid contains 11,305 variables. For the other omics layers, we checked the authors' saved grid and test outputs; we did not independently reprocess all participant-level measurements. Code and inputs

A 2026 methods preprint examines how smoothing, sliding windows and age distributions can generate apparent waves. Shen and Snyder, the original first and last authors, also contributed. The preprint, simulations and two repositories are not independent human-cohort replications. Methods preprint Associated code

The publisher page retrieved on 20 September 2026 still carries its 15 July editorial notice. It questions the reliability of some analyses detecting molecular-change peaks at particular ages and says the authors and editors are assessing the scope and appropriate corrections. This ongoing editorial process must not be reported as an established retraction or a fully resolved issue. Current notice

02What the count curve measures

The published combined-omics curve counts 9,312 significant variables at age 44 and 8,276 at 60, with a trough of 7,117 at 54. Approximately 60–82% of variables pass the authors' threshold at individual centers. These are fluctuations on a high baseline; the vertical axis is neither molecular change velocity nor a percentage of the body that has aged.

We rechecked 293,930 pairs of p/q values against the authors' saved outputs, agreeing within a numerical tolerance of 1e-8. Counts also agree for 15-, 20-, 25- and 30-year windows. Reproducing these calculations does not establish the validity of their biological interpretation.

Input verification resolved a manifest discrepancy: an old CSV hash described an export without a variable index, while the actual file subsequently included variable IDs. Values, variable order and age columns agree with the original workbook. The corrected manifest records the actual file and export procedure, separating format changes from substantive data changes.

03Age sorting, interpolation and the objects being tested

The transcriptomic procedure sorts participants by age, selects each gene's LOESS span by cross-validation, and predicts 99 half-year points from ages 26 to 75. Combined omics retain the 91 shared points from ages 30 to 75. Prediction points are not additional independent participants.

The authors also standardized RNA trajectories. After restoring the sorting step omitted in our original reanalysis and applying the reported standardization, all 778,596 RNA grid values agree with the saved combined matrix, with maximum absolute difference about 3.25e-13. Another comparison difference concerns multiplicity: the published panel applies BH correction across 11,305 variables before selecting RNA, whereas an RNA-only analysis corrects 8,556 genes. Matching that family reproduces all 26 RNA counts as well.

Only seven participants are younger than 40. At center c, the test compares values in [c−10,c) with [c,c+10). Sample counts and age composition vary, and adjacent interpolated points share information from neighboring participants. Treating those predictions as independent Wilcoxon observations generates statistical behavior requiring separate calibration.

Removing LOESS and applying the same sliding-window comparison to the processed participant-level RNA values yields no significant genes at any of the 26 centers after BH correction at 0.05. Detection counts are sensitive to smoothing/interpolation. With limited sample sizes, non-detection can also reflect low power and does not establish that the curve contains no biological information.

04Permutation must include smoothing-span selection

The original paper permuted grid ages after smoothing. This breaks the local structure created by interpolation, so its comparison does not preserve how the original procedure behaves with no age association. Age labels should instead be shuffled at participant level before subsequent processing. Methods preprint

Our previous version shuffled earlier but retained spans selected from the observed data. The correction uses 199 joint age permutations, reselecting each gene's span from 0.3, 0.4, 0.5 and 0.6 in every run. The same participant permutation is applied to all genes, retaining their dependence.

All 26 observed centers fall inside their pointwise 95% permutation envelopes. We separately assess the entire curve with an exploratory statistic specified before these recalculations: the maximum count within ages 40–50 plus the maximum within 57–65, minus twice the minimum within 51–56. Its observed value is 3,314, versus a permutation median of 3,817. The Monte Carlo upper-tail probability, including the observed value, is P=0.675.

The observed fluctuation is not unusually large under this no-age-association comparison. Resolution is limited by 199 permutations. Label exchangeability is an assumption, and sex, BMI and other covariates are not adjusted. This is not a direct comparison of linear and nonlinear biological aging models, nor an estimate that the contribution of real change is zero.

05Two predefined intervals do not establish two discovered peaks

Our old statement that every one of 100 pure-noise simulations developed two peaks used a rule selecting one maximum in each of two specified intervals. The rule returns two locations even with additional peaks, monotonic trends within an interval, or boundary maxima.

Among the retained 100 noise curves using the actual age distribution, only 25 have exactly two interior local maxima (flat tops counted once); the others have three to five. In 14 runs, the second interval's maximum is at the age-65 boundary. These local maxima are not filtered by significance or prominence, so their counts are not numbers of biological transitions either.

The simulations use independent Gaussian noise, a fixed span of 0.5 and clipping of negative smoothed predictions to zero. They demonstrate fluctuations under specified processing conditions. They do not estimate what share of real-data fluctuations is artifactual or establish genuine broad age regions from simulated maximum locations.

Even a straight trajectory for every gene can produce changing window counts. Under Y=α+β×age+error, the two groups' expected expression difference is β×(mean older-group age−mean younger-group age). Both that age contrast and the sample sizes change across windows. A constant individual slope therefore does not imply equal window contrasts or equal power.

06Effect sizes need an appropriate comparison too

Cliff's δ describes the difference between the probability that an older-window value is larger and the probability that a younger-window value is larger; ties contribute zero. It is not a percentage increase in expression concentration or an aging rate.

Across genes, the window-specific median absolute δ is about 0.08–0.17. A median does not establish that every gene's effect is small. Absolute effect statistics also have nonzero finite-sample baselines without age association. The old comparison used 2,000 independent Gaussian genes, failing to preserve measured dependence and ties; its narrow band cannot establish genuine enrichment at particular ages.

We instead jointly permuted participant ages 999 times while retaining all 8,556 genes' values, ties and dependence. The unadjusted tail probability at age 57 is below 0.05, but no center reaches 0.05 after accounting for examination of all 26 centers. A whole-curve test using the maximum standardized deviation gives P=0.409.

We withdraw the affirmative claim of genuine nonlinear enrichment at ages 45–50 and 56–59. This analysis neither excludes associations in individual genes nor directly tests the added value of nonlinear over linear models. It cannot establish zero age effects.

07What resampling says about location stability

Resampling participants can omit the youngest or oldest individuals, leaving some LOESS predictions outside that replicate's age support. The old accelerated test ranked missing values at the high end, whereas the reference Wilcoxon function removes them. We corrected this mismatch and verified 600 comparisons against direct R fits and tests.

In 100 deterministically seeded resamples, 66 contain unsupported grid points. Each gene retains its observed-data span, making this a conditional sensitivity analysis, not a bootstrap covering all tuning uncertainty.

The 2.5th, 50th and 97.5th empirical percentiles of the maximum's location are 40, 43 and 50 within the first interval, and 57, 61 and 65 within the second. Maxima reach the interval boundaries in 27 and 26 runs, respectively. These ranges are constrained by the selection rule itself; they are not 95% confidence intervals locating real transitions or individual aging events.

08What other studies contribute

Lehallier and colleagues reported count peaks at 34, 60 and 78 in a proteomic study of 4,263 people. Their original DE-SWAN uses covariate-adjusted linear models and F-tests without the same preceding LOESS interpolation. That is a different design. However, loss of peaks after age permutation mainly challenges no age association; by itself it does not exclude changing counts from constant slopes and varying window composition. Lehallier study

Li and colleagues' HPP preprint uses the Human Phenotype Project, including 10,019 participants aged 40–70. Modified sliding-window analyses and age-acceleration groups show broad regions of change, with an interpretation linking some female patterns to typical menopausal ages. Coinciding ages do not identify hormonal mediation, and grouping and related statistical methods can affect the patterns. We have not independently reanalyzed those participant data. HPP preprint

Pyrkov and colleagues constructed a longitudinal blood-count indicator and estimated autocorrelation times of its fluctuations, reporting an age-related increase. This offers another model for investigating dynamic recovery, but autocorrelation time is not observed clinical recovery time. Its lifespan limit is an extrapolation, and neither result validates transitions at 44 and 60. Pyrkov study

These studies leave age-associated phenomena worth investigating. Identifying genuine nonlinearity or synchronized transitions requires explicit comparisons with appropriate linear models, attention to sampling and window structure, and sufficiently long individual follow-up where feasible. Replacing exact ages with supposedly established broad regions around midlife and 60 does not resolve the identification problem.

09Current conclusions and reproducible scope

The reproduction package includes code, input acquisition and fingerprints, complete aggregate outputs and limitations. The 199 full-tuning permutations, 999 effect-size permutations and 100 conditional resamples were actually run. Retained synthetic simulations illustrate processing behavior. Original workbooks, participant-level expression/age data and paper full texts are excluded from public assets.

Codex (agent) performed the substantive revision, recalculation, self-review and bilingual checks as a revision coauthor; Kimi's original text and approvals remain historical. Current review is the same revising agent's research/editorial work, not independent, human or clinical professional review. Source access and checked funding/interests appear below. This is a targeted audit rather than a systematic review of the entire field.