Rapamycin’s status in the longevity field hardly needs elaborate setup. From the NIA Interventions Testing Program (ITP) first reporting mouse lifespan extension in 2006, to Harrison’s 2009 paper that electrified the field by showing benefit even when started at 20 months of age, to the 2025 PEARL trial’s signals of muscle and mood benefits in healthy older adults, this thread has stretched for nearly two decades and has been repeatedly framed as “the closest thing to a clinical anti-aging intervention.” Yet one question has long remained unresolved: is the ruler we use to measure whether rapamycin “works”—the parallel control group—itself reliable?
A preprint posted on bioRxiv in October 2026 uses 18 years of ITP data from 10,757 control mice to conduct a thorough calibration of that ruler. The conclusion is not gentle: at least one previously unflagged male control cohort was anomalous enough that, when excluded, two rapamycin regimens lost statistical significance. Meanwhile, the 2017 UT male control cohort—the very cohort that ITP publications had singled out as “unusually short-lived”—does not constitute a statistical outlier under a rigorous prediction framework.
This means that some of our past judgments about whether certain rapamycin regimens are “effective” or “ineffective” may rest on a reference frame that was never properly tested.
The ITP Control Record: Eighteen Years, Three Sites, One Unmeasured Variable
The ITP’s design is remarkably stringent by aging research standards: the same protocol runs in parallel at three independent sites—The Jackson Laboratory (TJL), the University of Michigan (UM), and the University of Texas (UT)—using genetically heterogeneous HET3 mice, both sexes, with adequate sample sizes. But even under such standardization, control lifespans fluctuate from year to year. The authors—LTM and GAC at UT San Antonio, among others—did not sidestep this awkward fact; they turned it into the object of analysis.
They first did something fundamental that had been missing: they decomposed the variance. At the individual level, residual standard deviations were 183.2 days in females and 253.4 days in males—males were inherently more dispersed. But when these individuals were averaged into cohort means, what truly dominated cohort-to-cohort fluctuation was the site × year interaction effect, not simple sampling noise. In females, year-to-year variation accounted for 50.9% of the variance in observed cohort means; in males, that figure reached 71.1%. Systematic differences among sites were comparatively small: the highest and lowest site means differed by only 20.4 days in females and 37.2 days in males.
The implication is direct: increasing the number of parallel controls per group can compress sampling noise, but it cannot eliminate real year-to-year shifts. In a given year, UM males may, for reasons of husbandry, microbiome, or environmental factors not yet fully identified, live longer or shorter as a group than in previous years. Every intervention tested against that control group in that year is measured with this “shifted ruler.”
Rejudging Two Old “Anomalies,” and One Missed New Signal
The paper handles historical data with two frameworks. The first is a “full-record in-sample” analysis: fitting prediction intervals using all 108 site × sex × year cohorts from 2004–2022. Under this framework, not a single historical cohort fell outside the 99% prediction interval. C2022 TJL females (948 days, the highest female median lifespan on record) and C2017 UT males (among the lowest) both fell within their intervals. This itself is a warning: searching a complete historical record for the “most extreme” cohort and then declaring it anomalous is statistically almost doomed to fail, because extreme values are bound to exist.
The second framework is the paper’s real contribution: prospective rolling prediction. For each newly completed cohort, the model is fitted using only prior historical data to generate a 99% prediction interval, and the cohort is then tested for whether it falls outside that interval. Among the 60 comparisons meeting the eight-year minimum window, 57 were contained (95.0%). Only three were flagged: C2015 UM males, C2018 UM males, and C2022 TJL females.
C2017 UT males were not flagged. The paper’s wording is restrained but clear: their unusually short observed lifespan fell entirely within the range expected from prior male year-to-year variation. The “anomaly” label attached to this cohort in earlier ITP publications lacks sufficient statistical support.
C2015 UM males, by contrast, were the only cohort flagged under both the primary model and the temporal-drift sensitivity model. This cohort had been noted in the original C2015 report for “site-specific survival patterns,” but had never been formally evaluated or subjected to a post-exclusion sensitivity analysis. The authors did exactly that: excluding all UM males (controls and treatment-assigned animals alike) and re-running the C2015 male intervention analyses.
The results are as follows:
- Rapa 20 (rapamycin starting at 20 months): median lifespan increase dropped from 10.7% to 6.9%, remaining significant (P from 6.57×10⁻⁴ to 0.0022).
- Rapa cycles (intermittent rapamycin): changed from a 9.3% increase (P=0.0019) to a 3.6% decrease (P=0.217)—significance vanished and the direction reversed.
- Rapa 20-23 (rapamycin window at 20–23 months): changed from a 10.7% increase (P=0.024) to a 6.0% increase (P=0.085), no longer significant.
So, is intermittent rapamycin still worth pursuing? Based on this dataset, the answer must be qualified. The intermittent regimen’s effect not only lost significance after excluding a formally flagged anomalous control cohort—it reversed direction, indicating that its previously observed benefit depended heavily on the choice of reference frame. The continuous regimen, while attenuated, retained significance. But this is, after all, only a sensitivity analysis; excluding an entire site also reduces sample size and statistical power, so one cannot conclude that the intermittent regimen is ineffective. A more reasonable conclusion is: the effect estimate for intermittent rapamycin is more sensitive to control-group quality than the continuous regimen, and until stronger evidence with better-calibrated controls is available, the intermittent regimen should not be treated as an equivalent alternative to the continuous one.
[1] Original paper: Quantifying historical variation in control lifespan across three sites and eighteen years of the NIA Interventions Testing Program. bioRxiv, 2026.10.21.760720. Specific statement: In the full-record in-sample analysis, none of the 108 historical cohorts fell outside the 99% prediction interval; the prospective rolling framework flagged C2015 UM males, C2018 UM males, and C2022 TJL females, but did not flag C2017 UT males; after excluding UM males, the Rapa cycles effect changed from +9.3% (P=0.0019) to −3.6% (P=0.217). Source link: https://doi.org/10.64898/2026.10.21.760720
[2] Paper Methods section: 10,757 control mice, 18 cohort years from C2004–C2022 (no C2008 cohort), three sites (TJL, UM, UT), HET3 mice. Source link: A prediction-interval framework for identifying outlier control cohorts in lifespan studies | bioRxiv
[3] NIA ITP 2006 cohort data: rapamycin at 14 ppm (starting at 9 months), female median lifespan 1052 days vs. control 892 days, an 18.1% increase; male 902 days vs. 790 days, an 8.2% increase. Source: Miller et al., 2011, cited in PMC7198054 Table 1. Link: Translational Geroscience: From invertebrate models to companion animal and human interventions - PMC
[4] PEARL trial: 114 healthy adults aged 50–85, randomized double-blind placebo-controlled, weekly 5 mg or 10 mg rapamycin for 48 weeks; females in the 10 mg group showed significant increases in lean mass and reduced pain, while the 5 mg group showed improved emotional well-being; safety profile was favorable. Source: Aging-US, 2025. Link: https://www.eurekalert.org/news-releases/1082933