Rapamycin: the mouse lifespan effect is real. What is still missing in humans?

01What we found

The mouse-level conclusion holds, and it has more structure than the usual summary. In genetically heterogeneous mice rapamycin does lower mortality risk, and 14 ppm given continuously replicated across three mutually independent cohorts (male hazard ratios 0.59, 0.59, 0.72; heterogeneity p=0.31). The dose relationship rises monotonically in males, while in females the lowest dose is already near saturation. Against every compound the same program has tested, the ten strongest arms in females are all rapamycin or rapamycin combinations. The human level is a different story entirely: the only long-term randomized trial in healthy adults was negative on its primary endpoint, the dose it actually delivered was about one third of the labelled amount, and its statistical power was only ever enough to exclude medium-or-larger effects. No randomized trial has ever tested rapamycin against human lifespan or healthspan.

The scope is rapamycin's survival effect in mice and the conditions on it, against what human randomized trials have actually measured and with what precision. This is an evidence description. It provides no dosing, scheduling, or discontinuation regimen of any kind. The milligram and ppm figures that appear below are records of what a given trial actually used; they are not reference doses. We also do not convert mouse lifespan percentages into human years, and we do not treat results for mTOR-related analogues (everolimus, RTB101) as results for rapamycin itself.

02The data: why we can recompute it ourselves

The ITP is a drug screening program the National Institute on Aging runs in parallel at three sites (the Jackson Laboratory TJL, the University of Michigan UM, and the University of Texas Health Science Center UT). It uses UM-HET3, a four-way cross in which every mouse is genetically distinct — which matters, because it avoids conclusions specific to one inbred strain. The program publishes the record for every individual mouse: birth cohort, site, sex, group, drug concentration in the food, age at which dosing began, outcome status and age in days. We downloaded all 17 cohorts from C2004 to C2021: 36,975 mice, 35,297 deaths, 1,678 removals, with a SHA256 fingerprint kept for every file. ITP per-mouse lifespan data

Rapamycin appears eleven times in this collection, and not as simple repetition. The arms differ in ways that matter:

One thing has to be clear before any of the numbers below can be read correctly: each cohort has a single control group, shared by every intervention arm in that cohort. Those 580 control mice in C2009 serve simultaneously as the comparison basis for 17α-estradiol, acarbose, methylene blue and all three rapamycin doses. This means several "significant" results within one cohort share the random fluctuation of one control group, and cannot be counted as several independent confirmations. Genuine independent replication exists only across cohorts, because that is a different set of control mice. Every claim of reproducibility in this article rests only on cross-cohort consistency.

03First, checking that we read the files correctly

Before estimating anything, we used the per-mouse data to recompute the numbers the cohort papers printed. This step is not a formality: if we had misread a code or misdefined the analysis population, everything downstream would be an artifact of our own making.

Matching requires getting two things right at once, and the literature is not uniform on either. First, which statistic. Harrison 2009 reports percentage change in mean lifespan; Miller 2011 and Miller 2014 report median lifespan. Second, which risk set. Rapamycin in C2005 did not start until 600 days of age, and mice that died before that never saw the drug.

Recomputed on each paper's own terms, 16 of 17 target values fall within tolerance. All eight of Harrison 2009's targets match — both pooled sexes plus six site-by-sex cells — with a maximum deviation of 1.3 percentage points. Harrison 2009 For Miller 2014's 42 ppm arm we get pooled median increases of +22.9% in males and +26.9% in females against printed values of +23% and +26%; for the 14 ppm arm the paper reports a 10% decrease in median lifespan in TJL males, and we get −9.8%. Miller 2014 Reproducing even that reversed, drug-unfavourable single-site result is the strongest confirmation we have that the data were read correctly.

The one value that does not match is a single-site effect size: 4.7 ppm males at UM, printed as 16%. We cannot produce it with the median (+7.1%), the mean (+8.6%), the 75th percentile (+6.4%) or the 90th (+12.8%). But the log-rank test for that same comparison gives us p=0.0230 against the printed P=0.02 — an exact match. A test statistic that matches while the effect size does not is not a reading error. That cohort was released on the public database in June 2026 and last updated in July 2026, twelve years after the paper, so the records may have been revised in between. We record it as an unresolved residual and do not explain it away; it is one exploratory single-site figure, and the original paper itself notes the severe power loss in site-specific analyses.

One number that is routinely restated incorrectly

Harrison 2009's well-known "+9% in males, +13% in females" is mean lifespan, not median. Secondhand accounts frequently render it as a median, including old notes in our own neighbouring research repository. The same cohort computed on medians gives different numbers, and the two are not interchangeable. This kind of drift between statistics looks minor, but in a judgment about reproducibility — "several studies keep finding similar results" — it makes different things look like the same thing.

A group difference that was already there before dosing began

Conditioning on 600 days is not just a technical choice. We ran a pre-specified check: deaths before dosing began should be unrelated to group assignment, because nobody had received the drug yet. In both cohorts that started at 20 months, males were imbalanced. In C2005, 17.5% of rapamycin-assigned males died before dosing against 29.9% of controls (p=0.0018). In C2015 the three rapamycin arms pooled give 29.9% against 37.3% (p=0.034), and split apart the imbalance sits in the continuously dosed arm (25.0% against 37.3%, p=0.0087) while the other two are not significant. The two cohorts that started at 9 months show nothing of the kind (0.0%–5.8% pre-dosing deaths across arms, none differing significantly from control).

The original paper found this itself, and gave a reason. Harrison 2009 notes that "mice assigned to the rapamycin-fed group at UT and perhaps at UM had lower mortality prior to 600 days than controls," and our site-by-site results land exactly where it points (UT p=0.0010, UM p=0.077, TJL p=0.86). More importantly, it goes on to say that control and rapamycin mice differed "not only in rapamycin exposure from 600 days of age, but also in the specific formulation of the mouse chows... used between weaning and 600 days." Harrison 2009

So this cohort's comparison has a pre-dosing difference mixed into it. That does not overturn the conclusion — the same dose on the same schedule in C2006, which starts at 9 months and has no such problem, still shows the effect — but it explains why the absolute numbers from any single cohort are unreliable on their own. One contrast worth remembering: on the hazard ratio scale, conditioning barely matters (0.59 against 0.57 in C2005 males); on an absolute scale like "percent increase in mean lifespan" it matters a great deal (+8.9% against +16.0%). Computing a mean-lifespan percentage over all mice nearly doubles this cohort's apparent male effect.

04Finding 1: the effect is real, and it replicates across cohorts

Each of the eight rapamycin monotherapy arms was compared against the control mice of its own cohort and sex, with the time origin set at that arm's own start of dosing, and the model stratified by site.

The same dose on the same schedule — 14 ppm continuous — appears in three independent cohorts, and this is the core of our reproducibility judgment. In males the hazard ratios are 0.59, 0.59 and 0.72, pooling to 0.63 with a heterogeneity test at p=0.31: consistent. In females they are 0.45, 0.38 and 0.30, pooling to 0.38 with heterogeneity p=0.066: a trend toward inconsistency. The female effect is larger but less stable across cohorts.

On an absolute scale, restricted mean survival time gives a more intuitive quantity. In the 42 ppm arm of C2009, females lived 209.5 days longer than controls (95% interval 174.8–244.2) and males 108.7 days longer (54.9–162.6). The proportional hazards assumption is violated in that female arm (p<0.05), consistent with Miller 2014's finding via Gompertz modelling that the slope of the mortality curve changed in this group — so that hazard ratio of 0.22 should be read as an average effect over follow-up, not as a constant multiplier.

One detail that should not be averaged away: the same dose can produce results in opposite directions at different sites. For 14 ppm in males, median lifespan was +21.4% at UT and −9.8% at TJL. Three sites, the same mice, the same food, the same protocol, and the spread is still that large. This is what "the same study was repeated at three sites" is actually worth: more reliable than a single centre, but not the same as results that agree everywhere.

05Finding 2: the shape of the dose relationship differs by sex

C2009 is the only design that tested three doses inside a single cohort. All three arms share those 580 controls, so the three contrasts are correlated — but the comparisons between doses are made against the same set of control mice.

In males the dose relationship rises monotonically and fits a log-linear form (departure test p=0.78): 4.7 ppm gives a hazard ratio of 0.85 (p=0.11, no statistical evidence), 14 ppm gives 0.71, and 42 ppm gives 0.60. In females it does not work that way: the lowest dose, 4.7 ppm, already reaches 0.41, then 0.26 at 14 ppm and 0.24 at 42 ppm, departing clearly from log-linearity (p=0.007). The curve is close to saturation at the low end, and further dose buys little. The dose×sex interaction is p=6×10⁻⁶.

So the common statement that "females respond more strongly" can be made more precise. It is not simply a larger magnitude: the shape of the dose-response curve differs between the sexes. The dose that does nothing in males is already near maximal effect in females.

06Finding 3: when to start, and whether to keep going

Starting age. At the same 14 ppm continuous dose, starting at 9 months (C2006) and at 20 months (C2005) are nearly identical in males: hazard ratio 0.59 in both, with restricted mean survival differences of +76 and +81 days. In females the 9-month start is stronger (0.38 against 0.45). Twenty months in a mouse is roughly middle-to-old age in a human. That a late start still works is one of rapamycin's most discussed properties, and it holds up in our independent calculation. But note that this is a cross-cohort comparison: the two cohorts have different control mice, and no mouse was randomized to "start early" versus "start late."

Schedule. C2015 compared three arrangements of 42 ppm within one cohort: continuous dosing, alternating cycles, and dosing from 20 to 23 months and then stopping. The stopped arm is the weakest (0.80 in males, 0.85 in females, neither with statistical evidence). In females, comparing the stopped arm directly against the continuous arm gives a clear difference (hazard ratio 1.86, p<0.0001); in males the same comparison is inconclusive (1.14, p=0.36), and this arm is not large enough to settle it. Alternating cycles are comparable to continuous dosing.

What this set of results means is limited but clear: in females at least, three months of exposure followed by withdrawal does not buy the survival benefit that continuous dosing does. It cannot answer what happens when a human stops, and it cannot be used to design any schedule.

07Finding 4: a problem that biases survival estimates, and how much it actually matters

Miller 2014 reports a higher proportion of cages in which males were removed from study for fighting at the highest dose: 11%, 18% and 22% across the three sites. Removed mice are censored at the date of removal in a survival analysis, and censoring that depends on treatment biases the estimate.

We tested this directly. In C2009 the proportion of males removed rises monotonically with dose: 1.7% of controls, 3.8% at 4.7 ppm, 9.0% at 14 ppm, and 23.1% at 42 ppm (both 42 ppm and 14 ppm against control at p<0.001). Females show no such pattern at all (11.8%–15.4% across arms, against 13.6% in controls).

A thirteen-fold difference, 23.1% against 1.7%, demands a bounding analysis. We took the most unfavourable assumption available: treat every removed male in the rapamycin arm as having died on the day of removal, while treating every removed control as having survived to the end of follow-up. Under that worst case the hazard ratio moves from 0.59 (0.48–0.74) to 0.79 (0.65–0.97) — the upper bound of the interval is still below 1. The worst case in the opposite direction gives 0.37.

The conclusion is therefore robust: the censoring imbalance weakens the effect estimate but is not enough to erase it. This is also why the check is worth running rather than assuming censoring is random — had the effect itself been only marginally significant, the same bias would have been enough to flip the conclusion.

08Finding 5: how much support "most promising" actually has in the mouse data

The old article on our legacy site is titled "the most promising longevity drug?" That claim can be put to the data rather than left as rhetoric. The ITP has tested dozens of compounds over twenty years. Run them all through one estimator — against their own cohort's same-sex control, from their own start of dosing, stratified by site — and where does rapamycin land?

In females, the ten strongest arms are all rapamycin or rapamycin-containing combinations, led by C2009's 42 ppm at 0.22. Of 89 estimable arms, only 17 reach p<0.05 in the protective direction.

In males the ranking is far more crowded. Of 93 arms, rapamycin-containing ones hold ranks 1, 3, 7, 8 and 9 — but rank 2 is acarbose alone (0.49) and rank 4 is 17α-estradiol (0.51), comparable to or stronger than rapamycin monotherapy at 0.59. In males 34 arms reach p<0.05, and that proportion is itself a hint that these nominal significances contain a multiple-comparison component (the shared controls within cohorts also make these tests correlated).

So "within this screening program's mouse survival data, rapamycin sits at the front of the effect-size distribution" is a statement the data support, and it is particularly striking in females. But it is a statement about a ranking on mouse survival. It does not carry over to "most promising for humans" — and when the two are written in the same paragraph with a milligram figure following, what the reader receives is the second one.

09The human level: what the one long-term trial measured

PEARL is the longest randomized trial of rapamycin in healthy, normally aging adults: 48 weeks, double-blind, placebo-controlled, run decentrally, registered as NCT04488601. Participants were randomized to placebo, 5 mg per week, or 10 mg per week of compounded rapamycin. The primary endpoint was visceral fat measured by dual-energy X-ray absorptiometry; secondary endpoints were blood markers, lean tissue mass and bone mineral content. PEARL trial report

The primary endpoint was negative: no significant change in visceral fat, η²ₚ=0.001, p=0.942.

What that negative result can support depends on two things the trial reported about itself and that are routinely dropped in retellings.

First, the dose actually delivered was below the label. During the trial the authors discovered that the compounded rapamycin they were using — chosen so that a matching placebo could be prepared — had lower bioavailability than commercial formulations. The trial was paused and the question was tested in an independent group: the compounded form reached about one third of the commercial 24-hour blood concentration. The paper states plainly that equivalent effective doses for the compounded form are "approximately 66% less." On that basis, a nominal 5 mg and 10 mg per week correspond roughly to 1.7 mg and 3.3 mg per week of a commercial formulation. This cuts two ways. Reading the negative result as "rapamycin does nothing at 5–10 mg per week" is inaccurate, because the exposure tested was below what the label implies; and these milligram figures cannot be compared directly against milligram figures from any other source.

Second, this trial only ever had the power to exclude medium-or-larger effects. 114 people completed and entered the analysis (39 placebo, 40 on 5 mg, 35 on 10 mg), and a further 11 discontinued and were excluded — so this is a completer analysis, not intention-to-treat, and in a trial whose main purpose is safety, excluding those who dropped out biases toward underestimating risk. At this sample size, a one-way analysis of variance needs η²ₚ≥0.080 for 80% power; against a conventionally "medium" effect it has about 65% power. The observed value was 0.001. "No significant change" therefore cannot be read as "no effect has been demonstrated," and still less as "equivalence has been demonstrated."

As for the positive findings that circulate widely, it is worth looking at where they come from. The lean tissue mass improvement occurs in the subgroup of "women on 10 mg" — an arm containing 8 women. The odds ratio for improvement is 28, with a 95% confidence interval from 2.4 to 324, spanning two orders of magnitude: this is what an odds ratio estimated on single-digit cell counts looks like. The same measure across all participants is 2.29 (p=0.12), and in men 1.09 (p=0.90). There is also a less obvious point: for all 40 women to detect an effect with 80% power, the effect would need to reach η²=0.207, and the printed female lean-mass effect is 0.202 — right at the detection threshold. An effect that can only be detected when it is that large is also precisely the kind most inflated by selection. Across the whole body-composition table there are 24 comparisons, 2 reach nominal p<0.05, and pure chance would produce 1.2; after Holm or Benjamini–Hochberg correction, none survive. The two nominally significant cells also point in opposite directions: lean mass improved in women, while the odds of bone mineral density improvement were significantly lower on 5 mg (odds ratio 0.24, p=0.04) — and on 10 mg the same measure is 1.00. A real dose effect does not usually appear at 5 mg and vanish at 10 mg.

Beyond that, the trial's self-reported health and pain instruments (SF-36, WOMAC) were, by the paper's own statement, "not listed as specific study endpoints," so the reported improvements in emotional wellbeing and general health are non-prespecified outcomes.

10Where the chain from mouse to human breaks

The mechanism is more complicated than "switching off mTOR." Rapamycin does not directly inhibit mTOR kinase activity. It binds FKBP12 and allosterically inhibits mTORC1; acute dosing leaves mTORC2 largely uninhibited, and even some mTORC1 substrates are resistant to it. Feldman 2009 The path from "partially inhibiting a nutrient-sensing complex" to "extending mammalian lifespan" is mechanistically coherent but not unique.

Extending lifespan and slowing aging are not the same thing, and the field has not settled which is happening. One side holds that rapamycin's lifespan effect is "largely separable" from aging itself, with much of the extra survival coming from suppression of the mice's main cause of death, cancer. Neff 2013 The other holds that it does slow aging across multiple domains in mice. Wilkinson 2012 Both are peer-reviewed primary literature reaching opposite conclusions. What our data can speak to is survival curves; it cannot adjudicate this dispute, because the ITP lifespan files contain no functional outcomes.

On the human side, the dose question is simply empty. A 2025 review of the evidence for off-label use in healthy adults states plainly that, as far as healthspan extension is concerned, "there is still no established dose response curve for rapamycin," while the recommendations circulating in practice (5–7 mg weekly, or 10–15 mg every other week) are not built on such a curve. The same review raises a theoretical concern: prolonged mTOR inhibition may be followed by rebound mTOR hyperfunction, which is directly relevant to intermittent schedules. 2025 clinical evidence review Set that against what we see in mice — a clear dose relationship whose shape differs by sex — and on the human side even the shape of the curve is unknown.

The confirmatory trial that came closest to a clinical endpoint failed, and the drug was not rapamycin. RTB101 is a mechanistically mTOR-related but distinct compound, and it missed its endpoint in a phase III trial using respiratory tract infection as the clinical outcome. Mannick 2021 This cannot be scored as a failure of rapamycin, and it cannot be waved away by rapamycin's advocates either: it is what the record for this class of intervention on a hard human endpoint currently looks like.

Real-world usage reports do not fill the gap. The "1–6 mg per week" range in the old article comes from a survey of 333 adults with a history of off-label use, plus 172 never-users as a comparison. The study's value is in describing this population's basic characteristics and self-reported side effects; the authors' own formulation is that it offers "preliminary evidence" that use is safe in otherwise healthy adults. 333-person self-report survey It is a self-selected sample with self-reported outcomes and no randomization, so it cannot support causal conclusions about efficacy or safety, and it cannot serve as a dose reference — least of all after PEARL showed that the same milligram figure can correspond to a threefold difference in blood concentration across formulations.

11Safety boundaries: what is known, and what this article cannot carry

Rapamycin is a prescription immunosuppressant with approved indications and an adverse-reaction profile stated on its official label. This article is not a safety assessment. It records what human studies actually observed, and the denominator behind those observations.

In PEARL the total number of adverse events was similar across arms (117 on 10 mg, 116 on 5 mg, 122 on placebo), with serious adverse events numbering 1, 2 and 3 respectively; one case of anemia occurred in the 5 mg arm, resolved after transfusion, and that participant completed the trial. Working out the denominator makes the point better: 114 people for 48 weeks each is about 105 person-years of exposure, and it excludes the 11 participants who discontinued. At that size, that duration, and with actual exposure below the label, similar event counts across arms cannot support any conclusion about long-term safety. The reduced odds of bone mineral density improvement on 5 mg (odds ratio 0.24) does not survive correction for multiple comparisons, but its direction is unfavourable, which makes it a signal to keep watching rather than one to drop. The 2025 review separately collects concerning changes including raised triglycerides, glycated hemoglobin and very-low-density lipoprotein. 2025 clinical evidence review

The mouse data contain one thread with human relevance: in one ITP cohort, metformin reduced the glucose intolerance caused by rapamycin. Metformin–rapamycin metabolic interaction That indicates metabolic disturbance is a genuine pharmacological consequence, but this evidence too is in mice.

Under this repository's publication policy, the scope of this article is fixed as evidence description. Providing actionable dosing, scheduling, or discontinuation advice requires matching review by a real human clinical or pharmacy professional, which this repository does not have. So we do not provide it, and we do not substitute a disclaimer for the review.

12What we know and what we do not

Known. In genetically heterogeneous mice, oral rapamycin lowers mortality risk, and the effect recurs across three mutually independent cohorts, three trial sites, both sexes, and start ages from middle age to near-old age. Among all compounds the same program has tested, its effect size sits at the front, particularly in females. The dose relationship is monotonic in males and near-saturated at low dose in females. In females at least, stopping the drug gives up most of the survival benefit. And these conclusions survive the worst case for the censoring imbalance.

Not known. Whether rapamycin extends human lifespan or healthspan: no randomized trial has ever tested those endpoints. The only long-term trial in healthy adults was negative on its primary endpoint, but its actual exposure was below the label and its power reached only medium-or-larger effects, so it supports neither efficacy nor its absence. There is no established human dose-response relationship; there is no reliable bridge between mouse ppm and human milligrams, and milligrams are not even comparable across formulations. How much of the extra survival amounts to slowed aging remains disputed in the primary literature. And long-term use in healthy populations has no trial long or large enough to support a safety conclusion.

What would change the judgment. A randomized trial in healthy adults with a functional or disease endpoint as its primary outcome, adequate power, and a known exposure level. The trials now under way cannot carry that question: ERAP, which images brain metabolism by positron emission tomography, is a six-month single-arm open-label phase IIa biomarker study, and REACH tests the effect of 1 mg daily on markers of Alzheimer's disease burden. 2025 clinical evidence review Single-arm, open-label, surrogate endpoints, six months — none of these designs are built to answer whether healthspan can be extended.

13On the revision to the legacy article

This article updates the legacy site's piece "Rapamycin and anti-aging: the most promising longevity drug?" The old draft got something right: it stated honestly that PEARL's primary endpoint did not reach statistical significance, which is more than many contemporaneous reports managed. What this revision changes:

• The unqualified characterization "the most promising longevity drug" is withdrawn. As shown above, that ranking has support in the mouse survival data and no counterpart in the human evidence. The animal and human levels are now stated separately.
• Two key qualifications on PEARL are added: the dose actually delivered was about one third of the nominal amount, and the trial's power boundary. The old draft mentioned neither.
• Subgroup results are returned to their sample size: the lean tissue mass improvement the old draft cited comes from 8 women, and does not survive correction for multiple comparisons.
• Actionable milligram figures are no longer presented alongside. The old draft placed the "most promising" characterization next to PEARL's 5/10 mg and the 333-person survey's "1–6 mg per week," which makes it easy for a reader to take the latter as a reference dose. This article keeps to the scope of evidence description.
• Branches of the old draft not verified in this round are not retained, including the details of autophagy mechanisms, topical skin use, several immune surrogate endpoints, and the metformin comparison section. That is not a judgment that they are wrong; this round did not check them, and this repository does not retain unverified content.

14Research notes, reproduction and updates

The animal-level conclusions here come from our own reanalysis of the public per-mouse data, not from restating published figures. The analysis plan — estimands, handling of left truncation, handling of shared controls, dose and interaction tests, censoring checks, whole-library reference — was written and locked before any effect was estimated, and deviations during execution are registered in section 9 of the plan file. There are no individual data at the human level; all recomputation there is at the summary level, as stated in the text.

Every source's identity was verified. This is not a formality: two reference IDs carried over from a neighbouring research repository turned out to point at completely unrelated papers (one a kidney cancer surgery trial, one a cryptography paper), and were corrected by search. Every source in this article therefore records whether the material actually retrieved matched the expected identity.

Reproduction covers all code, input fingerprints and outputs: download and fingerprints (A0), source retrieval and identity checking (A0b), search log (A0c), data construction and sample flow (A1), input validation (A2), main analysis (B1), structural analyses and whole-library reference (B2), human-level recomputation (C1), plotting data export (E) and figures (F). Reproduction package

Limitations worth stating. The ITP lifespan files contain only survival and censoring, with no functional, disease or pathology outcomes, so this article cannot judge from these data whether extended lifespan amounts to slowed aging. Food concentration (ppm) is not a blood level, and we obtained no matching exposure measurement. PEARL's individual data are not public. And the one unresolved input-validation residual is recorded as such.