🧬 Longevity
An AI-Designed Drug Reversed 3-4 Years of Biological Age in 12 Weeks. The Anti-Aging Dose Is Not the Lung Dose.
In a 42-patient trial, rentosertib, a drug whose target and molecule were both found by AI, reversed proteomic biological age by 3-4 years in 12 weeks, and the regimen with the strongest aging signal was not the one with the best lung results.
Three to four years, in twelve weeks. That is the headline from a September 7 Nature Biotechnology study in which rentosertib patients saw predicted biological age fall roughly 3 to 4 years over a 12-week course, measured by six proteomic aging clocks built independently at Harvard, Oxford, Peking University, and Insilico Medicine, using blood proteins rather than DNA marks. All six pointed the same way. 2013 Nobel laureate Michael Levitt said that agreement is what convinced him, since the models share neither features nor training data. And the detail that elevates this above another biomarker paper is a dosing paradox: the regimen that reversed aging best was not the one that fixed lungs best, which means the trial accidentally tested two different medicines under one name.
Rentosertib is the drug we covered in March and again in May: the first molecule whose disease target, the kinase TNIK, was identified by AI and whose structure was designed by generative AI, an 18-month target-to-candidate sprint followed by positive Phase 2a lung results in Nature Medicine in June 2025. This is the first clinical geroprotective readout for a de-novo AI-designed drug, extracted from blood already sitting in freezers. Unlike repurposed generics such as rapamycin or metformin, this is a novel molecule for a novel AI-found target showing an aging signal in humans. On the evidence, the claim stands.
Six clocks, one verdict
Proteomic aging clocks read biological age from blood proteins, the molecules doing the actual work, so proteome shifts can be tied to mechanisms. The six here, ProtAge, OrganAge in chronological and mortality flavors, PAC, ipfP3GPT, and PAOPAC, span classical machine learning and deep learning and were trained on chronological age or mortality risk. Applied to 42 trial participants (mean age 67.1, all Asian, all with idiopathic pulmonary fibrosis; serum drawn at baseline and weeks 2, 4, and 12), every treatment arm showed biological age falling while placebo barely moved.
Across 54 treatment-versus-placebo comparisons, 21 reached significance at Q below 0.10, concentrated at week 4, where 11 of 18 were significant, an early peak suggesting a fast proteomic response, then a plateau. A permutation test with shuffled patient labels gave a null expectation of 0.15. Twenty-one against 0.15 is 140 times the chance expectation. Founder Alex Zhavoronkov summarized the peak as 3 to 4 years reversed, 6 at the maximum on one clock.
Same 60 milligrams, split in two
For lungs, the winner was clear. 60 mg once daily raised forced vital capacity 98.4 mL over 12 weeks against a 20.3 mL placebo decline, while 30 mg twice daily won on aging with nine significant clock comparisons across both chronological and mortality clocks. Same drug, same 60 milligrams per day, split in two, and the anti-aging effect strengthened while the lung effect weakened. That divergence is the finding no single-biomarker artifact explains cleanly.
Pharmacokinetics makes it stranger: once-daily 60 mg delivered 2.5 times the total exposure of twice-daily dosing, 3,450 versus 1,390 h·ng/mL at week 12, with the highest peak concentration, yet trough levels were similar between the arms and the half-life ran 10.9 to 12.0 hours, so the twice-daily arm simply bathed tissues in steadier, lower-amplitude exposure. Were the aging signal tracking total exposure or peak, the once-daily arm would have dominated every clock, and it did not.
The signal tracks steady trough-level exposure, which saturates, while the lung benefit keeps climbing with dose, a fingerprint suggesting TNIK's aging effect maxes out at modest, steady levels, exactly what chronic dosing wants, which means the future longevity dose, if one ever exists, will likely sit well below the lung dose.
Two rulers, same answer
A second, independent check hides in the lung data. Healthy adults over 65 lose roughly 20 to 50 mL of forced vital capacity per year, so dividing the 98.4 mL gain by that range converts the lung improvement alone into 2.0 to 5.0 years of reversed age-related decline, a back-of-the-envelope calculation that assumes healthy-population rates apply to IPF lungs, which they only approximately do, and which carries a wide confidence interval of 10.9 to 185.9 mL.
On the same arm at week 4, the proteomic clocks read 2.71 to 3.46 years. One ruler measures exhalation while the other measures 2,841 serum proteins, and both land near three to four years. The paper never claims that convergence. Its own numbers deliver it, though part of the FVC gain in IPF patients is disease reversal rather than aging reversal.
A flagship result for the price of a re-analysis
Step back and the study design is the real story. Serum collection at four timepoints was already in the Phase 2a protocol for exploratory biomarkers, so all 168 samples were drawn, run on the Olink Explore 3072 platform, and paid for by the lung program years ago, and the September paper added only the re-analysis: six published clocks, a permutation test, pathway enrichment, and comparison against 55,319 UK Biobank profiles, with open-source Python code and data under accession OMIX008341.
Marginal cost of the aging readout was compute time plus analyst labor. A standalone 12-week longevity trial in 42 patients would run into the millions.
That is the dual-purpose blueprint: bank serum prospectively in every disease trial, run proteomics, apply open clocks, and let geroprotective candidates surface years before anyone tests them in healthy people. Forty-plus Insilico programs now work this way, and the company posted its first profitable half-year on $106 million in revenue, up 287 percent, which makes the business model and the science a single bet that aging biology is where the targets are.
How fast is 3-4 years in 12 weeks?
Annualized, 3.5 years in 12 weeks is roughly 15 years of biological age per treatment-year. That is illustrative only; nothing says the effect compounds linearly. For context, with the caveat that clocks, populations, and designs differ:
| Intervention | Clock type | Result | Implied rate |
|---|---|---|---|
| Rentosertib, 12 wk (n=42, IPF) | 6 proteomic clocks | 3-4 y reversed | ~13-17 y/yr |
| Semaglutide, 32 wk (UCSD) | 7 epigenetic clocks | 3.1 y reversed | ~5 y/yr |
| Semaglutide, HIV cohort | PhenoAge | 4.9 y/yr | 4.9 y/yr |
| Exercise, 12 wk (n=26 men) | ProtAge | 10 mo reversed | ~3.6 y/yr |
We priced semaglutide's effect in July at about $3,220 per biological year reversed. Rentosertib has no price and no longevity indication, so no per-year figure exists, which is the point: the drug is now in Phase III (GENESIS-IPF-3) for IPF, not headed to anyone's medicine cabinet.
The strongest case against
Here is the objection at full strength. These were sick people whose lungs improved, and IPF floods the body with inflammatory signaling that proteomic clocks read as old age, so calming the disease makes the clocks read younger without any aging being touched, which matches the mortality clocks barely correlating with chronological age at baseline (r of 0.16 to 0.23), as expected when disease burden drives the readout. On this reading, the paper measured good disease treatment and labeled it aging.
Its authors do not dodge this: "Full disentanglement of the two signals is not achievable within an IPF cohort," the paper states, "and requires validating the drug or its underlying mechanism in healthy volunteers." Levitt put it plainer: the healthy-volunteer experiment is the one he wants next.
Three findings push back, none decisive, starting with the dose divergence: were clocks only tracking lung improvement, the 60 mg once-daily arm, with the best FVC and 2.5 times the exposure, should have dominated the aging signal, and it did not. Second, the UK Biobank comparison showed reversal of normal age-related protein trajectories across 55,319 profiles, not just IPF signatures. Third, the mechanism reads senomorphic, with suppressed senescence drivers and dialed-down RTK-PI3K and RAS-ERK signaling. A skeptic can still note that senescence and fibrosis share pathways. Fair, and only the healthy-volunteer trial settles it.
Limitations
Forty-two patients, all Asian, mean age 67, all with IPF. This generalizes to approximately nobody yet.
The statistical bar is lenient at 21 significant comparisons under Q below 0.10, the signal lived mostly at week 4, and durability past 12 weeks is unproven, leaving open whether the plateau is a true ceiling or just a short observation window. Safety is not a footnote: treatment-related adverse events reached 61 percent twice-daily and 78 percent at 60 mg once-daily versus 29 percent for placebo, with liver-injury discontinuations in 4 of 18 and 3 of 18, ALT elevation in a third of the high-dose arm, and four of seven liver withdrawals also taking nintedanib, which complicates attribution without erasing the signal.
The FVC cross-check borrows healthy-population decline rates for IPF patients and carries a wide interval. Twelve weeks is a snapshot, and no clock has been validated as a surrogate for lifespan.
What to watch
First, the Phase III IPF trial: Insilico dosed the first patient in September in GENESIS-IPF-3, a 52-week trial of 320 patients across 47 centers in China with annual FVC decline as the primary endpoint. A replicated lung benefit validates the drug regardless of the aging story, and a bigger trial means a bigger proteomic dataset for the clocks.
Second, the healthy-volunteer experiment: until TNIK inhibition moves biological age in people without IPF, the aging claim stays provisional.
Third, the blueprint: any sponsor running a disease trial today can bank serum prospectively, run proteomics, and apply the open-source clock pipeline for a fraction of a percent of the trial budget, which means the next aging signal is probably already sitting in someone's freezer. For readers: no longevity indication, no price, and a liver signal, so follow the story and do not chase the prescription.
The Bottom Line
An AI-designed drug moved six independent aging clocks the same direction in a randomized, placebo-controlled trial, and the effect, 3 to 4 years in 12 weeks, survived a permutation test at 140 times the chance expectation.
The aging dose was not the lung dose, which is both a pharmacological clue and the strongest hint that something beyond disease treatment is happening. Honest accounting: 42 sick patients, a lenient threshold, a 12-week plateau, no healthy-volunteer data. But the method outlives this molecule. Every disease trial already collects the blood, the clocks are open source, and the data are public. Longevity drug discovery just got a cheap new front door.