Three to Six Years Younger, By Every Clock: Insilico's AI-Designed Drug Shows Biological Age Reversal in Phase IIa

The longevity field has spent a decade arguing about how to measure whether a drug actually slows aging. On September 7, 2026, Insilico Medicine offered the most striking answer yet: ask six independent clocks, and see if they agree.

Published in Nature Biotechnology, the new study analyzed longitudinal blood proteomics from 42 patients in the Phase IIa trial of rentosertib, Insilico’s candidate for idiopathic pulmonary fibrosis (IPF). Six proteomic aging clocks — ProtAge, OrganAge (chronological and mortality variants), PAC, ipfP3GPT, and PAOPAC — developed independently by leading groups at Harvard, Oxford, Peking University, and Insilico all reached the same verdict: patients who received rentosertib got biologically younger. At peak effect, Week 4 in the 30 mg twice-daily arm, predicted biological age fell by roughly 3–4 years, and by up to 6 years on one clock.

Why this drug is different

Rentosertib occupies a unique position in the history of pharmaceutical AI. It is the first drug candidate where both halves of the equation came from machines: the target, TNIK, was discovered by Insilico’s PandaOmics platform and scored highly across six of the “hallmarks of aging,” making it a dual-purpose target relevant to both aging biology and fibrosis; the molecule itself was designed by the generative chemistry platform Chemistry42. The program went from target identification to preclinical candidate nomination in about 18 months — a timeline traditional discovery pipelines measure in decades — and the drug’s name honors Insilico co-CEO Dr. Ren.

The contrast with prior geroscience is the point. Almost every celebrated “anti-aging” intervention to date — rapamycin, metformin, senolytics — is a repurposed drug designed for something else. Rentosertib is a de-novo molecule aimed at a de-novo aging-relevant target, now in Phase III for IPF in China. As 2013 Nobel chemistry laureate Michael Levitt put it in commentary accompanying the announcement: “What convinces me is not the size of the effect but the agreement, because these models share neither their features nor their training data.”

What the study actually measured

The trial protocol had prospectively built in longitudinal serum proteomic screening — a decision that made this analysis possible. Using Olink technology, the researchers profiled 2,841 proteins per patient across the 12-week study and applied the six clocks to each sample.

Three details make the result more than a statistical curiosity:

The clocks disagree on everything except the direction. Their methodologies range from traditional machine learning (OrganAge, PAOPAC, PAC) to deep learning (ProtAge, ipfP3GPT); their training targets include chronological age and mortality risk; they were built by six separate groups. That methodological diversity, all pointing toward age reversal in the treated arms, is what persuaded the reviewers and commentators.

The aging effect appears partially independent of the lung effect. The dose that produced the strongest age-reversal signal (30 mg BID) was not the dose with the greatest improvement in lung function (60 mg once daily, where patients gained a mean +98.4 mL of forced vital capacity versus a −20.3 mL decline on placebo). If the proteomic changes were merely a shadow of recovering lungs, the two would be expected to track together.

The signal runs against 55,319 real aging trajectories. Comparing patient samples against UK Biobank profiles, the team showed rentosertib directly reversed typical age-related protein-expression trajectories. Mechanistically, the drug behaved as a senomorphic agent — suppressing key senescence drivers including EREG, ESM1, IGFBP4, ITGA2, MMP10, MMP13, and SPP1, and dialing down growth-factor signaling through the RTK–PI3K and RAS–ERK pathways.

The honest caveats

The paper’s own authors are blunt about limits, and so was Levitt: “This trial cannot yet separate slower aging from a treated lung, and the authors say so plainly. The experiment in healthy volunteers is the one I want to see next.” Forty-two patients is small; twelve weeks is short; IPF itself is an age-associated disease, so a drug that treats it well will inevitably nudge aging biomarkers. Whether rentosertib makes healthy people younger — not just sick people less old — remains untested.

There is also a circularity trap inherent to all proteomic clocks: they are trained to predict age or mortality from protein profiles, so any intervention that shifts those proteins shifts the prediction. The multi-clock agreement mitigates but does not eliminate this. The definitive test, a dedicated trial with aging endpoints in non-diseased populations, has not yet run.

A blueprint, not just a result

Perhaps the study’s most durable contribution is procedural. It demonstrates, for the first time, a standardized workflow for embedding geroscience endpoints into ordinary disease trials: collect aging biomarkers prospectively as exploratory endpoints, run multi-clock analysis in parallel, cross-validate, then pursue formal biomarker qualification under the FDA’s Biomarker Qualification Program and the FDA–NIH BEST framework. That path could surface true geroprotective drugs years or decades earlier than waiting for post-approval repurposing.

Insilico has put its money behind openness here. All data are deposited at the China National Center for Bioinformation (accession OMIX008341), and the analysis pipeline is published as an open-source Python library on GitHub.

The commercial backdrop

The science lands on a company that has quietly become the counter-example to AI-drug-discovery skepticism. Listed on the Hong Kong Stock Exchange since December 30, 2025 (03696.HK), Insilico reported roughly $106 million revenue in the first half of 2026 — up 287% year-on-year — with its first profitable half-year and adjusted net profit above $51 million. Total contract value of 2026 transactions has reached about $7.3 billion, pushing cumulative major-deal value since 2021 to roughly $11 billion, across partners including Eli Lilly, Servier, Takeda, and Qilu Pharmaceutical. On the R&D side, the company nominated nine development candidates in the first nine months of 2026, a company record.

Founder and co-CEO Alex Zhavoronkov, who presents the results at the Sorbonne on September 8 at the Nature conference “Redefining Healthcare in the Age of AI,” frames the stakes in characteristic terms: adding three healthy years to everyone’s lifespan would amount to roughly 25 billion life-years globally — “more lifetimes than humanity lost in all the wars ever fought.”

The bigger picture

For AI observers, the significance extends beyond one molecule. Rentosertib is the clearest demonstration yet that the “AI designs the target, AI designs the drug, aging biology guides the strategy” loop can produce something that survives contact with human clinical data — and now, with a consensus aging-biomarker readout. If Phase III confirms efficacy in IPF, and a healthy-volunteer trial confirms genuine geroprotective effect, the longevity industry’s central bottleneck shifts from “can AI discover drugs” to “can regulators accept aging as an indication.” This paper is the first serious down payment on answering both questions.