Insilico Medicine has a real investigational pill in a late-stage lung-disease trial. The same programme also published a paper in which six blood-protein clocks scored treated patients younger than placebo. Those two facts travel together in headlines. They are not the same claim.
This page splits them. Cipher Projects is an Australian software studio. We build production agents, private data layers, and evidence trails. We do not make drugs, and this is not medical advice.
Quick answer
Rentosertib is in Phase III for idiopathic pulmonary fibrosis (IPF). It is investigational and not approved. In a 42-person proteomic slice of the earlier Phase IIa study, six independently trained ageing clocks moved younger versus placebo, with the peak around week 4 at 30 mg twice daily. That is a scored biomarker signal. It is not proof of longer life, and it is not longevity escape velocity.
Best for: founders and operators who need a plain split between the IPF registry trial and the longevity marketing. Honest limit: Cipher is not a biotech. We read the public papers and the trial listing. We did not run the clocks.
Last updated: 12 September 2026.
Is there a real AI-designed longevity drug in trials, or is it marketing?
There is a real drug candidate in a real IPF trial. The longevity story is a second layer built on clocks. Insilico identified the target with PandaOmics and designed the molecule with Chemistry42. The company calls rentosertib (also ISM001-055 / INS018_055) a potentially first-in-class oral TNIK inhibitor. That “first AI-designed drug this far” line is Insilico’s claim. Other programmes have used AI on known targets. The distinctive public package here is a novel target plus a generated molecule plus a Phase III listing.
The trial indication is IPF, a progressive scarring disease of the lung. Ageing is not the labelled endpoint. The US Food and Drug Administration has not treated ageing itself as an indication you can code and sell against. That gap is why dual-purpose language shows up: run a disease trial, then argue the same molecule also moved ageing biomarkers.
| Layer | What you can check | What it is not |
|---|---|---|
| IPF Phase III | NCT07687459: 320 patients, 47 centres in China, annual FVC decline over 52 weeks | A registered longevity trial |
| Clock paper | Six proteomic models on 42 Phase IIa patients and 2,841 proteins | Proof of extra years of life |
| Marketing | Dual-purpose / geroprotective language around those clocks | An approved longevity drug |
Use the table the way we use a quote sheet in how we price production agents: name each line so a blob cannot hide inside another blob.
What is rentosertib?
Rentosertib is an oral small-molecule TNIK inhibitor that Insilico is developing for IPF. TNIK is a serine/threonine kinase in fibrosis and inflammation pathways. Insilico says PandaOmics ranked it from multi-omics and ageing-hallmark scoring, then Chemistry42 generated and optimised the compound. Discovery through early clinic is in Nature Biotechnology (Ren et al., 2025). Randomised Phase IIa results are in Nature Medicine (Xu et al., 2025).
The Phase IIa study (NCT05938920) enrolled 71 people with IPF across 22 sites in China for 12 weeks. Arms were placebo, 30 mg once daily, 30 mg twice daily, and 60 mg once daily. The primary job was safety. Treatment-emergent adverse event rates were similar across arms. In the 60 mg once-daily arm, mean forced vital capacity (FVC) changed by +98.4 mL at 12 weeks, against −20.3 mL on placebo. FVC is a lung-volume measure. It is not a lifespan measure. The FDA granted Orphan Drug Designation for IPF in February 2023. None of that makes the pill approved.
What did the Phase III IPF trial actually start?
Insilico initiated a prospective, randomised, double-blind, placebo-controlled Phase III study of once-daily rentosertib in IPF in July 2026. ClinicalTrials.gov lists it as NCT07687459 (Insilico also cites CTR20262475 and later calls the programme GENESIS-IPF-3). Planned enrolment is 320 adults at 47 centres in China. The primary endpoint is the annual rate of FVC decline over 52 weeks. A key secondary endpoint is time to a first disease-progression event. The drug remains investigational. No regulator has approved it.
The registry record last updated on 7 July 2026 still said “not yet recruiting,” with an estimated start of 30 August 2026 and completion in October 2029. On 10 September 2026 Insilico posted that the first GENESIS-IPF-3 patient had been dosed at Peking Union Medical College Hospital. Treat the registry as the listing and the company note as a company note. If you need current site status, refresh NCT07687459 rather than a press release.
What did the six ageing clocks actually show?
They scored treated patients younger than placebo on all six models, in a 42-person proteomic subset of the Phase IIa trial. That subset is not the full 71-person safety cohort. Researchers ran 2,841 serum proteins through six independently trained clocks: ProtAge, OrganAge chronological, OrganAge mortality, PAC, ipfP3GPT, and PAOPAC. Insilico says those models came from groups including Harvard, Oxford, Peking University, and Insilico, on different features and training targets, and that all six pointed the same way versus placebo.
The peak sat at week 4 in the 30 mg twice-daily arm: about 3 to 4 years of predicted biological-age reversal on most clocks, and about 6 years on one clock. Those numbers are model outputs in years of predicted age. They are not calendar years added to anyone’s life. The dose with the strongest clock move (30 mg twice daily) was not the dose with the largest FVC move (60 mg once daily). The authors treat that split as a reason not to collapse the clock signal into “the lung got better, so the blood looked younger.”
The Nature Biotechnology paper (Zhavoronkov et al., 2026) also says the quiet part: proteomic clocks cannot, on their own, separate slower ageing from a treated disease. Pathway work suggested senomorphic and metabolic shifts beside anti-fibrotic activity. That is mechanism language. It is still not a survival curve.
Do the clocks prove people will live longer?
No. A proteomic ageing clock is a regression from blood proteins onto chronological age or mortality risk, trained on other cohorts. When the score drops, the model says the protein pattern looks like a younger or lower-risk training example. That is useful as an exploratory endpoint. It is not a substitute for years of follow-up, hard clinical events, or a regulator-qualified ageing indication.
“Longevity escape velocity” is a metaphor people paste onto any downward blip. Do not paste it onto this paper. The study is 12 weeks, 42 people with IPF, and six models that the authors themselves refuse to read as deconvoluted ageing. If someone sells you a supplement, a clinic package, or a “LEV is here” deck off this paper, they are selling. Cipher will not sell that.
How do clocks compare to production-agent evals?
A clock is a scored task. So is an agent eval. You do not “have AGI” because a model talks well, and you do not “have a longevity drug” because a clock moved. You have a number on a named test, collected under a protocol, with a stated failure mode. Production agents need the same discipline: a specified job, a score, a halt when the score is the wrong kind of win.
The transferable pattern for a health-adjacent founder is three named lines, not a vibe:
- Target specification: IPF and FVC, or a named workflow and a write that must halt.
- Clocks / evals: a model you can rerun, with the honest limit written next to the score.
- Regulatory or operational path: a trial you can list, or an evidence trail an auditor can read.
We already price that way on agent jobs: stamp, per agent, per connector. A blob quote hides the clock. A line-item quote makes the clock visible. If your “AI health” product has no scored task and no path, you are in the marketing column of the table above.
How is this like AlphaFold or AlphaGenome?
Both stories bulk-solve a finite list, then let operators work from a table. AlphaFold enumerated protein structures and published a database. DeepMind’s AlphaGenome Atlas precomputes molecular-effect scores for billions of single-letter DNA changes. Rentosertib’s clock paper does a smaller version of the same move: take a closed proteomic panel, run several independent scorers, and publish the disagreement as well as the agreement.
The Millennium Prize week is the same lesson at another scale. Ten thousand agents can grind a named mathematical target. They cannot invent the target. “Slow FVC decline over 52 weeks in 320 people” is a target. “Make ageing go away” is not. If you want the company analogue, write the question the way Clay wrote the prize, then attach an eval. Do not wait for a cleverer model to guess what you meant.
What should a founder do with this story?
Steal the method and leave the molecule. If your domain can be listed (mutations, evals, failure modes, connector states), someone will precompute it or score it. Your job is to own the proprietary outcomes and the evidence layer, or to plug into the public table fast. Cipher builds that software and data layer for Australian and Singapore teams. We do not synthesise compounds, run clinics, or sell longevity products.
When a client asks whether “the AI longevity drug” changes their roadmap, the useful reply is the split: a late-stage IPF programme is a clinical fact to watch; a 42-person clock signal is an eval result; a pitch that collapses those into immortality is marketing. Put the terms on the cluster glossary before they enter a board pack.
FAQ
Is there a real AI-designed longevity drug in trials, or is it marketing? A real investigational TNIK inhibitor is in Phase III for IPF. The longevity reading is marketing laid on top of six clocks from 42 people. Hold both sentences. Do not fuse them.
Is this marketing? Part of it is. The NCT07687459 listing, the Phase IIa Nature Medicine paper, and the clock paper are checkable. Headlines that call rentosertib “the first longevity drug” skip the indication, the sample, and the authors’ own limit on what clocks can prove.
What did the ageing clocks actually show? All six proteomic clocks trended younger versus placebo in the 42-person subset. The peak was about 3 to 4 predicted years at week 4 on 30 mg twice daily, and about 6 years on one clock. Predicted years are not lived years.
Is rentosertib approved? No. It is investigational. Phase III is underway as a listed study. Approval, if it ever comes, will be for a disease label, not for ageing as such.
Does this mean longevity escape velocity is here? No. A 12-week biomarker move in people with IPF does not establish that ageing has been outrun. Treat LEV talk as a metaphor someone failed to keep in the metaphor box.
Should I take it, or buy a clinic package built on this paper? This page cannot advise you on treatment. Rentosertib is not an approved medicine. Cipher does not sell health products. Talk to a clinician who can see your record, not a software studio.
Why is a software company writing this? Clients ask. The transferable lesson is target plus clock plus path: the same cut we use on production agents. If you need that layer built, applied AI engineering is the offer. The molecule is not.
Sources
- Insilico, 7 July 2026: Phase III initiation for rentosertib in IPF; 320 patients, 47 centres, FVC over 52 weeks; investigational, not approved.
- ClinicalTrials.gov NCT07687459: registry record. Last update posted 7 July 2026 still listed the study as not yet recruiting.
- Insilico, 10 September 2026: company announcement that a first GENESIS-IPF-3 patient was dosed. Not a registry update.
- Ren et al., Nature Biotechnology 43, 63–75 (2025): discovery-to-clinic TNIK inhibitor paper. DOI 10.1038/s41587-024-02143-0.
- Xu et al., Nature Medicine 31, 2602–2610 (2025): randomised Phase IIa in 71 people with IPF. DOI 10.1038/s41591-025-03743-2.
- Insilico, 7 September 2026: company summary of the six-clock analysis (42 patients, 2,841 proteins; peak ~3–4 years at week 4, 30 mg BID).
- Zhavoronkov et al., Nature Biotechnology (2026): “Integration of proteomic aging clocks in a phase 2a clinical trial supports simultaneous geroprotective assessment.” DOI 10.1038/s41587-026-03286-y. The paper states that clocks alone cannot deconvolute ageing from disease.
We could not fetch a New York Times URL that held up, so we did not cite one.
Related: Buzzword soup glossary · AlphaGenome Atlas · AI claimed a Millennium Prize · How we price production agents