Six proteomic ageing clocks built by six independent teams all measured a younger biological age in patients treated with rentosertib, a molecule whose target and structure were both found by AI systems. The reversal reaches 3 to 4 years at week four, and up to 6 years on some clocks. The sample is 42 patients, every one of them living with a severe lung disease.
Key Takeaways
- Rentosertib inhibits TNIK, a target picked by an AI platform then drawn by generative chemistry
- Six independent proteomic clocks converge on a measured biological rejuvenation
- The trial covered 42 idiopathic pulmonary fibrosis patients, not healthy volunteers
Have an AI Sum Up This Article
ChatGPTA molecule whose target the machine picked first
Rentosertib, filed as ISM001-055, inhibits TNIK, a protein involved both in fibrosis and in several mechanisms of ageing. The compound is described as potentially first in class, meaning nothing else on that target has reached the market.
What sets this candidate apart is the production chain behind it. The biological target came out of Insilico Medicine’s discovery platform, and the chemical structure was generated by Chemistry42, its generative chemistry engine. Both steps upstream of bench work were handed to learning systems.
That detail matters because it breaks with most announcements in the field. AI usually sorts existing molecules or predicts a protein fold, backing a hypothesis a human already had. Here the hypothesis itself came from the machine, and that hypothesis is what is now being tested in patients.
The same movement runs through the general-purpose labs on the scientific side. It looked much the same when Claude designed proteins that then worked in the lab, with the same generate-then-validate loop rather than a screening exercise.
What the six clocks are actually measuring
The data comes from a randomised, double-blind, placebo-controlled phase IIa trial run across Chinese sites in 2023 and 2024, in patients with idiopathic pulmonary fibrosis. Forty-two participants consented to longitudinal proteomic profiling, with a mean age of 67.1 years.
The protocol measured 2,841 serum proteins at baseline, then at weeks two, four and twelve. Those profiles were run through six separate proteomic ageing clocks, each developed independently by a different research group: ProtAge, OrganAge variants, PAC, ipfP3GPT and PAOPAC.
A proteomic clock estimates biological age from the signature of circulating proteins, independently of a birth date. The value of the exercise lies in convergence: six models built on different data and different methods returned the same directional verdict on the same cohort.
The measured reversal reaches 3 to 4 years at week four on the arm dosed at 30 mg twice daily, and climbs to 6 years on some clocks. Michael Levitt, 2013 Nobel laureate in chemistry, summed the finding up in one line: six clocks from six independent groups, applied to the same 42 patients, all reported a younger biological age in the treated arms.
The clinical endpoint moved as well. Forced vital capacity improved by a mean of 98.4 mL on the 60 mg once-daily arm, against a 20.3 mL decline on placebo. In a disease defined by the steady loss of lung function, that is the number that decides anything.
More articles on Horizon
- Mistral Raises 3 Billion Euros in Record Funding
- Artificial Analysis Rebuilds Its AI Ranking
- Astra Productivity Moved OpenAI Plans Up Six Months
Why 42 patients settle nothing yet
The authors state the central limitation themselves: the study cannot separate slowed ageing from a simple improvement in the treated lung disease. A patient whose lungs recover shows a shifting protein profile, and nothing in this design attributes that shift to ageing itself.
Scale adds its own caveats. Forty-two patients make an exploratory cohort, mean age sits above 67, and every participant carries a severe pathology. Nothing in this protocol says what the compound would do in a healthy population, and the authors make no such claim.
The programme itself stays indexed on the lung disease rather than on longevity. Rentosertib has moved into phase III in China on idiopathic pulmonary fibrosis, and that indication is what will decide its regulatory fate.
For competing labs, the signal sits somewhere other than the six-year figure. It sits in the fact that a machine-picked target survived all the way into phase III, which remains the only real referee for AI-assisted discovery. Campaigns that die in preclinical work prove nothing either way.
That stepwise validation is exactly what other players are chasing on neglected ground, as when Anthropic took on diseases big pharma skips alongside Novartis. These programmes share a bet on areas where the human hypothesis is missing, for lack of commercial pull or accessible data.
For teams tracking medical uses of these models, the lesson is methodological. A result that holds across six clocks beats a single spectacular score, and the same reflex applies elsewhere: we saw it when GPT-5.5 Instant beat doctors on health answers, where the robustness of the evaluation protocol mattered more than the headline gap.
Follow the story on Horizon.
