Six aging clocks agreed about an AI-designed drug, and the people who ran them say that is not enough

A clinical trial measures whether a drug helps with the illness it was given for. It does not normally ask whether the patient aged more slowly while taking it, because there is no agreed way to ask. A paper published in Nature Biotechnology on 7 September 2026 asked anyway, on blood already drawn for another purpose, and reports that six separate models all answered the same way.

The drug is rentosertib, formerly INS018_055. Its target, a kinase called TNIK, was picked out of the literature by software; the molecule was generated by software as well. It is developed by Insilico Medicine, whose first-listed address on this paper is in Abu Dhabi, with further offices in Shanghai and Cambridge, Massachusetts.

What was actually measured

The trial itself was a phase 2a study in idiopathic pulmonary fibrosis, run across 21 sites in China in 2023 and 2024. Of 71 patients randomised, 42 agreed to have their serum proteins read as well, at the start and at weeks 2, 4 and 12. That reading covered 2,841 proteins per sample.

Six published models were then applied to those profiles: ProtAge, two variants of OrganAge, PAC, ipfP3GPT and PAOPAC. Four are trained to guess a person's calendar age from their proteins, two to predict mortality risk. They disagree with each other considerably about individuals — the mortality models correlate with real age at only 0.16 to 0.23 — and that disagreement is the point. Six models built on different principles converging on one answer is a stronger reading than any one of them alone.

They converged. Every treated group came out younger than it started; the placebo group did not move, or drifted slightly older. At week 4, the four calendar-age clocks put the 60 mg once-daily group between 2.7 and 3.5 years below its own baseline. Across the whole design, 21 of the comparisons cleared the significance threshold where chance alone predicted 0.15 of them.

The finding the authors keep returning to

The dose that worked best on the lungs was not the dose that moved the clocks. 60 mg once daily produced the largest gain in lung capacity in the original trial report; 30 mg twice daily, the same total per day in two smaller pieces, produced the broader and more consistent aging signal. Lung-function change explained almost none of the variation in biological age — a median of 6% across the six models.

That dissociation is the strongest argument in the paper that the clocks are reading something other than a sick lung getting better. It is also, as the authors are careful to say, an argument and not a demonstration.

What it does not establish

Everyone in the cohort had the same serious lung disease, and fibrosis proteins carry a great deal of weight inside these models. The single most influential protein in the whole analysis, LTBP2, is a regulator of fibrosis, and it happens to be the only feature shared by all six clocks. Separating the drug's effect on the disease from any effect on aging cannot be done inside this cohort at all; the paper says so plainly and calls for studies in people who do not have the illness.

The declared interests belong in the same paragraph as the result. The corresponding author is the founder and chief executive of Insilico Medicine, which is developing the drug, and several co-authors are employees; the journal publishes this at the foot of the paper. The analysis is open access, the trial is registered as NCT05938920, and two of its reviewers are named.