An AI-designed drug was tested for a serious lung disease. When researchers went back and looked at the patients' blood, six different biological aging tests all pointed the same direction: younger.

Here's the full story, from a new paper in Nature Biotechnology.

Rentosertib (formerly INS018_055) inhibits an enzyme called TNIK. What makes it notable is how it came to exist: the paper describes it as "an artificial intelligence (AI)-designed TRAF2- and NCK-interacting kinase (TNIK) inhibitor." AI was used both to identify the target and to design the molecule. TNIK was originally flagged as a gene involved in six recognized hallmarks of aging, and idiopathic pulmonary fibrosis, a progressive scarring of the lungs, was chosen as the first indication precisely because it overlaps so heavily with aging biology.

The underlying study was a randomized, double-blind, placebo-controlled Phase 2a trial (NCT05938920), conducted across sites in China between July 2023 and June 2024 in adults over 40 with confirmed IPF. Of 128 patients screened, 71 were randomized into four groups: 30 mg once daily, 30 mg twice daily, 60 mg once daily, or placebo. Blood was drawn at baseline and at weeks 2, 4 and 12. Forty-three participants consented to serum proteomic screening; one was excluded for a missing measurement. That left 42 people, mean age 67.1, whose stored samples became this analysis.

Aging clocks estimate "biological age" from molecular data. The best-known ones read DNA methylation, but the authors argue those have produced "inconsistent, hard-to-interpret results" in trials. This study instead used proteomic clocks, which read protein levels directly, on the grounds that proteins are the immediate effectors of biological change.

They applied six independently developed models: ProtAge, OrganAge in both chronological and mortality-trained variants, PAC, ipfP3GPT and PAOPAC. The four chronological clocks tracked actual age well (Spearman's r of 0.70 to 0.84, error under 4 years after correction). The two mortality-trained clocks correlated far more weakly with calendar age, as expected given what they were built to predict and how sick this cohort was.

The core result, in the authors' words: "We measure the variance between the clocks and find that all six clocks consistently predicted lower biological age in treated arms." The placebo group showed minimal change or slight increases over the 12 weeks.

The statistics are worth stating precisely. Each arm generated 54 comparisons (6 clocks × 3 timepoints × 3 regimens). Twenty-one reached significance, clustered at week 4, where 11 of 18 comparisons showed significantly lower biological age in treated patients. A permutation test put the expected-by-chance number at 0.15. In the 60 mg once-daily arm, all four chronological clocks registered reductions of "−2.71 to −3.46 years" at week 4. The 30 mg twice-daily regimen produced the broadest agreement across clock types. Body mass index did not explain the effect.

Organ-specific versions of one clock went further: the artery clock showed lower predicted age across all treated arms at all timepoints, ranging from −6.95 to −16.57 years, with stomach, brain, pancreas and immune clocks also registering reductions in some arms. These are mortality-trained organ models, a different instrument from the six headline clocks, and should be read as a separate exploratory signal.

Of the 2,841 proteins measured, 326 changed significantly under treatment, compared with just 2 under placebo. Among the most strongly downregulated were drivers of fibrosis and tissue scarring: COL1A1, MMP10 and FAP. Senescence markers including EREG, IGFBP4, MMP10, MMP13 and SPP1 fell consistently across treated groups. The SenMayo senescence signature rose sharply in placebo patients, consistent with senescence driving IPF progression, and dropped substantially in every treatment arm.

This is where the paper is more disciplined than most coverage of it will be. The authors describe the results as "exploratory findings." They write plainly that "proteomic clocks alone cannot deconvolute aging- and disease-specific effects," meaning they cannot currently tell whether patients' biological age readings improved because the drug slowed aging or simply because it treated their lung disease. Settling that, they say, "is not achievable within an IPF cohort and requires validating the drug or its underlying mechanism in healthy volunteers."

Other limits: the effect peaked at week 4 and then plateaued rather than deepening. Establishing causality from clock outputs "remains a challenge." The protein panel covers under 3,000 proteins, and TNIK, the drug's own target, isn't on it, so target engagement could only be inferred indirectly. And several authors are affiliated with Insilico Medicine, the company developing the drug: the paper states, "We are developing rentosertib."

The headline is the aging result. The durable contribution is the trial design. Proving a drug extends human lifespan through conventional endpoints would mean waiting decades to count who lives longer, which is why almost no one attempts it. The authors propose something more practical, in their words: "This work supports the goal of dual-purpose clinical trial designs that integrate aging endpoints into studies for specific disease indications."

In plain terms: run the trial you were already running, for the disease you were already targeting, and measure aging biomarkers alongside. Candidates worth a closer look would surface years earlier, without asking anyone to wait a generation for an answer. Whether the clocks are trustworthy enough to carry that weight is exactly the open question this paper is trying to move forward.

Source: Nature Biotechnology: https://www.nature.com/articles/s41587-026-03286-y

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter