--- title: "An AI-designed lung drug just passed a human trial — and the hard part is next" date: 2026-10-01 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-app-insilico-ai-drug-rentosertib language: en --- # An AI-designed lung drug just passed a human trial — and the hard part is next > Insilico Medicine's (英矽智能) Rentosertib, a TNIK inhibitor whose target and molecule were both generated by AI, met its safety endpoint and showed a lung-function signal in a peer-reviewed Phase IIa trial published in Nature Medicine. A drug candidate usually takes four to six years just to reach the clinic. This one took eighteen months from a blank sheet to a preclinical nominee — and then it went into real patients. Whether that speed means anything for the drugs of the future still hinges on a much larger, much harder test. In June 2025, the peer-reviewed journal *Nature Medicine* published results that the field had been waiting for: the first Phase IIa trial of a therapy whose disease target and chemical structure were both produced by generative AI. The drug is Rentosertib (ISM001-055), developed by Insilico Medicine (英矽智能), a clinical-stage biotech with operations in Hong Kong and Shanghai. ## The disease, and why it matters Idiopathic pulmonary fibrosis (IPF) is a progressive, ultimately fatal scarring of the lungs. Median survival after diagnosis is commonly reported at two to four years. Only two drugs are approved, and both slow decline rather than reverse it. A treatment that could halt or improve lung function would be a genuine advance. ## What "AI-designed" actually means here Insilico did not start from a known target and ask AI to tweak a molecule. Its platform worked in two stages: - **PandaOmics**, the company's biology engine, sifted multi-omics, literature and network data to rank **TNIK** — Traf2- and NCK-interacting kinase — as a novel fibrosis target that had been relatively underexplored for IPF. - **Chemistry42**, its generative-chemistry engine, then designed a small-molecule inhibitor against that target. The *Nature Medicine* paper describes this as the first reported instance of an AI platform discovering both a disease-associated target and a compound for it. From project start to a preclinical candidate took about 18 months; first-in-human testing followed in under 30 months. ## What the trial showed The GENESIS-IPF study was a randomized, double-blind, placebo-controlled Phase IIa trial conducted at **21 sites in China**, enrolling **71 patients** with IPF across three dose arms and a placebo: - 30 mg once daily (n=18) - 30 mg twice daily (n=18) - 60 mg once daily (n=18) - placebo (n=17) The **primary endpoint — safety and tolerability — was met**, with treatment-emergent side effects mostly mild to moderate and serious events rare. On the exploratory efficacy measure, forced vital capacity (FVC, a standard lung-function gauge), the highest dose (60 mg daily) showed a mean change of **+98.4 mL** over 12 weeks, versus **−20.3 mL** on placebo. The authors note those are 95% confidence intervals that cross zero for the placebo arm, and the readout was exploratory, not the trial's main goal. ## The part that is genuinely new Plenty of AI has *optimized* existing drug programs. What makes Rentosertib a bellwether is that AI nominated the target itself — the slowest, most expensive step in drug discovery — and designed the molecule. If that holds up at scale, the bottleneck shifts from "can we find a target" to "can we prove it in patients," where the old rules still apply. Encouragingly, the company has since initiated a **Phase III trial** designed to enroll 320 patients over 52 weeks, led from Peking Union Medical College Hospital with co-investigators including respiratory-medicine leaders — a far more decisive test than the 71-patient Phase IIa. ## Why skeptics are right to pump the brakes Phase IIa is early and small. The lung-function gain was exploratory, the study ran only 12 weeks, and every patient was in China. Many drugs with positive mid-stage signals fail in larger trials. And "AI-designed" should not be read as "AI alone" — medicinal chemists and biologists were involved throughout. The fair claim is that AI compressed the discovery timeline; it did not remove the clinical and regulatory gauntlet that every drug must clear. ## Honest limitations - The headline figures (+98.4 mL vs −20.3 mL FVC) are exploratory secondary outcomes from a 71-patient, 12-week trial; they are not proof of efficacy. - We relied on the peer-reviewed *Nature Medicine* paper (vol. 31, pp. 2602–2610, 2025) and the company's own case materials; we did not re-analyze the raw dataset. - Site count is reported as 21 (company materials) versus 22 in some secondary summaries; the discrepancy does not change the trial's small size. - The Phase III initiation is reported by the company; its results are years away, and a positive readout is not guaranteed. - "First AI-designed drug" is the company's framing; the precise superlative depends on definitions of "AI-discovered" versus "AI-optimized." ## What readers can do now - **Read the primary paper, not the press release** — the *Nature Medicine* article (s41591-025-03743-2) lays out the confidence intervals that the headlines omit. - **Track the Phase III endpoint** — a 52-week, 320-patient FVC result is the number that will actually validate or bury the AI-discovery thesis. - **Separate discovery from approval** — when a company says "AI-designed," ask what was de-risked (target + molecule) versus what remains (safety at scale, efficacy, regulatory approval). --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-insilico-ai-drug-rentosertib Free to quote with attribution and a link to the original.