--- title: "An AI-Designed Drug Reached Phase III. Read the Endpoint, Not the Model" date: 2026-10-22 category: China AI Decoded site: NeuroAI canonical: https://neuroai.site/a/na-decode-ai-drug-phase-three language: en --- # An AI-Designed Drug Reached Phase III. Read the Endpoint, Not the Model > Rentosertib entered a 320-patient Phase III trial in China in September 2026. The mid-stage study that justified it tested safety, not efficacy. BEIJING — On 10 September, a patient at Peking Union Medical College Hospital received the first dose in a Phase III trial of a small molecule designed by software, aimed at a target that software also chose. The setting was ordinary: a respiratory ward, a consent form. That is the news. A decade of claims about artificial intelligence in drug discovery has produced mostly slides; this programme has produced a patient and a 52-week endpoint. The model is no longer the interesting part. The endpoint is. ## Key takeaways - Rentosertib's Phase III trial plans 320 patients at 47 centres in China over 52 weeks, with the annual rate of lung-function decline as the primary endpoint (company disclosure, 10 September 2026). - The Phase IIa study that justified it had safety as its primary endpoint; the lung-function gain was a secondary finding in 71 patients over 12 weeks (Nature Medicine, 3 June 2025). - Insilico Medicine has nominated 35 preclinical candidates since 2021 and holds 13 trial clearances; one asset is in Phase III (company disclosures, October 2026). - ByteDance's spun-out Anew Labs raised US$290 million at a US$1.5 billion post-money valuation in September 2026 (Jiemian News, 12 October 2026). - The next read-out is first-in-human safety data for a second AI-designed molecule, ISM6331, at ESMO in Madrid on 25 October 2026. ## What is actually being tested The trial is registered as NCT07687459 in the US database and CTR20262475 in China's. It is prospective, randomised, double-blind, placebo-controlled, and runs for 52 weeks. Xu Zuojun of Peking Union Medical College Hospital is the lead investigator, with Zhong Nanshan and Chen Chang as co-leads. Three design details matter: - **The primary endpoint** is the annual rate of decline in forced vital capacity (FVC, 用力肺活量), the volume a patient can force out of the lungs after a full breath. - **The key secondary endpoint** is the time to a first disease-progression event. - **The population** is idiopathic pulmonary fibrosis (特发性肺纤维化), a progressive scarring disease with median survival after diagnosis commonly reported at two to four years; approved antifibrotics slow it and do not reverse it. FVC is the standard measure of progression in this disease, which is why the sponsor chose it. A drug that flattens the slope has demonstrated something a drug that improves a twelve-week reading has not. ## The study that got it here measured safety Nearly every claim about this programme traces to one paper: the Phase IIa results in Nature Medicine, 3 June 2025. Read the design first. Seventy-one patients, four arms, twelve weeks. The primary endpoint was safety and tolerability, and it was met. The lung-function result — a mean FVC gain of 98.4 mL at the 60 mg once-daily dose, against a mean decline of 20.3 mL on placebo — was a secondary, exploratory finding in a small sample. Insilico's own description is careful: a dose-dependent trend. That is the correct phrase. A 71-patient study of twelve weeks can generate a hypothesis and can justify a Phase III; it cannot establish efficacy. Coverage of this field often collapses those two sentences into one. The novelty is specific. Both the target — TNIK, a kinase the investigators describe as not previously linked to fibrosis — and the molecule came out of the company's platform, and Insilico says the preclinical candidate was nominated in about 18 months after 78 molecules were synthesised. ## The pipeline arithmetic The company's disclosures say more than any single drug. Since 2021 Insilico has nominated 35 preclinical candidates, 13 of which have received or passed clinical trial applications; 11 were nominated in the first ten months of 2026 alone, and eight clinical milestones were reached by mid-August. One of those 35 is in Phase III. That ratio is the story: discovery has become cheap enough to industrialise, and clinical validation has not. The commercial figures need the same discipline. Insilico reported roughly US$7.3 billion of partnership value disclosed in 2026 through mid-August, taking the cumulative total since 2021 to near US$11 billion. These are "up to" figures — milestone schedules contingent on events that have not occurred — not revenue and not cash. Anew Labs, the drug-discovery team ByteDance spun out in June 2026, closed a US$290 million first round in September at a US$1.5 billion post-money valuation, with ByteDance retaining 56% (Jiemian News, 12 October 2026). ## Why the trial runs in China The 47 Phase III centres are all in China, per the company's announcement. That is an operational choice with a cost. A single-country design is simpler to run, but it gives regulators in Washington and Brussels a narrower basis for extrapolation, because background therapy and patient mix differ from the populations those agencies review. China's trial infrastructure is also why a company this size can run such a study at all. ## The next data point On 25 October at the European Society for Medical Oncology congress in Madrid, investigators present abstract 997: first-in-human data for ISM6331, an oral pan-TEAD inhibitor designed with the same chemistry platform, in mesothelioma and other advanced solid tumours. It is a Phase I: dose, pharmacokinetics, safety and at best an early response signal. It cannot show that the platform works. Read it as a manufacturing check on a second molecule, not as confirmation of the first. ## Honest limitations Pipeline counts and partnership values are company disclosures and are not audited; "up to" totals should never be read as booked revenue. The Phase IIa evidence is small, short and drawn from one country, and its lung-function effect was a secondary endpoint, so the Phase III can still fail. The "world's first" descriptions are Insilico's own characterisation, not a regulatory finding. This column found no independent replication of the Phase IIa result. The ESMO abstract had not been presented when this was written. ## What readers can do now - **Analysts:** check NCT07687459 quarterly against the 320-patient target. Enrolment pace is the earliest public signal on whether the 2029 read-out holds. - **Licensing teams:** when a deal value is quoted, ask for the split between upfront payment, near-term milestones and the tail. The headline total is a marketing number until the schedule is disclosed. - **Newsrooms:** establish which part of a programme was machine-derived — target, molecule, or only optimisation. Here it is the first two, which is still rare, and saying so is more useful than "AI drug". --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-decode-ai-drug-phase-three Free to quote with attribution and a link to the original.