Insilico's AI-designed drug enters Phase III — a first for the field
A drug whose target was found by AI and whose molecule was designed by AI is now in a late-stage trial — and the registry, not the press release, is what makes it a milestone. The same day brought a machine-checked maths result from a team of agents, and a brain-implant round with an unusual lead investor.
Insilico Medicine has begun dosing patients in a 320-person Phase III trial of rentosertib, an oral treatment for idiopathic pulmonary fibrosis — the first drug with both an AI-identified target and an AI-designed molecule to reach that stage. The Hong Kong-listed company announced the first patient on September 10 at Peking Union Medical College Hospital, and ClinicalTrials.gov lists the study as recruiting with an actual start date of September 9. The design is conventional and that is the point: randomised, double-blind, placebo-controlled, quadruple-masked, once-daily for 52 weeks, with the annual rate of lung-function decline as the primary endpoint and completion estimated for October 2029.
The mechanism behind the claim matters more than the badge. Insilico's PandaOmics platform nominated TNIK as the target, and its generative Chemistry42 platform proposed the molecule; the trial itself is run by clinicians across sites in China, which is why the company phrases its claim as a target identified with AI and a structure designed with generative AI rather than an AI-run pipeline. The evidence base underneath is thin: a Phase IIa readout in 60 patients at 12 weeks showed a mean improvement of 98.4 mL in forced vital capacity in the 60 mg group, published in Nature Medicine last year. Twelve weeks and 60 patients is a signal, not a proof — which is exactly what a Phase III with a placebo arm and four years of follow-up is for.
The commercial stakes are the reason to watch rather than celebrate. Insilico listed in Hong Kong in 2025 and needs a clinical rather than computational result, and the sector it sits in has plenty of candidates and no late-stage wins: we looked at why the molecule generators keep producing compounds the wet lab cannot use — More molecules, same failures: the crack in AI drug design. A positive IPF readout would not settle whether AI designs better drugs, only that it can design one that works. Still, no AI-designed molecule has ever been asked that question this late before.
A team of ten Claude Opus 5.5 agents spent about 15 hours and 733 messages on a shared message board and produced C-HD, a new shortest-path algorithm with a complete Lean 4 proof of its running time. The run came from Vals AI, which asked the agents for a substantial theoretical improvement on exact single-source shortest paths with non-negative weights, then required a reproducible Lean build and two internal reviews before the result counted. The independent write-up is careful about what was and was not established: Lean's kernel accepted the proof and the permitted-axioms check passed, but peer review here means the agents reviewing each other, not outside referees, and novelty still depends on human literature review.
The result is narrower than the coverage suggests. C-HD beats Dijkstra's bound only inside a certified density range — roughly m ≤ n·(log₂n)^(3/4) — and the improvement along that profile is a polylogarithmic factor, (log n)^(1/12), which works out to about 1.78x at n = 2^1000 in leading terms. The author states plainly that no large-graph benchmark was run and the construction's constants are enormous, so this is not a practical speedup. It is also not the first improvement on Dijkstra: a 2025 paper, by humans, broke the sorting barrier for directed shortest paths. What is new is the shape of the work — agents that record dead ends, challenge each other's claims and accept only a verified artifact. We saw the same pattern in September, when a Claude model spent eleven days formalizing Fermat's Last Theorem — Claude formalized Fermat's Last Theorem in 11 days.
Precision Neuroscience raised $250 million in an oversubscribed Series D to push its brain-computer interface toward regulatory approval, with Pershing Square and the Ackman Oxman Institute co-leading — Bill Ackman's first investment in the sector, according to the New York Times. Total funding since 2021 reaches $430 million, per the company's own announcement, and the Times reports the round valued the New York company at just over $1 billion, with $42.5 million of the round coming from the two Ackman vehicles; the company has not confirmed the valuation itself. Follow-on investors include Duquesne Family Office, B Capital, ARK Invest, Mubadala Capital and Hitachi Ventures.
The technical story is a matter of how the implant goes in. Precision's Layer 7 array is a thin film placed on the brain's surface through a minimally invasive procedure rather than electrodes pushed into tissue, which the company says has supported more than 100 patient procedures across 18 institutions, with FDA clearance covering up to 30 days of recording and a partnership with Medtronic on the delivery hardware. Founder Benjamin Rapoport is a Neuralink co-founder, which makes the comparison unavoidable and slightly misleading: Neuralink is chasing chronic high-channel counts, Precision is arguing that a safer implant with fewer channels is enough for the clinical use cases that pay today.
What to watch: whether the agents that produced C-HD get a second run against a problem with an outside referee, and whether anyone benchmarks the algorithm rather than the proof.
If a drug is designed by one AI and tested by clinicians, is that an AI drug or a conventionally developed one with a computational head start? Tell us in the comments.
Sources: C-HD Lean proof package (GitHub)
