In June 2025, a 71-patient trial out of 22 sites in China quietly became one of the most consequential data points in drug development this decade. The drug, rentosertib (INS018_055), is a TNIK inhibitor for idiopathic pulmonary fibrosis (IPF) discovered and designed by Insilico Medicine using generative AI models. In the Phase 2a GENESIS-IPF trial, patients on the 60 mg once-daily dose gained a mean 98.4 mL in forced vital capacity (FVC) over 12 weeks, while the placebo group declined 20.3 mL — a real, measurable slowing of lung function loss in a disease that usually only gets worse. The results were published in Nature Medicine and presented at the American Thoracic Society 2025 meeting, marking the first time an AI-discovered and AI-designed small molecule produced a positive, peer-reviewed proof-of-concept in a randomized human trial.
That is the surprising part: AI drug discovery has been promising "first-in-class in months, not years" for a decade, and for the first time there is a real, published, randomized dataset to check the claim against — not a press release.
What AI actually shortened
Rentosertib is instructive because Insilico has been explicit about which stage AI compressed. Their generative chemistry engines proposed and scored novel TNIK-binding scaffolds, and structure-prediction models (in the same family of technique popularized by AlphaFold, though Insilico uses its own Chemistry42/PandaOmics stack) helped prioritize which of those molecules were worth synthesizing. Insilico has stated the program went from target discovery to a Phase 1-ready candidate in roughly 18 months, versus an industry norm closer to 4–6 years for that stage. That is a real, mechanistically explainable acceleration: generative models search chemical space combinatorially and rank candidates computationally, cutting the number of physical synthesis-and-test cycles a medicinal chemistry team needs to run.
What AI did not shorten is the clinical trial itself. GENESIS-IPF still took 12 weeks of dosing per patient, still required 71 human volunteers, and still went through the same regulatory and ethics review any IPF trial would. Isomorphic Labs — the DeepMind spinout built on AlphaFold — illustrates the same limit from the other direction: CEO Demis Hassabis has publicly pushed back the company's own first Phase 1 filing from a 2025 target to end-of-2026, because better structure prediction does not remove the need to run an actual trial. Designing a better molecule faster does not make a human liver metabolize it faster, a tumor respond faster, or a regulator review faster.
The failures nobody puts in the pitch deck
The same window that produced rentosertib's win also produced a string of AI-discovered drugs that failed in the clinic. Recursion Pharmaceuticals took REC-994, an AI-informed candidate for cerebral cavernous malformation, through a Phase 2 trial that met its safety endpoint but showed no meaningful clinical benefit on the disease itself. Its NF2 candidate REC-2282 fared similarly in the POPLAR-NF2 study — only the lowest dose cohort cleared a futility threshold. In Q1 2025, Recursion discontinued both programs, along with REC-3964, citing the "totality" of the data. Exscientia, whose DSP-1181 (for OCD) was one of the earliest "AI-designed molecule enters the clinic" headlines back in 2020, completed Phase 1 safely but was quietly shelved before ever reaching Phase 2; its EXS-21546 program was discontinued in 2023. Exscientia was acquired by Recursion in November 2024 as both companies restructured around a smaller, more selective pipeline.
These are not edge cases — they are the base rate. Reaching Phase 1 with an AI-designed molecule has gotten easier and cheaper; getting that molecule to show efficacy in Phase 2, where most drugs of any origin fail, has not changed. AI accelerates the funnel's front door, not its historically brutal middle.
Why this matters
Update, July 2026: Rentosertib has since cleared the bar this piece describes as the industry's hardest test. In July 2026, Insilico Medicine initiated a Phase III trial — a randomized, double-blind, placebo-controlled study enrolling 320 IPF patients across 47 centers in China — making it the first AI-discovered and AI-designed small molecule to reach pivotal late-stage testing. Isomorphic Labs, by contrast, has not yet entered the clinic: Demis Hassabis reaffirmed at Davos in January 2026 that the company's first trial remains targeted for "end of 2026," even as Isomorphic closed a $2.1 billion Series B round in May 2026 to fund the effort. The pattern holds: AI is compressing the path to a trial, not the trial itself.
The honest read of 2026's data is that AI drug discovery has crossed from marketing claim to falsifiable science: there is now at least one randomized, peer-reviewed trial showing an AI-designed molecule doing something clinically meaningful in humans, which did not exist eighteen months ago. But the same period shows, just as clearly, that "designed by AI" says nothing about whether a molecule will survive Phase 2 — the stage where biology, not computation, has always made the final call. As Isomorphic Labs, Insilico, and the reconstituted Recursion-Exscientia platform all push toward more internally-originated candidates in 2026 and beyond, the metric worth watching is not how many AI-designed molecules enter trials, but how many clear Phase 2 — because that number, so far, still looks a lot like the rest of the industry's.