The first AI-discovered drug is now in Phase 3. A physician’s read

A milestone the AI-in-medicine field has promised for a decade is now in sight. Insilico Medicine’s Rentosertib, a drug whose target and molecule were both identified by AI, has begun Phase III trials for idiopathic pulmonary fibrosis, the final stage before regulators decide whether it becomes medicine. The X posts calling it an “AI-only discovered drug” oversimplify; the precise claim is remarkable enough.

Rentosertib’s TNIK target was proposed by Insilico’s discovery platform and the molecule generated by its chemistry models, with humans running the trials that every drug must pass. The company also announced a collaboration with Takeda this month worth up to roughly $600 million in upfront and milestone payments.

Why Phase III is the line that matters

Plenty of AI-designed molecules have entered Phase I, which mostly tests safety. Phase III tests whether the drug actually works at scale, and it is where most drugs die. As a physician, this is the first checkpoint I would call decisive: pass it, and AI drug discovery stops being a promising method and becomes a validated one. The economics of health AI’s funding boom are, in a real sense, priced on trials like this one.

The honest caveat

IPF is a graveyard for hopeful compounds; existing therapies slow decline rather than reverse it. A Phase III initiation is a bet, not a result, and readouts are years away. What has already changed is the timeline: target-to-Phase-III in this timeframe is dramatically faster than the industry’s decade-plus norm, and that speed advantage survives even if this particular molecule fails.

What “AI-discovered” precisely means here

The distinction that matters is which decisions the machines made. Drug discovery has used computational screening for decades, software ranking molecules humans already proposed. Rentosertib is categorically different on two counts: the platform proposed TNIK as a target for pulmonary fibrosis when it was not an obvious candidate in the field’s literature, and generative chemistry models produced the molecule itself rather than filtering a human-built library. Novelty of target plus novelty of molecule is the specific pairing that justifies the milestone framing, and it is the pairing the casual “AI-only drug” phrasing both overstates and undersells.

The Takeda collaboration is the commercial validation worth reading alongside the trial. Large pharmaceutical companies do not attach nine-figure milestone structures to compounds their own scientists cannot defend; the deal means Takeda’s diligence teams examined the AI-originated target thesis and priced it as credible. That is a different kind of evidence than a benchmark or a paper, and for the sector it may matter as much as the Phase III start itself: the industry’s incumbents are now underwriting AI-discovered biology with their own capital, exactly the pattern of buyer-side conviction we track across clinical AI adoption.

What to watch

Watch enrollment pace, any interim analyses, and whether other AI-native drugs follow into late-stage trials this year. One success reprices the whole category; so would one high-profile failure. For how the field works end to end, see our physician’s explainer on AI drug discovery.