Artificial Science

Biology

The First AI-Designed Drug Just Entered Phase III. Here's What That Actually Proves.

Insilico Medicine has pushed rentosertib, a drug whose target and molecule were both generated by AI, into a pivotal Phase III trial. It's the clearest test yet of whether AI drug discovery can produce medicines that work, not just headlines.

By Dr. Sreenu Vattipally 9 min read 1,771 words

For a decade, AI drug discovery has run on promises. Faster timelines. Novel targets. Molecules no chemist would have drawn. The pitch decks were beautiful. The clinical proof was thin.

That gap just narrowed. On July 7, 2026, Insilico Medicine announced it had begun a Phase III trial for rentosertib, a drug for idiopathic pulmonary fibrosis (IPF) whose biological target and chemical structure both came out of generative AI systems. It is the first program of its kind to reach late-stage testing, the last hurdle before a company can ask regulators for approval. (Insilico Medicine press release, July 7, 2026)

If you only remember one thing about the current state of AI in biology, make it this: the field has finally produced a candidate that has to succeed or fail on hard clinical endpoints, in front of regulators, with no room to hide behind a demo. That is a very different thing from a benchmark score.

What actually happened#

Rentosertib, formerly known as ISM001-055, targets a protein called TRAF2- and NCK-interacting kinase (TNIK), a kinase tangled up in the fibrotic and inflammatory signalling that scars lung tissue in IPF. TNIK wasn't a fashionable IPF target. Insilico's biology engine, part of its Pharma.AI platform, surfaced it by chewing through multi-omics data, biological networks, and the scientific literature, then ranking candidates by how strongly they connected to fibrosis and ageing biology. In the company's peer-reviewed account, TNIK came out as the top-ranked candidate in that discovery run. (Nature Biotechnology, 2024)

The molecule itself was generated and optimised by a separate generative-chemistry system. So both halves of the classic drug-hunting problem, "what should we hit?" and "what molecule hits it?", were handed to software. That end-to-end quality is what makes rentosertib different from most AI drug stories, which usually amount to using AI to find a better molecule against a target humans already knew about.

The clinical evidence so far comes from GENESIS-IPF, a Phase IIa trial in 71 patients across 22 sites in China, published in Nature Medicine in 2025. It was double-blind and placebo-controlled. The main goal was safety, and the drug cleared it: adverse-event rates looked similar across the treatment arms, and most events were mild or moderate. The efficiency signal was the part that got people leaning forward. In the 60 mg once-daily group, patients showed a mean improvement in forced vital capacity (FVC, a standard measure of lung function) of +98.4 mL at 12 weeks, versus a decline of -20.3 mL in the placebo group. (Nature Medicine, 2025)

For a disease where the goal of existing drugs is merely to slow the loss of lung function, a group that gained function is hard to ignore. It's also why the company felt it had earned a shot at Phase III.

The new trial is bigger and longer: 320 patients across 47 centres in China, dosed once daily for 52 weeks. The primary endpoint is the annual rate of FVC decline, which is the endpoint regulators actually care about in IPF. It's led by Professor Zuojun Xu of Peking Union Medical College Hospital. (Insilico Medicine press release, July 7, 2026)

Why it matters#

The AI drug discovery scorecard, as of early 2026, is sobering. Industry trackers count well over 150 AI-discovered or AI-designed molecules in human trials, and zero approvals so far. (IntuitionLabs pipeline analysis, 2026) A pipeline that broad with nothing across the finish line invites a fair question: is any of this working, or is it a very expensive way to generate press releases?

Rentosertib is the closest thing the field has to an answer in progress. A few reasons it carries more weight than the average milestone.

It attacks a novel target, not a crowded one. Plenty of AI wins so far are "me-better" molecules against proteins that pharma has hammered for years. Insilico's claim is bolder: the AI proposed a disease hypothesis, picked an underexplored target, and designed a drug for it. If that holds up in Phase III, it validates the biology-generation part of AI, not just the chemistry-optimisation part. That's the harder, more valuable claim.

The paper trail is unusually complete. This isn't a black box. The discovery-to-clinic story ran in Nature Biotechnology, the Phase IIa results in Nature Medicine, and the medicinal chemistry in the Journal of Medicinal Chemistry. For a field that critics accuse of overselling, having the target biology, the chemistry, and the human data all in peer-reviewed venues is exactly the kind of receipt that sceptics have been asking for. (Insilico Medicine press release, July 7, 2026)

The speed numbers are concrete and testable. Insilico says its platform reaches a preclinical candidate in roughly 12 to 18 months, against a more typical 2.5 to 4 years, while synthesising only 60 to 200 molecules per program instead of thousands (Insilico Medicine press release, July 7, 2026). Whether that speed translates into drugs that survive Phase III is the whole question. But at least it's a falsifiable claim, not a vibe.

There's a broader industry backdrop, too. 2026 has been a year of AI infrastructure landing hard in pharma, from Eli Lilly standing up a supercomputer built around more than a thousand Blackwell GPUs to churn through genomics and molecule design (R&D World, 2026), to DeepMind's drug spinout Isomorphic Labs publishing a next-generation "drug-discovery engine" that reportedly more than doubles AlphaFold 3's accuracy on a tough protein-ligand benchmark (Nature news, 2026). The compute and the models are getting serious. Rentosertib is the part where all of that has to meet a patient.

The limitations (read this part)#

Phase IIa was small and short. Seventy-one patients over 12 weeks is enough to spot safety problems and a hint of efficiency. It is nowhere near enough to prove a drug slows a disease that plays out over years. Small trials also produce noisier estimates, and encouraging early signals routinely shrink or vanish when the sample size grows. The +98.4 mL figure is a reason to run Phase III, not a result you should bank on.

The trials are China-only so far. Both the completed Phase IIa and the new Phase III enrol patients in China. That's fine scientifically, but IPF biology, standard-of-care, and regulatory expectations vary across regions. A single-country Phase III may not, on its own, satisfy the FDA or EMA for a global approval, which can mean additional trials and more years.

FVC is a surrogate, not survival. Forced vital capacity is the accepted regulatory endpoint for IPF, and slowing its decline is meaningful. But patients ultimately care about living longer and breathing better for longer. Phase III is powered for FVC decline over 52 weeks, not for a mortality benefit. A win here is real, but it's a win on the metric, not the final word on outcomes.

"AI-designed" doesn't mean AI did it alone. It's worth keeping the language honest. Humans designed the assays, ran the chemistry, chose the clinical protocol, recruited the patients, and interpreted the data. The AI proposed and prioritised; a large team of scientists did the rest. The interesting claim isn't that software replaced drug hunters. It's that software meaningfully changed where they started and how fast they moved.

And the field's base rate is brutal. Historically, only around one in ten drugs that enter Phase I ever reach approval, and IPF specifically is a graveyard of late-stage failures. AI origins don't exempt a molecule from biology. Rentosertib could still fail in Phase III for reasons that have nothing to do with how it was discovered. If it does, that won't disprove AI drug discovery, and if it succeeds, one drug won't prove the whole paradigm either. One data point is one data point.

The most useful framing: this is the first time the AI-drug thesis gets tested on the terms that actually matter to patients and regulators. That's progress worth taking seriously, and hype worth resisting.

FAQs#

Is rentosertib the "first AI drug"? It's the first drug where AI both identified the biological target and designed the molecule that reaches Phase III, according to Insilico. Other AI-involved drugs have entered or completed trials, but most used AI for a narrower slice of the process, such as optimising a molecule against a known target. No AI-originated drug has been approved by regulators yet. (IntuitionLabs, 2026)

What is IPF, and why does it matter here? Idiopathic pulmonary fibrosis is a progressive scarring of the lungs that gets worse over time and can't currently be reversed. It affects roughly 5 million people worldwide, with a median survival of about 3 to 4 years after diagnosis. Existing antifibrotic drugs slow the decline but don't stop it, so the unmet need is large. (Insilico Medicine press release, July 7, 2026)

What does TNIK do? TNIK is a kinase involved in several fibrosis- and inflammation-related signalling pathways. Insilico's AI flagged it as a high-priority IPF target that the field had largely overlooked, which is part of why the program is notable: it's a new mechanism, not a rework of an established one. (Nature Biotechnology, 2024)

When will we know if it works? The Phase III trial doses patients over 52 weeks, so meaningful readouts are years away, not months. Even a positive result would then need regulatory review, and quite possibly additional trials outside China before a global approval.

Should I read the Phase IIa numbers as proof it works? No. Treat them as a promising early signal from a small, short study, exactly the kind of result that justifies a larger trial and exactly the kind that sometimes doesn't replicate. The Phase III trial exists precisely because Phase IIa can't settle the question.

Were the papers peer-reviewed or preprints? The three key papers cited here, in Nature Biotechnology, Nature Medicine, and the Journal of Medicinal Chemistry, are peer-reviewed publications, not preprints. The industry pipeline counts and market context come from analyst reports and trade coverage, which are secondary sources.

Sources#

Note: This post concerns an investigational drug that has not been approved by any regulatory authority. Nothing here is medical advice.

Related observations

Adjacent work from the same lines of enquiry.