AI Drug Discovery Reaches Phase 3 as Protein Design Hit Rates Rise

The pipeline has caught up with the hype, mostly

AI drug discovery now has a molecule in late-stage testing. That is the clearest change this autumn. In September 2026, Insilico Medicine's rentosertib, a treatment for idiopathic pulmonary fibrosis, became the first drug to reach Phase 3 trials with both its biological target and its molecule produced by AI.2 The GENESIS-IPF-3 study plans to enroll 320 patients at 47 centers in China and follow them for 52 weeks. The first doses were given on 9 September.2 Coverage differs slightly on timing. One account dates the Phase 3 announcement to July 2026, while another ties the milestone to the September dosing.42 Either way, a program that published Phase IIa results in 2025 has now moved into the stage where most drug candidates fail.4

The AI-linked pipeline is wider than one drug. One tracker counted more than 173 AI-discovered programs in clinical development in early 2026: about 94 in Phase I, 56 in Phase II and 15 in Phase III.4 Generate:Biomedicines has GB-0895, a severe-asthma antibody, in two Phase 3 trials enrolling about 1,600 patients in more than 40 countries.2 Relay Therapeutics' zovegalisib, which was designed by modeling how proteins move, is also in Phase 3.3 Zasocitinib, a psoriasis drug that Nimbus developed on Schrödinger's computational platform and that Takeda now owns, is under FDA priority review. A decision is expected in the first quarter of 2027.2

Coverage agrees on the main caveat. As of October 2026, no drug discovered mainly by AI had won approval.24 Only about 6.7% of drugs that enter Phase 1 reach the market, based on 2014–2023 data.4 AI-found molecules appear to clear early safety tests more often, but proving they work in patients is still the hard part.4 My reading is that the industry has shown AI can fill a clinical pipeline. It has not yet shown AI can improve the odds of success in that pipeline. The zasocitinib decision and the rentosertib Phase 3 results will be the first real tests of the second claim.

Structure prediction moves from shapes to binding

The AlphaFold line of models is where the hype is highest and the public evidence is thinnest. In February 2026, Isomorphic Labs released its Drug Design Engine, IsoDDE. The company says it more than doubles AlphaFold 3's accuracy on protein–drug pairs that look very different from its training data, and that it can find hidden binding pockets using only a protein's sequence.2 On the hardest cases of one benchmark, the reported accuracy is about 50% for IsoDDE against about 23% for AlphaFold 3. The company also says its binding-strength predictions beat physics-based methods at a fraction of the cost.25 Unlike earlier AlphaFold models, whose code and database DeepMind released, IsoDDE is proprietary.21 Most of its performance claims come from Isomorphic's own publications, and independent testing has been limited.25

The money has followed regardless. In May 2026, Isomorphic raised $2.1 billion in a round led by Thrive Capital, bringing its disclosed outside funding to about $2.7 billion.3025 It has research deals with Eli Lilly and Novartis, and added Johnson & Johnson in January 2026.2 One profile counts 17 active programs across oncology, immunology and cardiovascular disease, though the company's own funding announcement does not give a program count.25

The clinical timeline is where coverage disagrees most. Reuters reported in January that Demis Hassabis now expects the company's first clinical trials by the end of 2026. That is a year later than his earlier target of end-2025.27 Reuters repeated that timeline when it covered the May funding round.30 A trade profile from August says Isomorphic's oncology and immunology programs are still in IND-enabling studies, with no human safety or efficacy data.28 One April headline said Isomorphic was "launching" human trials, but the article itself describes trials that are only set to begin.24 One outlet claimed an Isomorphic candidate entered Phase 2 in June 2026.26 That claim conflicts with every other account and should be treated as unreliable until the company confirms it. The best-supported position is that Isomorphic has not yet dosed a patient and has about ten weeks left to meet its revised deadline.228

Protein design: binder design is becoming routine

If Isomorphic is the story about drug-like small molecules, de novo protein design is the story about designing new proteins from scratch, and its evidence is in many ways stronger because so much of it is public. A May commentary in Communications Biology argued that biology is going through two "AlphaFold moments" at once: predicting the full range of shapes a protein can take, and routinely designing high-affinity protein binders.11 It pointed to independent competitions run by Adaptyv Bio. In a contest to design binders that block the Nipah virus, entries overall achieved a hit rate above 8%, with 26 binders at single-digit nanomolar strength or better.11 A Molecular Cell perspective published in October said binders are now often produced with success rates well above 10% straight from the computer, though not for every target.14 David Baker's lab has gone further, writing that the long-standing problems of designing new protein structures and binders are close to solved. In its view, the key question is now what to design, not how.15

The most notable result is an August preprint from Anthropic. In it, Claude models ran full binder-design campaigns on their own against 16 targets, choosing binding sites, installing open-source tools and returning ranked designs without any human design decisions.12 Two outside contract labs built and tested the designs. Of 1,320 designs with usable measurements, 354 bound their target, a 27% hit rate, with binders found for 14 targets.12 On RBX1, a target from an earlier open competition where 9 of 245 designs bound, 28 of Claude's 90 designs bound. The best reached 3.9 nM, compared with 45 nM for the competition winner tested on the same plate.19 The agent also produced binders for TNFα, where earlier studies had reported none.12 The authors also report a limitation: computational scores tracked poorly with how strongly successful designs actually bound, so lab testing is still essential.12

The broader set of tools points the same way. NVIDIA's Proteina-Complexa was tested on about a million candidates against 133 protein targets.16 The Molecular Cell review adds an important qualifier: in one large benchmark, only about 2% of its designs bound the intended target and nothing else.14 Biohub, the research organization tied to the Chan Zuckerberg Initiative, released open-source ESM-based models in May aimed at binder design for cancer and immune targets. Coverage noted that this is a research tool, not a clinical milestone.10 Results also vary by target. A Nature Communications study found plenty of strong binders for PD-L1 but far fewer for the related cancer targets B7-H3 and B7-H4.20

Agents enter the lab

The third trend is AI agents that coordinate whole workflows rather than single models. A quarterly review highlighted the Virtual Biotech framework, in which more than 37,000 agents analyzed the outcomes of 55,984 clinical trials. They found that drugs aimed at genes active in specific cell types were 48% more likely to reach the market and had 32% fewer adverse events.1 The New York Times profiled James Zou's lab at Stanford, where agent teams designed proteins against Covid variants (two of which worked) and recommended a way to target the lung-cancer protein B7-H3.6 Commercial products have followed, including OpenAI's GPT-Rosalind, Microsoft Discovery and Anthropic's Claude Science.5 So have deals: AstraZeneca agreed to pay insitro $40 million upfront, with up to $1.46 billion more in milestones, to find heart and liver disease targets.7

The review's advice is reasonable. The next sign of real progress would be agents used in live discovery programs, not only tested against historical data.1 A smaller study from Korea University shows what that looks like at small scale. All three AI-generated compounds predicted to block the antiviral target DHPS did so in the lab, and all three predicted to fail did not. A crystal structure confirmed one of them binding. Even so, these are weak early hits that still need selectivity and toxicity testing.8

The verdict

Across the coverage, early discovery has clearly changed. Binder design has moved from rare successes to dependable engineering, structure models now predict how drugs bind and not just protein shapes, and agents are starting to run parts of the process.111225 Late-stage evidence has not changed yet. The next six months will matter more than anything announced so far: Recursion's REC-4881 data in November, Isomorphic's year-end trial deadline, and the FDA's early-2027 decision on zasocitinib.2 Analysts still say an approval could come within the next few years, but that is a forecast, not something any drug has achieved.422