AI Drug Discovery Reaches Phase III While Protein Design Matures

The field's first late-stage test has arrived

AI-driven drug discovery now has a molecule in late-stage human testing, but it still has no approved drug. That gap explains most of what has happened in the field this year. On July 7, 2026, Insilico Medicine started a Phase III trial of rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis (IPF) that it describes as AI-empowered.2130 The study is reported to be randomized, double-blind and placebo-controlled, with 320 patients treated for 52 weeks.28 As of October 2026, no fully AI-discovered drug has been approved.28

Rentosertib is the leading example because its target and its molecule both came from AI. Insilico's PandaOmics platform flagged TNIK as an IPF target that conventional methods had overlooked.24 The company says the platform then designed a new molecule and brought it into clinical trials in about 30 months.30 Insilico also reports that it found the candidate after synthesizing fewer than 80 compounds.27

What the Phase IIa data showed, and where accounts differ

The Phase IIa trial, published in Nature Medicine, was designed mainly to test safety. On a secondary endpoint, patients taking 60 mg once daily gained an average of 98.4 mL in forced vital capacity over 12 weeks, while the placebo group lost 20.3 mL.25 One review in a Nature portfolio journal described the result more cautiously. It said the trial showed safety, target engagement and a trend toward slower functional decline, but was not powered to prove efficacy.22 Another report noted that tolerability at the highest dose was not entirely clean.25

The numbers are not reported the same way everywhere. Some industry analyses put the placebo decline at 62.3 mL, which makes the gap between groups look much larger than the 20.3 mL figure most other accounts use.262530 Readers should treat the larger gap with caution until the Phase III data arrive.

Counting the pipeline is harder than it sounds

Everyone agrees the clinical pipeline is growing. The estimates of its size vary by more than a factor of two:

• About 75 molecules in active human trials as of May 2026, according to a Boston Consulting Group analysis.27
• 117 AI-enabled assets from 63 companies in interventional trials as of December 2025, according to an abstract presented at ASCO 2026.21
• More than 173, or over 200, programs, according to other trackers, with roughly 94 in Phase I, 56 in Phase II and 15 in Phase III.2328

The spread comes from definitions. There is no official category for an "AI-discovered" drug, and FDA approval records do not track one.21 Each tracker decides for itself how much AI involvement a program needs to count.

The ASCO figures are the most useful because they show attrition. Of the 117 assets, 60 (51.3%) had completed Phase 1, but only 8 (6.8%) had completed Phase 2.21 The median time from a company's founding to its first Phase 1 trial was 6.5 years.21

The Phase I advantage is real; the Phase II advantage is not yet

The most quoted statistic in the field is that AI-discovered molecules pass Phase I 80% to 90% of the time, compared with a historical average of 40% to 65%.2329 That figure comes from a peer-reviewed analysis of only 21 completed Phase I programs.23 BCG researchers have warned that the absolute numbers are small and that early programs may have been chosen because their targets were easier to drug.27

Phase II, where drugs must show they work, tells a different story. Analyses put AI-discovered drugs' success rate there at about 40%, roughly the same as conventional drugs.25 The likely explanation is that algorithms are good at producing molecules with drug-like properties but have not yet shown they can pick the biological targets that matter in patients.25 One industry report claims a 68% Phase II success rate.26 That figure stands apart from the more cautious analyses and should not be relied on.

The careful conclusion is that AI makes early discovery faster and cheaper, and lead AI-native companies report 50% to 70% savings on both preclinical time and cost.27 There is not yet evidence that AI improves a drug's odds of surviving Phase II and Phase III.21 Saving time early is valuable. It is not the same as raising the share of drugs that succeed.

Beyond Insilico

Other companies have clinical programs, but most of the data so far comes from the companies themselves:

• Recursion: In an ongoing open-label Phase 2 trial, REC-4881 reduced polyp burden in familial adenomatous polyposis by a median of 43% at week 13, based on 12 evaluable patients.21 Separately, two of Recursion's lead programs have faltered in Phase 2 readouts.27
• Iambic: IAM1363 produced partial responses in 28% of 18 evaluable patients with HER2-altered cancers.21
• Schrödinger: SGR-1505 showed a 22% overall response rate in lymphoma.21
• Generate:Biomedicines: GB-0895 is in Phase 3 trials for severe asthma that enroll about 1,600 patients.21

Structure prediction moves from shapes to drugs

Protein structure prediction remains the foundation of the field. Isomorphic Labs, spun out of Google DeepMind in 2021, is its best-known commercial test. AlphaFold 3, built jointly by DeepMind and Isomorphic and released in 2024, extended prediction from single proteins to their interactions with DNA, RNA and small molecules.1 More than 3 million researchers in over 190 countries have used AlphaFold tools.4

On February 10, 2026, Isomorphic published a technical report on IsoDDE, its proprietary drug design engine.3 The report makes several claims:

• Binding prediction: On the hardest cases of a protein-ligand benchmark, where molecules differ most from the training data, IsoDDE reaches about 50% accuracy, versus about 23% for AlphaFold 3.5
• Binding strength: It predicts how tightly molecules bind more accurately than standard physics-based methods, in far less time.2
• New binding sites: It can find new binding pockets using only a protein's amino-acid sequence.2
• Antibodies: It modeled the CDR-H3 loop, a key part of an antibody's binding site, correctly 70% of the time, versus 58% for AlphaFold 3.38

The trade-off is openness. Unlike earlier AlphaFold systems, IsoDDE is proprietary, and its report gives few details that would let others reproduce the results.3 Its performance figures are company benchmarks that have not been independently validated.36 This marks a shift from AlphaFold's era of open publication.

Isomorphic: big money, no patients yet

Investors are pricing in success ahead of clinical evidence. Isomorphic raised a $2.1 billion Series B in May 2026, led by Thrive Capital, and says it has 17 active programs.5 By October, it was reportedly in talks to raise money at a valuation of at least $40 billion.32 It has partnerships with Eli Lilly, Novartis and Johnson & Johnson.33

It has not yet dosed a patient or released any clinical data.34 Demis Hassabis had originally aimed for trials by the end of 2025. The company now targets the end of 2026.632 In September it said its preclinical work was "on track" but gave no date.33

At least one outlet reported that Isomorphic entered Phase 2 in June 2026.7 Several other accounts flatly contradict this, saying no Isomorphic-designed compound has been given to a human.3134 The Phase 2 report should be disregarded. Until a first patient is dosed, Isomorphic's high valuation rests on its benchmarks, partnerships and reputation rather than on biology.35

Antibodies designed from scratch move from concept to practice

The most striking scientific progress this year is in designing new proteins from scratch, known as de novo design, especially antibodies.

• University of Washington (Baker lab): The team adapted its RFdiffusion model to design antibody fragments and full antibodies that bind chosen sites on a target with atomic-level precision. The work was published in Nature.11 The first designs bound only modestly, at tens to hundreds of nanomolar. Further lab-based optimization was needed to reach single-digit nanomolar binders.11
• Absci: Its Origin-1 platform designed antibodies against four human proteins in fewer than 100 attempts per target. It targeted "zero-prior" sites, meaning sites on proteins with no previously solved complex structures.17 Cryo-electron microscopy confirmed the designs closely matched the models. AI-guided optimization then produced IL36RA blockers that bind at sub-nanomolar strength.20
• Chai Discovery: Chai-2 reported a 16% hit rate in de novo antibody design.12

The field's own reviews are frank about the limits. A Molecular Cell review puts experimental success rates for designing antibody binding loops at 0% to 2% across hundreds to thousands of designs.13 Another group reported success rates of 1.5% to 4.4% and argued that lab screening is still essential.16 The immune response to non-natural protein sequences has not yet been studied in patients.15

The first clinical examples in oncology are AI-designed antibodies rather than entirely new protein structures. Generate:Biomedicines, for example, put an antibody that neutralizes the toxic payload MMAE into the clinic in 2026.15 By one observer's assessment, the field has not yet shown it can take an AI-originated biologic all the way from design to an approved manufacturing process.14

The reading

AI now reliably helps with discovery and gets molecules into Phase I efficiently. It has not yet shown that it produces better drugs. Three events over the coming months will test whether that is changing: Isomorphic's first patient dosing, rentosertib's Phase III results, and the first de novo-designed antibodies reaching patients.33 Until then, the most important number in the field is still zero AI-discovered drugs approved.