Why AI Keeps Winning Discovery and Losing Phase III

Artificial intelligence has done something remarkable to the front end of drug discovery. Targets that once took years to nominate are now proposed in weeks. Molecules that once took thousands of synthesis cycles are designed and ranked before a single flask is touched. Antibodies are being invented from scratch in a 24-well plate.

And yet, as of September 2026, no drug whose biological target and chemical structure were both discovered by AI has received full FDA approval.

That gap is the most important fact in computational biology right now. Not because AI has failed — it clearly has not — but because it reveals exactly which part of drug development the technology has solved, and which part it has barely touched.

What genuinely changed in 2025–26

Three shifts deserve the attention they got.

Structure prediction became co-folding — and learned to predict binding. AlphaFold 3 showed that protein–ligand complexes could be predicted directly. In June 2025, MIT and Recursion released Boltz-2, the first model to predict structure and binding affinity together, approaching free-energy perturbation accuracy at roughly one-thousandth the compute. It topped the blind CASP16 affinity challenge across 140 complexes. Open alternatives followed — OpenFold3 under Apache 2.0, and ByteDance’s Protenix-v1 in early 2026. The licensing matters as much as the accuracy: pharma can now retrain frontier co-folding models on proprietary internal chemistry.

Generative design started producing real molecules, not just candidates. The Baker lab’s fine-tuned RFdiffusion designed antibodies against chosen epitopes with atomic accuracy, confirmed by cryo-EM at 1.45 Å backbone RMSD (Nature, 2025). Chai Discovery’s Chai-2 reported a 15.5% hit rate across 52 novel targets using at most 20 designs each — against a traditional screening baseline closer to 0.1%. At MIT and the Broad, generative models designed 36 million candidate antibiotics; of 24 synthesised, two killed drug-resistant gonorrhoea and MRSA through an apparently novel mechanism (Cell, 2025).

Regulators wrote AI into the rulebook. The FDA issued draft guidance on AI in regulatory decision-making in January 2025 and a roadmap to replace animal testing with human-relevant and in silico methods in April 2025. In January 2026, the FDA and EMA jointly published ten Guiding Principles of Good AI Practice spanning the full product life cycle.

The clinical scorecard

In July 2026, Insilico Medicine registered a 320-patient, 52-week Phase III trial for rentosertib in idiopathic pulmonary fibrosis — the first pivotal trial of a drug whose target (TNIK) and molecule were both AI-derived. Its Phase IIa showed a mean forced vital capacity change of +98.4 mL at 12 weeks versus −20.3 mL on placebo (Nature Medicine, 2025). That result is genuine, and it is also small, short, single-country and measured on a surrogate endpoint. Many IPF drugs have looked good at that stage.

Takeda’s zasocitinib beat an approved rival head-to-head in Phase III in June 2026 and is heading to an FDA submission. It was widely called an “AI-designed pill”. In fact, Nimbus and Schrödinger began the programme around 2016 using physics-based free-energy perturbation against TYK2 — a target Bristol Myers Squibb had already validated. It proves computation can build a better molecule against a known target. It does not prove AI can pick a new one.

The headline statistics deserve the same care. AI-derived molecules report 80–90% Phase I success against a historical 40–65%, which sounds transformative until you notice Phase II success sits at about 40% — the industry average — and that the whole sample is roughly 67 molecules out of 6,147 in clinical development. Phase I tests safety and drug-likeness, precisely what AI optimises. Phase II tests whether the biology is right.

The Phase III wall is a target problem, not a molecule problem

Only about 6.7% of drugs entering Phase I reach approval — an all-time low in Citeline’s 2014–2023 data. Phase III itself passes barely 55%, even after a drug has already shown a Phase II signal. And when late-stage programmes die, they die of inefficacy: in a Harvard study of 640 novel therapeutics, 57% of failures were efficacy failures, 17% safety, 22% commercial (JAMA Internal Medicine, 2016).

A flawless inhibitor of the wrong target still fails. And target choice is where AI contributes least. Of the AI-associated clinical pipeline, only about 24 programmes involve an AI-discovered target at all; the rest apply generative chemistry to targets selected by conventional biology.

The reason is structural. These models are trained on crystal structures, in vitro assays and animal data. Phase III measures clinical outcomes in heterogeneous humans over months or years. No public dataset connects molecular features to those outcomes at scale, and pharma’s richest evidence — its unpublished failures — never enters the training distribution. The veteran chemist Derek Lowe puts it plainly: poor target selection and human toxicity prediction are what kill most programmes, and neither is solved.

Three further traps compound it. Small Phase II trials inflate effect sizes, so Phase III inherits optimism it cannot reproduce. Surrogate endpoints — a biomarker, a 12-week lung function reading — do not reliably predict survival or function. And retrospective benchmarks flatter models through data leakage and rarely test genuinely novel chemistry.

What would actually move the number

The single best-evidenced lever on late-stage success is not chemistry at all. Drug mechanisms supported by human genetic evidence are 2.6 times more likely to reach approval than those without (Minikel et al., Nature, 2024) — and the authors argue we are nowhere near exhausting that signal.

That suggests where the next generation of models should aim:

•             Rank targets by human causal evidence, not literature popularity — biobank genetics, Mendelian randomisation, loss-of-function data. Judge platforms by the genetic support of the targets they nominate.

•             Build human-relevant translational models. Organoids, organ-on-chip and in silico toxicology, now actively encouraged by the FDA’s NAMs roadmap, generate exactly the human data these models lack.

•             Stratify patients from Phase I. A drug that works in a subgroup fails in an all-comers Phase III. Find the responder biomarker before the pivotal trial, not after it.

•             Close the loop and publish the failures. Every design–make–test cycle should retrain the model, negative results included. Federated learning offers a way to pool industry failure data without exposing IP.

The honest summary

AI has substantially improved the probability of finding a good molecule. Phase III tests the probability of a correct biological hypothesis. Those are different problems, and only the first has been meaningfully solved.

The rentosertib readout and the zasocitinib submission will dominate the narrative over the next eighteen months, and a single trial — in either direction — will prove far less than the headlines claim. The durable lesson is simpler: speed in discovery only compounds if it is aimed at the right biology. Until AI helps choose the target and the patient as well as it designs the molecule, the wall stays where it is.

Sources

_{^{•             Xu Z. et al. Rentosertib phase 2a trial in IPF. Nature Medicine, 2025. https://www.nature.com/articles/s41591-025-03743-2}}

_{^{•             Bennett N.R. et al. De novo antibody design with RFdiffusion. Nature, 2026. https://www.nature.com/articles/s41586-025-09721-5}}

_{^{•             Krishnan A. et al. Generative de novo antibiotic design. Cell, 2025. https://www.broadinstitute.org/news/using-ai-researchers-design-new-compounds-kill-antibiotic-resistant-bacteria}}

_{^{•             Passaro S. et al. Boltz-2. bioRxiv, 2025. https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1}}

_{^{•             Minikel E.V. et al. Genetic evidence and clinical success. Nature, 2024. https://www.nature.com/articles/s41586-024-07316-0}}

_{^{•             Hwang T.J. et al. Failure of drugs in late-stage development. JAMA Internal Medicine, 2016. https://pubmed.ncbi.nlm.nih.gov/27723879/}}

_{^{•             Norstella/Citeline. Why are clinical development success rates falling? https://www.norstella.com/insight/why-are-clinical-development-success-rates-falling/}}

_{^{•             EMA & FDA. Guiding Principles of Good AI Practice in Drug Development, January 2026. https://www.ema.europa.eu/en/documents/other/guiding-principles-good-ai-practice-drug-development_en.pdf}}

_{^{•             IntuitionLabs. AI drug discovery FDA approvals: the 2026 reality check. https://intuitionlabs.ai/articles/ai-drug-discovery-fda-approvals}}

Disclaimer

_{^{This article is for general information and education only. It is not medical, clinical, regulatory, legal or investment advice, and is not a recommendation regarding any security, drug or therapy. Content reflects publicly available information as of 27 September 2026; trial status, guidance and pipelines change quickly, and readers should verify against primary sources. Some cited material comprises press releases, preprints not yet peer-reviewed, or industry commentary that may reflect the commercial interests of its authors. Figures are reported as published and have not been independently verified. Company and product names are used for illustration only and imply no endorsement or affiliation.}}