AI-Designed Drugs Are in Phase III. The Hard Part Is What Comes Next.

Three AI-designed drugs entered the most expensive stage of clinical testing in 2026, each one a different test of a different target. GB-0895, from Generate Biomedicines, is a long-acting antibody that blocks a signaling protein called TSLP; it is now in two global Phase III studies for severe asthma, planning to enroll roughly 1,600 patients in more than 40 countries. Rentosertib, an oral pill from Insilico Medicine that targets a fibrosis-driving enzyme called TNIK, began a randomized Phase III in July for idiopathic pulmonary fibrosis, with 320 participants planned across 47 centers in China. Zovegalisib, a Relay Therapeutics pill aimed at a mutated form of PI3Kα, is in Phase III for HR-positive, HER2-negative metastatic breast cancer. They are the first medicines designed largely by generative AI to face the stage where a drug either works in people or it does not.

At the Ai4 2026 conference this week, a source described generative biology as a move away from "heroic, one-off efforts" toward something "programmatic, repeatable, and scalable." The approach replaces screening molecules found in nature or built by hand with models that learn the rules of how a protein folds, binds, and behaves, then propose new structures that fit a target.

Each of the three programs is a different bet on the bottleneck. GB-0895 is a biologic, a manufactured antibody grown in cells and given by injection. Rentosertib is a small molecule, the kind of pill a factory can churn out by the kilogram, though its Phase III footprint is set by where idiopathic pulmonary fibrosis trials are most feasible, not by where the molecule was invented. Zovegalisib sits in between, a targeted pill aimed at a specific tumor mutation.

The common pattern is that the model has compressed the discovery step from years of medicinal chemistry to months. What has not compressed is everything that comes after: the regulators who must weigh a new class of evidence, the contract manufacturers who must scale biologic production, the clinical sites that must run multi-thousand-patient trials, and the chemists who must figure out why a designed molecule behaves differently in a human than in a screen.

2026 is being framed, in industry coverage, as the year those readouts arrive. If any of the three succeed, the bottleneck moves down the pipeline and the question becomes whether the next batch of programs can keep up. If all three fail, the discipline has a generation problem: it can produce candidates faster than it can prove them.

The same models that design a therapeutic can, in principle, lower the barrier to designing one that is dangerous. The source's argument for generative biology is that it is a programmable engineering discipline, and that argument cuts both ways: a tool that lets a researcher specify "an antibody that binds TSLP" can also let someone else specify a toxin. The biosecurity question is not a footnote to the science. It is part of what has to be built before the technology reaches patients at scale.

The 2026 readouts will answer one question: whether generative biology produces drugs that work. They will not answer the other: whether the regulations, factories, and biosecurity rules built around a slower, hand-designed drug pipeline can keep up with a discipline that moves at model speed. That is the part still being built, and it decides whether a discipline built on models can deliver a medicine, or only more candidates.