AI in Drug Discovery: What Has Actually Reached Patients

Drug discovery is where AI’s scientific case is strongest — and also where the gap between a press release and a treated patient is widest. So the useful question isn’t “can AI design molecules?” (it can) but what has actually reached human beings?

Why the problem suits AI

Bringing a drug to market classically takes 10–15 years and billions of dollars, and the early stages are a search problem over an impossibly large space: there are more possible drug-like molecules than atoms in the solar system. You must pick a target (a protein implicated in the disease), then find a molecule that binds it, doesn’t bind everything else, survives the body, and is safe.

That’s pattern-matching over chemistry and biology at a scale humans cannot brute-force — genuinely well matched to machine learning, unlike a lot of enterprise AI.

What AlphaFold changed

For fifty years, predicting how a protein folds from its amino-acid sequence was a grand-challenge problem. AlphaFold largely solved it, and the structures were released openly. If you want to design a molecule that fits a target, knowing the target’s shape is the starting gun — and that step went from years of lab work per protein to a prediction.

Successor work continues: Isomorphic Labs released IsoDDE in February 2026, roughly doubling AlphaFold 3’s accuracy on the hardest ligand-binding cases (targets with under 20% sequence similarity to anything in training data) — precisely the novel targets that matter.

The proof point: rentosertib

Here is the concrete result, and it’s worth knowing in detail because it is the most-cited evidence in the field.

Insilico Medicine used generative AI to identify a novel target (TNIK) for idiopathic pulmonary fibrosis — a brutal lung disease with few options — and then to design a molecule against it: rentosertib. Both steps came from the AI platform, and both advanced in under three years, against an industry norm measured in many more.

The Phase IIa results were published in Nature Medicine in June 2025: a double-blind, placebo-controlled trial of 71 patients across 22 sites. Patients on the 60mg daily dose showed a mean improvement of +98.4 mL in forced vital capacity, while the placebo group declined by −20.3 mL. In a disease defined by relentless lung-function loss, improvement is a meaningful signal. Rentosertib has since entered Phase III.

This is the field’s landmark: the first case where AI found both the target and the compound, and the result survived a randomized, placebo-controlled trial.

Where the frontier actually is

Worth holding alongside that: Isomorphic Labs, the DeepMind spin-off with partnerships with Eli Lilly and Novartis, has pushed its first-in-human trials to the end of 2026 — a year later than it originally committed to. It is prioritizing oncology.

That delay isn’t a scandal; it’s the normal friction of turning computational output into something you can put in a person. But it’s a useful corrective to “AI will design drugs in an afternoon.”

The honest limits

Discovery was never the main bottleneck. Roughly 90% of drugs that enter clinical trials fail — usually in Phase II or III, on efficacy or safety in real humans. AI has compressed the first few years of a 10–15 year process. It has not touched the expensive, slow, failure-prone part: the trials themselves.

No AI-designed drug has been approved. Rentosertib in Phase III is genuinely exciting and genuinely not the same thing as an approved medicine. Phase IIa was 71 patients; Phase III is where most hopes die.

Biology is not a solved simulation. Predicting a structure is not predicting what a molecule does inside a living system with an immune response, a liver, and a decade of comorbidities. The models are strong at binding, weak at organism-level consequence.

Data is the constraint. Pharma’s most valuable data — including failures — is proprietary, fragmented, and often not machine-readable. Public structural data is abundant; clinical outcome data is not.

What to take from it

AI in drug discovery is one of the few applications where the scientific case is real, the evidence is peer-reviewed, and the upside is measured in lives rather than margin. It has already done something remarkable: taken an AI-found target and an AI-designed molecule into a successful randomized trial in under three years.

It has not yet produced an approved drug, and the honest framing is that it makes the first act of a long play dramatically faster. That’s worth a great deal — it’s just not the same as curing disease on a GPU.

For the broader clinical picture, see our AI in healthcare use case; for how these systems are governed, browse AI safety and policy.