AI Drug Discovery: What Happens Between a Promising Molecule and a Medicine?

The headline writes itself: “AI invents a new drug.” It is a satisfying story — a machine sifts billions of possibilities overnight and hands scientists a cure. It is also a myth, or at least a badly compressed version of the truth. AI can now help design a promising molecule faster than before. But a promising molecule is not a medicine. Between the two lies the longest, most expensive and most failure-prone stretch of biomedical science.

This explainer follows that stretch: what AI is genuinely good at, where the evidence stands, and why the hardest problems in drug development are the ones software has not yet solved.

The long road from molecule to medicine

The U.S. Food and Drug Administration describes five steps in developing a drug:

1. Discovery and development — research begins in the laboratory, identifying a biological target and molecules that act on it.
2. Preclinical research — laboratory and animal testing to answer basic questions about safety.
3. Clinical research — testing in people, typically in Phase 1, 2 and 3 trials, to establish safety and effectiveness.
4. FDA review — regulators examine all the data and decide whether to approve.
5. Post-market safety monitoring — surveillance continues once the drug is in use.

The journey is punishing. A widely cited 2022 review in Acta Pharmaceutica Sinica B puts the figures starkly: drug development takes more than 10 to 15 years and costs over $1 to 2 billion per new drug on average, and around 90% of drug candidates that enter clinical trials fail. The review attributes those failures mainly to lack of effectiveness (40–50%) and unmanageable toxicity (about 30%), with poor drug-like properties and commercial decisions making up the rest.

That breakdown is the key to judging AI’s promise. Most drugs do not fail because chemists couldn’t design a molecule. They fail because, in real patients, the molecule doesn’t work well enough or causes too much harm.

Plain English: A target is the molecule in the body — often a protein — that a drug is meant to act on. Pick the wrong target and even a perfectly designed drug will fail.

Where AI genuinely helps

“AI drug discovery” is really a bundle of different techniques, each aimed at a different bottleneck:

• Finding targets. Machine learning can mine genetic, protein and patient data for proteins linked to a disease.
• Predicting protein shapes. A drug usually works by fitting into a pocket on a protein, so knowing the shape matters. DeepMind’s AlphaFold2 predicted the structure of virtually all of the roughly 200 million proteins researchers have catalogued, and Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for protein structure prediction, alongside David Baker for computational protein design.
• Generating and ranking molecules. Generative models propose new chemical structures; predictive models estimate which are likely to bind the target, be absorbed by the body and avoid obvious toxicity.
• Designing smarter trials. AI tools are being explored to help select patients and sites, and to analyse trial data.

The payoff, where it works, is mostly speed and efficiency in the early, preclinical phase: fewer molecules made and tested before a candidate emerges. That is valuable. But it is the cheapest, fastest part of the journey.

AI can shorten the road to the starting line. It has not yet shown that it can shorten the race.

A case study: from algorithm to a Phase 2a trial

One of the most closely watched examples is rentosertib, a drug candidate from the company Insilico Medicine for idiopathic pulmonary fibrosis (IPF) — a disease in which lung tissue becomes progressively scarred and stiff. The company says both the target, a protein called TNIK, and the molecule itself were identified using its generative AI platform.

In June 2025, results of a Phase 2a trial were published in Nature Medicine. According to the company’s announcement, the double-blind, placebo-controlled trial enrolled 71 patients at 22 sites in China for 12 weeks. It met its primary goal, which was safety: side-effect rates were similar across groups. In an exploratory measure of lung function, forced vital capacity (FVC — how much air a person can forcibly exhale), patients on the highest dose showed an average gain of 98.4 mL, while the placebo group declined by 20.3 mL.

That is an encouraging signal — and exactly the kind of result that is easy to overread. The trial was small, short and designed chiefly to test safety; efficacy measures in such studies are hypothesis-generating. The researchers themselves noted that group sizes were limited and the findings need validation in larger studies. Whether rentosertib becomes a medicine depends on trials that have not yet reported. For a guide to what each trial phase can and can’t tell you, see Inside a Clinical Trial.

Is AI improving the odds?

The honest answer: too early to say. A 2024 analysis in Drug Discovery Today looked at how AI-discovered molecules have fared in trials. In Phase 1, which mainly tests safety, 21 of 24 molecules succeeded — an 80–90% success rate, above historical averages. In Phase 2, the first real test of whether a drug works, 4 of 10 succeeded, around 40% — comparable to historical industry averages.

The authors stressed that the sample is small and the numbers are likely to change substantially as more data arrive. Read carefully, the pattern fits the failure breakdown above: AI may be good at producing molecules that behave well in the body, but the leap from “safe” to “effective in patients” — where most drugs die — remains as hard as ever. That leap depends on biology: choosing the right target, in the right patients, for the right disease.

How regulators are responding

Regulators are paying attention to AI across the drug lifecycle, not only in discovery. In January 2025 the FDA issued draft guidance on using AI to support regulatory decisions about drugs and biological products. It proposes a risk-based framework for establishing whether an AI model is credible for a particular “context of use” — the specific question the model is being used to answer. As a draft, it contains non-binding recommendations. The underlying principle matters, though: a molecule’s origin does not change the standard of evidence. An AI-designed drug must pass the same trials as any other.

How to read the next “AI drug” headline

• What stage is it? A computer prediction, a lab result, an animal study and a Phase 3 trial are different universes.
• What did AI actually do? Pick the target, design the molecule, or just speed up one step?
• What was measured? Safety, a biomarker, or an outcome patients feel?
• How many people, for how long? Small, short trials generate hypotheses rather than proof.
• Who is saying it? Company announcements are useful but are not independent evaluations.

AI is making the early stages of drug discovery faster and more inventive, and that could matter a great deal. But the medicine is made in the clinic. Follow the full path in our interactive Drug Development Journey, learn to spot overreach with How to Read a Biotech Breakthrough, and see how young companies navigate the gauntlet in The Biotech Startup Map.