Drug discovery has a famous failure rate. Most molecules that enter human trials never reach patients. If AI could pick better molecules, that rate should improve, and the first test is simple: do AI-discovered drugs survive clinical trials more often than other drugs?

A 2024 paper in Drug Discovery Today gave the first answer. It looked at the pipelines of AI-native biotech companies and found a result that is easy to misread.

Do AI-discovered drugs do better in trials?

The authors, led by Madura Jayatunga, analyzed molecules from AI-native biotech companies that had finished clinical trials by December 2023. In Phase I, the first test in humans, they report a success rate of 80 to 90%, against a historic industry range of roughly 40 to 65%. In Phase II, where a drug must show it works, the rate was about 40%, similar to historic averages.

Source: Jayatunga et al., Drug Discovery Today, 2024. Counts (21 of 24 and 4 of 10) are those reported for completed trials as of December 2023. The authors do not give a single Phase II industry figure in the abstract, so none is shown.

View data

group	rate
AI-discovered, Phase I (21 of 24)	87.5
Industry history, Phase I, upper end	65
Industry history, Phase I, lower end	40
AI-discovered, Phase II (4 of 10)	40

The first bar is the headline: 21 of 24 AI-discovered molecules that completed Phase I succeeded, or 87.5%. The last bar is the part that gets less attention: 4 of 10 completed Phase II trials succeeded, which is 40%.

How much should we trust these numbers?

Less than the bar chart suggests. The samples are small. A 95% confidence interval on 21 of 24 runs from about 69% to 96%. On 4 of 10 it runs from about 17% to 69%. The Phase II rate could plausibly be well below or well above the industry average, and ten trials cannot tell us which.

Phase I also measures something narrower than Phase II. It mainly tests whether a molecule is safe and behaves in the body as expected. That is where a model that designs drug-like molecules should help. Phase II asks whether the drug treats the disease, which depends on whether the target was the right one in the first place.

The authors conclude that AI is highly capable of designing or identifying molecules with drug-like properties. Phase II tests the target as much as the molecule, and on that the data so far come from a limited sample, as they note.

What has happened since?

In 2025, Nature Medicine published a randomized Phase 2a trial of a generative-AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis. One trial does not change a rate, but randomized trials of this kind are the evidence that will.

Limits

The analysis covers AI-native biotech companies, not every drug that used machine learning somewhere. The cutoff is December 2023, so newer trials are not included. The comparison with industry history is rough: historic averages differ by therapeutic area and by how early or late a molecule enters a company's pipeline, and AI-native companies may choose different targets than the industry as a whole.

Sources

1. Jayatunga et al., How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons, Drug Discovery Today, April 2024. Abstract via Europe PMC.
2. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial, Nature Medicine, 2025.
3. Confidence intervals are Wilson 95% intervals computed by the editors from the reported counts.