How AI Designs a Drug: From Target to Clinical Trial

In AI drug discovery, software takes on two early jobs. It ranks which biological target to attack, and it proposes molecules likely to hit it. Chemists still have to make those molecules. Preclinical studies and three phases of human trials decide whether a candidate is safe and effective. As of October 2026, one candidate whose maker credits generative AI has entered Phase 3. Evidence that AI improves the odds of success remains limited.

How does AI drug discovery work?

The US Food and Drug Administration (FDA) says thousands of compounds may be in play at the outset of drug discovery. It says only a few still look promising after early testing. AI tools aim to narrow that field by computer.

A well-documented example is rentosertib, an experimental pill for idiopathic pulmonary fibrosis (IPF). IPF is a progressive scarring disease of the lungs. Rentosertib's developer, Insilico Medicine, described the process in a 2024 paper in Nature Biotechnology:

1. Pick a target. A target is usually a protein involved in a disease. Insilico's software ranked candidates using molecular data from patient tissue. It also used text from publications, grant applications and clinical trial records. An enzyme called TNIK ranked first in the company's kinase-focused search.
2. Generate molecules. A second system used known crystal structures of TNIK to generate a virtual library of compounds designed to block it.
3. Make and test. Compounds were picked for novelty, ease of synthesis and drug-like properties. They were then made and tested in the lab and in mouse and rat models of fibrosis.

Insilico reported that going from target discovery to a preclinical candidate took roughly 18 months.

What happens after the computer?

Everything a conventional drug goes through. In the United States, a sponsor must file an investigational new drug application before testing in people. The FDA describes three trial phases:

• Phase 1: 20 to 100 volunteers, focused on safety and dosage. About 70% of drugs move on.
• Phase 2: up to several hundred patients, focused on efficacy and side effects. About 33% move on.
• Phase 3: 300 to 3,000 patients over one to four years. About 25% to 30% move on.

Rentosertib first went through a Phase 1 study in 78 healthy volunteers. A Phase 2a trial then assigned 71 IPF patients to one of three dosing regimens or placebo for 12 weeks. The trial's primary measure was the share of patients with at least one adverse event. Results published in Nature Medicine in June 2025 showed that this share was similar across groups. Adverse events judged related to treatment were more common on the drug. Seven patients stopped taking it because of liver-related problems. Forced vital capacity is a measure of lung function. It rose by an average of 98.4 milliliters in the 60 mg once-daily group and fell by 20.3 milliliters on placebo. The authors cautioned that the groups were small, every participant lived in China and follow-up was short.

In September 2026, Insilico said it had dosed the first patient in a Phase 3 trial. The ClinicalTrials.gov record lists 320 planned patients at 47 sites in China. It also lists 52 weeks of treatment and an estimated primary completion date of October 2029. The company calls it the first Phase 3 trial of a generative-AI drug. It says rentosertib has not been approved by any regulator.

Does AI make drug development faster or more successful?

Faster at the start, by the companies' own accounts. Whether it raises success rates is unproven. The rentosertib trial authors wrote in 2025 that AI-discovered drugs had failed in Phase 2 about as often as other drugs. They also wrote that none had yet completed Phase 3. A perspective article in Nature Reviews Drug Discovery in August 2026 concluded that evidence of clinically relevant impact was still limited. It urged the field to judge AI tools by whether they improve decisions.

STAT reported in January 2026 that the industry is divided over what counts as an AI-designed antibody. Trade publication BioSpace has called Takeda's psoriasis pill zasocitinib an AI-designed molecule. The pill is under FDA priority review with a target decision date in the first quarter of 2027, according to Takeda. Takeda's announcement does not use that label.

How do regulators treat AI-designed drugs?

The FDA's patient guide says a developer needs adequate data from two large, controlled trials before filing for approval. It does not mention how a molecule was discovered. The agency's draft guidance on AI was issued in January 2025 and was still a draft as of October 2026. It covers AI used to produce evidence about a drug's safety, effectiveness or quality. Its text says it does not address AI used in drug discovery. An FDA page dated January 2026 lists 10 guiding principles for AI in drug development, developed with the European Medicines Agency.

What to watch

Two things: rentosertib's Phase 3 results and the FDA's decision on zasocitinib. Neither drug appears on the FDA's list of novel drugs approved in 2026, which was current to September 28. Until late-stage results are in, speed claims remain company claims, and no AI-derived candidate should be described as a proven treatment.