Researchers at Stanford Medicine have built a drug company that exists only as software. It has no human employees, no lab space and no payroll1. Its “staff” of 37,000 artificial intelligence (AI) agents, which are AI programs that can look things up, use tools and pass work to each other, reviewed roughly 50,000 past clinical trials. They looked for traits shared by drugs that succeeded, and then proposed a new drug idea for lung cancer1. The work is described in a peer-reviewed paper in the journal Science1.
The results are patterns found in old data, plus a drug idea on paper. No medicine was made or given to anyone1.
What happened
The project was created by James Zou, an associate professor of biomedical data science at Stanford Medicine, and Harrison Zhang, a graduate student who is the paper’s lead author1. One account says the Science paper came out in September 20262. The latest coverage of the work appeared on October 26, 20261.
The team set the AI up like a real biotech firm. An AI “chief science officer” oversees specialist teams that work at the same time on different parts of making a drug1. Those roles include finding drug targets, judging safety, choosing what kind of drug to make, and planning trials3.
This builds on an earlier Stanford project. In 2025, Zou’s group described a “virtual lab,” a small team of AI scientists led by an AI principal investigator, in the journal Nature4. In that project, the AI team proposed a new basis for a COVID-19 vaccine using nanobodies, which are smaller, simpler pieces of antibodies4. Zou said then that he or his lab members stepped in about 1% of the time, and that every AI meeting was recorded in a transcript so humans could track progress4.
What the review of 50,000 trials found
Instead of setting general AI programs loose on the scientific literature, the researchers gave each agent one specific trial. Each agent pulled and analyzed the information on that trial’s safety and effectiveness1. The system catalogued roughly 50,000 trials in less than a week, work the researchers said could have taken people years1. Another report gives the exact count as 55,984 trials3.
The agents matched each trial’s outcome to lab measurements of the genes the drug targeted, including data on gene activity in single cells where it existed1. From this, they built two scores1:
- Focus: how much a drug acts on one particular cell type rather than many1.
- Switch-like behavior: whether the targeted gene acts more like a light switch, clearly on or off, or like a dimmer with a smooth range1.
Drugs that scored high on both tended to do better in trials1. Compared with drugs that act more broadly, drugs aimed at switch-like genes were reported to be 40% more likely to move from early safety testing to the next stage, 48% more likely to reach the market, and to have 32% fewer side effects1. The pattern showed up across several kinds of disease, including cancers, brain diseases, heart disease, and kidney and lung conditions1.
The two reports differ on which score these figures belong to. One ties the 40%, 48% and 32% figures to switch-like genes1. The other ties the 48% and 32% figures to drugs aimed at genes active in specific cell types3.
One proposed explanation is that genes with sharply separate on and off states, and with effects limited to certain cells, may be easier to control. That could help both how well a drug works and how safe it is1. That explanation is a hypothesis, not something the study tested.
The lung cancer drug idea
To see whether the system could go beyond analysis, the researchers asked the agents to work on B7-H3, a protein studied in lung cancer that is also called CD27613. Drawing on published studies and public databases, the AI found high levels of B7-H3 in fibroblasts, connective-tissue cells that often sit next to tumor cells1. It also looked at how cells signal each other and at maps of where genes are active in tissue1.
From this, the agents inferred that fibroblasts carrying B7-H3 might send signals that hold back nearby immune cells, which could help tumors escape the body’s defenses1. They then designed an antibody-drug conjugate aimed at B7-H31. This kind of drug joins an antibody, which homes in on a target, to a chemotherapy payload, so the toxic drug is delivered mainly to cells carrying lots of the target1.
The system used only information available before January 2025 for this design1. In August 2025, an established private drug company, not named in the coverage, independently arrived at the same general approach against B7-H31. That company’s drug later received the US Food and Drug Administration’s breakthrough therapy designation, a program meant to speed up promising treatments after signs of effectiveness in a human study1. Zou described the match as independent, third-party support for the virtual biotech’s reasoning and design1.
The team also pointed the system at a failed trial in ulcerative colitis, a bowel disease. The agents suggested the drug, which targeted a protein called OSMRβ, may have fallen short because the body has backup signaling routes. They proposed ways to choose patients by biological markers in future studies3.
What it does not show yet
- Cause and effect. The trial review is a look back at existing data. It shows a link between gene traits and trial success, not proof that picking such genes will make a drug succeed13.
- A working drug. The virtual company did not test a drug in people or produce trial data1. Its role was to generate and rank an idea that still needs normal lab and clinical testing3.
- That the AI’s version would work. A drug company choosing a similar approach is a point of outside comparison. It is not evidence that the AI-designed drug itself would be safe or effective13.
- How often the AI gets it right. The coverage describes one lung cancer example and one failed-trial example, so it is not yet clear how reliable the system’s ideas are in general13.
Who is behind it
The work comes from Zou’s group at Stanford Medicine, with Zhang as lead author1. The sources reviewed for this article do not say who funded the project, and they do not name the drug company whose B7-H3 drug was used as the comparison. Speaking to NBC Bay Area, Zou said the aim is not to remove human scientists but to give them more time for higher-level creative and scientific work3.
How it fits the field
Making a drug is expensive and risky, because many candidates that look good in the lab fail once they are tested in people3. Trial information is scattered across journal articles, regulatory records and registries, so pulling out patterns by hand takes a long time3. The pitch for a system like this is that it could weigh many kinds of evidence before a company commits money, dropping weaker ideas sooner3.
Biohack recently covered another side of AI in medicine: a company’s AI-designed pill that shrank a rare cancer in 2 of 5 people at the top dose. That drug reached human testing. The Stanford project has not. It is a test of whether AI agents can do the early thinking of a drug company.
What to watch next
- Whether other groups check the trial-success patterns in separate data, and settle which score the reported percentages belong to.
- Whether any AI-proposed drug idea from the virtual biotech is made and tested in a lab.
- Progress of the drug company’s B7-H3 drug, which serves as the outside comparison for the AI’s design1.
- Whether the suggested patient-selection approach for the failed ulcerative colitis drug is tried in a future study3.
