AI Virtual Biotech Hunts for Cancer Drug
An AI Workforce That Never Sleeps
A pharmaceutical-company executive’s dream has long been a workforce that never eats, never sleeps, and never gets distracted. Tens of thousands of employees laboring around the clock to find the next blockbuster drug. That dream now has a working prototype, and it is not made of people. An artificial-intelligence system called the Virtual Biotech, described in the journal Nature on 17 September 2025. [1] These are AI systems that autonomously interact with large language models, or with each other, and that can carry out multistep tasks. Large language models are software systems trained on vast amounts of text that can read instructions and produce reasoned responses.
The system uncovered a molecular signal that could help predict whether a clinical trial will succeed. With some human oversight, it also identified a promising treatment for lung cancer. “We want to see how far these agent teams of AI scientists can help us to really accelerate drug discovery and development,” says James Zou, a computer scientist at Stanford University in California who led the effort. [2] The comparison to a living company is not decorative: it is the architecture.
To build it, Zou assembled agents that mirror the staff of a biotechnology company. A chief scientific officer agent directs employees in different divisions, each with its own subspecialties, such as target identification and clinical-trial design. For the study, Zou’s team used versions of Claude, developed by Anthropic in San Francisco, California, as the underlying language model powering the agents. [1] Zou says any advanced language model will do, including open-source models that researchers can run on their own computers. The result is a system that mirrors the division of labor, chain of command, and specialized roles of a real research organization.
Where the Machine Still Needs a Human Hand
The Virtual Biotech has not been vetted in the crucible of real-world drug discovery. Its predictions were not validated through experiments, let alone clinical trials. The system’s output is a hypothesis, not a proven medicine. That distinction matters because the distance between a promising target and a pill a patient swallows is measured in years and in failures. If AI systems can generate drug candidates faster than laboratories can test them, the question of who decides which hypotheses deserve scarce experimental resources becomes urgent?
To test the system’s capabilities, Zou’s team tasked it with analyzing the published results of thousands of clinical trials run for drugs across a wide range of conditions. In recent years, increasingly capable AI systems have taken hold in various fields. In biomedicine, they have shown aptitude for complex tasks ranging from genomic data analysis to hypothesis generation and experimental design. Genomic data analysis means reading the complete set of genetic instructions active in a cell. The Virtual Biotech pushes that aptitude from single tasks toward an entire pipeline.
Other virtual employees searched for predictors of success in data sets showing which genes were active in different cell types. This analysis found that drugs targeting proteins active in specific cell types were likelier to reach market, compared to other drugs. [2] That is a concrete, checkable pattern drawn from thousands of trials — not a guess about one molecule.
One Protein, One Antibody, One Open Question
In a second demonstration, Zou’s team directed the chief scientific officer to investigate whether a protein would make a good therapeutic target for lung cancers. Previous work had suggested that CD276 dampens immune responses and is highly expressed in lung tumours. [2] With this tip-off, the system confirmed CD276 as a candidate using previously collected data and developed a strategy to target it: a CD276-recognizing antibody tethered to an anticancer drug. An antibody is a protein that latches onto a specific molecular shape, and here it serves as a delivery vehicle.
What changes is the scale of hypothesis generation: a single research team once weighed one target at a time, while a virtual company of tens of thousands processes tens of thousands of trials in parallel. What remains open is everything downstream. The lung-cancer strategy has not been tested in a laboratory dish, in an animal, or in a person. The market-likelihood pattern was mined from past trials, not confirmed by a new one.
Sources
1. DOI: 10.1038/d41586-026-02954-y
2. PubMed/NCBI — Quote source (study)