Stanford Medicine researchers have created Virtual Biotech, a system that simulates the work of an entire biotech company. Instead of humans, it employs AI agents.
Stanford Medicine researchers have created Virtual Biotech, a system that simulates the work of an entire biotech company. Instead of humans, it employs AI agents.
Phys.org reports this.
AI agents can search for potential targets for new drugs—for example, proteins or genes that can be affected by a drug to alter the course of a disease. They can also analyze molecular data and clinical trials, and can even suggest promising directions for new drugs.
Virtual Biotech is structured somewhat like a real pharmaceutical company. It is headed by an AI agent, the Chief Scientific Officer. He receives tasks, distributes them among specialized teams, and collects their results. And separate groups of agents deal with genetics, disease biology, molecules, or clinical trials. The study used Anthropic’s Claude models for this, although the authors claim that the system can be built on other large language models.
To test what such a “company” is capable of, the researchers gave it one of the most laborious tasks in pharmacology — to figure out why some drug candidates pass clinical trials and others fail. The system collected data on almost 56,000 trials. And for 37,075 phase II and III trials, it ran a separate agent for each.
Each agent had to dissect a single trial: check its results on ClinicalTrials.gov, search for scientific publications, and if that wasn’t enough, reach out to press releases and official announcements from agencies like the FDA. With tens of thousands of agents working in parallel, this phase was completed in about six hours. The authors estimate that a single agent would have spent more than 76 days on the same task.
After that, the AI workers began looking for patterns that could tell whether a drug had a good chance of passing the next phases of testing. One of them concerned how accurately the drug “hit” a particular type of cell, as well as the activity of the target gene. The agents noticed that the best results were often shown by drugs that targeted genes with unstable activity. In some cells, such a gene was very active, and in others - almost inactive.
In the data analyzed, drugs with such more “pointed” targets were 40% more likely to advance from phase I to phase II clinical trials and 48% more likely to eventually reach the market.
The researchers also decided to test whether the virtual company could create a new drug. To do this, Virtual Biotech was asked to study the protein B7-H3, or CD276, which has long been of interest to lung cancer researchers. The agents analyzed genetic data, gene activity in individual cells, and how the cells are arranged within the tumor.
The system focused on fibroblasts near the tumor. It concluded that cells with high B7-H3 activity could help the tumor suppress the immune system. So the agents proposed attacking B7-H3 with an antibody-drug conjugate — an antibody to which a toxic drug is “attached.” The antibody finds the right cells and delivers the toxic substance directly to them.
For this experiment, the agents used only information available up to January 2025 to see if they could come up with such an approach on their own. A few months later, a pharmaceutical company independently proposed a similar therapy against B7-H3.
The study authors call this independent confirmation that Virtual Biotech came up with this approach on its own. The researchers' next step is to take other targets that Virtual Biotech found and test in real laboratories how many of its ideas will "survive" an encounter with real biology.
Previously, dev.ua talked about OpenAI's GPT-Rosalind for biomedical research and drug development,
