“Virtual Biotech,” introduced in a study published in Science, was designed as a virtual biotechnology company made up of AI agents working on drug discovery. The system has the capacity to support up to 37,000 agents that can interact autonomously with large language models or with one another and carry out multistage tasks.
The team, led by Stanford University computer scientist James Zou, modeled the company structure with different departments managed by a “chief science officer” (CSO) agent. These departments included areas of expertise such as target identification and clinical trial design. The study used Claude models developed by Anthropic; Zou stated that any advanced large language model, including open-source models, could be used.
The system analyzed the published results of more than 55,000 clinical trials conducted for different diseases. A total of 37,075 agents were assigned to examine each late-stage clinical trial. The analysis showed that drugs targeting proteins active in specific cell types were approximately 50% more likely to reach the market than other drugs.
In a second test, the system investigated whether the CD276 protein could be a treatment target in lung cancers. In line with previous data, the agents confirmed CD276 as a candidate target and developed a treatment strategy in which an antibody recognizing this protein was attached to a cancer drug. With contributions from external evaluators, the approach was found to be promising.
However, the researchers emphasized that the system had not been tested in real-world drug development processes and that its predictions had not been validated through experiments or clinical trials.
Why it matters
The study’s main significance lies in its attempt to support the prioritization of which biological targets in drug discovery should be investigated further, using a system that examines the results of numerous clinical trials together. The findings offer a metric that could be used in research design by showing that focusing on proteins active in specific cell types may be associated with the likelihood of candidate drugs reaching the market. This approach could add a new layer of analysis to decision-making processes for researchers involved in target selection and clinical trial planning. However, the fact that the system has not been tested in real-world development processes shows that the results need to be tested through laboratory experiments and clinical trials; therefore, the open question is whether the targets identified by artificial intelligence can be validated in practice.
Term: agent
An artificial intelligence agent is software that calls tools and carries out multi-step tasks to achieve a goal rather than producing a single response.