The matching of the cancer treatment design proposed by the AI agents with a real drug that received medical approval represents a historical proof of the machine’s ability to drive biological discoveries.
Article Index:
- Virtual biotech system in Science journal
- Company structure: Scientific leadership and 37,000 independent agents
- Analyzing 55,000 clinical trials in one week
- Innovating a lung cancer treatment via the B7-H3 protein
- Independent validation and medical approval for Glaxo
- Operational boundaries and the necessity of real-world laboratory experiments
- Frequently asked questions
Virtual biotech system in Science journal
In an unprecedented scientific breakthrough paving the way to reshape global drug research mechanisms, a research team from Stanford University in the United States has successfully built an integrated artificial intelligence system that fully simulates the structure and procedures of a real biotechnology company. Details of this innovation were published on Thursday in the prestigious scientific journal “Science”.
This system — which researchers named “Virtual Biotech” — handles all complex tasks in drug discovery pipelines; ranging from identifying molecular targets and disease-causing proteins to the chemical design of therapeutic compounds and planning clinical trials with extreme precision and without direct human intervention.
Company structure: Scientific leadership and 37,000 independent agents
This revolutionary system was developed by Associate Professor of Biomedical Data Science James Zou, in collaboration with graduate student Harrison Zhang. The system distributes up to 37,000 independent intelligent agents operating across specialized functional departments that mimic real companies, including a Chief Scientific Officer and specialized teams in biochemistry, bioinformatics, and clinical trials.
The system relies on the “Claude” language model developed by Anthropic as a knowledge and reasoning foundation to simulate scientific thinking; agents debate biological hypotheses, review colleagues’ results, and make collective research decisions that accelerate the pace of medical innovation.
Analyzing 55,000 clinical trials in one week
To test the system’s efficiency, researchers directed the virtual company’s agents to analyze a massive database containing over 55,000 historically registered clinical trials. In less than a single week, the system completed precise inferential analysis that would have taken human researchers long decades, uncovering vital patterns critical for predicting drug success.
The agents concluded that drugs targeting active proteins in specific cell types increase their chance of reaching commercial markets by 50 percent compared to other drugs, while drugs targeting genes with switch-like on-off behaviors increase their chance of moving from phase one to phase two by 40 percent and reduce their side effects by 32 percent.
Innovating a lung cancer treatment via the B7-H3 protein
Researchers then moved on to the greater challenge: testing the virtual company’s ability to design a real, viable biological treatment to combat malignant lung tumors. The team directed agents to investigate a specific protein known as “B7-H3” or “CD276”, which plays a pivotal role in helping cancer cells evade the immune system.
Relying exclusively on medical knowledge and research papers published prior to January 2025 only, the artificial intelligence agents invented an advanced therapeutic strategy relying on an “antibody-drug conjugate” to target this protein and destroy cancer tissues with surgical precision.
Independent validation and medical approval for Glaxo
The resounding scientific surprise occurred when global pharmaceutical giant “GlaxoSmithKline” announced several months later its independent development of the exact same therapeutic strategy; its scientists invented a treatment relying on the same antibody conjugate targeting the exact same protein, and the company’s drug subsequently received “Breakthrough Therapy” designation from the U.S. Food and Drug Administration.
Professor James Zou stated in remarks to “Stanford Medicine”: “This match served as an extremely exciting, independent validation by an external third party confirming that the design and results suggested by our virtual company match real-world advanced pharmaceutical industry standards.”
Operational boundaries and the necessity of real-world laboratory experiments
Despite this exceptional success, scientists and specialists emphasized that the Virtual Biotech still undergoes academic evaluations and has not yet been tested under complex actual drug discovery conditions, warning against relying on algorithmic predictions without biological examination and real clinical trials on patients.
Zou himself acknowledged that physical laboratory tests and human intervention remain an indispensable cornerstone, confirming that his team’s next step consists of transferring the new results and hypotheses generated by the agents to live laboratories to test their endurance and effectiveness in complex biological environments.
Frequently asked questions
Question: How many artificial intelligence agents work in Stanford’s virtual biotech company?
Answer: The system includes up to 37,000 independent intelligent agents working across specialized departments mimicking global pharmaceutical companies.
Question: How did GlaxoSmithKline prove the accuracy and success of the virtual model?
Answer: Glaxo independently arrived at the same lung cancer treatment design proposed by the agents, which received breakthrough therapy approval from the FDA.
Question: What language model did Stanford researchers base this system on?
Answer: The team relied on Anthropic’s Claude model as a core of thought and cognitive reasoning to simulate scientific discussions and decisions.
