AI-Powered Virtual Biotech Uses 37,000 Agents to Explore New Lung Cancer Treatment
Science & Artificial Intelligence: Researchers have created an experimental “virtual biotechnology company” powered by thousands of artificial intelligence agents that has identified a potential therapeutic strategy for lung cancer while also uncovering biological patterns linked to the success of medicines in clinical development.
The system, developed by a team led by Stanford researcher James Zou, uses a network of specialized AI agents designed to perform different stages of the drug-development process. The research was published in Science on September 17, 2026.
Thousands of AI Agents Work as a Virtual Company
The Virtual Biotech can deploy as many as 37,000 AI agents. Instead of assigning every task to a single AI system, the researchers organized the agents into specialized divisions resembling those found inside a conventional biotechnology company.
A virtual chief scientific officer coordinates activities involving areas such as biological target identification, safety assessment, drug design and clinical-development analysis.
The underlying AI systems were powered by large language models, including versions of Anthropic’s Claude.
Analysis of Nearly 56,000 Clinical Trials
To test the system, researchers assigned AI agents to examine the results of 55,984 clinical trials.
The analysis identified a relationship between the biological characteristics of drug targets and the likelihood of a medicine progressing successfully through development. According to the research, medicines targeting genes that are particularly specific to certain cell types were 48% more likely to reach the market and were associated with 32% fewer adverse events compared with the comparison group.
The agents also examined whether targeted genes behaved more like an “on-off switch” or a gradual “dimmer.” The researchers found that these biological characteristics were associated with differences in drug-development outcomes.
AI Explores B7-H3 as a Lung Cancer Target
The researchers then asked the virtual company to investigate potential therapeutic opportunities in lung cancer.
The AI system focused on B7-H3, a protein found at high levels in certain lung tumors. By combining information from different biological datasets, the agents proposed that B7-H3 could be targeted through an antibody-drug conjugate.
Such a treatment combines an antibody designed to recognize a particular target with a drug payload intended to deliver an anti-cancer compound to cells carrying that target.
Independent Development Provided an Interesting Comparison
One notable aspect of the research was that a pharmaceutical company later independently developed a similar antibody-drug-conjugate strategy targeting B7-H3.
The researchers said this provided an external comparison supporting the plausibility of the AI-generated strategy. However, this does not mean that the Virtual Biotech itself developed an approved cancer medicine or demonstrated clinical effectiveness.
Human Oversight Remains Necessary
Despite the scale of the AI operation, the research does not suggest that human scientists can be removed from the drug-development process.
The researchers emphasize that the conclusions produced by the Virtual Biotech depend on the quality and availability of existing biological and clinical data. Its predictions also require experimental testing before they can be considered reliable evidence for patient treatment.
Independent researchers have similarly cautioned that the system has not yet undergone the full testing required in real-world drug discovery, and its predictions have not themselves been validated through clinical trials.
A New Model for AI-Assisted Drug Discovery
The study demonstrates how multiple specialized AI agents can be coordinated to examine large volumes of biomedical information and connect evidence from different stages of drug development.
Rather than using AI for a single task, the Virtual Biotech attempts to create an integrated research environment capable of moving from biological analysis to therapeutic hypotheses and clinical-development questions.
The researchers say such systems could eventually help scientists explore a much larger number of potential drug-development ideas. For now, however, laboratory experiments and clinical research remain essential steps before any AI-generated therapeutic strategy can be considered an established medical treatment.
