In Experiment, A.I. Agents Quickly Found New Drug

By the Editorial Team · Tuesday, September 22, 2026

Researchers led by Stanford's James Zou used coordinated AI agents to rapidly design novel proteins targeting COVID-19 variants and lung cancer, producing validated candidates in days at minimal cost. This demonstration highlights AI’s potential to accelerate drug discovery by independently generating and vetting treatments, though human oversight and further development remain essential.

In a recent breakthrough, researchers demonstrated that artificial intelligence (AI) agents could rapidly and efficiently design novel proteins for medical applications, including COVID-19 vaccines and cancer treatments. The work, conducted in early 2024 by a team led by Stanford University scientist James Zou and senior platform lead John Pak at Biohub, highlights the potential for AI to function as autonomous co-scientists in drug discovery.

The experiment involved hundreds of AI agents engaging in simultaneous, rapid-fire conversations to brainstorm improvements to a COVID-19 vaccine. Unlike human researchers, these AI agents did not experience distractions such as fatigue or breaks, enabling continuous, high-volume idea generation and evaluation. Within a few days, the system produced dozens of new protein designs predicted to bind effectively to emerging variants of the virus. Laboratory testing validated that two of these proteins successfully performed as intended, achieving results in a fraction of the typical time and cost.

The AI agents, organized as virtual teams with specialized roles spanning immunology, computational biology, and machine learning, operated under the guidance of a so-called “professor AI” that coordinated their efforts. This structure allowed the agents to independently formulate hypotheses, prioritize promising avenues, and specify the research tools needed for further investigation. The researchers noted that early on, the agents’ polite communication style limited the rigor of their debates, prompting the team to introduce more critical “devil’s advocate” personas. Such mechanisms ensured thorough vetting of ideas and helped mitigate risks of unintended consequences.

The project extended beyond COVID-19 research, applying the AI system to study B7-H3, a protein implicated in lung cancer. Thousands of agents formed a virtual biotechnology company that analyzed scientific literature and conducted computational experiments to identify effective strategies to target the protein. Remarkably, their findings aligned closely with a drug candidate developed by Merck, whose recent clinical success reflected a similar therapeutic approach. Dr. Zou emphasized that the AI agents had no access to the latest clinical trial data due to internet restrictions and a knowledge cutoff date, indicating an independent convergence on the effective strategy.

While experts, including Dr. Charles Rudin of Memorial Sloan Kettering Cancer Center, recognize the promise of such AI-driven approaches to accelerate drug development, they caution the work remains a proof of concept. Key challenges persist, such as the AI agents’ inability to perform physical laboratory experiments and questions about their performance on less-studied molecular targets. Additionally, human oversight remains critical at all stages to ensure safety and feasibility.

Despite these limitations, the experiment underscores the growing role of AI in biomedical research. The AI system accomplished its tasks within a single day at a computing cost of just $46, demonstrating significant gains in speed and efficiency. As the research community continues to refine these virtual labs, the integration of AI co-scientists may become a transformative force in the future of drug discovery.

This article was reported by synthesizing coverage from multiple independent news outlets to present a balanced, unbiased account. Read about our editorial standards. Spotted an error? Let us know.