Stanford's Virtual Lab of AI agents designed 92 nanobodies, and lab tests found two strong candidates
A team of LLM agents, led by an agent in the role of principal investigator, built a design pipeline for SARS-CoV-2 nanobodies. A human gave high-level feedback, and human scientists ran the lab tests.
- Organization
- Stanford University and Biohub
- Industry
- Life Sciences
- When
- Jul 2025
- Evidence
- Research paper
- Agents and tools
- Virtual LabESMAlphaFold-MultimerRosetta
Reported outcome
The agents designed 92 new nanobodies. In lab tests, two showed improved binding to the JN.1 or KP.3 variants while keeping strong binding to the ancestral spike protein.
- 92Nanobodies designed
- 2Highlighted candidates after lab tests
Caveats
Described in a paper with a stated method. Peer review status varies.
- The paper calls the two nanobodies suitable candidates for further investigation. They are not treatments.
- Humans ran every lab experiment. The agents cannot do wet-lab work.
- A researcher told The Stanford Daily that the agents were too agreeable with each other and can suggest impractical experiments.
- The speed figures (a few days, one to two hours of agent discussion) come from a press article, not from the paper's abstract.
What the agent does
The Virtual Lab is a team of LLM agents. One agent plays the principal investigator and guides a group of scientist agents through a series of research meetings. A human researcher gives high-level feedback.
The team was asked to design nanobody binders for recent variants of SARS-CoV-2. The agents built a new computational pipeline that uses the protein language model ESM, the folding model AlphaFold-Multimer and the Rosetta software. The pipeline produced 92 new nanobodies.
What happened
Scientists at Biohub tested the designs in the lab. The paper reports a range of working nanobodies. Two showed improved binding to the JN.1 or KP.3 variants and kept strong binding to the original spike protein.
The Stanford Daily reported that the design work took a few days, with the core planning done in one to two hours of agent discussion. The same article gives the limits. The agents do not know what equipment a lab has, so the wet-lab scientists must choose among their suggestions.
What to copy
- Give each agent a clear role, and let one agent run the meeting.
- Keep a person in the loop for direction, not for every step.
- Test the output in the real world before you call it a result.
Starter prompt
Written by us, not by the paper's authors. The paper's own code and agent discussions are on GitHub (zou-group/virtual_lab).
You are the lead of a small research team: {roles}.
Goal: {goal}
Constraints (equipment, budget, time): {constraints}
Run a meeting. Each role gives its view in turn, and must name one weakness in the plan so far.
End with: the chosen approach, the tools to use, and what a human must test.
Flag any step that the listed equipment cannot do.
Sources
- The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies — Nature, Jul 29, 2025
- Stanford researchers develop AI scientists for therapeutic discovery — The Stanford Daily, May 6, 2026
Added Nov 16, 2026 · Sources last checked Nov 16, 2026.
Get new use cases weekly
One short email. Sources included.
The newsletter is paused. Past issues stay in the archive.
Related use cases
FutureHouse's Robin agents proposed ripasudil as a dry AMD candidate; humans ran the lab tests
Robin's analysis agent reported ripasudil raised retinal cell phagocytosis 7.5-fold over controls in a lab assay; a human analysis of the same data found 1.75-fold.
Google's AI co-scientist proposed the same superbug hypothesis an Imperial College team spent years on
The system's top hypothesis matched the Imperial team's unpublished finding about how bacteria spread resistance genes.
Edison's Kosmos AI scientist: independent scientists rated 79.4% of its report statements accurate
Independent scientists found 79.4% of statements in Kosmos reports accurate. Collaborators said one 20-cycle run did the equivalent of 6 months of their own research time on average.