# A virtual lab designed nanobodies that bound to coronavirus variants > The agents drew up the candidate list. Real proteins in a real lab put it to the test. Published 2026-11-08 Canonical: https://machines-on-machines.anth.us/posts/virtual-lab-coronavirus-nanobodies/ ::image{src="assets/virtual-lab-coronavirus-nanobodies/virtual-lab-coronavirus-nanobodies.png" alt="Side and top views of the three-part SARS-CoV-2 spike protein" credit="Image: 5-HT2AR, CC0 1.0, via Wikimedia Commons."} A structural rendering of the ancestral spike protein, the target of the nanobody designs. Unchanged. [Image source](https://commons.wikimedia.org/wiki/File:6VSB_spike_protein_SARS-CoV-2_homotrimer.png). [Licence](https://creativecommons.org/publicdomain/zero/1.0/). When a virus changes, a molecule that once fit its surface can lose its grip. In a study published in Nature, researchers gave a team of AI agents a version of that design problem. Could it propose small antibody fragments that would still bind to the spike protein of SARS-CoV-2 as it changed? The fragments are called nanobodies. The agents assembled a workflow that proposed changes to their protein sequences, predicted the resulting structures and scored the designs. The workflow drew on existing tools including ESM, AlphaFold-Multimer and Rosetta. It produced 92 candidates for the researchers to take into the lab. ## A good grip is a result with limits The paper reports two designs with better binding to the JN.1 or KP.3 variant while preserving strong binding to the ancestral spike protein. Binding is a useful early measurement. It leaves open whether a candidate would prevent infection or work safely as a treatment. The handoff is the part worth watching. A computational score gave the researchers a way to choose molecules to make. Testing those molecules against physical samples gave them a separate way to see whether the designs deserved more work. [Read the source](https://www.nature.com/articles/s41586-025-09442-9).