Google researchers used the Antigravity agent platform to find a previously uncharacterized peptide that may act against bacteria involved in respiratory infections. The result is a computational lead, not a validated antibiotic: Google says the molecule and several other candidates are now undergoing laboratory tests.
The workflow combined ProtNLM, a model that predicts protein function from amino-acid sequences, with searches across protein, structure and literature databases. Starting from millions of poorly characterized UniProt records, the agent narrowed the field to a 97-amino-acid protein from the Lactococcin 972 family.
Researchers then directed the system to look for independent evidence. It found no meaningful sequence match among manually reviewed proteins, compared the predicted three-dimensional fold with known structures and inspected nearby genes. The candidate sits just six base pairs before transporter-related genes and near another annotated bacteriocin, evidence consistent with a bacterial defense system.
Antigravity also drafted a bench-testing plan covering protein production, antimicrobial activity, target engagement and toxicity. A new Biomolecular Structure Viewer lets a scientist and agent inspect the same three-dimensional model and exchange annotations. The central limitation remains experimental: structural and genomic clues can support a hypothesis, but only wet-lab work can establish whether the peptide is effective or safe.