An AI agent and a researcher hunt new antibiotics together

Google DeepMind described a working session in which a researcher and an AI agent together searched for a new antibiotic. The target is a class of molecules called antimicrobial peptides, and specifically bacteriocins, which bacteria evolved to kill each other in ways that make resistance harder to develop. The starting pool was the "dark matter" of protein databases: millions of UniProt entries whose function is unknown. A Google model called ProtNLM predicts function straight from a protein's amino acid sequence, which the team used to flag promising candidates in the bacterium Streptococcus pseudopneumoniae.

What is new is that the whole workflow ran in one place. Inside Google Antigravity, the agent (shown as Gemini 3.8 Flash) handled database queries, ran a BLAST search to confirm the candidate had no known close relative, compared the AlphaFold-predicted structure against experimentally determined ones, inspected the surrounding genes, and pulled together the literature, while the researcher set strategy. A shared 3D structure viewer let both sides work on the same visual: the researcher could highlight a particular strand and ask the agent how it might affect binding. The result was a 97-amino-acid bacteriocin, UniProt ID A0A0U0D591, sitting right next to the transporter genes you would expect of a real system, which the authors call a biologically plausible lead. It and several others are now in wet-lab validation. The honest limit: there is no potency data yet, so the activity is predicted, not shown.

Why it matters

For anyone doing early-stage biology, the interesting part is not the molecule but the shape of the work: database triage, structure comparison, genomic context, and literature review happening in one agentic session instead of across separate tools and specialists. If that holds up once the lab results come back, it changes how fast a small team can move from a database to a testable hypothesis.