An AI Agent Goes Hunting for a New Antibiotic

Google DeepMind has published a step-by-step account of using an AI agent to go looking for a new antibiotic, and the result is a single promising molecule that still has to prove itself at the bench. The work, described in a technical walkthrough, put the company's Antigravity agent to work sifting protein databases for an overlooked antimicrobial peptide, the kind of small protein bacteria use to kill their rivals.

The hunt centred on Streptococcus pseudopneumoniae, a relative of the bacterium behind many lung infections. The agent's job was to find a bacteriocin, a natural antibacterial compound, hiding among protein entries that had never been properly characterised. It settled on one candidate, catalogued only by its database identifier, from the same family as a known antibacterial peptide.

How the search worked

The pipeline strung together tools DeepMind has built over several years. A model called ProtNLM, which predicts a protein's function straight from its amino-acid sequence, was used to put tentative labels on entries that had sat in the UniProt database without one. A standard sequence search then checked the candidate against a curated reference set and found no close matches, a sign it might be genuinely novel. AlphaFold, DeepMind's protein-structure predictor, modelled the molecule and returned a shape resembling a known bacteriocin, which is what you would hope to see if the function labels were right.

It is a tidy demonstration of what these systems can do when chained together, and a reminder of how much is still missing. Everything so far is computational. The candidate has not been made in a lab, let alone shown to kill anything. Wet-lab validation is ongoing, and plenty of promising sequences fall apart at exactly that stage.

A pattern, not a cure

The more interesting signal is the method. An agent ran a multi-step scientific search, from hypothesis to structural check, with far less hand-holding than such work once required. That fits a run of recent results in which models are turned loose on biology. Anthropic's Claude designed protein binders against a scorpion toxin in a multi-day campaign, and other systems have begun mapping chemical reactions at scale. The promise is a faster path to drugs the world badly needs as resistance spreads. The catch is the same every time: a prediction is a starting point, and the laboratory still gets the final say.