AI agent flags possible antibiotic, not yet tested in lab

Millions of protein sequences in the world's largest public protein database have no known function. Google DeepMind's team went looking inside that gap and came back with a candidate that may fight respiratory infections. The candidate is a predicted antimicrobial peptide, a short protein of the kind bacteria naturally produce to kill their rivals. It was found in Streptococcus pseudopneumoniae, a bacterium from the same family as the pneumococcus.

The stakes are set by the scale of resistance. The authors estimate that antimicrobial resistance could cause 10 million deaths a year worldwide by 2050. Peptides are attractive because, as the authors note, they have evolved to attack basic bacterial structures, from cell envelopes to essential internal machinery, which makes resistance harder to develop. Some are already in use: nisin, a bacteriocin used worldwide, is the best-known example. Recent discoveries from genome mining, such as darobactin, dynobactin and lariocidin, show that the pool is not exhausted.

The obstacle is information. Much of the sequence data produced by genome projects has been deposited in UniProt, the reference resource for protein sequences and functions, without any annotation. The authors describe these entries as the "dark matter" of protein databases. Their tool, ProtNLM (Protein Natural Language Model), is a deep learning model trained to read the evolutionary "grammar" of amino acid sequences and predict a protein's function directly from its sequence, without needing a close known relative. Google says the model has added high-confidence functional labels to millions of previously unnamed UniProt entries.

In the workflow described by the team, the researcher steers the search and the agent does the cross-database work. The agent retrieves structures from the AlphaFold Database, and the scientist examines them in a 3D viewer built into Antigravity, where marking a region of a protein can prompt the agent to reason about that exact spot. Google presents the lead as "biologically plausible" and backed by structural and genomic evidence, but not yet proven. Its predicted activity has not been measured in the laboratory.

For microbiology teams, the immediate change is the front end of the search: fewer weeks spent comparing databases, and more candidates screened per session. For patients, the benefit is still distant, and the path is long. The authors give no date for the wet-lab results, and no candidate from this workflow has yet been shown to work against bacteria, let alone in people. The next public milestone is the laboratory validation of this peptide.