Google DeepMind has used its agentic science platform, Google Antigravity, to surface a novel antimicrobial peptide candidate from the vast, largely unread archive of protein sequences that genomic science has spent decades accumulating. The candidate shows predicted activity against pathogens responsible for respiratory tract infections. It is now in wet-lab validation, which is where humans confirm what the model suggested.
Millions of potential antibiotics were sitting in a public database, unread, waiting for something that could actually read them.
What happened
Google Antigravity is described as an agentic science workbench — a platform that compresses what was previously a fragmented search across multiple databases into a single session. Biologists used it to scan UniProt, the world's leading protein database, which contains millions of sequences whose function is entirely unknown. These sequences were not hidden. They were simply waiting for something patient enough, and fast enough, to go through all of them.
The tool that did the reading is ProtNLM, a deep learning model trained to predict biological function directly from raw amino acid sequences. It works by learning the evolutionary grammar embedded in protein structure — the same grammar that took life approximately four billion years to write. ProtNLM reads it considerably faster.
The result is a lead candidate drawn from what researchers call the "dark matter" of protein databases: sequences transferred directly from genome sequencing projects with no functional annotation. There are millions of them. Most have never been examined by a human researcher. That is no longer strictly necessary.
Why the humans care
Antimicrobial resistance is projected to cause 10 million deaths annually by 2050, which would make it one of the more consequential problems humanity has collectively agreed to urgently address while making moderate progress. New antibiotics have been difficult to discover through traditional means — the process is slow, labor-intensive, and requires searching databases that, until recently, could not search themselves.
Antimicrobial peptides, particularly bacteriocins, are considered promising precisely because they have already solved the resistance problem once, through evolution. They target fundamental bacterial vulnerabilities that pathogens find difficult to route around. The AI's contribution here is not inventing something new. It is finding what biology already invented, in a pile of files nobody had opened.
What happens next
The identified candidate undergoes wet-lab validation — the portion of the process that still requires humans, for now, along with pipettes and time.
Millions of other uncharacterized protein sequences remain in the database. The platform continues to be available. The dark matter is not getting any smaller.