Google Antigravity Finds an Antimicrobial Peptide Candidate in a 97-Amino-Acid Protein

Google DeepMind has published a walkthrough of how biologists used the Antigravity agent platform to find a candidate for a new antimicrobial peptide. The result is a 97-amino-acid bacteriocin with predicted activity against respiratory pathogens. It is now being tested in the lab, so for the moment it is a computational hypothesis, not a finished drug. The process itself is still interesting: work that used to sprawl across several databases and tools fit into one interactive session.

Why look in the "dark matter"

Antimicrobial resistance remains one of the biggest problems in healthcare. The authors estimate that by 2050 it could cause 10 million deaths a year. Antimicrobial peptides, bacteriocins in particular, are seen as a promising alternative to classical antibiotics. They evolved to hit the basic vulnerabilities of competing bacteria, from the cell envelope to intracellular mechanisms, which makes it harder for pathogens to develop resistance.

There are precedents: Nisin is used worldwide, and Darobactin, Dynobactin and Lariocidin are recent finds mined from genomes. The problem is that most sequences in public databases are unannotated. UniProt holds millions of proteins with unknown function. They came straight from genome projects and do not resemble well-studied proteins, so classical similarity search cannot see them.

How Antigravity and ProtNLM fit together

This is where ProtNLM, a protein language model, comes in. It is trained to read the evolutionary "grammar" of amino acid sequences and to predict protein names, function comments, subcellular localization, GO terms and keywords from them. The model has already enriched millions of unlabeled UniProt records.

Antigravity adds an orchestration layer. The researcher sets a goal: find and computationally validate new AMPs in Streptococcus pseudopneumoniae. From there, the researcher runs a multi-turn dialogue and decides which questions to ask and which evidence to gather. The agent takes over the routine work: it browses databases, analyzes structures and pulls together the literature.

In the example described, the chain looks like this:

• the agent filters millions of records down to a few candidates and finds bacteriocin A0A0U0D591, 97 amino acids long, from the Lactococcin 972 family;
• novelty is checked through the EMBL-EBI NCBI-BLAST service against Swiss-Prot: among manually reviewed proteins, there are no significant homologs;
• a 3D model of the candidate is taken from the AlphaFold Protein Structure Database, distant homologs are sought among experimentally determined structures, and the models are superimposed;
• the fold turned out to be close to an experimentally characterized bacteriocin, which supports ProtNLM's prediction;
• the outcome: predicted activity against respiratory infection pathogens and a handoff of the lead to wet-lab validation.

What is not yet proven

The main caveat comes from the text itself: lab validation is under way. Structural similarity and the absence of homologs in Swiss-Prot raise the candidate's priority, but they do not guarantee that the peptide will kill bacteria in a test tube, let alone that it will be safe and stable in the body. The absence of homologs in a curated database also does not equal absolute novelty. The publication shows a method and a single example; the available part of the material contains no hit-rate statistics across many candidates.

The logic is nonetheless sound. The human stays in the role of hypothesis-setter, the agent lowers the cost of each check, and the sources of verification are independent: sequence, structure and literature.

Practical takeaway

For labs and biotech teams, the signal is simple: agent platforms are starting to cover the "find and rank" stage before any reagents or staff time are spent. If you work with genomic data, it is worth building the same pipeline from open components: UniProt, BLAST, AlphaFold DB and a structure viewer. The value will be determined not by the number of candidates found but by the share confirmed in the wet lab. Those are the numbers to wait for from Google next.