Google's Antigravity agent flags 97-residue antimicrobial peptide candidate in UniProt; wet-lab validation underway

Google DeepMind researchers used Antigravity, the company's agentic development platform, as a science workbench to identify a previously uncharacterized bacteriocin candidate in the UniProt database. According to Google, the peptide has predicted activity against pathogens involved in respiratory tract infections. The finding is computational only. The candidate is now in wet-lab validation, along with several others from the same pipeline.

The post was published October 15, 2026, and is authored by Pengfei Tian, Maxwell L Bileschi, Zoi Tsangalidou, Sukhdeep Singh and Lucy J. Colwell.

The candidate

• Target organism: Streptococcus pseudopneumoniae
• UniProt ID: A0A0U0D591
• Length: 97 amino acids
• Family: Lactococcin 972 bacteriocin
• Novelty check: No meaningful sequence homolog among manually reviewed Swiss-Prot proteins, tested via EMBL-EBI NCBI-BLAST
• Structure: The AlphaFold Protein Structure Database model closely matches the fold of an experimentally determined bacteriocin, which Google says corroborates the ProtNLM prediction
• Genomic context: The gene sits 6 base pairs upstream of an ABC transporter permease and its potential ATP-binding partner. The same operon contains another annotated bacteriocin and an XRE transcriptional regulator, a pattern Google calls strong contextual evidence of a working bacteriocin system
• Hypothesized mechanism: Based on literature review by the agent, action against bacterial cell division, with a predicted spectrum covering several drug-resistant pathogens

How the workflow ran

A researcher set the goal of finding and computationally validating novel antimicrobial peptides (AMPs) in S. pseudopneumoniae. Over a multi-turn session, the researcher chose the questions and evidence while the agent handled cross-database queries, structural analysis and literature synthesis. The steps were:

1. Filter millions of UniProt entries to a short candidate list
2. Check novelty against Swiss-Prot
3. Retrieve and superimpose AlphaFold structures against experimental ones
4. Inspect flanking genes for operon evidence
5. Propose a mechanism from the literature
6. Generate a bench protocol covering protein expression, antimicrobial susceptibility testing, target engagement and in vitro toxicity testing

The discovery relies on ProtNLM, Google's protein language model, which predicts biological function from amino acid sequences. Google says it has been used to add functional labels to millions of unnamed UniProt entries. The labels cover protein names, function comments, subcellular locations, GO terms and keywords.

New tool: Biomolecular Structure Viewer

Google is introducing a 3D Biomolecular Structure Viewer embedded in Antigravity. The agent can pull structures from the AlphaFold Database or from a scientist's own predictions and drive the viewer, for example by highlighting functional regions to explain a hypothesis. Google says the shared visual state works in both directions. The source text is cut off at that point, so further details are not confirmed here. Pricing and availability for the viewer were not disclosed in the source material.

Confirmed vs. claimed

Google states that the candidate "still requires wet lab validation." No experimental activity data, hit rates or benchmark scores were published. The source does not name the underlying language model that powers the Antigravity agent. The figure of an estimated 10 million annual deaths from antimicrobial resistance by 2050 is the post's cited projection, not a new finding.

What this means

This is a workflow demonstration, not a validated discovery. The meaningful claim is about time: Google says the agent compresses days of fragmented database triage into a single session. The individual steps (BLAST, AlphaFold lookups, operon inspection) are standard bioinformatics, so the value lies in orchestration rather than new analytical methods.

The real test is the wet-lab data. If the 97-residue candidate and the others in validation show activity in susceptibility assays, the pipeline gains credibility as a way to mine UniProt's unannotated sequences. If hit rates are low, the speedup mainly benefits hypothesis generation. The post also positions Antigravity, which Google describes as an agentic development platform, as a general-purpose scientific tool rather than only a coding one. That is a notable product direction, but no usage or performance numbers support it yet.