DeepMind shows an Antigravity agent finding a new antimicrobial peptide candidate

What DeepMind showed on 15 October

Google DeepMind published “Accelerating the search for antimicrobial peptides with Google Antigravity” on 15 October 2026. The post, by Pengfei Tian, Maxwell L. Bileschi, Zoi Tsangalidou, Sukhdeep Singh and Lucy J. Colwell, is a technical walkthrough rather than a paper: it follows one session in which a biologist and an AI agent hunted for new antimicrobial peptides inside Antigravity, Google's agentic development platform.

The session ended with what DeepMind calls “a novel antimicrobial peptide candidate which has predicted activity against the pathogens involved in respiratory tract infections.” The candidate is a 97-amino-acid bacteriocin from the Lactococcin 972 family, found in the bacterium Streptococcus pseudopneumoniae. DeepMind says it is “now undergoing wet-lab validation,” along with several other candidates found through the same pipeline.

The post is also a product announcement. It presents Antigravity as an “agentic science workbench” and introduces a new 3D Biomolecular Structure Viewer built into it. It follows DeepMind's push to put its science models in researchers' hands, from the AlphaGenome Atlas to compute credits for European labs.

Why antimicrobial peptides, and why the protein “dark matter”

The motivation is antimicrobial resistance, which the post says could cause an estimated 10 million deaths a year worldwide by 2050. Antimicrobial peptides, and bacteriocins in particular, are small proteins that bacteria make to kill competing bacteria. DeepMind argues they are promising because they evolved to hit fundamental weak points, from cell envelopes to essential internal machinery, which makes resistance harder to develop.

Nature has already supplied useful examples. The post cites nisin, a bacteriocin used worldwide, and recent genome-mined finds such as darobactin, dynobactin and lariocidin. Genomic and metagenomic sequencing has filled public databases with billions of natural protein sequences, so more candidates are almost certainly sitting in them.

The catch is that most of those sequences have never been studied. They come straight from genome sequencing projects and do not resemble any characterised protein, so ordinary similarity searches cannot say what they do. DeepMind calls this the “dark matter” of protein databases, and UniProt, the main public protein resource, holds millions of such entries with unknown function.

ProtNLM, the model that names unknown proteins

Google's answer is ProtNLM, short for Protein Natural Language Model. It is a deep learning model that reads raw amino acid sequences and writes natural-language annotations: protein names, function comments, subcellular locations, Gene Ontology terms and keywords. Its labels have been added to millions of unreviewed UniProtKB/TrEMBL entries that previously had no name.

The candidate in this walkthrough shows how that looks in practice. Its UniProt entry, which AI Lookout checked on 15 October, is unreviewed, first appeared in 2016 from a whole-genome shotgun project, and carries the name “Lactococcin 972 family bacteriocin” with Google's ProtNLM listed as the evidence source. In other words, the clue that started the search was itself a machine prediction.

That is why the rest of the session matters. A name predicted from sequence is a hypothesis; the workflow DeepMind describes is a series of independent checks meant to raise or lower confidence in it before anyone spends money on lab work.

Six steps from a database query to a lab plan

The researcher set the goal: find and computationally validate new antimicrobial peptides in Streptococcus pseudopneumoniae. DeepMind describes a multi-turn session in which the researcher steers strategy and decides what evidence to collect, while the agent handles cross-database queries, structural analysis and literature synthesis.

First, the agent filtered millions of UniProt entries down to a handful worth attention and surfaced A0A0U0D591. Second, at the researcher's direction, it ran the sequence through the EMBL-EBI NCBI-BLAST service against Swiss-Prot, UniProt's manually reviewed section, and found no meaningful homolog, supporting the claim that this is a new peptide rather than a known one.

Third, it fetched the candidate's predicted structure from the AlphaFold Protein Structure Database and compared it with experimentally solved structures to find distant relatives. DeepMind says the fold closely matches an experimentally determined bacteriocin, which corroborates ProtNLM's label. AlphaFold DB gives the model a mean confidence score of 88.5 out of 100.

Fourth, the agent read the gene's neighbourhood. The candidate sits at the start of a tightly coupled operon, six base pairs ahead of an ABC transporter permease and a likely partner ATP-binding protein, next to another annotated bacteriocin and an XRE transcriptional regulator. That arrangement looks like a working bacteriocin system, with machinery to export the peptide and control its production.

Fifth, from the literature the agent proposed that the peptide may act on bacterial cell division and could work against several drug-resistant pathogens. Sixth, it drafted a bench protocol covering protein expression, antimicrobial susceptibility testing, target engagement and in vitro toxicity testing. DeepMind says that plan is already in motion.

A shared 3D structure viewer for scientist and agent

Alongside the walkthrough DeepMind introduced the Biomolecular Structure Viewer, a 3D visualisation tool embedded in Antigravity. The agent can load structures from AlphaFold DB or a scientist's own predictions and drive the viewer itself, for example highlighting regions it thinks matter to explain a hypothesis.

The awareness runs both ways. If the researcher highlights residues or pins a note on a binding pocket, that state is synchronised back to the agent. DeepMind's example is a scientist marking a region and asking, “How might this β-strand affect Lipid II binding?”, so the agent reasons about the exact structural context on screen instead of a description typed into chat.

The argument is about friction. DeepMind says this kind of work used to require custom database scripts, exporting files into separate 3D modelling programs, disconnected genomics tools and several experts to design assays. Putting the pieces in one workspace is meant to turn “days of fragmented database triage into a single collaborative session.”

How researchers can try the workflow

Antigravity's science page says researchers can install a Science plugin in one click from the app's Customizations marketplace, giving agents skills for more than 30 life-science databases, including the AlphaGenome Atlas, UniProt, PubChem and ChEMBL. The same page advertises free access to Gemini 4 Argon at the highest usage tier for 10,000 scientists, an offer that launched with Gemini 4 Argon's general availability.

The plugin's skills are open source in the google-deepmind/science-skills repository under Apache-2.0. When AI Lookout checked, it listed skills for UniProt, AlphaFold DB fetching and analysis, Foldseek structural search, protein sequence similarity search, PyMOL rendering and literature search across PubMed, Europe PMC, bioRxiv, arXiv and OpenAlex. Version 1.3.0 was committed at 14:22 UTC on 15 October.

DeepMind stresses that the walkthrough covered a narrow amino acid window in one species. Researchers can change the filters, look at other sequence lengths, target other priority pathogens, or widen the search from human-associated bacteria to environmental microbes.

Limits: a prediction, not yet an antibiotic

DeepMind is careful about the result, and readers should be too. The post calls the candidate “a biologically plausible lead backed by structural and genomic evidence” that “still requires wet lab validation.” Every piece of evidence so far is computational: a predicted name, a predicted structure, a predicted mechanism and a predicted spectrum of activity.

Most antimicrobial candidates fail somewhere between the screen and the clinic, on potency, toxicity, stability or manufacturing. The post does not share sequences for the other candidates, measured data, or how many were rejected along the way. That contrasts with recent AI biology claims that arrived with lab results, such as DeepMind's AlphaProtein Novo enzymes and the enzyme system Claude helped identify, which scientists then verified.

What the walkthrough does show is a credible, inspectable workflow: each step uses a public database or service a biologist could rerun. The questions to watch are whether DeepMind publishes the wet-lab results, positive or negative, and whether outside labs report finding leads with the Science plugin.