Google Antigravity and ProtNLM Surface a Novel Antimicrobial Peptide Candidate
Google DeepMind published a technical walkthrough showing that Google Antigravity, its agentic development platform, can carry out an end-to-end antimicrobial peptide (AMP) discovery workflow — database search, novelty checking, structural analysis, genomic context, literature synthesis and wet-lab planning — inside one session. The workflow produced a novel bacteriocin candidate with predicted activity against the pathogens involved in respiratory tract infections, and that lead is now undergoing wet-lab validation.
What was announced
The announcement is a demonstrated scientific workflow rather than a new model release. Google Antigravity is described as providing an "agentic science workbench" that lets biologists discover, structurally evaluate and prioritize targets quickly, compressing what used to be a fragmented search over multiple databases into a single session. The post also introduces the Biomolecular Structure Viewer, a 3D visualization tool embedded directly into Antigravity.
The work was published on 15 October 2026 under DeepMind's Science category by Pengfei Tian, Maxwell L. Bileschi, Zoi Tsangalidou, Sukhdeep Singh and Lucy J. Colwell.
Why antimicrobial peptides matter
Antimicrobial resistance is framed in the post as one of the most pressing global health crises, with an estimated 10 million annual deaths globally by 2050 if it is not addressed. Antimicrobial peptides, particularly bacteriocins, are presented as a promising alternative to traditional antibiotics because they have naturally evolved to target fundamental vulnerabilities in competing bacteria — from cell envelopes to essential intracellular machinery — making it difficult for pathogens to develop resistance.
Genomic and metagenomic sequencing has populated global databases with billions of natural amino acid sequences. That pool has already produced globally used bacteriocins such as Nisin, alongside recent genome-mined breakthroughs including Darobactin, Dynobactin and Lariocidin. Many candidates, however, remain locked in the unannotated "dark matter" of protein databases because they are not similar to known proteins.
ProtNLM: predicting protein function from raw sequence
ProtNLM (Protein Natural Language Model) is a deep learning model developed by Google that is trained to read the underlying evolutionary "grammar" of amino acid sequences and predict biological function directly from raw protein sequences. It is used to enrich millions of previously unnamed UniProt entries with high-confidence functional labels.
The model's annotation scope covers protein names, function comments, subcellular locations and GO terms, in addition to keywords. In the pipeline, an input protein sequence passes through a sequence-to-sequence transformer for functional annotation, followed by an "evidencer" pipeline that adds supporting evidence such as sequence and structural similarity. UniProt, described as the world's leading public resource of protein sequence and functional information, holds millions of sequences with unknown functionality transferred directly from genome sequencing projects — the gap ProtNLM is designed to bridge.
The walkthrough: finding a candidate in Streptococcus pseudopneumoniae
The session began with a focused biological goal: find and computationally validate novel AMPs in Streptococcus pseudopneumoniae. The researcher steers the scientific strategy — deciding which questions to ask and what evidence to collect — while the agent handles cross-database orchestration, structural analysis and literature synthesis. Screenshots in the post show the model selected as "Gemini 3.8 Flash."
Step 1 — Navigating a vast search space
The agent queried UniProt and filtered millions of entries down to a handful of candidates. It surfaced a bacteriocin with UniProt ID A0A0U0D591, 97 amino acids long, from the Lactococcin 972 family.
Step 2 — Validating novelty
To verify the discovery was novel, the researcher guided the agent to interrogate the candidate against curated knowledge in Swiss-Prot via the EMBL-EBI NCBI-BLAST service. The result confirmed the peptide has no meaningful sequence homolog among manually reviewed proteins.
Step 3 — Reasoning over 3D structure
The agent retrieved the candidate's 3D model from the AlphaFold Protein Structure Database, queried it against experimentally determined structures to identify remote homologs, and superimposed the models to assess similarity. The structural evidence corroborated ProtNLM's prediction: the candidate's fold closely matches an experimentally determined bacteriocin. The structure viewer in Antigravity allows this structural analysis to be performed in one place.
Step 4 — Reading genomic context
The agent inspected the candidate's genomic neighborhood to test whether surrounding machinery supports the AMP hypothesis. It found the previously uncharacterized candidate sitting upstream of a tightly coupled operon, just 6 base pairs ahead of an ABC transporter permease and its potential partner ATP-binding protein, alongside another annotated bacteriocin and an XRE transcriptional regulator. The post calls this strong contextual evidence that the candidate is part of a working bacteriocin system.
Step 5 — Building a mechanistic hypothesis
Reviewing existing literature, the agent proposed a mechanism of action targeting bacterial cell division and a predicted potency spectrum spanning several drug-resistant pathogens.
Step 6 — Generating a wet-lab plan
The agent translated its in silico findings into an actionable bench protocol spanning protein expression, antimicrobial susceptibility testing, target engagement and in vitro toxicity testing. The candidate, along with several others discovered through the same pipeline, is now in wet-lab validation.
The Biomolecular Structure Viewer
A protein's 3D structure helps explain its biological function, and examining a candidate's predicted fold in space is a key part of the workflow. The Biomolecular Structure Viewer is a 3D visualization tool embedded directly into Antigravity so that this end-to-end workflow lives in a single workspace.
The researcher and agent work over a shared visual state. As the agent retrieves structures from the AlphaFold Database or the scientist's own predictions, it can dynamically drive the viewer — for example, highlighting key functional regions to visually explain its hypotheses. Awareness flows both ways: actions such as highlighting specific residues or pinning notes directly onto a binding pocket are synchronized back to the agent. A researcher can mark a region of interest in 3D space and ask, for example, "How might this β-strand affect Lipid II binding?", and the agent reasons over the exact structural context the scientist is looking at.
What this replaces
Historically, insights like those in this workflow required fragmented and specialized software: writing custom scripts to query databases, manually exporting files to 3D modelling programs, managing disconnected genomic tools, and working with multiple experts to design wet-lab assays. Because wet-lab validation remains resource-intensive, narrowing the search space to only the highest-confidence leads upfront is critical. Antigravity is presented as executing this end-to-end scientific reasoning in one place so scientists can spend more time evaluating high-quality hypotheses, running experiments and solving research challenges.
Caveats and scope
Google DeepMind states plainly that the candidate uncovered in the walkthrough is a biologically plausible lead backed by structural and genomic evidence, but that it still requires wet-lab validation. The session focused on a narrow amino acid window in Streptococcus pseudopneumoniae; the post notes researchers can take the conversation in other directions, such as layering in new filtering criteria, exploring different sequence lengths, or targeting other priority pathogens. Broadening the search from human commensals to environmental microbes and beyond is described as a way to guide the agent across UniProt to discover novel AMP candidates at scale.
No benchmark numbers, hit rates or validation results are reported in the announcement — the wet-lab results are described as still pending and as providing the ultimate test of the agent's analysis.