Google Antigravity helps biologists find an antimicrobial peptide candidate

Google DeepMind has published a technical walkthrough of Google Antigravity, its agentic development platform, used as what the post calls an agentic science workbench for biologists. The example is a hunt for new antimicrobial peptides (AMPs). The post says the session uncovered a novel AMP candidate with predicted activity against pathogens involved in respiratory tract infections. The lead is now undergoing wet-lab validation, and the post says it still requires it.

The motivation is antimicrobial resistance. The post says it could result in an estimated 10 million annual deaths globally by 2050. It describes AMPs, particularly bacteriocins, as a promising alternative to traditional antibiotics, because they evolved to target fundamental weaknesses in competing bacteria, which makes it hard for pathogens to develop resistance. It points to Nisin as a widely used bacteriocin and to the genome-mined finds Darobactin, Dynobactin and Lariocidin. Many more candidates, it says, sit in the unannotated "dark matter" of protein databases. UniProt, which the post calls the world's leading public resource of protein sequence and functional information, holds millions of sequences with unknown function.

To read that dark matter, Google built ProtNLM (Protein Natural Language Model), a deep learning model trained to read the evolutionary "grammar" of amino acid sequences and predict biological function directly from raw sequences. The post says it has enriched millions of previously unnamed UniProt entries with high-confidence functional labels. Antigravity lets researchers scan these entries using ProtNLM.

In the walkthrough, the researcher and the agent start from a focused goal: find and computationally validate novel AMPs in Streptococcus pseudopneumoniae. Over an interactive, multi-turn session, the researcher steers the strategy, deciding which questions to ask and what evidence to collect. The agent does the cross-database orchestration, structural analysis and literature synthesis. The post claims this compresses a fragmented search over multiple databases into a single session, and turns days of database triage into one collaborative session. This session used a narrow amino acid window in one species. The post says researchers could add filtering criteria, try other sequence lengths or target other pathogens, and could widen the search from human commensals to environmental microbes.

The post also introduces the Biomolecular Structure Viewer, a 3D visualization tool embedded in Antigravity. The researcher and agent share one visual state. The agent can pull structures from the AlphaFold Database or from the scientist's own predictions and drive the viewer, for example by highlighting functional regions to explain a hypothesis. Awareness runs both ways: when the scientist highlights residues or pins notes on a binding pocket, those actions are synchronised back to the agent. A researcher can mark a region and ask, for example, "How might this β-strand affect Lipid II binding?", and the agent reasons over that exact structural context.

The post contrasts this with earlier practice: custom scripts to query databases, manual exports to 3D modelling programs, disconnected genomic tools, and several experts to design wet-lab assays. Because wet-lab validation is resource-intensive, it argues, narrowing the search to the highest-confidence leads up front is critical. The post closes by saying Antigravity should let scientists spend more time evaluating hypotheses and running lab experiments. The candidate itself is described as biologically plausible and backed by structural and genomic evidence.

Key facts

Google DeepMind's walkthrough uses Google Antigravity, its agentic development platform, with the ProtNLM model to surface a novel antimicrobial peptide candidate in Streptococcus pseudopneumoniae.
The candidate has predicted, not measured, activity against pathogens involved in respiratory tract infections; the lead is now in wet-lab validation.
ProtNLM predicts protein function from amino acid sequences and, per the post, has added high-confidence functional labels to millions of previously unnamed UniProt entries.
Google introduces the Biomolecular Structure Viewer, a 3D tool inside Antigravity where highlights and notes made by the researcher are synchronised back to the agent.
The post claims the workflow turns days of fragmented database triage into a single collaborative session, without giving measured time savings.

Why it matters

Antimicrobial resistance is the backdrop: the post says it could result in an estimated 10 million annual deaths globally by 2050. Finding new antibiotics is slow, and the post argues that peptides hidden among millions of uncharacterised protein sequences are a promising source. The story is mainly a showcase of an agent working across databases, a function-prediction model and a shared 3D view in one workspace. The scientific result is early: a candidate that still needs lab testing.

Who it affects

The audience is biologists and drug-discovery researchers who search protein databases such as UniProt for antimicrobial peptides and who now juggle separate scripts, databases, modelling programs and genomic tools. It also concerns anyone who wants to examine predicted structures, for example from the AlphaFold Database, alongside an AI agent. The post names no specific biologists who ran the session.

How to use it

The post describes a workflow rather than a recipe. A researcher sets a focused goal, such as finding novel AMPs in one species, then steers a multi-turn session while the agent queries databases, analyses structures and summarises literature. In the Biomolecular Structure Viewer the researcher can mark a region in 3D and ask a question about it. Filters, sequence lengths and target pathogens can be changed along the way. The post gives no availability, pricing or access details for the viewer or the science workbench.

How solid is it

This is a vendor's own technical walkthrough, not a paper. The candidate's activity is predicted, and the post itself says it still requires wet-lab validation, which is under way. The post describes it as biologically plausible and backed by structural and genomic evidence. The peptide is not named and its sequence is not given. No number of candidates screened or shortlisted is reported, and nothing compares the approach with other methods. The time saving is stated only as days compressed into one session. The 10 million deaths figure is a projection with no cited source in the text.

Risks and caveats

No laboratory results are reported, so the candidate may not show activity when tested, and the post gives no outcome or timeline for the validation. The session covered a narrow amino acid window in a single species; scaling to other filters, lengths or pathogens is described as possible, not shown. The speed-up and "barrier lowered" claims are the authors' own and unmeasured. The post does not say which AI model powers the Antigravity agent.

“How might this β-strand affect Lipid II binding?”

— Example question a researcher could ask the agent, from the Google DeepMind post