Accelerating the search for antimicrobial peptides with Google Antigravity
Google DeepMind published a technical walkthrough describing how biologists used Google Antigravity, its agentic development platform, to identify a novel antimicrobial peptide (AMP) candidate with predicted activity against the pathogens behind respiratory tract infections. The work, credited to Pengfei Tian, Maxwell L. Bileschi, Zoi Tsangalidou, Sukhdeep Singh and Lucy J. Colwell, centers on a Streptococcus pseudopneumoniae search. The lead the researchers surfaced is now undergoing wet-lab validation, so its activity remains a prediction rather than a confirmed result.
The motivation is antimicrobial resistance, which the post says could cause an estimated 10 million deaths globally per year by 2050. AMPs, and bacteriocins in particular, are attractive because they evolved to attack vulnerabilities in competing bacteria, from cell envelopes to essential intracellular machinery, which makes resistance harder to develop. Genomic and metagenomic sequencing has filled public databases with billions of natural amino acid sequences. That pool has already yielded globally used bacteriocins such as Nisin and genome-mined discoveries including Darobactin, Dynobactin and Lariocidin. Many sequences remain uncharacterized because they resemble no known protein, including a large share of UniProt entries with unknown function.
ProtNLM (Protein Natural Language Model) is Google's answer to that gap: a deep learning model trained to read the evolutionary grammar of amino acid sequences and predict biological function directly from raw sequence. According to the post, it enriches previously unnamed UniProt entries with high-confidence functional labels covering protein names, function comments, subcellular locations, GO terms and keywords. In practice, this is what lets a researcher scan millions of uncharacterized proteins without needing a homolog to already exist in a curated database.
The Antigravity session itself is framed as collaborative rather than automatic. A researcher sets a goal, in this case finding and computationally validating novel AMPs in Streptococcus pseudopneumoniae, then steers strategy across multiple turns while the agent handles cross-database orchestration, structural analysis and literature synthesis. The workflow compresses what used to be a fragmented search across several databases into one session.
The concrete steps are worth following. The agent filtered millions of entries down to a shortlist and found a bacteriocin, UniProt ID A0A0U0D591, 97 amino acids long, from the Lactococcin 972 family. To test novelty, the researcher directed the agent to check the candidate against curated knowledge in Swiss-Prot via the EMBL-EBI NCBI-BLAST service, confirming no meaningful sequence homolog among manually reviewed proteins. For structural evidence, the agent pulled the candidate's 3D model from the AlphaFold Protein Structure Database, queried it against experimentally determined structures to find remote homologs, and superimposed models to assess similarity. The fold closely matched an experimentally determined bacteriocin, corroborating ProtNLM's prediction.
The agent also reasoned about genomic context. Inspecting flanking genes, it found the uncharacterized candidate sitting 6 base pairs upstream of a tightly coupled operon containing an ABC transporter permease and its potential ATP-binding protein partner, alongside another annotated bacteriocin and an XRE transcriptional regulator. Tight coupling to transporter machinery is a pattern that supports the bacteriocin hypothesis, since such systems typically need dedicated export and immunity components.
From the literature, the agent proposed a mechanism targeting bacterial cell division and a predicted potency spectrum spanning several drug-resistant pathogens. It then converted the in silico findings into a bench protocol covering protein expression, antimicrobial susceptibility testing, target engagement and in vitro toxicity testing. Several candidates from the same pipeline, not just this one, are now in wet-lab validation.
Google also introduced the Biomolecular Structure Viewer, a 3D visualization tool embedded directly in Antigravity. The agent can drive the viewer as it retrieves structures from the AlphaFold Database or the scientist's own predictions, highlighting functional regions to explain hypotheses. The awareness runs both ways: a researcher can highlight residues or pin notes onto a binding pocket and those actions synchronize back to the agent, enabling questions about a specific structural context, for example how a β-strand might affect Lipid II binding. That shared state is the interesting design choice, because it removes the file-export step between a language agent and a molecular graphics program.
The trade-offs are worth stating plainly. All of the validation described here is computational. Structural similarity to a known bacteriocin, a plausible operon neighborhood and literature-derived mechanism make a biologically plausible lead, but predicted activity is not measured activity. The post itself notes the candidate still requires wet-lab confirmation. A second limitation is scope: the session explored a narrow amino acid window in a single organism, and the authors describe broadening to human commensals, environmental microbes and other priority pathogens as future direction rather than completed work.
Why it matters: drug discovery teams and computational biologists get a template for compressing database triage, novelty checks, structure retrieval and protocol drafting into one agent session, with a 3D viewer that keeps human and agent looking at the same structure. If the wet-lab results hold, the same pattern could extend to other target classes where unannotated sequence space is large. For platform teams, the architectural point is the shared visual state between agent and scientist, which is a more general pattern than antimicrobial peptides. Whether that translates into faster validated leads is an open question the current evidence cannot answer.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Accelerating the search for novel antimicrobial peptides using Google Antigravity for Science.