A 97-amino-acid protein in a database has become a candidate for antimicrobial testing after a researcher-guided search with Google Antigravity. The workflow connected predictions about protein function with sequence matches, structural relatives and neighbouring genes. DeepMind’s 15 October walkthrough says the lead is now undergoing laboratory validation.[1]
Scientists used the Antigravity workbench with ProtNLM, a model that predicts biological function from amino-acid sequences. Their demonstration focused on Streptococcus pseudopneumoniae and identified a possible bacteriocin, a peptide that may inhibit competing microorganisms. The proposed activity remains a hypothesis awaiting experimental results.[1]
Finding a function for an unreviewed protein
Modern sequencing can reveal vast numbers of genetic sequences without revealing what their encoded proteins actually do. A protein can lack close matches among well-characterised examples and still play an important role in its organism. This creates an enormous triage problem: researchers need to search among millions of unreviewed entries, select those with promising functional clues and decide which deserve expensive experiments. ProtNLM is intended to supply one type of clue by translating sequence patterns into proposed functions. Its output remains a model-generated inference.[1]
The candidate is UniProt accession A0A0U0D591, associated with the Lactococcin 972 family. The team checked for meaningful sequence matches among curated Swiss-Prot proteins, then compared predicted structures with known bacteriocins. The structural search offered a way to investigate related function even where a simple sequence match was weak.[1]
The genomic neighbourhood supplied another clue. The candidate sits near transport and regulatory machinery, including an ABC transporter permease and a possible partner ATP-binding protein. DeepMind highlights a separation of six base pairs between the candidate and nearby transport machinery. Such proximity can support a hypothesis about production and export, without establishing the peptide’s activity.[1]
Following the evidence across databases
The scientist chose the organism, challenged apparent novelty and steered the searches. The agent moved among protein records, predicted structures, genome annotations and literature before proposing an experimental protocol. Its contribution was continuity: evidence from one tool informed the next question, with identifiers that could be checked along the way.[1]
From a candidate to an experiment
The laboratory work must establish whether the peptide can be produced and whether it inhibits the relevant bacteria at measurable concentrations. Selectivity, stability and toxicity would shape further development. Those results will determine whether the computational clues converge on useful biological activity.[1]
The workflow turns an uncharacterised sequence into an explicit, testable proposal. That is a valuable intermediate product: a researcher can see why the candidate was selected and identify which predicted relationship a positive or negative experiment would resolve.[1]
Sources & further reading
- Accelerating the search for antimicrobial peptides with Google Antigravity Google DeepMind · 15 Oct 2026 · Accessed 26 Oct 2026