Google DeepMind has used its Antigravity workbench to assemble evidence for an antimicrobial protein candidate that is now undergoing wet-lab validation. The 15 October demonstration follows a researcher-guided search from a database entry through sequence, structural and genomic checks.[1]
The selected candidate is a 97-amino-acid protein, UniProt accession A0A0U0D591, from Streptococcus pseudopneumoniae and associated with the Lactococcin 972 bacteriocin family. Its proposed activity remains a computational hypothesis pending experimental results.[1][2]
Several kinds of evidence contribute to the hypothesis
The workflow starts with ProtNLM, a model that predicts biological function from amino-acid sequences. Antigravity helps the researcher narrow a large collection of entries to candidates that merit closer attention. A predicted functional label supplies a lead to investigate, rather than an experimental observation about what the protein does.[1]
The researcher then directs a comparison against manually reviewed Swiss-Prot proteins. Google reports no meaningful reviewed sequence match for the candidate. That suggests it is less characterised in that reference set; it does not establish that every aspect of the molecule or its biological role is novel.[1]
Structural analysis adds a different view. The agent retrieves an AlphaFold model and compares it with experimentally determined structures, finding a fold similar to a known bacteriocin. Sequence and structural resemblance can support a functional hypothesis even when either comparison alone would be inconclusive.[1]
The surrounding genes provide further context. Google describes nearby transport machinery, another annotated bacteriocin and a transcriptional regulator. The arrangement is consistent with a functional bacteriocin system, strengthening the reason to investigate this particular protein. It remains contextual evidence rather than a measurement of antibacterial activity.[1]
How the workflow builds a case for testing
| Evidence | Contribution |
|---|---|
| Sequence and predicted annotation | Identifies a plausible candidate and its proposed function |
| Reviewed-protein comparison | Checks how closely it resembles already characterised sequences |
| Predicted structure and structural matches | Looks for relationships that sequence similarity may miss |
| Neighbouring genes | Tests whether the biological setting fits the proposed role |
The researcher directs the scientific questions
Google’s walkthrough describes an interactive session in which the scientist chooses the biological goal and decides what evidence to seek next. The agent handles movement among databases, structural analysis and literature. That orchestration can reduce the effort of maintaining context across tools while leaving the interpretation open to challenge.[1]
This is a useful distinction in AI-assisted science. A list of plausible proteins may be difficult to act on; an inspectable chain of evidence gives a researcher a reason to prioritise one and a record to revisit if the experiment fails. The workbench’s contribution is to connect those steps coherently.[1]
The announcement says laboratory validation is underway but reports no completed efficacy, selectivity or toxicity result for this candidate. It should therefore be described as a lead under investigation. The value demonstrated so far is a documented route from an uncharacterised database entry to a biologically motivated test.[1]
Sources & further reading
- Accelerating the search for antimicrobial peptides with Google Antigravity Google DeepMind · 15 Oct 2026 · Accessed 27 Oct 2026
- UniProtKB protein entry A0A0U0D591 UniProt · Accessed 27 Oct 2026