Google’s Antigravity follows sequence, structure and neighbouring genes to an antimicrobial candidate

Google DeepMind has demonstrated an agent-assisted search that follows a protein from a database entry to a proposed antimicrobial experiment. The candidate is a 97-amino-acid sequence associated with the Lactococcin 972 bacteriocin family. Researchers used Antigravity to connect sequence searches, predicted structure, neighbouring genes and literature into a case for testing it in the laboratory.[1]

In the 15 October demonstration, the team focused on proteins from Streptococcus pseudopneumoniae. ProtNLM, a protein-language model, supplied functional predictions, and the workflow selected UniProt accession A0A0U0D591 for closer examination. Its proposed antimicrobial role is a computational hypothesis; the announcement did not report a validated treatment or completed efficacy result.[2][1]

A scientific detective story in several databases

Finding a short candidate sequence is relatively easy to describe and difficult to trust. Familiar-looking amino acids may belong to proteins with very different functions. To test novelty, the researchers directed their agent to compare the sequence against curated proteins in Swiss-Prot. The lack of a meaningful reviewed sequence match suggested a less-characterized candidate, not proof of a new mechanism. They then considered predicted three-dimensional structure, drawing on AlphaFold database entries and structure-search tools to look for remote relatives that sequence matching alone might miss.[1]

Structural resemblance can strengthen a hypothesis about a protein's role, but two similar folds do not always have the same biological activity. The workflow therefore looked beyond the protein itself and inspected neighbouring genes. According to the demonstration, the candidate lies near an operon containing an ABC transporter permease, a possible partner ATP-binding protein, another annotated bacteriocin and a transcriptional regulator. The article highlights a separation of only six base pairs between the candidate and nearby transport machinery. That genomic neighbourhood is an additional piece of biological context: organisms often arrange related functional genes together.[1]

The agent used the assembled evidence to propose a mechanism and an experimental plan. Those are useful outputs when they expose a hypothesis that can fail cleanly in a test. The announcement had not established antibacterial efficacy, selectivity or toxicity for the candidate, so those measurements remain the basis for deciding whether to pursue it.[1]

Connecting tools while the researcher steers

A researcher performing the same search without an integrated workbench would move among sequence databases, similarity-search services, structural viewers, literature, genome annotations and experimental plans. Each hand-off creates an opportunity to lose context or misinterpret a result. The agent's visible contribution is orchestration: assembling and cross-checking evidence across tools while the scientist decides whether each new query makes biological sense. Google's description repeatedly emphasizes researcher steering rather than an autonomous machine discovering a validated antibiotic alone.[1]

The workflow’s appeal is the continuity between tools. A result from one database informs the next query, while the researcher can challenge the interpretation and change direction. Preserving that sequence of evidence also helps explain a negative experiment: the team can identify which predicted relationship failed instead of starting again from an opaque recommendation.[1]

What the first experiment needs to answer

Useful follow-up reporting would specify which candidate proteins were synthesized or expressed, how many were actually tested, the organisms and strains involved, the inhibitory concentrations and independent controls. It should distinguish active compounds from those merely predicted to be active. If the candidate works, the next questions are stability, resistance, toxicity, manufacturing and whether it adds anything beyond known antimicrobials.[2][1]

The candidate’s value now lies in the experiment it motivates. Sequence, shape and genomic context point towards a biological function, while the laboratory can determine whether the molecule has the predicted activity. Antigravity’s contribution is to make the route between those two stages more coherent and easier for the scientist to inspect.[1]

Report an error or suggest a correction