A protein sequence can sit in a database without anyone knowing its function. In a Google DeepMind walkthrough, a researcher uses Antigravity to investigate one such sequence as a potential antimicrobial peptide, following the evidence across protein annotations, three-dimensional structures and neighbouring genes. The result is a candidate selected for laboratory validation.[1]
The workflow addresses a familiar research bottleneck: the evidence needed to assess a molecule is spread across different tools. A scientist may find a promising annotation in one database, inspect a predicted fold in another and search the literature separately. Antigravity coordinates those steps in a shared workspace while the researcher directs the questions.[1]
From a sequence to a biological hypothesis
The search begins with ProtNLM, a model that predicts protein functions from amino-acid sequences. In the demonstrated session, the researcher looks for antimicrobial candidates in Streptococcus pseudopneumoniae. A predicted function provides a lead to investigate, rather than an experimental measurement of what the protein does.[1]
The next step checks structure. The agent retrieves a predicted three-dimensional model from the AlphaFold Protein Structure Database and compares it with known structures. In this case, the candidate’s fold resembles an experimentally characterised bacteriocin—a molecule bacteria use against other bacteria. Structural resemblance adds a different kind of evidence to the sequence-level prediction.[1]
The genomic neighbourhood supplies a further clue. The candidate lies close to genes associated with transport and regulation, alongside another annotated bacteriocin. A gene’s neighbours can help explain the system it belongs to: a molecule used outside a bacterial cell needs machinery that can produce, move or regulate it. The context makes the bacteriocin interpretation more coherent.[1]
Keeping the scientific reasoning in one place
The workbench includes a biomolecular structure viewer. A researcher can mark a region of a protein, inspect how it aligns with another structure and ask the agent about the selected area. That shared visual context reduces the need to describe a location in words or move files repeatedly between separate programs.[1]
The scientific value is the assembled chain of evidence. A predicted annotation, a similar fold and a compatible genomic neighbourhood each support a particular hypothesis. Keeping them together makes it easier to see why a lead was prioritised and which part of the reasoning a laboratory experiment should test.
DeepMind says the selected candidate is undergoing wet-lab validation. The walkthrough does not report measured antimicrobial potency, toxicity or clinical efficacy. The next experimental task is to establish whether the proposed activity occurs and at what concentration, before any claim about a therapeutic application.[1]
The workflow can be adapted to a different organism, sequence length or biological target. Its immediate contribution is to organise a search that would otherwise require many separate queries, scripts and visual inspections. That gives researchers a more developed case for the molecules they choose to take to the bench.
Sources & further reading
- Accelerating the search for antimicrobial peptides with Google Antigravity Google DeepMind · 15 Oct 2026 · Accessed 25 Oct 2026