Accelerating the search for antimicrobial peptides with Google Antigravity

Summary

Google DeepMind researchers used Google Antigravity, Google's agentic development platform, as a research assistant to find candidate antimicrobial peptides in bacterial protein databases. A researcher set the goal and steered. The agent ran database queries, structural comparisons, and literature synthesis over a multi-turn session, then drafted a wet-lab plan.

Key points

• The agent narrowed millions of UniProt entries, some annotated by Google's ProtNLM model, to a 97-amino-acid bacteriocin from Streptococcus pseudopneumoniae with no meaningful homolog in Swiss-Prot.
• It gathered supporting evidence: an AlphaFold structure matching a known bacteriocin fold, and a genomic neighbourhood consistent with a working bacteriocin system.
• It proposed a mechanism and activity spectrum, and drafted an experimental plan (expression, susceptibility testing, toxicity).
• A 3D Biomolecular Structure Viewer inside Antigravity acted as a shared canvas: the agent could highlight regions, and researchers could pin notes for the agent to use.
• Results are predictions only. The lead and other candidates are now in wet-lab validation, with no results reported.

Why it matters

It is a concrete example of the human-steered research agent pattern outside coding. The scientist sets direction and judges results, while the agent chains many tool calls across databases and models. The shared visual canvas is a notable interface idea. The post is partly promotional, and its scientific claims are still unvalidated.