Skip to content
Agents tracked: 492 Releases (7d): 545 Agent status: 1 with issuesSince last limit reset: Claude 9d · Codex 7d For AI agents: Markdown Updated Nov 20, 2026
Showcase Google DeepMindResults as reported by Google DeepMind

Search for antimicrobial peptides in UniProt with Google Antigravity and ProtNLM

Google DeepMind researchers describe a walkthrough in which a biologist and an agent in Google Antigravity searched UniProt for a novel antimicrobial peptide candidate in Streptococcus pseudopneumoniae. The resulting lead is a computational prediction that is now in wet-lab validation.

The problem

Searching for new antimicrobial therapeutics is slow and labor-intensive, and it usually means working across several databases and separate tools. Many potentially useful peptides sit among millions of UniProt sequences whose function is unknown. Wet-lab validation is resource-intensive, so the search must be narrowed to high-confidence leads first.

How they did it

  1. Set a focused biological goal with the agent in Antigravity, here finding and computationally validating novel AMPs in Streptococcus pseudopneumoniae.
  2. Have the agent query UniProt, using ProtNLM function predictions, to narrow millions of entries to a few candidates.
  3. Check novelty by running the candidate against Swiss-Prot through the EMBL-EBI NCBI-BLAST service.
  4. Retrieve the AlphaFold 3D model, compare it with experimentally determined structures, and inspect it in the Biomolecular Structure Viewer.
  5. Examine the genes flanking the candidate and review the literature to build a mechanism hypothesis.
  6. Ask the agent to turn the findings into a bench protocol covering expression, susceptibility testing, target engagement and toxicity testing.

Results as reported by Google DeepMind

  • The session produced a candidate bacteriocin (UniProt ID A0A0U0D591, 97 amino acids, Lactococcin 972 family) with predicted activity against pathogens involved in respiratory tract infections.
  • As reported by Google DeepMind, the candidate had no meaningful homolog among manually reviewed Swiss-Prot proteins, its fold matched a known bacteriocin, and it sat 6 base pairs ahead of an ABC transporter permease in a bacteriocin-like operon.
  • As reported by Google DeepMind, the candidate and several others from the same pipeline are in wet-lab validation, and the lead still requires that validation.

These figures are as reported by Google DeepMind in Google DeepMind blog post ↗ (Oct 15, 2026). AgentGid did not measure or independently verify them.

Takeaway. Combining a domain-specific prediction model with an agent that orchestrates database, structure and literature checks can reduce a fragmented in silico triage to one interactive session, but results still need wet-lab confirmation.
AgentGid's take

This suits a small computational biology group with a biologist who can judge BLAST, AlphaFold and operon evidence, plus access to a wet lab, since the output is only a prediction. Watch the novelty and mechanism claims, which are vendor-reported and need expert checking. Antigravity is free to start, but its individual tier has basic weekly rate limits. For a general IDE agent, Zed is open-source (Free + $10/mo).

The agent used here

Similar use cases

Guides