Foldseek Structural Search (Claude Skill)

Take a protein 3D coordinate file and find structurally similar proteins across large structure databases, returning ranked matches with probability, coverage, E-value, and identity scores.

   
Type Claude Skill
Supplier Google DeepMind
Availability GA
Pricing Free / OSS skill (Apache-2.0 code, CC-BY-4.0 docs); Foldseek Search API is a public web service, no key
Capabilities Read-only — Claude runs the skill’s Python locally (uv run) and submits your structure to the Foldseek web API
Verified degraded · 2026-07-20 — removed stale scienceskillscommon copy line; foldseek skill dir resolves
Security cleared · 2026-07-20 — provenance matches google-deepmind, Apache-2.0, maintained, no OSV advisories

How to install

The google-deepmind/science-skills collection follows the Agent Skills SKILL.md spec. The repo’s primary npx skills add path targets Gemini/Antigravity; for Claude the followable path is a manual copy of the skill directory.

  • Claude Code / Claude Desktop — clone and copy the skill into your skills directory: 
    git clone https://github.com/google-deepmind/science-skills
cp -r science-skills/skills/foldseek_structural_search ~/.claude/skills/
  • Prerequisite — the skill runs its scripts/search.py via uv run; install uv first if absent: curl -LsSf https://astral.sh/uv/install.sh | sh. Python deps install into an isolated environment on first run.

What it does

Submits a structure to the Foldseek web search API and interprets the hits:

  • Accepts a physical coordinate file (.cif, .mmcif, or .pdb) — it will not accept a bare sequence, gene name, or accession.
  • Searches a chosen target database from the allowlist: afdb50, afdb-swissprot, pdb100, BFVD, mgnify_esm30, cath50, gmgcl_id, bfmd, afdb-proteome.
  • Writes two outputs: a JSON file with the full API results for downstream analysis and a Markdown file with a formatted match table.
  • Interprets matches via probability, query coverage, E-value, and sequence identity to infer likely function from target annotations.

Primary use cases: structural-homology search for an uncharacterized model (e.g., an AlphaFold prediction), function inference from structure, finding remote homologs that sequence search misses.

Notes

Requires a real 3D structure file as input — pair it with AlphaFold or the RCSB PDB skill to obtain coordinates first. The skill validates inputs strictly and halts on sequence-only input or an unsupported database. It calls the public Foldseek web service, so results depend on that service’s availability; large queries are subject to its server-side limits. The npx skills add google-deepmind/science-skills/ command documented upstream is oriented at Gemini/Antigravity (it writes to ~/.gemini/config/skills/); for Claude, the manual copy into ~/.claude/skills/ shown above is the equivalent path. Foldseek itself is GPLv3 (steineggerlab/foldseek); this skill uses the hosted search API rather than a local Foldseek install.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.