| name | foldseek-structural-search |
| description | Find structurally similar proteins by submitting a 3D coordinate file (.pdb, .cif, or .mmcif) to the Foldseek web API and search databases such as PDB, AlphaFold, CATH, and MGnify. Use ONLY when the user provides a physical structure file and wants structural homologs. Public data, no credential needed. |
| license | Apache-2.0 |
Foldseek Structural Search (Codex-native)
Codex-native adaptation of Google DeepMind's foldseek-structural-search skill.
Attribution: adapted from
google-deepmind/science-skills
(Apache-2.0 + CC-BY-4.0). Foldseek: https://search.foldseek.com/search and
https://github.com/steineggerlab/foldseek
Hard requirement
Foldseek searches by 3D shape only. It cannot search by sequence, protein
name, or accession. If the user gives anything other than a path to a .pdb,
.cif, or .mmcif file, halt and tell them to obtain a structure first
(e.g. via $cx-alphafold-structure-analysis for a UniProt ID).
Gates (ask before proceeding)
- Terms notice: on first use in a workspace, tell the user to review the
Foldseek terms, and record the notice in provenance.
- Network / install: submission is read-only public HTTP. Ask before
installing any client package; run Python via
uv run, not bare python3.
Workflow
- Validate input. Confirm the file exists and ends in
.pdb, .cif, or
.mmcif. Otherwise halt (see hard requirement).
- Validate databases. Default to the standard set. If the user names
databases, allow only:
afdb50, afdb-swissprot, pdb100, BFVD,
mgnify_esm30, cath50, gmgcl_id, bfmd, afdb-proteome. Any other
database → halt and show this allowlist.
- Submit the structure to the Foldseek API, poll until the job completes,
and save two outputs under the user's project with descriptive names:
<name>_foldseek_results.json (full payload) and <name>_foldseek_results.md
(a Markdown hits table).
- Read the
.md table for your summary; do not hand-parse the raw JSON or
the 3D coordinates.
- Interpret the top 3-5 meaningful matches:
- Prob near 1.0 → high confidence of a true structural homolog.
- Q-Cov high → the match covers most of the query shape, not just a
local motif.
- E-value / Seq identity → evolutionary context.
- Functional synthesis: report the specific protein names/functions of the
top homologs (from the target annotations) and summarize the variety of
families/functions across the hit list.
- Tell the user where both output files are so later steps can reuse them.
- Provenance & review: record the query file, databases, and outputs with
$science-provenance; check claims with $science-review.
Boundaries
- Structural similarity suggests, but does not prove, functional or evolutionary
relationship; state confidence honestly.
- On API error or missing file, report it plainly and ask the user to verify the
path — do not fabricate results.