| name | alphafold-structure-analysis |
| description | Retrieve an AlphaFold predicted structure from the AlphaFold Database for a given UniProt accession and analyze its confidence (pLDDT), intrinsically disordered regions, and rigid domain boundaries (PAE). Use when the user has a UniProt ID and wants structural confidence, disorder assessment, or domain layout. Public data, no credential needed. |
| license | Apache-2.0 |
AlphaFold Database Fetch and Analyze (Codex-native)
Codex-native adaptation of Google DeepMind's
alphafold-database-fetch-and-analyze skill.
Attribution: adapted from
google-deepmind/science-skills
(Apache-2.0 + CC-BY-4.0). AFDB terms: https://alphafold.ebi.ac.uk/
When to use / not use
- Use when the user provides a UniProt accession (e.g.
P04637) and
wants confidence metrics, disorder, or domain boundaries.
- Do not use when the user has only a protein/gene name or a raw sequence
(ask them to find the UniProt ID first), wants to search structural homologs
(use
$cx-foldseek-structural-search), needs an experimental structure (use a
PDB skill), or wants to run a new folding prediction.
Gates (ask before proceeding)
- Terms notice: on first use in a workspace, tell the user to review the AFDB
terms, and record the notice in provenance.
- Network / install: fetching from AFDB is read-only public HTTP. Ask before
installing any analysis package; run Python via
uv run, not bare python3.
Workflow
Fetch current model metadata first with science_search_alphafold using a
canonical UniProt accession. Prefer its returned versioned URLs over guessing a
model version.
- Fetch the structure for the UniProt ID from the AlphaFold Database:
the mmCIF model, the Predicted Aligned Error (PAE) JSON, and the entry
metadata JSON. Endpoint pattern (v6 example):
https://alphafold.ebi.ac.uk/files/AF-<ACC>-F1-model_v6.cif and the matching
-predicted_aligned_error_v6.json; entry metadata via the AFDB API
https://alphafold.ebi.ac.uk/api/prediction/<ACC>. Save all files to an
output folder under the user's project (absolute path), rate-limited and
polite. For very large proteins handle fragment fallback (-F2, ...).
- Analyze pLDDT from the per-residue confidence (embedded in the mmCIF
B-factor column, and in metadata). Report the global mean and the fraction in
each band: very low (<50), low (50-70), confident (70-90), very high (>90).
- Analyze domains / PAE. Detect rigid domain boundaries with a
sliding-window heuristic over the PAE matrix; report the number of distinct
domains and their residue ranges. Do not eyeball domains from coordinates —
derive them from PAE.
- Synthesize one cohesive summary: overall fold confidence, disordered
regions, and rigid-domain layout, plus supporting metrics.
- Mandatory warnings:
- If no canonical entry was found and an isoform/fragment was used, or the
protein is very large (>2700 aa), relay that prominently.
- If the protein is largely disordered (high <50 fraction / no rigid
domains), warn against whole-protein downstream analysis (Foldseek,
docking); if small ordered domains exist, advise restricting analysis to
those residue ranges.
- Remind the user that per-residue pLDDT lives in the mmCIF B-factor column.
- Provenance & review: record the accession, file versions, and outputs
with
$science-provenance; check claims with $science-review.
Boundaries
- pLDDT and PAE are confidence heuristics, not correctness guarantees; a
confident fold can still be biologically wrong.
- Report the model version explicitly; AFDB re-releases change coordinates and
metrics.
