alterlab-alphafold-db — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited alterlab-alphafold-db (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
AlphaFold DB is a public repository of AI-predicted 3D protein structures maintained by Google DeepMind and EMBL-EBI. Release v6 (October 2025, synced to UniProt 2025_03) holds ~241 million predictions, including ~40k isoforms and the input MSAs; since March 2026 it also serves precomputed homodimer and heterodimer complex predictions. Access structure predictions with confidence metrics, download coordinate files, retrieve bulk datasets, and integrate predictions into computational workflows.
This skill should be used when working with AI-predicted protein structures in scenarios such as:
| Scenario | Use Instead |
|---|---|
| Experimental X-ray / cryo-EM / NMR structure by PDB ID | alterlab-pdb |
| Folding a new or mutated sequence, or a custom protein–protein complex, yourself (ColabFold / AF2-Multimer) | alterlab-alphafold |
| Protein–ligand or protein–nucleic-acid co-folding | alterlab-boltz |
| Protein sequence, functional annotation, or accession ID mapping only | alterlab-uniprot |
Worked, copy-paste Python recipes for every capability below live in references/code_examples.md. Load it when you need runnable code; the summaries here give the routing and the key decisions.
Three entry points, in order of preference:
Bio.PDB.alphafold_db.get_predictions(accession),download_cif_for(...), get_structural_models_for(...) — simplest path.
GET https://alphafold.ebi.ac.uk/api/prediction/{uniprot_id}.The response is a list of every model for that accession — the canonical sequence plus isoforms (P00520-2, …) and, for some entries, third-party models — so select the record whose uniprotAccession equals your query rather than trusting [0]. The model ID is modelEntityId (e.g. AF-P00520-F1); entryId is the legacy name that passed its announced 2026-06-25 sunset, so don't build new code on it. Complex models are excluded unless you pass ?include_complexes=true (or call /api/complex/{id}).
use the UniProt ID-mapping job API (get_uniprot_ids helper in code_examples.md §1; valid db names at https://rest.uniprot.org/configure/idmapping/fields).
The /prediction response carries version-stamped file URLs — use those, don't hand-build a `_v{N}` suffix. The DB version advances (currently v6) and old _v4 file URLs now 404:
cifUrl / pdbUrl / bcifUrl — atomic coordinates (mmCIF / PDB / binary CIF).plddtDocUrl — per-residue pLDDT scores (0-100).paeDocUrl — PAE matrix.Download recipe (resolve URLs from the API, write bytes) in code_examples.md §2.
plddtDocUrl, read confidence['confidenceScore'] (keys:residueNumber, confidenceScore, confidenceCategory); thresholds in "Confidence Interpretation Guidelines" below.
paeDocUrl. The endpoint returns a single-element JSON array ofone object, so index [0] before the key (pae[0]['predicted_aligned_error']). Visualization recipe in code_examples.md §3.
proteome at https://ftp.ebi.ac.uk/pub/databases/alphafold/latest/ (e.g. UP000005640_9606_HUMAN_v6.tar; the index is download_metadata.json in the parent directory). Lower-confidence complex predictions are bulk-only, under .../alphafold/collaborations/nvda/.
gs://public-datasets-deepmind-alphafold-v4/proteomes/with gsutil, or query bigquery-public-data.deepmind_alphafold.metadata to filter by organism/confidence. The species-download helper validates the taxonomy ID and uses list-form subprocess.run (never shell=True).
See code_examples.md §4 and references/api_reference.md (Bulk Downloads).
Parse mmCIF with Bio.PDB.MMCIFParser; pLDDT is stored in the B-factor column (residue['CA'].get_bfactor()). Contact-map and B-factor extraction recipes in code_examples.md §5.
Loop accessions → predictions → confidence stats → summary DataFrame. Full example in code_examples.md §6.
uv pip install biopython requests # core: structure access + API
uv pip install numpy matplotlib pandas scipy # analysis + PAE plots
uv pip install google-cloud-bigquery gsutil # optional: bulk GCP access3D-Beacons alternative: AlphaFold is also reachable via the 3D-Beacons federated API (https://www.ebi.ac.uk/pdbe/pdbe-kb/3dbeacons/api/uniprot/summary/{id}.json), filtering entries where structures[i]['summary']['provider'] == 'AlphaFold DB'. Recipe in code_examples.md (3D-Beacons section).
UniProt Accession: Primary identifier for proteins (e.g., "P00520"). Required for querying AlphaFold DB.
AlphaFold model ID (`modelEntityId`): AF-[UniProt accession]-F[fragment number] for DeepMind monomer models (e.g., "AF-P00520-F1"; isoforms look like "AF-P00520-2-F1"). Complex and third-party models use opaque numeric IDs (e.g., "AF-0000000365776990"); both forms are accepted by /api/prediction/{id}.
pLDDT (predicted Local Distance Difference Test): Per-residue confidence metric (0-100). Higher values indicate more confident predictions.
PAE (Predicted Aligned Error): Matrix indicating confidence in relative positions between residue pairs. Low values (<5 Å) suggest confident relative positioning.
Database Version: The REST API and the FTP latest/ archives serve v6 (the response reports latestVersion / allVersions); the GCS/BigQuery datasets lag at v4. File URLs include a version suffix (e.g., model_v6.cif, while newer third-party models start at _v1) — read them from the prediction response rather than hardcoding the suffix.
Fragment Number: Large proteins may be split into fragments. Fragment number appears in AlphaFold ID (e.g., F1, F2).
pLDDT Thresholds:
PAE Guidelines:
Worked, copy-paste Python recipes for every Core Capability: prediction retrieval (Biopython / REST / UniProt mapping), file downloads, pLDDT + PAE analysis, GCP/BigQuery bulk access, mmCIF parsing, batch processing, and the 3D-Beacons alternative.
Load this when you need runnable code.
Comprehensive API documentation covering:
Consult this reference for detailed API information, bulk download strategies, or when working with large-scale datasets.
latest/ serve v6 (latestVersion); GCS/BigQuery bulk datasets lag at v4/prediction response — never hardcode the _v{N} suffix_v4 file URLs now 404; superseded versions are removed from /files (older releases remain on the FTP site under v1/–v6/)entryId→modelEntityId, uniprotStart/End→sequenceStart/End, uniprotSequence→sequence, isReviewed→isUniProtReviewed) passed their 2026-06-25 sunset; paeImageUrl is slated for removal — use paeDocUrl/prediction results are single chains. Precomputed complexes (≈1.7M high-confidence homodimers and ≈80k heterodimers, added March–May 2026) come back only with include_complexes=true or /api/complex/{id}; judge them by interface metrics (ipTM, pDockQ) as well as pLDDT. For a complex that is not in the DB, fold it yourself (alterlab-alphafold)scripts/query_alphafold.py — runnable helper for the AlphaFold REST API (no key):
python scripts/query_alphafold.py prediction P00520
python scripts/query_alphafold.py confidence P00520 --summary
python scripts/query_alphafold.py download P00520 --fmt cif -o ./structures~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.