Skip to content

Recipe: Trusting an AlphaFold model

You are about to design a point mutation, dock a small molecule, or write a result paragraph that hinges on the predicted structure of a protein. The first question is always: how much should I trust this AlphaFold model?

uniprot_get_alphafold_confidence(accession) answers that without parsing the structure file.

Example: TP53 (P04637)

> What is the AlphaFold pLDDT confidence summary for P04637?

Tool call: uniprot_get_alphafold_confidence("P04637").

Output (captured 2026-10-15 from AlphaFold DB; excerpt, verbatim):

## AlphaFold confidence — AF-P04637-F1: Homo sapiens

**Gene:** TP53
**Residues modelled:** 1-393
**Model version:** v6
**Global pLDDT (mean):** 75.1  (confident)

**pLDDT band distribution:**
- Very high (≥ 90):  52.7%
- Confident (70-90):   7.1%
- Low (50-70):  10.4%
- Very low (< 50):  29.8%

The full response, footer and source bytes are in examples/captures/2026-10-15_tp53_r175h/.

What the bands mean

Band pLDDT range Trust level Typical region
Very high ≥ 90 Often reliable local geometry; not a guarantee of experimental accuracy Well-modelled regions
Confident 70-90 Generally good backbone confidence; side-chain uncertainty remains Structured regions
Low 50-70 Local model uncertainty; interpret cautiously Loops and less certain regions
Very low < 50 Coordinates should not be interpreted; may indicate disorder or missing structural context Flexible or poorly modelled regions

Source: AlphaFold-DB FAQ.

How to use the bands

A protein is rarely uniform. The four-band distribution is more useful than the global mean:

  • TP53 (above): 59.8 % confident-or-better and 29.8 % very low. The summary does not say which residues fall in each band; to judge a specific region (e.g. the DNA-binding domain versus the termini), read per-residue pLDDT from the model file before relying on it.

  • A protein with a large low-confidence fraction may contain disordered or context-dependent regions. Do not use coordinates of those regions for residue contacts or docking without independent validation. The global fraction alone does not identify which residues are affected.

  • A high-confidence model can still have uncertain loops, interfaces, ligands, or domain orientations. A high global fraction alone does not establish publication readiness or justify a binding site, docking, or mutational-effect claim.

Pairing with the per-residue tools

uniprot_get_alphafold_confidence gives the global picture. To zoom into a specific residue, pair with:

> What features are at residue 175 of TP53? Is that residue in a
> high-confidence region of the AlphaFold model?

Tool calls: uniprot_features_at_position("P04637", 175) plus the confidence summary above. The annotation places residue 175 in the DNA-binding region, but the global pLDDT distribution cannot tell which confidence band residue 175 occupies. Inspect the model's per-residue pLDDT before drawing any residue-specific structural conclusion.

These sources provide annotation context and a model-wide confidence summary. Residue-specific or mechanistic claims require additional position-level evidence and, where appropriate, experimental validation.

What this tool does not do

  • Does not download the structure file. Structures are large (~50-500 kB CIF; bigger PDB). The agent gets the URL (https://alphafold.ebi.ac.uk/files/AF-<acc>-F1-model_v<N>.cif) and can fetch or display it separately.
  • Does not return per-residue scores. The metadata endpoint aggregates pLDDT into the four bands; per-residue scores live in the structure file itself. For most decisions the bands are enough; if you need per-residue, parse the CIF (the B-factor column carries pLDDT in AlphaFold-DB convention). For any residue-specific or docking inference, do not substitute the global mean for these values.
  • Does not run AlphaFold. This is purely a confidence-summary retriever for models AlphaFold-DB has already published. Custom predictions live in AlphaFold-3 or ColabFold.

Provenance for structural claims

The returned confidence summary records the AlphaFold DB provider, retrieval URL, timestamp, and response digest; where available, it also records the prediction model version. The October 15, 2026 capture contains dated source bytes for independent inspection.

uniprot_provenance_verify currently verifies UniProt REST URLs only. It does not re-fetch or validate the AlphaFold DB prediction endpoint. To check a later AlphaFold result, retrieve the same accession from AlphaFold DB, compare the response, model identity, and version with the recorded source, and separately evaluate the scientific interpretation. A digest match proves source-byte consistency, not structure accuracy.

Operational note

This tool calls a different origin from the main UniProt surface (alphafold.ebi.ac.uk). The cross-origin allowlist is documented in the threat model and the third-party listing is in PRIVACY.md.