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 (illustrative):

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

**Gene:** TP53
**Residues modelled:** 1-393
**Model version:** v6
**Global pLDDT (mean):** 76.4  (confident)

**pLDDT band distribution:**
- Very high (≥ 90):  32.0%
- Confident (70-90): 40.0%
- Low (50-70):       20.0%
- Very low (< 50):    8.0%

What the bands mean

Band pLDDT range Trust level Typical region
Very high ≥ 90 Position-level accuracy comparable to a 1.5 Å crystal structure Folded domains with reliable templates
Confident 70-90 Backbone-level accuracy; sidechain placement reasonable Folded domains with weaker templates, or surface
Low 50-70 Topology may be right; do not trust local geometry Surface loops, flexible linkers
Very low < 50 Treat as disordered; structural reasoning is unsafe Intrinsically disordered regions, signal peptides

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): 72 % confident-or-better. The DNA-binding core is very high; the N- and C-terminal regulatory regions sit in the low and very low bands. → Structural reasoning about the core is sound; reasoning about the tails is not.

  • An intrinsically disordered protein (e.g. some transcription factors, prion proteins): may show 60-90 % very low. → The AlphaFold model is essentially a guess; do not run docking on it, do not interpret a single-residue contact.

  • A typical enzyme (e.g. CYP450, kinase): often 90 %+ very high. → Publication-grade. Structural reasoning, sidechain analysis, surface-pocket identification are all on solid ground.

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. Cross-reference: residue 175 sits in the DNA-binding domain (positions 102-292), which from the band distribution falls in the very high confidence range.

The combination — feature context + structural confidence — is the evidence packet any structural-biology decision should rest on.

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).
  • 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

Every confidence-summary response carries the standard provenance footer:

_Source: AlphaFoldDB release v6 (2024-09-01) • Retrieved 2026-04-25T12:00:00Z_
_Query: https://alphafold.ebi.ac.uk/api/prediction/P04637_
_SHA-256: …_

A year later, you can call uniprot_provenance_verify with those fields. If AlphaFold-DB has issued a new model version (v7, v8), the verifier returns release_drift — useful, because new AlphaFold releases sometimes change structural predictions materially (e.g. when a new template lands).

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.