AlphaFold Database for annotation review

Using predicted monomers and complexes to test GO claims. Scoped, not yet built.

AI Gene Review · projects/ALPHAFOLD · 2026

Bottom line

  • Scoped, not started as a pipeline: 0 of 6 action items done; no AFDB fetch step, no schema field.
  • Idea: check each GO or ARBA claim against the predicted model (pocket, active site, TM helices, disorder, complex interface).
  • One worked use so far: in the BGC project, AF3-predicted complexes corroborated three known enzyme complexes.

Why predicted structures

  • AFDB now includes proteome-scale quaternary predictions, not only monomers.
  • 1,580 of the 2,529 pipeline genes have no deposited PDB structure (PDB project inventory).
  • Five use cases: binding-site checks, transfer confidence for ARBA rules, interface evidence instead of protein binding, pLDDT disorder context, flagging implausible rules.

Proposed workflow

The one worked example: BGC complexes

Used as corroboration, not to drive the call. The screen's own validation set shows real complexes can score low, so a missing prediction is not evidence against a complex.

Next steps

  • ⬜ Add an AFDB lookup to the bioinformatics pipeline (fetch by UniProt ID)
  • ⬜ Script to extract features relevant to GO validation
  • ⬜ Pilot on 5–10 ARBA rule reviews: does structure change the outcome?
  • ⬜ Test quaternary predictions for complex membership; test pLDDT for domain claims
  • ⬜ Consider a structural-evidence field in the review schema

Read more: projects/ALPHAFOLD.md · projects/BGC.md · projects/PDB.md