Over-annotation patterns

Recurring ways a GO annotation can be defensible and still say nothing, or say something false

AI Gene Review · projects/OVER_ANNOTATION_PATTERNS · 2026

Bottom line

  • Many GO rows come from HTP screens, domain signatures, broad mappings or family propagation and tell a reader little, or assert an activity the protein lacks.
  • We catalogued eight recurring patterns, each tied to worked examples in real reviews; first presented at the GO Consortium meeting, October 2025.
  • Nine exemplar reviews are complete and their recorded actions match the patterns. The catalogue is qualitative: frequency across the repo is not yet measured.

Why a catalogue

  • The same over-annotation shapes recur gene after gene.
  • Naming a pattern lets a curator or pipeline author recognise it once instead of rediscovering it.
  • Over-annotations dilute informative rows, create a false impression of understanding, skew enrichment, and propagate by IBA to other species.

Eight patterns and their sources

A pseudo-enzyme in practice: Epe1

Fission yeast Epe1 has a JmjC domain with Tyr at position 370, where canonical JmjC demethylases have the third iron-binding His; the IBA demethylase row is REMOVE with no replacement term, and Swi6 binding is a separate NEW row (GO:0070087).

Shared folds, shared subunits

  • Pattern 7: the KAS-III fold is common to fatty-acid and polyketide synthases, so IEA puts fatty-acid terms on polyketide enzymes.
    • actI-ORF1: GO:0030497 fatty acid elongation → REMOVE; GO:0006633 → MODIFY to GO:1901112 actinorhodin biosynthesis.
    • pqsC: GO:0006633 → REMOVE; the product is a quinolone signal.
  • Pattern 8: a catalytic MF lands on the partner without an active site.
    • pqsB, actI-ORF2: acyltransferase → MARK_AS_OVER_ANNOTATED; eryCII (heme-less P450): monooxygenase, heme, iron → REMOVE.

Actions in the exemplar reviews

Curation principles

  1. Specificity over breadth: prefer the most specific accurate term.
  2. Remove uninformative rows: generic protein binding from HTP screens.
  3. Validate domain predictions, especially enzymatic activity.
  4. Direct vs indirect: annotate the proximal function, not downstream consequences.
  5. Consider pseudo-enzymes: a domain does not guarantee activity.

Status and next steps

  • ✅ Eight patterns documented; nine exemplar reviews complete.
  • ⬜ Measure how often each pattern occurs across the repo (e.g. share of GO:0005515 rows demoted).
  • ⬜ Turn the most mechanical patterns (HTP protein binding, fold-based pathway terms) into flags a curator can filter on.
  • Related: projects/PSEUDOENZYMES.md, projects/PROTEIN_COMPLEX_FUNCTIONS.md, projects/REVIEW_QUALITY_AUDIT.md

Read more: projects/OVER_ANNOTATION_PATTERNS.md