ISO: annotation by orthology

Separating source defects from propagation defects

AI Gene Review · projects/ISO · 2026

Bottom line

  • ISO copies a GO annotation to an ortholog. It can fail at the source or at the orthology edge.
  • 4,345 ISO rows in 201 reviews: 2,073 kept as non-core, 1,456 accepted, only 179 removed. ISO is mostly true but contextual.
  • A failure taxonomy, now the structured propagation_review field, says where each defect lives.

Two questions for every ISO row

Three worked cases

Gene Pattern ISO outcome
Calm3 (mouse) Positive control: Calm1/2/3 encode the same protein 16 ACCEPT, 25 non-core, 3 over-annotated
Ghr (mouse) Direct donor support mixed with circular and stale transfers; GH-binding-protein isoform trap 18 ACCEPT, 11 REMOVE
Ang2 (mouse) Divergent angiogenin paralog: RNase kept, angiogenesis lost 41 of 46 REMOVE

What the structured review looks like

Ang2 angiogenesis (IBA row shown; the ISO row gets the same REMOVE): PROPAGATION_BAD + FUNCTIONAL_DIVERGENCE, each source marked "supports source but not target".

Results across the corpus

Findings

  1. ISO is not mostly garbage. The risk is a cloud of true-but-contextual rows hiding a few real defects.
  2. Where it breaks: paralogs and diverged members (Ang2), stale or circular donor chains (Ghr), isoform context (Ghr GH-binding protein).
  3. Most structured failures are propagation or scoping, not bad sources: 119 PROPAGATION_BAD and 101 TERM_SCOPING_PROBLEM vs 5 SOURCE_BAD.

Status and next steps

  • Done: case reviews, taxonomy, reviewer checklist, propagation_review schema.
  • Open: a reusable ISO donor-trace script generalising the Ang2 trace.
  • Related: projects/IBA_REVIEW.md (same taxonomy), projects/IEP.md.

Read more: projects/ISO.md · genes/mouse/Ang2/Ang2-bioinformatics/RESULTS.md · genes/mouse/Ghr/Ghr-iso-donor-trace.md