Propagation by homology

Reviewing GO annotations that a gene received from other gene products: ISO, ISS/ISA, IBA, Ensembl Compara, TreeGrafter, InterPro2GO

AI Gene Review · projects/HOMOLOGY_PROPAGATION · reviewed 2026-10-04

Bottom line

  • A transferred annotation can fail at the donor, at the orthology or tree relation, or because the term does not travel to the target's context.
  • We review these transfers method by method (ISO, IBA/PAINT, TreeGrafter, InterPro2GO, NCBIFam) and show every row in one propagation browser next to the gene review's verdict.
  • As of 2026-09-26, reviewers judged 34,253 of 34,265 propagated annotations and flagged 13% as REMOVE, MARK_AS_OVER_ANNOTATED or MODIFY: 9% of PAINT IBA, 20% of Ensembl Compara, 28% of TreeGrafter.

Every transfer has three parts

Donor in WITH/FROM, the relation that carries the term, the gene that receives it. A review asks one question of each part.

ISO: orthology transfers between species

  • 4,232 ISO annotations in the statistics; reviewers flagged 15% of Alliance human→mouse rows and 23% of RGD rows from other mammals.
  • 383 rest on a donor that no longer carries the term (stale transfer).
  • Rows where the donor has a different symbol (paralog or multi-locus sourcing) were flagged 48% of the time, against 16% for a namesake donor.
  • 3,000 (71%) have no related IBA on the target, so ISO mostly adds assertions PAINT does not make.
  • Worked cases: mouse Calm3 (paralog donors), Ghr (stale sources), Ang2 (41 of 46 ISO rows removed).

projects/ISO.md · propagation-stats.md

IBA and PAINT: one node, every descendant

  • 14 recurring failure patterns plus one positive control: pseudo-enzymes keeping a catalytic term, neo-functionalized subfamilies, wrong-paralog and cross-kingdom transfers.
  • The structured propagation_review vocabulary is now in use: 3,580 blocks in 691 gene reviews.
  • The gap runs both ways: in 1,015 reviewed human genes, 511 core molecular functions (423 genes) have no IBA support.
  • PAINT no-IBA genes: 756 of 7,593 listed human genes have a completed review.
  • A corpus-wide IBA re-review (3,427 genes, 11,829 annotations) started 2026-09-20; 81 genes reviewed so far.

projects/IBA_REVIEW.md · projects/PAINT.md

Electronic pipelines: no curator in the loop

Project Key finding
TreeGrafter (GO_REF:0000118) 43.2% accepted, 26.4% down-graded and 11.1% removed on the 1,074-row f81b9f300 snapshot; 77% accepted when another pipeline reproduces the call
InterPro2GO (GO_REF:0000002) 3,652 judged records on 1,706 genes; 36 proposed mapping edits across 15 entries, including 10 removals
NCBIFam / CDD NCBIFAM backs 705 (13%) of 5,549 InterPro2GO rows in the repo; a 250-row ncbifam2go SSSOM seed
Ortholog conjecture Scoped only: literature summary, draft open-world metric, five divergence cases; no metrics computed

projects/TREEGRAFTER.md · INTERPRO.md · NCBIFam.md · ORTHOLOG_CONJECTURE.md

How often reviewers flag a transfer

Why transfers were flagged

Failure mode Annotations
GRANULARITY_MISMATCH 329
CONTEXT_OR_TISSUE_MISMATCH 257
FUNCTIONAL_DIVERGENCE 170
ROLE_CONFLATION 169
COMPARTMENT_OR_COMPLEX_MISMATCH 139
WRONG_ORTHOLOG_OR_PARALOG 118
  • Most recorded failures are about the term: too coarse, the wrong context, or a role the target does not play.
  • Wrong orthologs or paralogs are a smaller share.
  • Root causes: 433 PROPAGATION_BAD and 407 TERM_SCOPING_PROBLEM, against 43 SOURCE_BAD.

Structured review.propagation_review records, all methods; propagation-stats.md

The propagation browser

One row per transfer: donor(s) → intermediate → target, donor support today, and the review verdict. app/propagation/index.html · docs/propagation_browser.md

Status and next steps

  • In progress: IBA corpus re-review (projects/IBA_REVIEW/rereview-2026-09-20/); PAINT no-IBA reviews.
  • Proposals for InterPro curators, not adopted: InterPro2GO mapping edits (projects/INTERPRO/interpro2go.sssom.yaml); the worklist past the first dozen entries is not yet assessed.
  • Not started: the ortholog-conjecture analysis.
  • Umbrella tracker: #4095.
  • Rebuild the browser and statistics offline from the committed donor cache:
    • just refresh-propagation-sources (network)
    • just deploy-propagation-browser → app/propagation/
    • just propagation-stats → projects/HOMOLOGY_PROPAGATION/propagation-stats.md