Inferred from Sequence Orthology (ISO) Evidence Code Review

Bottom line: ISO transfers a GO annotation from a gene to its ortholog in another species, so an ISO row can fail because the source annotation is weak or because the term should not cross that orthology edge. We read the ISO rows in existing gene reviews, worked through three contrasting cases in depth (mouse Calm3, Ghr and Ang2), and built a failure taxonomy that separates source defects from propagation defects, now recorded in the structured review.propagation_review field. We did this so that reviewers stop treating ISO as either trustworthy or garbage and instead say where each defect lives. As of 2026-09-26 the repo holds 4,345 reviewed ISO rows in 201 gene reviews (3,258 mouse, 962 rat, and 125 in human, fission yeast and bacterial reviews): 1,456 ACCEPT, 2,073 KEEP_AS_NON_CORE, 435 MARK_AS_OVER_ANNOTATED, 179 REMOVE, 111 MODIFY, 35 UNDECIDED, plus 56 NEW. The usual problem is a cloud of true but contextual transfers; outright failures cluster in cases like Ang2, a divergent angiogenin paralog where 41 of 46 ISO rows were removed. The corpus snapshot below (2026-06-29) predates these counts.

Of the 304 ISO rows that already carry a structured propagation_review, the commonest root causes are PROPAGATION_BAD (119) and TERM_SCOPING_PROBLEM (101), and the commonest subtype is CONTEXT_OR_TISSUE_MISMATCH (113). One action item remains: a reusable donor-trace script.

Part of Propagation by Homology.
Browse ISO rows ·
Current statistics

What an ISO annotation asserts

ISO (Inferred from Sequence Orthology, ECO:0000266) says: gene product X has
this function because its ortholog Y was shown experimentally to have it. Every
ISO row therefore names three things, and a review should name them too:

Almost all ISO in the corpus comes from three automated pipelines, which move
only experimental annotations (IDA, IMP, IPI, IGI, EXP) from the donor
(GO_REF definitions):

The validity of an ISO row is then three separate questions:

1. Is the source sound? Does the donor still carry the term, with
   experimental evidence, from a paper about that gene? A donor that has since
   lost the term leaves a stale transfer; a donor whose own support is inferred
   makes the ISO a transfer of a transfer.
2. Is the relation the right one? Is the donor the target's one-to-one
   ortholog, or a paralog, or one member of a one-to-many call?
3. Is the term safe to move? Conserved biochemistry usually is; tissue,
   developmental, compartment, and regulatory context often is not.

The propagation browser shows all
three for every ISO row: the donor resolved to symbol and species, whether the
donor currently carries the term (Donor support), whether the donor is the
target's namesake (Donor vs target symbol), and the review verdict.

Worked examples

Calmodulin: an orthology pipeline that also transfers across paralogs

Target: mouse Calm3. It receives ISO from four donors:

So yes — part of the Calm3 ISO set is paralogy transfer. The two pipelines
behave differently: human→mouse (GO_REF:0000119) pairs namesakes (CALM1→Calm1,
CALM2→Calm2, CALM3→Calm3), while mouse↔rat (GO_REF:0000096) donates from all
three rat loci to each mouse locus, so rat Calm3 also donates to mouse Calm1 and
Calm2 (browse).
Five Calm3 terms, including chromatin, are donated by rat Calm1 alone. Here it
is harmless for protein-level terms: Calm1, Calm2, and Calm3 encode identical
proteins in mouse, rat, and human, and differ only at the locus level (UTR
regulation and tissue expression).
The donors all still carry their terms experimentally. What does not transfer
is locus-level biology — Calm3's Stau2-dependent dendritic mRNA localization is
Calm3-specific — so synaptic, cardiac, spindle, and sarcomere rows are kept as
non-core rather than read as Calm3-defining. (An earlier version of the Calm3
review named the human donor as CALM1; P0DP25 is CALM3, and the review has
been corrected.)

Ang2: the right source, the wrong target

Target: mouse Ang2 (Angrp). Every ISO row comes from human ANG
(UniProtKB:P03950) via GO_REF:0000119. The donor is well characterised and
most of its terms are still experimentally supported, but mouse Ang2 is a
divergent member of the expanded mouse angiogenin family, not ANG's namesake
ortholog: direct mouse
evidence supports RNase/tRNA cleavage but not angiogenesis, receptor binding,
signaling, nuclear trafficking, or immune-effector roles. Most rows were
removed as PROPAGATION_BAD + WRONG_ORTHOLOG_OR_PARALOG +
FUNCTIONAL_DIVERGENCE. The donor check also finds rows whose human ANG term
has since gone or was itself only inferred, so source and propagation defects
compound. The manual trace is in
genes/mouse/Ang2/Ang2-bioinformatics/RESULTS.md;
browse the rows.

Ghr: sound orthologs, stale sources

Target: mouse Ghr. Donors are its orthologs, human GHR (UniProtKB:P10912,
GO_REF:0000119) and rat Ghr (RGD:2687, GO_REF:0000096), so the relation is
not the problem. Receptor activity, hormone binding, membrane localization and
JAK-STAT signaling transfer cleanly. The problems are on the source side:
several transferred terms (growth factor binding, receptor internalization,
response to estradiol, response to cycloheximide) are no longer on the human
donor, and one (cytosol) is supported on the donor only by IBA. A stale source
is not automatically a wrong term — the JAK-STAT rows are also stale on the
donor but were kept because the biology is well established for mouse Ghr. The manual trace is
genes/mouse/Ghr/Ghr-iso-donor-trace.md; the browser's Donor support = ABSENT
filter now finds the same pattern automatically
(browse).
Ghr also shows an isoform trap: extracellular-space rows fit the GH-binding
protein, signaling rows fit the full-length receptor.

What ISO adds on top of IBA

For each ISO row the browser records the closest IBA on the same target, using
the GO is_a/part_of closure
(statistics):

• IBA to the same or a more specific term — the ISO row is already implied
  by PAINT. Same-term rows are rarely rejected; ISO here is corroboration, not
  new information.
• IBA only to a more general term — ISO adds specificity below a PAINT
  term.
• No related IBA — ISO adds a new assertion. This is the majority of ISO
  rows, and mostly biological process: the tissue, pathway and physiological
  context that experimental work in human or rat establishes and PAINT does
  not propagate. It is also where most of the non-core, over-annotated and
  removed ISO rows sit.

So ISO's value beyond IBA is mainly mammal-specific process and context
annotation carried from the best-studied species, with conserved core
functions largely duplicated by PAINT. The review burden follows the same line:
the rows ISO uniquely contributes are the ones that most need checking.
Browse ISO rows with no related IBA.

Where ISO fails

The failure patterns seen so far, in rough order of how often they change the
verdict:

• Context, not function. Most ISO rows are true but contextual — the
  common outcome is KEEP_AS_NON_CORE, not REMOVE. The risk is that a cloud
  of true-but-contextual rows hides the few real defects.
• Donor is not the namesake. ISO annotations whose donors all have a
  different gene symbol from the target (paralogs, expanded families) are
  rejected or reduced about three times as often as annotations with a
  namesake donor (see statistics).
• Stale source. The donor no longer carries the term. The row outlives the
  evidence that justified it.
• Transfer of a transfer. The donor's own support is inferred.

Failure Taxonomy

Use two labels when reviewing IBA or ISO propagation:

1. A root-cause class: where the defect lives.
2. A biological subtype: what kind of transfer mistake it is.

Record these in review.propagation_review. Keep the narrative rationale in
review.reason; the structured object should stay mechanical.

Example:

review:
  action: REMOVE
  reason: Mouse Ang2/Angrp retains RNase activity but lacks the angiogenic function of human ANG.
  propagation_review:
    root_cause: PROPAGATION_BAD
    failure_modes:
      - WRONG_ORTHOLOG_OR_PARALOG
      - FUNCTIONAL_DIVERGENCE
    source_entities:
      - source_id: UniProtKB:P03950
        source_label: ANG
        source_status: SUPPORTS_SOURCE_BUT_NOT_TARGET
        comment: Human ANG supports angiogenesis; mouse Ang2/Angrp is non-angiogenic.

Root-cause classes

Biological subtypes

Examples:

• Ang2 human ANG-to-mouse Ang2 angiogenesis rows:
  PROPAGATION_BAD + WRONG_ORTHOLOG_OR_PARALOG + FUNCTIONAL_DIVERGENCE.
• Ghr receptor internalization:
  SOURCE_STALE_OR_MISSING + EVIDENCE_CIRCULAR_OR_REDUNDANT.
• Ghr hormone-mediated signaling:
  TERM_SCOPING_PROBLEM + GRANULARITY_MISMATCH.
• Calm3 calcium ion binding:
  NO_FAILURE_CORE.
• Calm3 spindle/centrosome contexts:
  NO_FAILURE_NON_CORE + CONTEXT_OR_TISSUE_MISMATCH.

Reviewer Checklist

Use this checklist before making a strong REMOVE call on ISO or IBA.

Source annotation checked

• Record the target row: gene, species, GO term, qualifier, aspect, evidence code,
  reference, WITH/FROM, and assigned-by source.
• Verify the GO term definition, not just the label.
• Resolve the source gene/product(s) named by WITH/FROM.
• Check whether the current source record still carries the same GO term.
• Record the source evidence code and original source reference.
• If the source evidence is experimental, verify that the cited paper supports
  the source gene and term. If only abstract text is cached, avoid claiming a
  curator misread full text.
• Check for source-side NOT annotations or contradictory direct evidence.

Propagation checked

• For ISO, confirm whether the donor-target relation is one-to-one orthology,
  one-to-many, in-paralog, identical protein sequence, or broader family
  similarity.
• For IBA, inspect the PANTHER family/node, seed genes, WITH/FROM proteins, and
  whether seeds map to the same functional subfamily as the target.
• Ask whether the exact term should transfer, not only whether the broad family
  function is conserved.
• Check active-site residues, key domains, isoforms, compartment targeting,
  lineage constraints, and target-species direct evidence when relevant.
• Distinguish source defects from propagation defects. Do not hide a bad source
  annotation under "orthology failure," and do not blame a source annotation when
  the real problem is target-specific divergence.

Disposition recorded

• Assign propagation_review.root_cause and any relevant
  propagation_review.failure_modes.
• Add propagation_review.source_entities entries for donor genes, seed
  proteins, PANTHER nodes, family nodes, or other source entities that were
  checked.
• State whether the recommended fix is source-side, propagation-side, term
  replacement, or non-core retention.
• Prefer UNDECIDED when the source full text or family-node evidence cannot be
  checked.

Action Items

• [x] Review mouse Calm3 ISO annotations as the initial positive-control case.
• [x] Summarize existing ISO reviews across the repository.
• [x] Add an ISO/IBA failure taxonomy that distinguishes source defects from
  propagation defects.
• [x] Add a source/family annotation checklist for future ISO and IBA reviews.
• [x] Add structured review.propagation_review schema fields for the taxonomy
  and per-source entity comments.
• [x] Add a reusable donor trace: just refresh-propagation-sources resolves every
  ISO/ISS/ISA/Compara donor and records whether it still carries the term;
  surfaced in the propagation browser.
• [x] Regenerated statistics (species pairs, donor support, ISO vs IBA) in
  propagation-stats instead of
  hand-maintained counts.
• [ ] Map automated donor support onto propagation_review.source_entities[].source_status
  suggestions for reviewers (e.g. ABSENT → SOURCE_STALE_OR_MISSING).
• [ ] Distinguish one-to-one from one-to-many orthology calls in the browser
  (needs Alliance/HCOP orthology type, not yet cached).

Slides

• Slides (Marp source: ISO-slides.md) — AI generated