IBA Annotation Quality Project

Bottom line: IBA annotations transfer GO terms along PANTHER family trees
from experimentally studied proteins to their relatives, and they make up a
large share of GO for most genomes. Working from gene reviews, we catalogued
where those transfers go wrong and why: 14 recurring failure patterns (such as
pseudo-enzymes that keep a catalytic term, neo-functionalized subfamilies,
wrong-paralog and cross-kingdom transfers) plus one positive control, 53 worked cases in the
table below, and a structured propagation_review vocabulary (root cause,
failure modes, per-source status) that reviews now use; 691 gene reviews
carried 3,580 such blocks as of 2026-09-26. We did this because an IBA error at a family node
spreads to every descendant, so one bad call can mislabel hundreds of
proteins. The work covers both directions: in 1,015 reviewed human genes, 511
curated core molecular functions (across 423 genes) have no IBA support at
all. A corpus-wide re-review started on 2026-09-20 over 3,427 genes and 11,829
propagated annotations; 81 genes are reviewed and 65 await adjudication
(rereview-2026-09-20).

Overview

Project log, per-pass verification narrative, and lessons learned: IBA_REVIEW/HISTORY.md.

Consistency audit of the structured propagation_review blocks across all 1843 rows:
IBA_REVIEW/propagation-review-audit.md.

This project examines the quality of IBA (Inferred from Biological Aspect of Ancestor) annotations discovered through AI-assisted gene review. IBA annotations use phylogenetic trees to transfer function from characterized proteins to uncharacterized orthologs.

While IBA is a powerful curation tool, it can produce problematic annotations when:
1. Functional divergence - Orthologs have evolved different functions
2. Pseudo-enzymes - Catalytic function lost despite domain retention
3. Context-specific function - Function differs between organisms/tissues
4. Over-generalization - Broad terms transferred when specific functions differ

This project covers both directions of IBA quality: most of the page catalogs
where IBA is wrong (over-annotation, patterns 1–15), while
IBA Incompleteness
quantifies where IBA under-calls established biology — 511 curated human core
molecular functions that IBA alone would miss.

Propagation Taxonomy and Checklist

The patterns below are biological failure modes, but IBA review also needs a
root-cause call: is the source annotation bad, or is the propagation bad? Use the
same review.propagation_review structure as the
ISO review project: root_cause, optional
failure_modes, and per-source source_entities with short source-specific
comments.

Root-cause classes

These are the schema values for propagation_review.root_cause.

Biological subtypes

These are the schema values for propagation_review.failure_modes.

Source status values

These are the schema values for propagation_review.source_entities[].source_status.

How many sources to enumerate

source_entities is a curated subset, not an exhaustive mirror of the row's
WITH/FROM. Enumerate the sources the review's argument actually rests on; where a
column carries dozens of donors, characterise the span in prose and name the ones that
matter. In mouse alone, Mapk1 GO:0035556 has 53 donors, Ccnb1 GO:0005634 has 45 and
Egfr GO:0005886 has 67 — listing all of them is noise, not rigour, and it buries the
two or three that carry the reasoning.

The rule that does bind: no sentence may claim more than the enumeration shows.
A block that names one seed and calls it "the IBD seed" asserts a sole-donor fact the
WITH/FROM may contradict, and that is a defect regardless of how many sources are
listed. Definite singulars — "the IBD seed", "the sole donor", "the only seed" — must be
checked against the donor count before use; prefer "one of the IBD seeds" whenever more
than one gene-level donor exists.

Two traps when counting donors:

• The PANTHER:PTN… entry is the node, not a donor. A row whose WITH/FROM is
  MGI:MGI:95407|PANTHER:PTN002571322 has exactly one donor, so "the IBD seed" is
  correct there.
• The same entity can appear twice under different identifiers — RGD:2275 and
  UniProtKB:P55213 are both rat caspase-3. Two entries, one donor.

Two mouse blocks were corrected under this rule, both by rewording the claim and adding
the co-donors: Grpel2 GO:0051082 said "the IBD seed" while the row carries a human
GRPEL1 co-donor, and Gulo GO:0016491 said the same while carrying seven gene donors
spanning fungi, plants and bacteria. Both now enumerate their full WITH/FROM.

Resolving a donor before calling it unresolvable

interpro/panther/<FAMILY>/<FAMILY>-entries.csv is keyed on UniProt accession and
covers the whole family, not just representative_members. It gives the protein name,
gene symbol and source organism, so it resolves donors that a grep over genes/ will
not:

$ grep '^P9WIT3,' interpro/panther/PTHR43762/PTHR43762-entries.csv
P9WIT3,"L-gulono-1,4-lactone dehydrogenase",protein,83332,
       Mycobacterium tuberculosis (strain ATCC 25618 / H37Rv),…,Rv1771,…

Check it before writing "the local index does not resolve this". The CLAUDE.md
restraint is assert nothing you cannot establish — not assert nothing about UniProt
ids, and a resolved donor usually strengthens the block's argument rather than being
mere bookkeeping. Note the limits: it is accession-keyed, so FB:/SGD:/RGD:/MGI:
identifiers are not lookups in it, and a family whose directory is absent from
interpro/panther/ resolves nothing.

Search every family's index, not just the node's own. A seed can sit in one family's
IBD while PANTHER classifies the protein into a different family, so the node's own
entries.csv will not contain it and a family-scoped lookup files it unresolvable when it
is in fact resolvable. Acadl's bacterial seeds are the worked case: both are seeds of
PTN002535634 in PTHR43884, but only P60584 (caiA) is in PTHR43884-entries.csv —
Q47146 (fadE) resolves in PTHR48083-entries.csv. The reason is one column away in
that row: Q47146 is classified into PTHR48083:SF18 ("ACYL-COENZYME A DEHYDROGENASE"),
so PANTHER's classification and the node it seeds simply are not the same family. So the
lookup is

grep -l '<ACC>' interpro/panther/*/*-entries.csv

rather than a single family's file. This is the same widening the WITH/FROM route already
gets below: "the family has no index" and "the node's family has no index" are both
weaker than "no index has it".

A MOD-id seed can therefore only be name-matched, and a name match returns every
species' ortholog — so check the taxon column. PTHR10836 carries Gapdh2 for
D. melanogaster (P07487), D. pseudoobscura (O44104) and D. subobscura
(O44105); matching on the gene name alone picks whichever comes first. A name match
corroborates that the family contains such a gene; it does not establish that the
FB:/SGD: id in the WITH/FROM is that entry. Say "corroborated", not
"resolved", when that is what happened.

In this corpus "corroborated through the UniProt cross-reference for this MOD id
(ACC)" is the standard phrasing for that situation: the family index confirms ACC
is the named protein, while the MOD-id-to-accession step is an inference rather than a
lookup. Reserve "resolved" for a step some file in the repository actually performs. And there
is a third state the substitution above can fall through: the family has no local directory
and the accession appears nowhere, so nothing corroborates it either. Say that
explicitly — "asserted from external knowledge, not corroborable here" — rather than
reaching for the weaker verb, which still implies a check that did not happen.

The mechanical test for which of the three verbs applies is whether any family's index
resolves the accession — the widened grep -l above, not the node's own family, for the
reason that paragraph gives — so check that before choosing the wording, not after. The
narrower question (does the node's family have a directory under interpro/panther/?)
settles it only when the answer is yes; a no leaves the seed still possibly resolvable
elsewhere, which is exactly the case Acadl's Q47146 turned out to be.

Before re-checking any existing "the local index does not resolve it" comment, work out
which seed it is about, and scope whatever you then claim to the comments you actually
identified. Two ways of assembling that set have already failed here, in opposite
directions, so the instruction comes before the criterion rather than after it:

• Grepping the prose under-counts and mis-classifies. The wording varies ("could not be
  resolved to a named gene product", "do not resolve against the local caches", "not
  resolved to a protein"), and three reasonable patterns return three different sets. That
  is how a claim that every such comment cites a MOD id got written here, when two of them
  cite accessions.
• Keying on the comment's own source_id misses the node-level ones. A comment can
  make an unresolvability claim about seeds it is not attached to: Gulo's self-seed
  comment sits under an MGI: id while naming UniProtKB:Q57ZU1 among donors that "do not
  resolve against the local caches", Ghr's sits under a PANTHER:PTN… id while saying
  "several" of the row's further accessions could not be resolved, and Notch1's sits
  under a PANTHER:PTN… id while naming four donors — a WB:, a FB: and two ZFIN:
  ids — that "resolve nowhere in the local caches". Three blocks, so this is the pattern
  rather than a pair: read the comment's content, not just its key.

The criterion itself is a property, not a list of the prefixes anyone has happened to see:
a seed is an index lookup iff it is a UniProtKB: accession. Nothing else is a key in
an accession-keyed file, so running the widened grep on anything else returns nothing
for a reason that says nothing about the seed — MOD identifiers (MGI:, SGD:, FB:,
RGD:, PomBase:, ZFIN:, CGD:, TAIR:, WB:, dictyBase:, AGI_LocusCode:) and
the non-MOD namespaces a WITH/FROM also carries (PANTHER:, InterPro: and ensembl:
among many more) alike. No count or per-directory scope is given for those prefixes on
purpose: an earlier revision said which seven genes/mouse used and was wrong by one
(Cftr cites a TAIR: id) — a list offered as illustration, read back as the set to
match against, which is the failure this whole passage exists to prevent. Match on the
property. ensembl: is the one worth a second sentence, not as an exception but because
it is the case a reader cannot settle from the prefix: an Ensembl protein accession
looks like something an accession-keyed index might carry, and what actually excludes it
is that no ENS…P… identifier is a key in any *-entries.csv. For anything that is not a
UniProtKB: accession the routes are instead the name match described above and the
target's own UniProt DR cross-reference line. The consequence for re-checking: such a
comment is correct under the widening without being re-run, while an accession-cited one
is exactly what the widened grep can move. Ccnt1's are the accession-cited ones this
sweep found, and they still resolve nowhere.

That is what separates Gulo (P9WIT3 resolves to M. tuberculosis Rv1771) from Bcl2
(PTHR11256) and Ednra (PTHR46099), whose accessions resolve in no family's
index, so no amount of grepping can corroborate the MOD-id-to-accession step. A bare grep
for the accession is not a substitute: a hit may be an unnamed ECO:0000250 or WITH/FROM
reference, or a substring collision with an EMBL id (AAP97287.1 matches P97287), and a
miss tells you only that the repository is silent. Check the route, not the grep.

The middle verb has a worked example too, and it is the one place all three states
appear in a single claim. genes/mouse/Bcl2 cites Q64373 as mouse Bcl2l1: the
accession is named in no cached record (so "resolved" is unavailable), and the
identification is nonetheless corroborated, because BCL2L1-goa.tsv carries it as
the mouse-ortholog donor (UniProtKB:Q64373|ensembl:ENSMUSP00000105445) in the
GO_REF:0000107 Ensembl-compara rows on human BCL2L1 — an orthology assertion is not a
name, but it establishes which gene the accession is. Note the route: an accession absent
from every entries.csv can still be corroborated by a GOA WITH/FROM column in
another gene's file, which is why "the family has no index" settles the first verb but
not the second.

Finally, do not assert the lookup in the same breath as disclaiming it. "The UniProt
cross-reference for this MOD id points to ACC" states the mapping as fact — which is
the step the sentence exists to say is uncheckable. Write "the accession ACC is asserted
from external knowledge" instead.

Universal quantifiers over a seed set need the same care as a definite singular:
"seeded entirely by X" or "every seed is an X" claims something about donors the block
may not have identified. Scope it — "every seed this review identified" — or name the
unidentified ones and say no claim is made about them.

When a REMOVE overrules a curator, and when it does not

CLAUDE.md forbids using REMOVE on an experimental annotation (IDA/IMP/IPI/IGI/IEP/EXP)
just because the cached title or abstract is about a different gene — the full text the
curator read is usually not in publications/. Applying that across genes/mouse needed a
test sharper than "is the reason biological?", and the two ways of getting it wrong are both
on the record here.

The condition, not the phrasing — but that is necessary, not sufficient. A first pass
selected rows by matching phrases in the reason and missed a row whose immediate neighbour it
caught — same gene, same PMID, same evidence code, same argument, differing only in that one
said "The paper concerns" and the other "This physiology belongs to". Key on the condition
instead: an experimental evidence code, a cited PMID whose cache is
full_text_available: false, and a reason that turns on what the paper contains or which
gene it studies. That took the class from the 25 the phrase list found to 42.

The class is actually 48 rows across 17 genes. The last six were routed into the keep
bucket because their shared reason cited biology — and it took reading each row's qualifier
to get them out. So the residual error does not live in the sweep; it lives in the triage of
what the sweep keeps, which is what the table below is for.

Then ask whether the cited biology contradicts this annotation, given its qualifier and
aspect. This is the step that separates a sound REMOVE from one that only sounds sound:

The last row is the trap. Six Bcl2 rows carried "…or describes an enzymatic/molecular
activity not enabled by Bcl2" — true of Bcl2, and irrelevant to every row it was applied to.
Four were qualified acts_upstream_of_or_within, which asserts regulation, not
catalysis; two were located_in cellular-component calls, which assert neither. Read the
qualifier and the aspect before accepting that a biological argument bites.

A shared reason is suspicious only under a corrective action. Reuse alone is not a
signal: a reason string whose rows span ≥3 distinct terms occurs in 201 groups across
genes/mouse, up to 243 rows on one string. Where those groups sit is the whole point:

(Neither column sums to the unscoped total — 200 groups against 201, and 5409 rows against
5432. Qualification is per-scope: a group can clear ≥3 terms overall and not within one
action, or the reverse. The clearest case is Secondary or downstream function, 243 rows
that split 228 KEEP_AS_NON_CORE + 15 ACCEPT, qualifying in one scope and not the other.)

Reproduce every figure here, as of the commit that ships the script
(git log --oneline -1 -- projects/IBA_REVIEW/shared_reason_groups.py):

S=projects/IBA_REVIEW/shared_reason_groups.py
python3 $S
for a in ACCEPT KEEP_AS_NON_CORE MARK_AS_OVER_ANNOTATED MODIFY REMOVE UNDECIDED; do
    python3 $S --action "$a"
done
python3 $S --action REMOVE --list
python3 $S --action MARK_AS_OVER_ANNOTATED --list

REMOVE is not the whole of "corrective", and scoping to it hid the larger half. An
earlier revision of this paragraph reported only the REMOVE row — 10 groups over 93 —
and called the rest "nearly all KEEP_AS_NON_CORE or ACCEPT". MARK_AS_OVER_ANNOTATED is
corrective too, and carries 587 rows, six times as many, with a single string on 202 of
them. "Nearly all" was also wrong on its own terms: KEEP_AS_NON_CORE and ACCEPT together
are 85% of scoped rows and 72% of groups, not nearly all.

The lesson is about operationalization, not arithmetic. The rule names a property
("corrective"); the measurement named one action; and because the narrower figure was the
one that made the point cleanly, nothing prompted a check that the two matched. When a rule
and its measurement are worded differently, the gap is where the finding hides.

The REMOVE breakdown, for reference, is Casp3 23; Ang2 18, 8, 4 and 3; Uox 12; Hsp90aa1 8;
Cyp1a1 7; Agtr1a 5; Ednra 5.

A caveat that the MARK_AS_OVER_ANNOTATED figure does not settle. The two actions are
not equally exposed to the objection. REMOVE withdraws an annotation, so its reason has to
carry a per-annotation verdict; MARK_AS_OVER_ANNOTATED withdraws nothing and records a
scoping judgement, where one class rationale across many terms is more often the honest
description. So the 587 is not 587 defects. But it is 587 rows the rule as stated reaches and
the measurement did not — and the second-largest group is a clear instance: Bcl2's 35 rows
over 19 terms on "this term is too indirect, broad, or mechanistically ambiguous", a
three-way disjunction that never says which disjunct applies, which is the shape dismantled
under REMOVE in that same file.

Apply this section's own qualifier-and-aspect test and the disjunction visibly breaks. Those
19 terms span all three aspects and five qualifiers in Bcl2-goa.tsv: 3 enables MF, 21
involved_in and 8 acts_upstream_of_or_within BP, 2 located_in and 2 part_of CC (36
GOA rows for 35 review rows — one term carries a duplicate). So "mechanistically ambiguous"
cannot be what is wrong with GO:0043209 myelin sheath, a curated located_in IDA on
PMID:7953633; and "too indirect" cannot be what is wrong with GO:0015267 channel activity,
an enables MF claim. Each row probably has one fitting disjunct and the reader is never told
which. That is the argument for the follow-up, recorded here so the deferral is not
mistaken for a judgement that the group is fine. The Bcl2 group is measured and recorded,
not repaired: doing it properly is per-term work across its 19 terms, and this
paragraph's own history is a warning against opportunistic rewrites (see the Ang2 note
below).

The largest group, 202 rows on "This overstates the direct role of the gene product; the
curated model…", is a separate case — a cross-file class rationale spanning far more terms,
and mostly though not wholly outside this review: Edn1 114, Ednra 32, Tnfrsf1a 31, Agtr1a 25,
so 88 of the 202 sit in files edited here.

That script exists because these figures had been wrong twice before it did, the second
time in a way no reader could have caught. The published pair — 160 groups, and 7 over 63
under REMOVE — turned out to come from three mutually inconsistent predicates in a single
sentence: the REMOVE breakdown reproduces only with a ≥40-character minimum on the reason
string, the "177 rows" figure only with rows deduplicated by (file, term), and 160 reproduces
under neither. grep-able counts of a folded scalar need a parser, and a figure a reader
cannot reproduce does not go wrong quietly — it goes wrong invisibly. State the predicate,
ship the command, and quote no number you cannot re-run.

The length cutoff was not a harmless tuning choice either. It hid Ang2's three largest
REMOVE groups — 20, 8 and 4 rows on "No direct mouse evidence.", "Stale ISO transfer." and
"Do not stack inference on inference." — the most boilerplate-looking rows in the corpus and
exactly the ones the paragraph wants counted. So no minimum length is applied here, and
counting those 32 rows is what turned up the finding below.

What the count is for: a pointer, not a verdict

All 32 are ISO or IEA — no experimental row among them — and inspection split them in two.
The question that separates them is what the reason is about:

• 12 rows (8 + 4) shared one string in reason and one in summary, and are
  legitimately shared: "the current human source no longer carries this term", "the
  current human source is itself inferred-only (IBA)". Those are claims about the
  transfer, and one GO_REF:0000119 transfer from one human source fails the same way
  for every term it carried. Restating it eight times would be duplication, not diligence.
• 20 rows were the real defect, and not the one a shared string suggests. Their reason
  said only "No direct mouse evidence." — which is not an argument against an ISO at all,
  since transferred-without-mouse-data is precisely what ISO asserts. The row-specific
  material sat in summary ("Transferred heparin binding from human ANG, not shown for mouse
  Ang2"), which restates the same absence. So 20 corrective verdicts rested on a
  tautology, while the file's neighbouring rows carried the real argument — the functional
  divergence PMID:8633065 documents. Those 20 now carry it (18 ISO rows share the
  transfer-level wording; the 2 IEA rows name the automated import instead). No action
  changed; what changed is that the reason now says something that could be wrong.

And the first repair got the paper wrong, which is the part worth keeping. It said
PMID:8633065 showed Angrp "lacks Ang's angiogenic activity, the one function directly
compared". The abstract compares three things: Angrp is not angiogenic; that is not a
catalytic deficiency, because Angrp's ribonucleolytic activity toward tRNA is somewhat
greater than Ang's; and an inability to bind cellular receptors is implicated, with poor
conservation of the receptor recognition sequence 58-69. The middle finding is quoted as
supporting_text six times in the same file — on four ACCEPTed rows (GO:0004540 as
IBA, IEA and ISO, and the narrower GO:0004549 tRNA-specific ribonuclease as ISO), in
core_functions, and in the reference_review.findings — so one document was
simultaneously accepting that Angrp is the better RNase, on a more specific RNase term,
and removing RNase-driven terms on the ground that it had lost activity. The reasons now
state all three findings and rest on the receptor-binding defect, which is what actually
blocks the receptor-mediated uptake that ANG's nuclear and stress-response biology depends
on.

Note where the right reading already lived: the file's own references[].findings for
this PMID recorded all three results, including that the angiogenic loss is "attributed to
defective cellular receptor binding rather than loss of RNase catalytic capacity" — while 20
reason fields citing the same paper said otherwise. A reference_review is the curator's
reading of the paper; when a reason citing that paper disagrees with it, the reason is
the thing to check first.

The summary fields were left as they are, deliberately. "Transferred heparin binding from
human ANG, not shown for mouse Ang2" is an accurate description of the annotation and its
status; it was only a defect while it was also the whole argument. Summary describes, reason
argues — and conflating the two is what produced the tautology in the first place.

Note what the two groups have in common after the fix: the repaired reason is also one
shared string across 18 rows, and belongs there. Sharing was never the defect. A claim about
the term must be per-term; a claim about the source or the transfer covers every term that
transfer moved; and a claim that merely restates the evidence code is not a claim.

source_label is perspectival, and the three-verb standard does not reach it. The verbs
(resolved / corroborated / asserted from external knowledge) govern provenance claims in
prose, where the comment says how an identity was established. source_label is a readable
handle for the source as it stands to this target — the same role preferred_term plays
beside a PANTHER term.label, which CLAUDE.md already separates from the checked field.

That makes a cross-file consistency check on it actively wrong. Of 64 distinct MOD ids
carrying a label in genes/mouse (89 labelled sources; MOD = MGI/RGD/SGD/FB/WB/
ZFIN/TAIR/PomBase/dictyBase/CGD/Xenbase), three are labelled differently in
different files, and all three are correct:

The identity half agrees in all three; only the relationship clause differs, because the
relationship differs. So check the symbol, leave the parenthetical alone.

Where to check it needs saying, because the obvious answer does not reach most rows. When
this was first measured only 26 of the 89 labelled sources had a comment carrying a
provenance verb (32 at the time of writing, after the repairs below); on the rest the
comment is biological commentary ("A mammalian D-type cyclin seed at the same node") and the
symbol is asserted in the label alone, with nothing to check it against. So the rule
is not check the label against the comment — it is that the symbol is an identity claim
wherever it is written, and is owed the same standard there: establish it from the local
index (PAINT seeds, *-entries.csv, the GOA WITH/FROM) or write no label. A label whose
symbol you cannot establish is a reason to write no label, not to invent one; the
SGD:S000004812 seed in Ccnb1 GO:0005737 is cited by bare identifier for exactly that
reason, three lines below a labelled row whose comment makes no identity claim at all.

A self-label is not an identity claim, and does not need establishing. Of the 89, 27
name the review target itself — "mouse Ccnb1 (this gene)", "the review target itself", or the
bare symbol — where the symbol is the file's own gene_symbol and there is nothing to look
up. The rule reaches the other 62: labels naming a different gene, where the symbol
asserts something a reader could not otherwise check. 31 of those carry a provenance
clause; 31 do not, across ten files: Sox2 6, Mapk1 5, Ifi204 4, Mapk3 4, Drd1 3, Frmpd2 3,
Nf1 3, Agtr1a 1, Aldh2 1, Ccne1 1.

Reproduce the split, and the residual by file:

python3 projects/IBA_REVIEW/shared_reason_groups.py --classify-labels
python3 projects/IBA_REVIEW/shared_reason_groups.py --classify-labels --list

The classifier has to be stricter than "contains the target's symbol", and the first
version was not. An earlier revision matched the gene_symbol as a word anywhere in the
label and published 32/57/28/29 — five rows too many on the self side, every one of them a
cross-species ortholog: "rat Casp3" ×2, "rat Aldh2", "Drosophila Nf1", and a Fbxo2 label
whose whole point is "not mouse Fbxo2, whose own record is MGI:2446216". Those are exactly
the identity claims the rule exists for — "rat Casp3" in a mouse file asserts something about
a rat gene a reader cannot check — and the loose match filed them under "nothing to
establish". A label is a self-label only when it carries a self-marker ("this gene", "the
review target itself") or is the bare symbol (optionally "mouse "-prefixed and with a
parenthetical), and carries no other-species qualifier and no negation. Ednra's "zebrafish
ednraa" is why the symbol test must be an exact match rather than a substring.

That predicate is now is_self_label() in the script rather than prose, with the boundary
cases as self-test arms — because the bare-symbol form is the only one classified by
comparison rather than by a visible marker, and so the first thing a refactor breaks.
Hsp90aa1's "Hsp90aa1" and Grb2's "Grb2" are the self side of that boundary; Ccnb1's
"budding-yeast CLB5" is the foreign side. Substituting the old word-boundary predicate now
fails the self-test rather than silently republishing 32/57.

Two of the five were unprovenanced, so the residual gained a file. Aldh2 RGD:69219 is the
highest-leverage entry: genes/rat/Aldh2 does not exist, the accession appears nowhere
name-carrying in the repository, and it is the ISO donor for nine further rows in that same
file — one establish would ground ten rows.

The label is the weaker signal, and the accession is the stronger one. Two further
misclassifications of the same shape turned up outside mouse — ACRBP and ADAMTSL5 each
cite a mouse ortholog whose label the species list did not cover — and chasing them
through that list was the wrong layer. A MOD accession names a gene in exactly one
species, so MGI: in genes/human is the mouse ortholog however the label reads — and
because MOD_PREFIXES is derived from a MOD_ORGANISM map rather than listed separately,
every prefix in that map resolves. mod_source_is_foreign() applies that first; of the
825 labelled species-scoped sources across the whole corpus it settles 708, leaving 117
same-species rows to is_self_label(), whose live remaining job is separating the target
from its own paralogs ("mouse Bax" in Bcl2, "Dictyostelium cAR1-type paralog" in
carD: 77 of the 117). The word list is now the fallback, not the decision.

That matters because the label can be silent. Four rows move corpus-wide, all in
genes/human, and two of them no word list could ever have reached:
genes/human/FEN1 cites MGI:MGI:102779 as bare "Fen1" — case-insensitively equal to
its own gene_symbol FEN1, with no species word present to match. ACRBP's "Acrbp (Mus
musculus)" and ADAMTSL5's "Adamtsl5 (M. musculus)" the word list does now catch, but the
accession catches them too and would have without the patch. Afterwards genes/human has
0 MOD self-labels, which is right by construction: no MOD namespace maps to human, so a
human review's own annotation can never arrive under one.

Deriving one list from another checks it against nothing. MOD_PREFIXES is derived
from MOD_ORGANISM, which makes the two agree with each other and neither with the
corpus — and MOD_ORGANISM was itself written from a roster of model-organism databases.
It was wrong in both directions. AGI_LocusCode is used 22 times, 15 of them on labelled
sources, and was absent: because the derived MOD_PREFIXES gates classify_labels'
startswith filter, those rows were discarded before mod_source_is_foreign() could
see them, so they were missing from all four buckets and from the denominator. It is not
an exotic namespace — it is the Arabidopsis locus spelling of the species TAIR already
maps to (AGI_LocusCode:AT2G17800 where GOA writes TAIR:locus:2827916), i.e. the same
both-written-forms problem as Mus musculus / M. musculus, one field over. In the other
direction Xenbase sat in the map with 0 uses and a genes/XENLA value that is not a
directory — the identical invented-entry defect described below, in the map the gate
derives everything from, and the arm written to catch that class checked only the sibling
map. Adding the one and dropping the other moves the corpus denominator 714 → 729
(gate 597 → 612) and leaves mouse untouched.

The same grep then caught the roster one level up. ensembl: was excluded on the stated
ground that it "names no gene in exactly one species" — true of UniProtKB, PANTHER,
InterPro, GO, RHEA and EC, and false of ensembl: all 312 of its sources are
ENSMUSP (216) or ENSRNOP (96), so the species is readable from the accession string,
which is more than UniProtKB offers. That left 96 labelled sources outside the gate —
and the hole is precisely the FEN1 case, since a bare-symbol ensembl:ENSMUSP… label in
genes/human would have no namespace to catch it. Not live today (none of the 96 is filed
self by the label predicate), but closed rather than documented, because all 96 sit in
genes/human and none in genes/mouse, so nothing published moves: denominator 729 →
825, gate 612 → 708, the 117 same-species rows and their 77 paralogs unchanged.

--check-coverage now also prints the prefixes it declines instead of skipping them.
The roster deciding what counts as a gap was itself hand-written, so it could silently
narrow what the checker was allowed to find — which is exactly how ensembl stayed outside
the gate while the checker reported zero gaps. Printing the declined list puts that
judgement in front of the reader on every run.

--check-coverage now runs the corpus greps for all three lists and reports gaps in both
directions, so "the list matches the corpus" is something to run rather than something to
trust:

python3 projects/IBA_REVIEW/shared_reason_groups.py --check-coverage

It reports 0 gaps, and re-introducing any of the four original defects makes either it or
the self-test fail.

Its scan is always the whole corpus, not the glob, unlike every other mode here.
The maps it audits are global objects, and one of its two directions — "the roster carries
an entry the corpus never uses" — is simply false over a subset: under this script's default
mouse glob it reported four such gaps for AGI_LocusCode, WB, dictyBase and ensembl,
every one of which genes/ARATH, genes/worm, genes/DICDI or genes/human does use. A
roster audit whose answer depends on which files you happened to pass is not an audit, so
this mode does not scope its scan by the glob. (A glob matching nothing is still an error,
as in every mode — that catches the typo, it does not narrow the audit.)

ORGANISM_WORDS is keyed by directory, and its scope is the directories that carry a
source_label, enumerable with

grep -rl 'source_label:' genes/ --include='*-ai-review.yaml' | cut -d/ -f2 | sort -u

— every one of which must be a key, which --check-coverage enforces on each run. The
count is deliberately not written down here: it was, and it went stale the first time a
new organism gained a labelled source (POPTR, added as a key for exactly that reason). A
first pass at that list was assembled from a taxonomy rather than from the corpus and got
it wrong in both directions: it omitted VIBCH, PSEPK, NEUCR and ANOGA, which do
carry labels (PSEPK is the second-largest gene directory in the repo), while adding
PIG and XENLA, which are not directories in this repository at all. A self-test arm now requires every key to name a real genes/<key>
directory, which is what catches the invented half; the species names for the four real ones
are read off each directory's own taxon.label. CHICK is kept as a key with no labels
yet, because that directory does exist.

None of this moves mouse. Every mouse self-label is MGI-sourced — mouse's own MOD — so
the gate is a no-op there and the 27 / 62 / 31 / 31 figures above are unaffected. The
composition is named is_target_source() rather than inlined, so a self-test arm can bite on
the gate being applied; testing mod_source_is_foreign() alone would stay green if the call
were dropped from the caller.

Beware two numerical coincidences in the 27 / 62 / 31 / 31 split above, since they are the
kind that hide an error. 31 appears twice and means different things — 31 third-party
labels carry a provenance clause and 31 do not. And 32 is the count of all labelled
MOD sources whose comment carries one of the three verbs, which is 31 third-party plus
one self-label (Gulo), so 31 and 32 are both right about different sets rather than
one being a correction of the other.

The reason the split matters is that it stops the count from being read as a defect count.
Sixty-two unprovenanced labels would be alarming; 31 third-party labels in ten files is a
chore. Measuring the wrong denominator turns a chore into an alarm, or the reverse — and the
32/57 revision above did the second thing to itself.

A propagation_review with no source_entities is the norm in at least one file, not an
anomaly. 16 of the 76 blocks carry none — 15 in Mapk1, 1 in Scgb1a1 — so "no sources at
all" is not the tell for the stale-block defect this document catalogues. The tell was the
accepting action under a failure-asserting root cause, which is what the auditor's third
arm keys on; a corrective action with no sources trips nothing, and should not.

A propagation_review block is not owed symmetrically across paralogs; the action is.
Mapk1 carries 18 blocks against Mapk3's 2, and 11 of Mapk1's ISO-row blocks sit
on a (term, evidence) pair whose Mapk3 twin row has none — the ciliary trio
(GO:0005929, GO:0036064, GO:0097542) is the clean case: same term, same ISO evidence,
same GO_REF:0000119, same MARK_AS_OVER_ANNOTATED in both files, diagnosis recorded on one.

That asymmetry is not a divergence and --pair is right to report clean. A block is optional
on an ISO row by construction — 528 such rows corpus-wide carry none, because ISO is pairwise
ortholog transfer with no PANTHER node behind it — so its presence tracks which rows a
reviewer happened to work, while the verdict is what must agree. Mirroring a subset would
make things worse rather than better: filling the three ciliary blocks leaves eight twins
still bare, and a reader who saw the file made symmetric on three would reasonably infer it
was symmetric on all. Either do all 11 or state the convention. This states it — and names
the 11, so choosing the other option does not start with a re-derivation.

Mapk1's 18 blocks are 2 IBA + 16 ISO, and the 16 split 11 with a bare Mapk3 twin
plus 5 whose term has no Mapk3 row at all, which is what lets the paragraph be checked
against the files:

• the 11: GO:0005929, GO:0010759, GO:0032206, GO:0032991, GO:0034198,
  GO:0036064, GO:0046697, GO:0051403, GO:0061514, GO:0097542, GO:0120041
• the 5: GO:0010800, GO:0042307, GO:0043627, GO:0045542, GO:0045893

Every figure above comes from a parse of propagation_review.source_entities, not a grep:
159 of the 234 sources in genes/mouse carry a label, of which 89 are MOD-prefixed, split
27 self / 62 third-party, and 32 of the 89 carry a comment matching one of the three
provenance verbs (resolved / corroborated / asserted from external knowledge) — the
same regex that returned 26 before the repairs. Where a count here moves between revisions it
is usually the corpus that moved; where the split moved, it was the classifier. An
earlier revision of this paragraph said 66 ids over 91 sources, which is the same sweep run
as not-UniProtKB, not-PANTHER — a filter that also admits the two InterPro: sources in
Serpinh1. Same finding either way; different set.

A self-seed marks node membership always; it marks independent grounding only when the
target's own experimental annotation behind the IBD is itself RETAINED by this review.
The operative word is retained, not same-term. An earlier revision of this sentence said
"same-term", which is narrower than both PAINT semantics and the practice underneath it: an
IBD seed is a gene whose experimental evidence the curator judged relevant to the node, which
routinely means a descendant or mechanistically adjacent term rather than the propagated
one. Of the eleven blocks in genes/mouse carrying the canonical gloss
(grep -rn "grounding exists on the target" genes/mouse), only three ground on the block's
own term; six legitimately ground on a different one — Mapk1 GO:0035556 and GO:0007166
on GO:0000165, Mapk3's pair on GO:0070371, Nf1 GO:1902531 on GO:0005096 and
GO:0007265 (its comment says so outright: "which are among the descendant evidences behind
the IBD"), and Sox2 GO:0030182 on GO:0045665. A literal same-term check would report all
six as defects, and a rule that mostly flags correct work is a rule that gets ignored.

Acadl GO:0005737 is the cleanest case: it has no experimental GO:0005737 row at all —
the only one is the IBA itself — and its grounding is the located_in GO:0005739 mitochondrion
IDA (ECO:0000314, PMID:26767982), which QuickGO confirms has GO:0005737 among its
is_a/part_of ancestors and which this review retains as KEEP_AS_NON_CORE. Textbook under
the descendant reading; a defect under "same-term".

CLAUDE.md glosses a self-seed as "a marker that experimental grounding exists on the target
itself", and that is the usual case — but it presumes this review still stands behind that
annotation. Where the experimental row the self-seed rests on is one this review removes, the
self-seed still shows the target is inside the clade and SUPPORTS_TRANSFER is still right
(CIRCULAR_OR_REDUNDANT is forbidden for a self-seed), yet citing it as grounding leans on
evidence the same file rejects. Say node membership and say explicitly that grounding is
not claimed. One block in genes/mouse needed this — Agtr1a GO:0006954, whose own IGI
is REMOVEd on full text that is available. Its comment asserted the canonical gloss
until 9ce63aa88, so the block as it now stands is the corrected form, not the defect.
The other twelve self-seed claims cite annotations their files retain, so the canonical
gloss holds for them.

Note what carries the Agtr1a example: the IGI being removed, not the term matching.
The retained test also catches a case a same-term test would wave through. Grb2's block
sits on GO:0007165, and the experimental row nearest it by term is GO:0008180 COP9 signalosome (IDA, PMID:22561606) — which this very file marks MARK_AS_OVER_ANNOTATED,
so it is precisely what must not be cited as grounding. What does ground the block is
GO:0005091 guanyl-nucleotide exchange factor adaptor activity (IDA ×3, all ACCEPT), a
different term: the SOS1-recruitment adapter activity that couples receptors to Ras,
matching the block's own GO:0007165 → GO:0007265 refinement.

Finally, keep the row's own evidence pointing the same way as its verdict. When an action
is withdrawn, the summary and the supported_by have to move with it; otherwise the block
asserts a conclusion its action has given up. Both have been missed here — 42 summaries that
still argued for removal, and six supported_by entries citing the very argument the new
reason withdrew. A propagation_review self-seed comment is the same hazard in another
place: do not cite, as the target's experimental grounding, an annotation this review
elsewhere removes. Of 13 such grounding claims in genes/mouse, one (Agtr1a GO:0006954)
did exactly that, until 9ce63aa88 rewrote it to the node-membership form above.

What an IBA actually asserts, and two ways to misread the WITH/FROM

An IBA is not a similarity transfer from one gene to another. Behind every IBA is a
PAINT curator's IBD (Inferred from Biological aspect of Descendant): the curator
inspected the family tree and the multiple sequence alignment, read the experimental
annotations of all extant members, decided at which node the function arose — sometimes
recent, sometimes as deep as LUCA — and placed the assertion there. The IBA rows are then
the mechanical consequence of descent from that node. So an IBA carries a considered
phylogenetic judgment that no pairwise transfer does, and reviewing it means arguing with
that judgment, not with a similarity score.

Two corollaries, both of which are easy to get backwards:

A small donor list does not mean weak support. A node seeded by a single well-characterized
MOD or human gene can be perfectly strong, because the curator's claim is about where the
function arose, and they made that call with the whole alignment and the whole tree in view.
Counting donor genes is not a measure of evidential strength. If you want to challenge an IBA,
challenge the node placement — is the target inside or outside the clade that inherited the
function, and is there target-specific evidence of loss or divergence?

The target appearing in its own WITH/FROM is correct and expected. When a gene has its
own experimental annotation for the term, that annotation is one of the descendant evidences
the curator used to place the IBD, so the gene legitimately appears among the sources of the
IBA it later receives. This is not circular and not self-citation. It means the target's
own experimental data helped establish that the function is ancestral, and the IBA is then
saying something additional: that the function is inherited rather than lineage-specific.
Do not mark such a source CIRCULAR_OR_REDUNDANT and do not describe it as inflating
support.

CIRCULAR_OR_REDUNDANT is for genuine circularity — a propagation whose source is itself
a propagated annotation with no experimental grounding anywhere in the chain, or a source
that adds nothing because the target already has stronger direct evidence for the same claim.
Target-in-own-WITH/FROM is the opposite situation: it is a marker that experimental
grounding exists, and on the target itself.

Before a strong REMOVE on an IBA row, record that these checks were done:

• The GO term definition, aspect, qualifier, and taxon constraints were checked.
• The GOA WITH/FROM field was read and the PANTHER PTN... node and seed
  proteins were recorded.
• The target and seed proteins were placed in their PANTHER family/subfamily
  context. Cross-subfamily propagation is triage evidence, not a verdict.
• The source annotation or family-node assertion was checked separately from the
  propagation decision.
• Target-specific evidence was checked where relevant: direct experiments,
  curated NOT rows, active-site residues, domain architecture, localization,
  isoforms, organismal context, and lineage constraints.
• review.reason states the biological rationale, while
  review.propagation_review records the mechanical root cause, failure modes,
  and source entities.

Slides

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

IBA Quality Issues

1. Pseudo-Enzyme Propagation

The Problem: Enzymatic activity transferred to proteins that have lost catalytic function.

Example - Epe1 (S. pombe):
- IBA annotation: GO:0032452 (histone demethylase activity)
- Source: Related JmjC domain proteins with characterized demethylase activity
- Reality: Epe1's Fe(II)-binding triad is H297-E299-Y370 (Tyr370 replaces the third, His, iron ligand), and no demethylase activity is detectable in vitro
- Impact: Misleading annotation propagated via phylogenetic inference

2. Ubiquitin-Like Modifier Specificity: IBA as a Positive Control

The contrast: UBA7 shows the opposite of an IBA failure. The IBA rows correctly
capture the conserved ISGylation pathway, while naive domain propagation and some
non-IBA annotations blur ISG15-specific E1 activity into generic ubiquitin-like or
ubiquitin terms.

Example - UBA7 (human):
- Correct IBA annotations:
- GO:0019782 (ISG15 activating enzyme activity)
- GO:0032020 (ISG15-protein conjugation)
- GO:0045087 (innate immune response)
- Over-generic or wrong non-IBA rows:
- GO:0008641 (ubiquitin-like modifier activating enzyme activity) - IEA
from InterPro2GO; technically related but too general.
- GO:0004842 (ubiquitin-protein transferase activity) and
GO:0016567 (protein ubiquitination) - non-IBA ubiquitin terms that should
be redirected to ISG15 activation/conjugation.
- Reality: UBA7 specifically activates ISG15, not ubiquitin, in mammals.
- Impact: IBA provides a useful specific annotation that corrects the less
discriminating domain/keyword propagation.

3. Substrate Specificity Transfer

The Problem: Specific substrate preference may differ between orthologs.

Example - LPL1 (C. albicans):
- IBA annotation: GO:0004622 (phosphatidylcholine lysophospholipase activity)
- Source: S. cerevisiae ortholog with characterized PC specificity
- Reality: Enzyme has broad specificity for ALL glycerophospholipids
- Impact: Over-specific annotation from ortholog doesn't capture full activity

4. Neo-Functionalization: Opposite Function in Subfamily

The Problem: When a subfamily undergoes neo-functionalization to catalyze the opposite reaction, IBA from the family root propagates the wrong function.

Example - Cds1 (M. tuberculosis, V. cholerae) - PTHR10314 SF135:
- IBA annotation: GO:0019344 (cysteine biosynthetic process)
- Source: Root node (PTN000034104) of PTHR10314 family
- Reality: Cds1 catalyzes cysteine CATABOLISM (EC 4.4.1.1), the exact opposite!
- Impact: IBA propagates biosynthesis when the enzyme actually degrades cysteine to H2S + pyruvate

Why this is severe: The annotation isn't just wrong - it's directionally opposite. The enzyme produces H2S from cysteine degradation; the annotation says it synthesizes cysteine.

Root Cause Analysis:
| Evidence | SF135 (Cds1) | SF194/SF162 (Synthases) |
|----------|--------------|-------------------------|
| EC Number | 4.4.1.1 (lyase) | 2.5.1.47 (transferase) |
| Reaction | L-Cys → H2S + pyruvate | O-Ac-Ser + H2S → L-Cys |
| Branch length | 0.528 (longest) | 0.402-0.458 |
| Seq identity to synthases | ~24% | 40-43% between each other |
| Active site motif | ASSGST | PTSGNTG |

The subfamily SF135 shares only 24% identity with synthases - less than synthases share with each other (43%). The longest branch length indicates maximum divergence and neo-functionalization.

5. Paralog/Secondary Activity Transfer

The Problem: A real activity in one paralog or source subfamily can be
promoted to a related but distinct target whose closest characterized comparator
supports a different substrate class. This should not be called
KEEP_AS_NON_CORE unless there is evidence the target actually has the activity
as a secondary function.

Example - LPL1 (C. albicans):
- IBA annotation: GO:0047372 (monoacylglycerol lipase activity)
- Source trace: PANTHER:PTN000773837|SGD:S000003112, corresponding to the
S. cerevisiae ROG1 monoacylglycerol lipase source
- Closest characterized LPL1 comparator: S. cerevisiae LPL1, a lipid-droplet
phospholipase B acting on glycerophospholipids
- Action: UNDECIDED for C. albicans LPL1 pending source-tree and substrate
review; do not mark as non-core without evidence that Candida LPL1 hydrolyzes
monoacylglycerols

6. Organism/Tissue Context Transfer

The Problem: Annotations derived from organism-specific experimental systems carry that context to orthologs where it doesn't apply.

Example - RIMBP2 (human):
- IBA annotation: GO:0007274 (neuromuscular synaptic transmission)
- Source: Drosophila ortholog (FB:FBgn0262483) where NMJ is the primary synapse model
- Reality: Human RIMBP2 functions mainly at CNS synapses (hippocampal, auditory)
- Impact: Term implies NMJ function when actual function is at central synapses
- Root cause: IBA quality limited by organism-specific biases in source annotations

7. Pseudo-Enzyme Propagation Is a Recurring Human Pattern

The Problem: The Epe1 pseudo-demethylase case is not isolated. Catalytic-residue loss with fold retention recurs across human families, and IBA repeatedly transfers the ancestral enzymatic activity to the catalytically dead member. These are the most defensible REMOVE calls, because the catalytic deficiency is independently documented in UniProt.

Examples (REMOVE — each verified against the UniProt record, not just the review):
- DPYSL2 / CRMP1 / DPYSL3 — GO:0016812 (metallo-hydrolase activity, cyclic amides): the CRMP/dihydropyrimidinase-like proteins are explicitly flagged by UniProt CAUTION — "Lacks most of the conserved residues that are essential for binding the metal cofactor and hence for dihydropyrimidinase activity." Confirmed first-hand by MSA (msa/RESULTS.md): aligned against active dihydropyrimidinase (DPYS), all five CRMP/DPYSL paralogs have lost the carbamylated catalytic Lys (K159→L/M/Q) plus multiple Zn-coordinating His/Asp. They are non-catalytic cytoskeletal regulators.
- AGO4 — GO:0004521 (RNA endonuclease activity): UniProt FUNCTION states directly that AGO4 "Lacks endonuclease activity and does not appear to cleave target mRNAs" — only AGO2 is the catalytic slicer in humans. MSA confirms (msa/RESULTS.md) AGO4 carries two substitutions in the catalytic tetrad (D669G, H807R) vs intact AGO2; usefully, the same alignment shows AGO3 retains the tetrad, so the "non-slicer" story is residue-specific, not a blanket family claim.
- UBAC2 — GO:0004252 (serine-type endopeptidase activity): UBAC2 has a rhomboid-like fold (a known inactive-rhomboid/pseudoprotease clan) but UniProt attributes no protease function — its curated roles are as an ERAD/ER-phagy adaptor. Now confirmed by MSA (msa/RESULTS.md): against active GlpG/RHBDL2/PARL, UBAC2 has lost both residues of the Ser-His catalytic dyad (S→L131, H→A183). (This supersedes the earlier caveat that the residue loss was only inferred from the inactive-rhomboid classification.)
- AKTIP — GO:0061631 (ubiquitin-conjugating enzyme activity): UniProt CAUTION states it "Lacks the conserved Cys residue necessary for ubiquitin-[conjugating]" activity. It is in the E2 PANTHER family (PTHR24067) and the IBA is inferred from many genuine UBE2 enzymes, but as a UEV-domain protein it is a catalytically dead pseudo-E2 (a component of the FTS/Hook/FHIP complex). A second annotation even records the NOT form.
- DPYSL4 — GO:0016812: a fourth CRMP-family member (Dihydropyrimidinase-related protein 4) carrying the same metallo-hydrolase IBA on the same basis as DPYSL2/3/CRMP1 — the whole CRMP/dihydropyrimidinase-related clade is non-catalytic (see the MSA above); only true dihydropyrimidinase (DPYS) retains the activity.
- CASP12 (human) — GO:0004197 (cysteine-type endopeptidase activity): UniProt RecName is literally "Inactive caspase-12." Most humans carry a truncated, catalytically dead variant; the caspase-family protease IBA is over-propagated. Note the mechanism is truncation, not site degeneracy — MSA shows the full-length Csp12-L variant retains an intact His-Cys dyad and the canonical QACRG motif (msa/RESULTS.md). The inactivity comes from a nonsense variant at codon 125 in the reference allele, upstream of the catalytic domain.
- Serpinh1/HSP47 (mouse) — GO:0004867 (serine-type endopeptidase inhibitor activity): the textbook non-inhibitory serpin — UniProt describes a collagen-specific ER molecular chaperone ("Collagen-binding protein"), not a protease inhibitor. The serpin-fold IBA propagates inhibitory activity HSP47 does not have.
- Lesson: a degenerate/absent active site, independently documented (UniProt CAUTION/FUNCTION, an "Inactive"/non-inhibitory RecName, or — best — the missing catalytic residue seen in an MSA), is the strongest single signal that an enzymatic IBA is wrong. Residue-verified by alignment: Epe1 (Fe-ligand His370→Tyr, against its own KDM2A/KDM2B IBA donors), the CRMP family DPYSL2/3/4/CRMP1 (amidohydrolase), AGO4 and worm wago-4 (slicer), UBAC2 (protease), AKTIP (E2 ligase), ADGB (calpain), ADPRHL1 (ARH2), SEPHS1 (selenophosphate synthetase), CPS1 (GATase half-reaction), KDX1 (pseudokinase), SSZ1 (Hsp70 ATPase), cts2 (GH18 chitinase), PGRPLC (PGRP amidase), and Arabidopsis CRY1 (photolyase). HSP47 (protease inhibitor) rests on the non-inhibitory RecName.
- Counter-lesson — check the mechanism, not just the conclusion. A systematic residue pass over this pattern found 4 of 17 targets with a fully intact catalytic site: CASP12 (truncation), LPA (blocked zymogen activation junction), AZIN1 (loses substrate contacts, keeps the catalytic Cys — while its paralog AZIN2 does the opposite), and HSPA13 (intact nucleotide site; the issue is the substrate-binding domain). Each conclusion survives, but the stated reason did not. "Lacks the catalytic residues" is cheap to assert and was wrong about a quarter of the time here — see msa/RESULTS.md.

8. Partial Sub-Activity Loss Within a Multidomain Family

The Problem: Distinct from full pseudo-enzymes — the protein retains part of the ancestral activity but lost a specific sub-activity, and IBA transfers the lost sub-activity.

Examples (REMOVE — verified):
- CAPG (human) — GO:0051014 (actin filament severing): CAPG caps but does not sever. The original characterization (cached PMID:1322908) states verbatim that CAPG "reversibly blocks the barbed ends of actin filaments but does not sever preformed actin". The severing term over-extends from the gelsolin/villin family; CAPG retains capping only.
- human/CRYAA — GO:0042026 (protein refolding): αA-crystallin is an ATP-independent holdase that prevents aggregation but cannot refold clients. This one is corroborated inside GOA itself: there is an explicit curated NOT|involved_in (ISS) annotation to GO:0042026, which the IBA directly contradicts. UniProt describes only aggregation-prevention chaperone activity, no refolding.
- Worm small heat-shock proteins — GO:0042026 (protein refolding): a stronger version of the CRYAA case. hsp-16.2 is a holdase, not a foldase; and hsp-12.3 / hsp-12.6 have no chaperone activity at all — the title of PMID:9744800 is literally "…Hsp12.2 and Hsp12.3 form tetramers and have no chaperone-like activity." They are pseudo-sHSPs (tetramers/monomers rather than the large oligomers holdase function needs), so the family-level refolding IBA is fully refuted.
- Lesson: capping≠severing and holdase≠foldase are sub-activity distinctions that family-level IBA flattens. The CRYAA and hsp-12.3 cases are especially clean because a curator negated the term (CRYAA) or a paper demonstrated zero activity (hsp-12.3).

9. Regulatory-Sign Inversion Within a Family

The Problem: When a protein family contains members with opposite regulatory signs (activators vs inhibitors of the same process), a family-node IBA can transfer the wrong sign.

Example - BCL2 (human and mouse):
- IBA annotation: GO:0043065 (positive regulation of apoptotic process)
- Reality: BCL2 is the prototypical anti-apoptotic guardian; it inhibits MOMP and cytochrome c release (PMID:9027314, PMID:9219694).
- Root cause (verified from GOA WITH/FROM): the IBA for GO:0043065 is inferred from a PANTHER node (PTN000135648) whose WITH/FROM list mixes pro- and anti-apoptotic BCL2-family members — pro-apoptotic BAX (Q07812) and BAK1 (Q16611) alongside anti-apoptotic members. The shared BH-domain fold unites activators and inhibitors of apoptosis under one family, so the "positive regulation" sign can leak onto BCL2.
- Caveat (why this is "non-core" rather than flatly "wrong"): BCL2 does have documented context-dependent pro-apoptotic behavior (e.g. caspase-cleaved BCL2), and GOA additionally carries a separate NAS annotation (PMID:14634621, ComplexPortal) to the very same GO:0043065. So the honest framing is that the IBA family-node inference is unreliable for sign (verified mechanism), not that positive regulation is impossible for BCL2. The companion review more confidently re-points related terms to their negative-regulation children (e.g. GO:0001836 release of cytochrome c → GO:0090201 negative regulation of release of cytochrome c).

10. Complex / Compartment / Pathway Membership Over-Transfer

The Problem: A family-level IBA asserts membership in a specific complex, compartment, or pathway that the target protein does not actually occupy, even though the catalytic fold or sequence homology is real. Compartment-split paralogs are the classic trap: they share a fold but route their product to different destinations.

Examples (verified across multiple lines of evidence):
- EIF4E2 (human) — GO:0016281 (eIF4F complex): 4EHP/EIF4E2 binds the cap but UniProt states it "is unable to bind eIF4G" and "Does not interact with eIF4G"; it is a translational repressor (4EHP-GYF2 complex), never an eIF4F subunit. A clear family-level over-transfer from EIF4E.
- ALDH1L1 (rat) — GO:0005739 (mitochondrion): UniProt names it Cytosolic 10-formyltetrahydrofolate dehydrogenase with SUBCELLULAR LOCATION: Cytoplasm, cytosol and a cytosol IDA. Mitochondrial one-carbon oxidation is the job of the distinct paralog ALDH1L2.
- HMGCS2 (rat) — GO:0010142 (farnesyl-PP biosynthesis, mevalonate pathway): a paralog-pathway conflation. The IBA comes from a PANTHER node (PTN000222418) that lumps the HMGCS paralogs. The cytosolic paralog HMGCS1 feeds mevalonate→FPP→sterol/isoprenoid synthesis; mitochondrial HMGCS2's HMG-CoA is cleaved by HMG-CoA lyase to acetoacetate (ketogenesis). The shared HMG-CoA-synthase reaction is correctly classified under mevalonate biosynthesis (UniProt UniPathway tag), but assigning HMGCS2 to FPP/isoprenoid biosynthesis follows the wrong paralog's flux — an over-annotation. (Nuance: the enzymatic step is real, so this is paralog over-annotation, not a fabricated activity.)
- PEX2 (human) — GO:0016593 (Cdc73/Paf1 complex): PEX2 is a peroxisomal RING E3 ligase for PEX5 retrotranslocation (UniProt) and has no role in RNA Pol II transcription elongation, so membership in the Cdc73/Paf1 complex is clearly wrong. (Caveat: the cause is unconfirmed — the IBA WITH/FROM is a PANTHER node, not a PAF1 gene. A legacy synonym collision — PEX2's old name "PAF1"/Peroxisome Assembly Factor 1 vs the unrelated transcription factor PAF1 — is a plausible but unverified explanation.)
- CIRBP (human) — GO:0005681 (spliceosomal complex) + GO:0000398 (mRNA splicing, via spliceosome): the cold-inducible RNA-binding protein CIRBP shares only the N-terminal RRM with the transformer-2/RBMX splicing factors that anchor these terms. PANTHER PAINT shows the splicing IBD sits at ancestral node PTN000391532 (seeded by TRA2A, TRA2B, RBMX, Drosophila tra2, rat Tra2 — all bona fide splicing factors), while CIRBP's own subfamily node PTN008729690 carries only mRNA binding. CIRBP has no experimental splicing evidence; its function is 3'-UTR binding, mRNA stabilization and translational control. A non-enzyme instance of complex-membership over-transfer across a functional-divergence boundary. (See Featured Example and interpro/panther/PTHR48034/PTHR48034-review.md.)
- Lesson: complex membership, compartment, and downstream pathway are not conserved across paralogs even when the fold/reaction is; verify the protein actually occupies the annotated complex/compartment and that its product reaches the annotated pathway.

11. Substrate Over-Propagation From a Multi-Specificity Enzyme Family

The Problem: A PANTHER family lumps enzymes of different substrate specificities; a substrate-specific term is then propagated to a member that experimentally lacks that activity. Unlike LPL1 (a real but secondary/broader specificity), here the propagated activity is absent in the target.

Example - AGK (human) — GO:0001729 (ceramide kinase), GO:0046513/GO:0046512 (ceramide/sphingosine biosynthesis):
- AGK sits in PANTHER PTHR12358, which lumps "ACYLGLYCEROL KINASE" with "SPHINGOSINE KINASE." The ceramide-kinase IBA is propagated from a family node (Drosophila + mouse + node PTN008994514); the IEA/ISS variants both trace to one rodent ortholog (Q9ESW4).
- Three independent lines refute the activity in human AGK, outweighing the family inference:
1. The direct human enzymology paper (the one UniProt itself cites): "Significant phosphorylated products were only detected with monoacylglycerols and diacylglycerols as substrates, but not with any other lipid tested, including ceramide and sphingosine" (PMID:15939762).
2. A second study: "No evidence for phosphorylation of ceramide by the recently described multiple lipid kinase was found" (PMID:16269826).
3. UniProt's own FUNCTION line states "Does not phosphorylate sphingosine (PubMed:15939762)"; its sole ceramide claim is a weak "By similarity" tag (propagated from Q9ESW4), not direct evidence.
- The dedicated ceramide kinase is the separate enzyme CERK. AGK's verified activity is MAG/DAG kinase (plus a kinase-independent TIM22 structural role).

Example - SAMD8/SMSr (human) — GO:0033188 (sphingomyelin synthase activity):
- SAMD8 is in the sphingomyelin-synthase PANTHER family (PTHR21290), but UniProt's experimentally-supported FUNCTION says it makes ceramide phosphoethanolamine (CPE), transferring a phosphoethanolamine head group from PE to ceramide — explicitly not the phosphocholine-from-PC reaction that defines sphingomyelin synthases SMS1/SMS2: "The larger PC prevents an efficient fit in the enzyme's catalytic pocket." So the family-level sphingomyelin-synthase term is the wrong product/substrate; this is also accompanied by mislocalization IBAs (Golgi, plasma membrane) that belong to SMS1/SMS2, whereas SMSr is ER-retained.

Example - CPT1C (human) — GO:0006631 (fatty acid metabolic process), GO:0009437 (carnitine metabolic process):
- A neofunctionalization case: UniProt's RecName is literally "Palmitoyl thioesterase CPT1C." Although it sits in the carnitine O-acyltransferase family (PTHR22589) with CPT1A/B, experimental work shows CPT1C lacks the canonical carnitine palmitoyltransferase activity (it binds malonyl-CoA but does not catalyze carnitine-dependent acyl transfer). The IBA propagates the ancestral CPT1A/B fatty-acid/carnitine metabolism that CPT1C no longer performs.
Example - cao-1 / CAO-1 (Neurospora crassa) — GO:0010436 (carotenoid dioxygenase activity), GO:0016121 (carotene catabolic process):
- CAO-1 sits in PANTHER PTHR10543 subfamily SF89 (labelled "carotenoid 9,10(9',10')-cleavage dioxygenase 1"), a node that is functionally heterogeneous: it lumps genuine carotenoid cleavers (Arabidopsis CCD1, Synechocystis apocarotenoid oxygenase, M. tuberculosis Rv0654), stilbenoid/resveratrol cleavers (U. maydis RCO1, Botrytis rco1, and CAO-1 itself), and phenylpropanoid cleavers (Pseudomonas isoeugenol monooxygenase). The carotenoid-dioxygenase IBA propagates from node PTN001631894, whose WITH/FROM includes the M. tuberculosis carotenoid cleaver UniProtKB:P9WPR5.
- Direct experimental evidence refutes carotenoid activity: heterologously expressed CAO-1 did not convert β-carotene or any carotenoid/apocarotenoid tested, while it cleaves the interphenyl Cα–Cβ double bond of resveratrol and piceatannol (PMID:23893079). GOA already carries the corrective experimental NOT carotenoid metabolic process (GO:0016116, IDA, PMID:23893079). Crystal structures show the conserved four-His non-heme Fe(II) center but a stilbenoid-adapted substrate cleft (PMID:28493664).
- A blinded OpenScientist function-assignment run — given only the neutral hypothesis "cao-1 has carotenoid dioxygenase activity" — independently returned "REFUTED (over-annotated)" and likewise attributed the error to CCO/RPE65 (PTHR10543) family IBA, an independent confirmation of the manual call.
- Positive-control paralog (the decisive contrast): the N. crassa paralog CAO-2 (A7UXI1, NCU11424) carries the identical family IBAs — GO:0010436 and GO:0016121, both from GO_REF:0000033 — but for CAO-2 they are correct: CAO-2 is a genuine torulene dioxygenase (EC 1.13.11.59, KEGG KO K17842) and an integral step of the carotenoid biosynthesis pathway (KEGG ncr00906), whereas cao-1 (KO K28521) is mapped to no pathway. Same family term, same GO_REF, one genome — right for one paralog, wrong for the other, distinguishable only by target-specific experimental evidence (the direct assay + curated NOT on cao-1). The two sit in different PTHR10543 subfamilies (cao-1 SF89, cao-2 SF24) yet inherited the same ancestral carotenoid annotation.
- Action: REMOVE the carotenoid MF IBA; MODIFY the carotene-catabolic BP IBA to GO:0046272 (stilbene catabolic process). The accurate MF (GO:0016702 dioxygenase) is already present by IDA, and a class-level stilbenoid α,β-dioxygenase activity grouping term is proposed. root_cause: PROPAGATION_BAD, failure_modes: [FUNCTIONAL_DIVERGENCE].

Example - NQO2 (human) — GO:0003955 (NAD(P)H dehydrogenase (quinone) activity):
- NQO2 sits with NQO1 in the NAD(P)H:quinone oxidoreductase family (PANTHER PTHR10204). The family IBA transfers GO:0003955, which maps to EC 1.6.5.2 and specifies NAD(P)H as the electron donor — but NQO2 characteristically does not use NAD(P)H; it uses dihydronicotinamide riboside (NRH) (PMID:10945627). The divergence here is in the cofactor / co-substrate, not the cleaved-substrate class or the fold.
- A blinded OpenScientist run (neutral hypothesis "NQO2 has NAD(P)H dehydrogenase (quinone) activity") independently returned over-annotated → the NAD(P)H term is substrate-incorrect, citing Wu et al. 1997 (PMID:9367528) that NQO2 uses NRH "rather than NAD(P)H," and noting the correct term GO:0001512 is already annotated by IDA.
- Action: MODIFY to the NRH-specific GO:0001512 (dihydronicotinamide riboside quinone reductase activity; EC 1.10.5.1, RHEA:12364), already supported by IDA. root_cause: PROPAGATION_BAD, failure_modes: [FUNCTIONAL_DIVERGENCE].

• Lesson: a "By similarity"/propagated annotation is weak evidence; when direct experimental papers in the target species report the activity is absent or different in product (AGK no ceramide; SAMD8 makes CPE not SM; CPT1C is a thioesterase not a transferase; CAO-1 cleaves stilbenes not carotenoids; NQO2 uses NRH not NAD(P)H), the substrate/activity-specific IBA is an over-propagation. A family node that mixes substrate specificities (acylglycerol+sphingosine kinases; SM+CPE synthases; carotenoid+stilbenoid+phenylpropanoid cleavage oxygenases) leaks substrate terms across specificity boundaries — and the leak can be in the cleaved substrate (CAO-1) or the cofactor/co-substrate (NQO2). Where a subfamily label itself names one specificity (PTHR10543:SF89 = "carotenoid … cleavage dioxygenase") while spanning several, that label is the mechanical origin of the leak.

12. Mis-Grouping Revealed by the WITH/FROM Column

The Problem: The IBA WITH/FROM field names the exact source proteins the function was transferred from. Reading it frequently reveals the error directly — the source is either the wrong family entirely or the wrong paralog. This is the single most useful diagnostic in this whole catalog.

Tier A — wrong family / over-broad superfamily (egregious; the source proteins are functionally unrelated):
- NTN1 / NTN3 (human) — GO:0000981/GO:0006357/GO:0000978 (DNA-binding transcription-factor activity, Pol II transcription regulation, cis-regulatory DNA binding): Netrins are secreted axon-guidance cues (UniProt: extracellular; PANTHER PTHR10574 Netrin/Laminin) with no DNA-binding domain — yet they carry nuclear POU-domain transcription-factor IBAs. The WITH/FROM proves it: the source list is POU-domain TFs (POU2F1 P14859, POU1F1 P28069, POU4F1 Q12837, POU4F3 Q15319, …). A secreted protein cannot be a Pol II transcription factor; this is a phylogenetic grouping error.
- NOTCH1 (human) — GO:0007411 (axon guidance): the WITH/FROM is SLIT1/2/3 (O75093, O94813, O75094). NOTCH1 signals in neurogenesis but axon guidance is a SLIT function transferred across an over-broad node.
- IL23R (human) — GO:0004925 (prolactin receptor activity), GO:0017046 (peptide hormone binding): the WITH/FROM is PRLR (P16471). IL23R is a type-I cytokine receptor that binds the cytokine IL-23, not the hormone prolactin; the superfamily node is too broad.

Tier B — wrong paralog (subtle; the source is a close relative with a different function):
- ABRAXAS1 (human) — GO:0090307/GO:0008608/GO:0008017 (mitotic spindle assembly, spindle–kinetochore attachment, microtubule binding): every one of these IBAs traces via WITH/FROM to UniProtKB:Q15018 = ABRAXAS2 (ABRO1, the BRISC-complex paralog). ABRAXAS1 is a nuclear BRCA1-A DNA-damage scaffold; the spindle/MT biology belongs to ABRAXAS2.
- HINT2 (human) — GO:0005737 (cytoplasm): HINT2 has a mitochondrial targeting sequence and is mitochondrial; the cytoplasm term reflects the HINT1 paralog.
- CPT1C (above) similarly inherits CPT1A/B metabolism it no longer performs.
- opa1 (zebrafish) / eat-3 (worm) — GO:0016559 (peroxisome fission): both are UniProt "Dynamin-like GTPase OPA1, mitochondrial" inner-membrane fusion proteins. Peroxisome fission is done by the DRP1/DNM1L branch of the dynamin superfamily; the term is a within-superfamily mis-transfer (wrong organelle and wrong direction).
- YAR1 (yeast) — GO:0045944 (positive regulation of transcription): YAR1 is an RPS3-binding 40S-ribosome-biogenesis factor (UniProt: interacts with RPS3), not a transcription activator; ACL4 (yeast) likewise gets mitochondrial-import terms by TOM70-family over-transfer despite being an Rpl4 chaperone.
- Lesson: always read the WITH/FROM before flagging. It tells you whether the IBA is a defensible family-level transfer or a traceable mis-grouping — and if a single paralog or out-of-family protein is the source, that is strong, near-mechanical evidence of error.

13. Generic / Mutually-Exclusive Compartment Over-Propagation

The Problem: Localization is one of the most frequently over-propagated IBA categories — but whether a flag is valid depends entirely on the GO compartment hierarchy, which makes this a two-sided pattern. Mutually-exclusive compartments are valid REMOVE grounds; broad subsuming terms are not.

Tier A — valid REMOVE: a mutually-exclusive specific compartment on a protein that lives elsewhere. GO:0005634 nucleus is the one compartment the cytoplasm definition explicitly excludes, and plasma membrane / peroxisome / a specific organelle are likewise non-overlapping — so these are genuine errors:
- Cytoplasmic PIWI/Argonaute & germ-granule proteins given GO:0005634 nucleus — PIWIL1 (human), and worm prg-1, wago-1, glh-1. All are cytoplasmic nuage/P-granule/chromatoid-body proteins (UniProt: cytoplasmic granule, no nucleus). The WITH/FROM nodes include nuclear-acting Piwi orthologs (e.g. Drosophila Piwi is nuclear; nuclear PIWIL4/MIWI2), so the nuclear compartment leaks onto the cytoplasmic members.
- EIF2AK3/PERK → nucleus (UniProt: ER membrane kinase) and BIRC6 → nucleus (UniProt: TGN/endosome/cytoskeleton/midbody — no nuclear pool).
- Ribosome-associated chaperones SSB2 / SSZ1 (yeast) → GO:0005886 plasma membrane (UniProt: cytoplasmic, ribosome-associated) — PM propagated across the HSP70 family node.
- BAIAP2L2 → GO:0005654 nucleoplasm (UniProt: plasma membrane / cell junction; I-BAR family) and PIK3C3/VPS34 → GO:0005777 peroxisome (UniProt: autophagosome/endosome/midbody).
- Inverse — strictly nuclear proteins given GO:0005737 cytoplasm: rqh1 (RecQ helicase) and HDA1 (HDAC) are nucleus-only, and nucleus is excluded from cytoplasm, so cytoplasm is wrong. And the genuinely extracellular SCGB1A1 given cytoplasm (secreted = outside the cell).

Tier B — anti-pattern (do NOT flag; these reviewer REMOVEs were over-reaches). GO:0005737 cytoplasm subsumes mitochondrion, ER, Golgi, and lysosome (all part_of cytoplasm), so "cytoplasm" is defensible — if imprecise — for an organellar protein:
- "cytoplasm" REMOVE on Aga / GLA (lysosome), DHCR24 (ER membrane, catalytic domain faces the cytosol), ISCA1 / ATP5IF1 / gtpbp3 (mitochondrion) — all should be UNDECIDED/KEEP, not REMOVE.
- "membrane" (GO:0016020) REMOVE on flvcr2a is wrong — it is a multi-pass membrane transporter.
- Self-correction: HINT2's "cytoplasm" flag (added in the WITH/FROM pass) belongs here too — HINT2 is mitochondrial, but mitochondrion ⊂ cytoplasm, so cytoplasm is not strictly wrong; downgraded from the findings.

Lesson: before a localization REMOVE, place both compartments in the GO hierarchy. Mutually-exclusive (nucleus vs cytoplasm; PM vs internal; one organelle vs another) → valid. A broad subsuming term over a more specific true location (cytoplasm over any organelle; membrane over a membrane protein) → leave it.

14. Lineage-Inappropriate (Cross-Kingdom) Process Transfer

The Problem: Phylogenetic inference crosses kingdom/clade boundaries and lands a biological-process term on an organism where the process does not exist — or where the homolog was repurposed into a different system. The WITH/FROM typically names a vertebrate or Drosophila source. This is one of the largest and cleanest BP error classes.

Examples (REMOVE — verified via UniProt + WITH/FROM):
- TOLL9 (mosquito, ANOGA) — GO:0006954 (inflammatory response): the IBA traces to human TLR4 (O00206) and other mammalian TLRs. Insects have Toll-pathway innate immunity but no vertebrate inflammation (no vasculature, immune-cell infiltration); the term is lineage-inappropriate.
- ndhA / ndhD / ndhK (poplar, POPTR) — GO:0009060 (aerobic respiration): UniProt labels these "NAD(P)H-quinone oxidoreductase, chloroplastic" (OG Plastid; Chloroplast). They are photosynthetic plastid NDH subunits homologous to mitochondrial complex I — an organelle-system swap, not mitochondrial respiration.
- che-3 (worm) — GO:0060294 (cilium movement involved in cell motility): che-3 is cytoplasmic dynein-2 (retrograde IFT motor); C. elegans sensory cilia are non-motile. The motility term comes from axonemal-dynein orthologs in organisms with motile cilia.
- D7 salivary proteins (mosquito, ANOGA: D7r2/D7r4/D7r5/D7L1) — GO:0007608 (sensory perception of smell): UniProt calls D7r4 a "salivary protein… modulates blood feeding," female-saliva-specific. The OBP/PBP-GOBP fold was repurposed for binding biogenic amines/eicosanoids in saliva — these proteins are not expressed in antennae and have no olfactory role.
- sta-2 (worm) — GO:0007259 (JAK-STAT signaling): transferred from fly/mammalian STATs, but C. elegans has no JAK kinases; STA-2 is activated via SNF-12/hemidesmosomes.
- fshr-1 (worm) — GO:0009755 (hormone-mediated signaling): C. elegans lacks gonadotropins (FSH/LH/TSH); FSHR-1 functions in innate immunity/stress.
- HEN1 (Arabidopsis) — GO:0034587 (piRNA processing): piRNAs are metazoan; plant HEN1 methylates miRNA/siRNA duplexes. Over-transfer from the metazoan HEN1/HENMT1 context.
- Lesson: check taxon appropriateness — does the process even occur in this lineage? GO taxon constraints catch some of these; the WITH/FROM naming a vertebrate/insect source is the tell. Watch especially for organelle-system swaps (plastid↔mitochondrion; cytoplasmic↔axonemal dynein).

15. Regulator / Effector and Direct / Downstream Conflation

The Problem: IBA (and curation generally) can blur the line between a regulator or effector of a process and the core machinery, or between a downstream consequence and the direct function.

Examples (verified):
- lys-7 (worm) — GO:0007165 (signal transduction): LYS-7 is an antimicrobial effector whose expression is regulated by signaling; it is not itself a signaling component. (Same logical error as arnF's "response to iron" — see §arnF.)
- SIR3 (yeast) — GO:0006270 (DNA replication initiation): SIR3 represses origin firing (negative regulation of MCM loading), the opposite of being part of the initiation machinery (ORC/CDC6/CDT1/MCM2-7).
- UBP3 (yeast) — GO:0031647 (regulation of protein stability): a downstream consequence of its deubiquitinase activity (GO:0004843, the direct function), not a separate function.
- sigF / sigG / sigK (B. subtilis) — GO:0003899 (DNA-directed RNA polymerase activity): these are sigma initiation factors (UniProt: "initiation factors that promote…" promoter recognition) — they confer promoter specificity to RNA polymerase but have no catalytic polymerase activity, which belongs to the core enzyme (RpoB/RpoC). A regulatory subunit assigned the holoenzyme's catalytic activity.
- Lesson: distinguish "does X" from "regulates/enables/results-in X." Effectors are not signal transducers; repressors are not part of the machinery they inhibit; a specificity subunit does not carry the catalytic activity of the complex it joins; downstream consequences are not direct molecular functions.

Featured Examples

Epe1 - Pseudo-Demethylase

Species: pombe
Status: COMPLETE

IBA Annotations Flagged:
| Term | Issue | Action |
|------|-------|--------|
| GO:0032452 histone demethylase activity | Pseudo-enzyme - no activity | REMOVE |
| GO:0006338 chromatin remodeling | Too broad | MODIFY |
| GO:0006357 regulation of transcription by RNA pol II | Valid | ACCEPT |
| GO:0003712 transcription coregulator activity | Valid | ACCEPT |

Lesson: IBA can propagate enzymatic activity to pseudo-enzymes that retain the domain fold but lack catalytic function.

UBA7 - ISG15-Specific E1

Species: human
Status: COMPLETE

Annotations reviewed:
| Term | Issue | Action |
|------|-------|--------|
| GO:0005737 cytoplasm | IBA, correct | ACCEPT |
| GO:0019782 ISG15 activating enzyme activity | IBA, correct and specific | ACCEPT |
| GO:0032020 ISG15-protein conjugation | IBA, correct | ACCEPT |
| GO:0045087 innate immune response | IBA, correct | ACCEPT |
| GO:0006974 DNA damage response | IBA, correct | ACCEPT |
| GO:0008641 ubiquitin-like modifier activating enzyme activity | InterPro2GO IEA, too general | MODIFY to GO:0019782 |
| GO:0004842 ubiquitin-protein transferase activity | Non-IBA ubiquitin term, wrong activity class | MODIFY to GO:0019782 |
| GO:0016567 protein ubiquitination | UniPathway IEA, wrong modifier process | MODIFY to GO:0032020 |

Lesson: UBA7 should be highlighted as a positive-control case for IBA
specificity. The bad rows are not IBA propagation errors; they are generic or
ubiquitin-biased non-IBA mappings that the IBA annotations help correct.

LPL1 - Phospholipase Specificity

Species: CANAL
Status: COMPLETE

IBA Annotations Flagged:
| Term | Issue | Action |
|------|-------|--------|
| GO:0006629 lipid metabolic process | Correct | ACCEPT |
| GO:0004622 PC lysophospholipase activity | Too narrow | MODIFY |
| GO:0005811 lipid droplet | Correct | ACCEPT |
| GO:0047372 monoacylglycerol lipase activity | ROG1-paralog/substrate-specificity transfer; target evidence absent | UNDECIDED |

RIMBP2 - Context-Specific Term Transfer

Species: human
Status: COMPLETE
PANTHER Family: PTHR14234 (RIM BINDING PROTEIN-RELATED)

IBA Annotations Flagged:
| Term | Issue | Action |
|------|-------|--------|
| GO:0007274 neuromuscular synaptic transmission | Wrong synapse type context | MODIFY |

Lesson: IBA transferred a term specific to Drosophila NMJ context to a human gene that functions primarily at CNS synapses. The IBA source (FB:FBgn0262483) is the fly ortholog where neuromuscular junctions are a major experimental system. However, human RIMBP2 functions at hippocampal (mossy fiber, CA3-CA1), auditory ribbon, and other central synapses - not primarily at neuromuscular junctions. This illustrates how IBA can propagate organism-specific or tissue-specific contexts that don't apply to the target species.

Root Cause Analysis: This is a case where IBAs are only as good as the manual annotations on orthologs. The Drosophila RIMBP ortholog is well-characterized at the NMJ because that's the major accessible synapse type in flies. When this annotation gets transferred to human via phylogenetic inference, it carries the fly-specific context with it.

Family-Level Context: The PANTHER family analysis reveals additional IBA quality concerns:
- The representative structure (PDB 4z8a) is from Drosophila RIM-BP bound to Cacophony (fly NMJ Ca2+ channel)
- The family contains functionally divergent subfamilies:
- RIMBP1/2 (SF18, SF20): Neuronal synaptic scaffolds
- RIMBP3 (SF21): Non-synaptic - testis-specific, spermiogenesis/manchette function
- The family research explicitly warns: "Avoid propagating 'regulation of neurotransmitter release' to RIMBP3 paralogs"
- This demonstrates that IBA issues can affect entire subfamilies, not just individual genes

arnF (PTHR30561) - Functional Divergence Within SMR Superfamily

Species: ECOLI (Escherichia coli K12)
Status: COMPLETE
Family: PTHR30561 (Small Multidrug Resistance / Drug-Metabolite Transporter superfamily)

IBA Annotations Flagged:
| Term | Issue | Action |
|------|-------|--------|
| GO:0022857 transmembrane transporter activity | Too generic; conflates drug efflux with lipid flipping | MODIFY → GO:0140303 |
| GO:0005886 plasma membrane | Correct | ACCEPT |
| GO:0055085 transmembrane transport | Acceptable broad term | ACCEPT |

Lesson: The PANTHER family PTHR30561 groups four functionally distinct protein families that share the SMR/DMT fold:
- ArnE/ArnF — undecaprenyl phosphate-L-Ara4N flippases (intramembrane lipid translocation)
- EmrE (P23895) — multidrug efflux pump (exports lipophilic cations)
- MdtI/MdtJ (P69210/P69212) — spermidine export
- Gdx/SugE (P69937) — guanidinium export
- Mmr (P9WGF1) — M. tuberculosis multidrug resistance

The IBA inference of "transmembrane transporter activity" comes from propagating annotations from these bona fide solute exporters (EmrE, MdtI/J, Gdx, Mmr) to ArnF. But ArnF does something mechanistically different: it flips a lipid-linked sugar between membrane leaflets, not export a solute across the membrane. The correct MF term is GO:0140303 (intramembrane lipid transporter activity).

Why this is instructive: Unlike the cds1 case (opposite reaction) or Epe1 (lost activity), arnF retains transport function — the IBA isn't wrong in kind, just wrong in specificity. The SMR superfamily is a case where sequence homology correctly identifies the structural fold (small multidrug resistance-like) but the functional annotation doesn't track the divergence from drug efflux to lipid flipping. This is a moderate severity issue because the parent term "transmembrane transporter" is not false, but it's misleading about mechanism.

EcoCyc vs. IBA comparison: EcoCyc contributed 5 annotations for arnF, including two "response to iron(III) ion" annotations (IGI from PMID:12139617, IEP from PMID:15322361) that were marked as over-annotated — they conflate transcriptional regulation of the arn operon by BasS-BasR with direct gene function. The IMP annotations from EcoCyc (carbohydrate derivative transport/transporter activity, plasma membrane IDA) are well-supported. The IBA annotations from GO_Central are reasonable at the superfamily level but miss the flippase specialization.

Root Cause Analysis:
| Feature | ArnE/ArnF | EmrE | MdtI/J | Gdx |
|---------|-----------|------|--------|-----|
| Function | Lipid flippase | Drug efflux | Spermidine export | Guanidinium export |
| Substrate | Lipid-linked sugar | Lipophilic cations | Spermidine | Guanidinium |
| Mechanism | Intramembrane flip | Transmembrane export | Transmembrane export | Transmembrane export |
| PANTHER SF | SF9 | (family-level) | SF6 | (family-level) |

Cds1 (PTHR10314) - Neo-Functionalization (Opposite Reaction)

Species: MYCTU (Mycobacterium tuberculosis), VIBCH (Vibrio cholerae)
Status: COMPLETE
Family: PTHR10314 (Cysteine Synthase/Cystathionine Beta-Synthase)

IBA Annotations Flagged:
| Term | Issue | Action |
|------|-------|--------|
| GO:0019344 cysteine biosynthetic process | Opposite function - enzyme is catabolic | REMOVE |

Lesson: This is the most severe type of IBA error - the annotation isn't merely imprecise, it's directionally wrong. Cds1 (subfamily SF135) catalyzes cysteine degradation (EC 4.4.1.1), producing H2S + pyruvate from L-cysteine. The IBA annotation GO:0019344 says it participates in cysteine biosynthesis - the exact opposite reaction!

Root Cause Analysis:
1. GO:0019344 is annotated at the PANTHER family root node (PTN000034104)
2. Root annotation propagates to ALL descendants, including subfamily SF135 (Cds1)
3. SF135 underwent neo-functionalization: same fold, opposite reaction
4. Evidence of neo-functionalization:
- Longest branch length (0.528) from root = most divergent
- Different EC class: 4.4.1.1 (lyase/catabolism) vs 2.5.1.47 (transferase/biosynthesis)
- Only 24% sequence identity with synthases (synthases share 40-43% with each other)
- Distinct active site motif: ASSGST (Cds1) vs PTSGNTG (synthases)

Experimental Evidence:
- PMID:34439535 (M. tuberculosis): Cds1 produces H2S, pyruvate from cysteine; KM=11.26 mM, kcat=78.71/s
- PMID:34283874 (V. cholerae): VC1061/cds1 is principal enzyme for cysteine-derived H2S production

Recommendation for PANTHER Curators:
- Remove GO:0019344 propagation to SF135
- Add SF135-specific annotations: GO:0019450 (L-cysteine catabolic process to pyruvate), GO:0080146 (L-cysteine desulfhydrase activity)

See detailed family analysis: interpro/panther/PTHR10314/PTHR10314-notes.md

CIRBP (PTHR48034) - Splicing Terms Over-Propagated to a Cold-Shock mRNA-Stability Subfamily

Species: human (Q14011); applies equally to the RBM3/CIRBP branch (mouse Cirbp P60824, RBM3, Xenopus cirbp-a)
Status: COMPLETE
Family: PTHR48034 (RNA-binding motif / RBM; InterPro IPR050441)

IBA Annotations Flagged:
| Term | Issue | Action |
|------|-------|--------|
| GO:0000398 mRNA splicing, via spliceosome | No splicing evidence for CIRBP; seeded by transformer-2/RBMX splicing factors | MARK_AS_OVER_ANNOTATED |
| GO:0005681 spliceosomal complex | CIRBP is not a spliceosome component; co-purification only | MARK_AS_OVER_ANNOTATED |

Lesson: PTHR48034 is a heterogeneous RRM family that lumps two functionally divergent groups sharing only the N-terminal RRM: (i) transformer-2/RBMX/SR-type splicing regulators (RS-domain C-terminus) and (ii) cold-inducible mRNA-stability/translation proteins CIRBP and RBM3 (glycine-rich/RGG C-terminus). The splicing terms are real for group (i) but were propagated across the divergence boundary into group (ii). This is the RNA-biology analogue of the arnF case — homology correctly identifies the fold, but the annotation does not track the functional split (pre-mRNA splicing → mRNA stabilization/translational control). CIRBP's verified activities are 3'-UTR binding (IDA: RPA2, TXN), mRNA stabilization, and translational control, with stress-granule recruitment — none of which is splicing.

Root Cause Analysis (PANTHER PAINT, interpro/panther/PTHR48034/PTHR48034-paint.tsv):
1. Splicing IBD is anchored at internal node PTN000391532, seeded only by splicing factors: GO:0000398 from TRA2A (Q13595), TRA2B (P62995), Drosophila tra2 (FBgn0003742), rat Tra2 (RGD:1306751, RGD:1565256); GO:0005681 from TRA2B (P62995), RBMX (P38159), rat Tra2 (RGD:1306751).
2. The node's annotations descend as IBA to all descendants, including the cold-shock branch.
3. CIRBP's own subfamily node PTN008729690 carries only GO:0003729 mRNA binding (and lists CIRBP, Q14011, as a seed) — the correct, generic call.
4. CIRBP's GOA WITH/FROM names the source node explicitly: PANTHER:PTN000391532|...|UniProtKB:P62995|UniProtKB:Q13595 — a textbook case of mis-grouping revealed by the WITH/FROM column (pattern 12).
5. PAINT already prunes other branches of this family (node PTN001924395 carries IRD/NOT records blocking GO:0003729, GO:0000381, GO:0016607), so the gap is the absence of an equivalent pruning on the CIRBP/RBM3 branch.

Seed (mod) genes verified as bona fide splicing factors (no change needed):
- TRA2B (P62995): UniProt "participates in the control of pre-mRNA splicing"; direct IDA GO:0000398 (PMID:9546399); controls SMN2 exon 7 / MAPT exon 10. Falcon: "Belongs to the splicing factor SR family."
- TRA2A (Q13595), RBMX (P38159), Drosophila tra2 (P19018), rat Tra2b (P62997): all UniProt-confirmed pre-mRNA / alternative splicing regulators.

Recommendation for PANTHER Curators:
- Add an IRD/NOT (or restrictive re-annotation) for GO:0000398 and GO:0005681 on the CIRBP/RBM3 subfamily branch (node PTN008729690 and descendants), so the splicing terms stop descending to the cold-inducible mRNA-stability members.

See detailed family analysis: interpro/panther/PTHR48034/PTHR48034-review.md

DICDI cAMP / STAT developmental families — Stage-Specific Paralog & Lineage Over-Propagation

Dictyostelium discoideum development is driven by several paralog families whose
members do the same molecular job at different developmental stages (the cAMP
receptors cAR1–4, the adenylate cyclases ACA/ACG/ACR, the Ras GTPases RasC/RasG,
the STATs Dd-STATa/c). Reviewing one representative per family alongside its
sisters exposed IBA transferring a family-node consensus onto the wrong member,
stage, compartment, or lineage — the same failure classes catalogued above, now
in a social amoeba:

• Stage/paralog leakage (PROPAGATION_BAD · WRONG_ORTHOLOG_OR_PARALOG).
  The aggregation-stage "adenylate cyclase-activating cAMP receptor signaling"
  role (GO:0007189) is propagated by the cAR family node onto the later,
  lower-affinity receptor cAR4/carD, whose characterised output is the
  PTP/GSK3 axis, not adenylate-cyclase activation. Likewise rasC carries
  GO:0000281 mitotic cytokinesis, but rasC-null cells divide normally — cytokinesis
  is RasG's job in this family.
• Lineage-inappropriate process transfer (LINEAGE_OR_TAXON_MISMATCH). The STAT
  family node PTN000927860 transfers metazoan STAT roles — GO:0006952 defense
  response and GO:0042127 regulation of cell population proliferation — onto both
  Dd-STATa and Dd-STATc from all-metazoan seeds (human/mouse/rat/fly/worm
  STAT1/2/5…). This mirrors the existing worm sta-2/fshr-1 cross-kingdom row.
• Term-scoping across a lineage gap (TERM_SCOPING_PROBLEM). Both Dictyostelium
  STATs are annotated GO:0007259 "signaling via JAK-STAT", yet Dictyostelium
  has no JAK (they are activated by the TKL kinases Pyk2/Pyk3); the term is
  rescoped to GO:0097696 STAT signaling.
• Functional divergence with fold retained (FUNCTIONAL_DIVERGENCE). regA
  inherits GO:0047555 cGMP-phosphodiesterase activity from the cyclic-nucleotide
  PDE family node, but RegA is cAMP-specific (>200-fold selectivity). yakA
  (a dual-specificity DYRK) carries the family's generic GO:0004713 protein
  tyrosine kinase activity.
• Compartment mismatch (COMPARTMENT_OR_COMPLEX_MISMATCH). pten inherits
  GO:0005634 nucleus (a mammalian-PTEN behaviour) although all Dictyostelium
  evidence places it at the membrane/cortex/cytosol; spiA (a demonstrated
  spore-coat protein) inherits the canonical SCAMP trans-Golgi/recycling endosome localisations.

Each of these rows carries a structured review.propagation_review
(root_cause + failure_modes + the real GOA WITH/FROM PANTHER PTN… node
and a representative seed) in the corresponding
genes/DICDI/<gene>/<gene>-ai-review.yaml. Full module and paralog context:
Dictyostelium Development Project.

Connection to pathway satisfiability. The JAK-STAT case is the clearest
instance of a broader, automatable rule. GO:0007259 "signaling via
JAK-STAT" names an obligate component — a Janus kinase — that the
Dictyostelium genome does not encode (the Dd-STATs are activated by the TKL
kinases Pyk2/Pyk3). Read as a boolean formula over required components under a
genome-content oracle, the annotation is unsatisfiable and the IBA
transfer is unsupportable — so it is rescoped to the JAK-independent parent
GO:0097696 STAT signaling. This is the signaling-domain analogue of the
genome-content check in the
Pathway satisfiability project: a process/pathway
term whose definition entails a component absent from the target's genome
is a candidate LINEAGE_OR_TAXON_MISMATCH over-propagation. A runnable
prototype of exactly this check —
taxon-absent-component detector
— confirms JAK is genome-absent in Dictyostelium (GO:0007259 unsatisfiable
→ GO:0097696). It uses two oracles: an InterPro domain signature and,
primarily, PANTHER family (PTHR…) membership — the divergence-robust one,
since IBA propagates along the PANTHER tree. That distinction matters: an
InterPro-only screen falsely calls STAT and P2X absent in Dictyostelium
(both diverged past their metazoan domain signature), whereas PANTHER correctly
recovers the 4 Dd-STATs and the ~5 divergent P2X receptors — so the organism
does have (ionotropic) P2X, and only the metabotropic P2Y (GPCR)
component that GO:0035589 specifically requires is genuinely absent. Even at
HIGH confidence an ABSENT verdict is a strong lead for review, not an
automatic REMOVE.

Genes with IBA Issues

IBA Incompleteness: core function that IBA fails to propagate

All the patterns above concern IBA being wrong (over-annotation). The opposite
failure mode is just as real: IBA is frequently incomplete — it under-calls
well-established biology. Phylogenetic propagation is conservative by
construction (it only transfers what a curated ancestor already carries, at the
granularity the ancestor was annotated), so a great deal of experimentally
defined molecular function never reaches the leaf.

We quantified this with a generic evidence-subtraction tool
(ai-gene-review subtraction-report; see
docs). Running it in
"keep only IBA" mode over the 1015 reviewed human genes — i.e. asking if IBA
were the sole evidence, what curated biology would we lose? — and applying
ontology closure so that an IBA call to a more general parent still counts as
covering its ancestors:

• 62% of annotation-grounded core_functions terms (4516 / 7278) would be
  lost if IBA were the only evidence.
• Restricting to molecular function and excluding low-information binding
  terms (GO:0005488, incl. protein binding): 511 curated core molecular
  functions across 423 genes have no IBA support at all, 401 of them
  grounded by experimental/traceable evidence (IDA/IMP/IPI/EXP/TAS).

Because a term is only counted when it sits in a gene's core_functions — the
curator's distilled, highest-confidence judgement of what the protein does —
these are not annotation noise; they are the central activities a leaf-level
review would lose by trusting IBA alone. Two mechanisms recur:

A. Activity absent from IBA entirely

The experimentally characterised activity is simply not propagated to the leaf —
no IBA annotation touches that branch — even though it is the protein's defining
biochemistry. These are clean, single-line losses (each verified: strong
non-IBA evidence, zero IBA at the term or any descendant):

B. IBA stops at a general parent (true "too conservative")

Here IBA does annotate the gene with a broad term, but the experimentally
established specific activity — the exact substrate, regioselectivity, or
sub-activity — is never propagated. Closure confirms the IBA term is a strict
ancestor of the missed term, so this is genuine loss of resolution, not absence:

Caveat — IBA is not always the laggard. Many of these genes are exceptionally
well studied; IBA legitimately covers their canonical function and only misses
secondary or recently characterised activities. PTEN is the clearest example:
its textbook PIP3 3-phosphatase activity (GO:0016314) is carried by IBA; what
IBA misses is the secondary protein-serine/threonine phosphatase (GO:0004722,
IDA PMID:9256433) and PI(3,4)P2 3-phosphatase (GO:0051800) activities. So
"incompleteness" should be read as resolution and coverage gaps, not as IBA
being useless — the same tool's forward direction shows IBA is the sole
support for 66 human core molecular functions (e.g. AKIRIN2 transcription
coregulator, GET1 protein-membrane adaptor, ATG14 PI3K regulator), so the two
analyses bound IBA's value from both sides.

Reproduce: just subtraction-report-iba-conservative-core-mf (writes
reports/iba-too-conservative-core-mf.md, the full ranked, evidence-enriched
table for all 423 genes); the raw keep-only TSVs come from
just subtraction-report-iba-only-tsv.

Lesson for curators: a leaf with only IBA annotations is very likely
under-annotated, not fully annotated. When IBA supplies only a broad
molecular-function term, treat it as a prompt to look for the specific
experimentally defined activity rather than as a finished call.

Recommendations for IBA Curation

1. Flag pseudo-enzyme candidates: Proteins with enzyme domains but missing catalytic residues
2. Flag neo-functionalized subfamilies: Long branch lengths and different EC numbers indicate potential opposite function
3. Validate substrate specificity: Don't assume identical specificity across orthologs
4. Distinguish core vs. secondary: Mark promiscuous activities as non-core
5. Prefer specific terms: Replace generic IBA with specific terms when evidence exists
6. Check for functional divergence: Especially in rapidly evolving families
7. Consider organism-specific biases: Source annotations may reflect experimental systems (e.g., NMJ in flies) that don't apply to target species
8. Validate annotations at family root: Root-level annotations propagate everywhere - ensure they're truly universal to ALL subfamilies
9. Synthesize multiple lines of evidence before flagging — never a single keyword: a UniProt keyword (especially "By similarity"), a PANTHER node label, and a review assertion are each individually weak. Cross-check the term definition, direct experimental papers in the target species, the IBA WITH/FROM provenance, the MSA/active-site residues, and phylogenetic placement, and reason over the whole picture. The strongest REMOVE cases pair an explicit UniProt CAUTION/NOT with direct enzymology (DPYSL2, AGO4, CRYAA, AGK)
10. Watch for opposite-sign family members: When a family contains both activators and inhibitors (e.g., BCL2 family), a family-node IBA can transfer the wrong regulatory sign — inspect the WITH/FROM list for mixed members (but check whether independent non-IBA evidence also supports the term before calling it flatly wrong)
11. Distinguish sub-activities: capping vs severing, holdase vs foldase, slicing vs non-slicing — family-level IBA flattens these distinctions
12. Don't inherit a paralog's compartment/complex: family members share folds but not localization or complex membership — verify the protein actually occupies the annotated complex/compartment (EIF4E2, ALDH1L1, PEX2)
13. Check the GO term's definition, not just its label, before calling an IBA directionally wrong: e.g. "copper ion import" (GO:0015677) covers movement into a cell or organelle, so a Golgi-loading copper exporter can still satisfy it — a label that looks opposite may not be
14. Read the WITH/FROM column first: it names the exact source proteins. If they are the wrong family (NTN1←POU TFs; NOTCH1←SLITs) or a single wrong paralog (ABRAXAS1←ABRAXAS2; HINT2←HINT1), that is near-mechanical evidence of error. If they are a broad, coherent set of true orthologs, the transfer is probably defensible — slow down before flagging

Quality Indicators

Signs of problematic IBA:
- Enzymatic activity for proteins with degenerate active sites
- Long branch lengths from family root (indicates divergence/neo-functionalization)
- Different EC numbers between subfamilies (may catalyze opposite reactions!)
- Generic terms when specific function is known
- Process annotations that don't match organism biology
- Multiple conflicting IBA annotations
- Superfamily contains members with different transport mechanisms (e.g., solute export vs lipid flipping in SMR family)
- Enzymatic terms on proteins with a UniProt-documented degenerate/absent active site (the strongest signal; e.g. DPYSL2, AGO4)
- Family unites opposite-sign regulators (activators + inhibitors of the same process; check WITH/FROM for mixed members — but confirm against non-IBA evidence)
- Complex-membership or compartment terms on a protein whose paralog/relative occupies it instead (cytosolic vs mitochondrial; eIF4F vs 4EHP repressor)

NOT reliable grounds for flagging — verify with reasoning, not a single keyword:
- A label that merely looks opposite — check the term definition. GO:0015677 "copper ion import" covers movement into a cell or organelle, so a Golgi-loading copper exporter (ATP7B) still satisfies it. (This is why ATP7B was not flagged.)
- A shared reaction that is classified under a pathway does not prove pathway membership in vivo — but it does mean the activity is real, so prefer "over-annotation/non-core" over "absent" (HMGCS2: the HMG-CoA-synthase step is genuine; only the FPP-pathway flux belongs to the other paralog).
- Neither a UniProt keyword nor a review assertion is sufficient on its own. Weigh all lines: a UniProt "By similarity" tag is weak and can be overturned by direct experimental papers (AGK: "ceramide By similarity" is refuted by two papers reporting no ceramide/sphingosine phosphorylation — so AGK was flagged); an explicit UniProt CAUTION or a curated NOT annotation is strong (DPYSL2, CRYAA).
- A broad subsuming compartment is not wrong just because a more specific location is known. GO:0005737 cytoplasm includes mitochondrion, ER, Golgi, and lysosome (part_of cytoplasm), so "cytoplasm" is defensible for an organellar protein; only nucleus, plasma membrane, the extracellular space, or a different organelle are mutually exclusive enough to justify a localization REMOVE (see Pattern 13).

Signs of reliable IBA:
- Core metabolic enzymes with conserved mechanism
- Well-defined pathways with conserved components
- Structural proteins with conserved function
- Process/location terms that are organism-agnostic
- Subfamilies with high sequence identity and same EC number

Writing the reason field — things nothing enforces:
- A quotation inside reason is not validated — nor inside comment or summary. The
substring check runs over supporting_text only (linkml_reference_validator's
supporting-text validator; grep -rn reason src/ai_gene_review/validation/*.py returns
one unrelated docstring). So a quote in any free-prose field — a review.reason, a
propagation_review.source_entities[].comment, a summary — can be paraphrased,
mis-attributed or invented and still pass. comment is the easiest to forget, because
it sits inside a structured block that looks validated and is not. Sweep all three
together, and take one of three paths for each span, saying which: mirror it into
supported_by, where the substring validator checks it; name its source with a locator in
the prose itself (projects/FOO.md:14, Ccne1-uniprot.txt:96), which is checkable by a
reader even though no tool enforces it; or state in the text that it is unchecked. The
middle path is the right one for a term label or a file this review does not cite as a
reference, and it is not the same as saying nothing - a source named without a locator is
the third path wearing the second one's clothes.
- Never state what an abstract-only paper "records". Check
full_text_available: first. A reason on this project asserted that PMID:12492473
"records Casp3 proteolytically cleaving iPLA2"; that cache is abstract-only, contains no
cleavage assay (grep -ic cleav -> 0), and names the enzyme a plasmalogen-selective
phospholipase A2, never iPLA2 — both the mechanism and the enzyme identity were inferred
and shipped. The conclusion survived on grounds the abstract does support (inhibitor
epistasis placing the protein upstream, typed as an enables MF), which is the form to
reach for. A sentence that claims what a paper contains and claims to need no full
text is self-refuting; one half has to go.
- Normalize before you decide a quote is fabricated. Every layout assumption
listed below - whitespace and markup alike - has produced a confident false negative on
this project, and every one fails toward "this quote is invented" -- the most expensive wrong answer available here.
A single-line grep misses any quote that wraps. A flat string match misses text inside a
folded YAML scalar. tr '\n' ' ' misses text whose source lines carry trailing spaces
(squeeze with tr -s instead). And a match against a flat-file source misses text carrying
structural line prefixes -- UniProt CC/FT, a # comment. Join and squeeze
(re.sub(r'\s+', ' ', ...)) AND strip the source's line prefixes before concluding
anything is absent. Better still, put the quote in supported_by and let the substring
validator answer.
Markdown emphasis is the widest of them, because every
file:*-deep-research-*.md citation in this corpus is markdown: Predominantly **nuclear** transcription factor will not match the quote predominantly nuclear
under any prefix-stripping, because the markup sits INSIDE the phrase rather than at
the line start. Strip [*_]too, or match permissively (predominantly[^A-Za-z]{0,40}nuclear`) before concluding anything.

• Quoted spans in prose are content-verbatim; strictness lives in supporting_text.
  Terminal sentence punctuation may sit inside the closing quote (American convention),
  and an elision may be marked with ... - both are ordinary typography, not claims about
  the source, and the corpus uses them: [Dnaja3](../../genes/mouse/Dnaja3/Dnaja3-ai-review.html) closes a quote on a period where its
  source writes a comma, [Sox2](../../genes/mouse/Sox2/Sox2-ai-review.html) closes one where the source sentence runs on past a
  semicolon, [Fbxo2](../../genes/mouse/Fbxo2/Fbxo2-ai-review.html) elides mid-quote. What is not allowed is dropping
  source words silently: the same [Fbxo2](../../genes/mouse/Fbxo2/Fbxo2-ai-review.html) row quoted GO:1990756's definition with the
  parenthetical "(including ubiquitin ligase and UFM1 ligase)" removed and no ellipsis,
  which reads as the whole definition and is not - that one is a defect and was fixed. So
  a sweep must strip a trailing [.,] and split on ... before calling a span unmatched,
  or it will report the convention as fabrication; and any span that must be
  machine-checked belongs in supporting_text, where the substring validator applies the
  strict reading. This holds for reason, summary and comment alike — the surrounding
  bullets are phrased around reason because that is where the first instances were found,
  but the exemplars here are spread across all three, and the ISO backlog will write far
  more comment and summary than reason. One span type has no local source at all: a
  GO term definition. The only tracked .obo here is interpro/panther/panther.obo,
  cache/go/terms.csv and cache/ontologies/go.tsv carry labels but no definitions, and
  cache/ontologies/*.obo is gitignored - so a definition verified against a working-tree
  obo is not reproducible from a fresh checkout. Quote it if it earns its place, but take
  the third path explicitly and say the check needs a GO lookup.

• "Nothing local says this" is a claim about where you looked. Before writing that an
  assertion cannot be checked from the repository, enumerate the places the repository
  keeps that kind of fact. A reason on this project stated that GO:0051082's obsoletion
  was "not checkable from this repository" on the strength of two ontology caches —
  while projects/UNFOLDED_PROTEIN_BINDING.md records it explicitly, verified live
  against QuickGO and OLS, and the same review file asserted it plainly twelve lines
  further down. projects/ is where this repo records ontology decisions postdating a
  cache snapshot, so a cache-only search will systematically miss them; cache/ontologies/README.md
  says the caches are sparse by design, holding only terms used in annotations. Absence
  from a cache means "not cached", never "not real" and never "not recorded here". Search
  the gene directory, cited publications, the caches, interpro/, projects/, and the
  file's own origin/main text before concluding anything is unprovable.

• A quoted-span sweep is only as good as its corpus — and a bad corpus fails toward
  "fabricated". Three rules, each learned by getting it wrong on this project's mouse
  pass. (i) Exclude the file under test. A corpus that walks the gene directory picks up
  GENE-ai-review.yaml itself, so every quote proves itself and the sweep reports a clean
  pass it did not earn; dropping that one file turned a zero-miss run into a list of misses.
  (ii) Every miss that survived was a corpus gap, not a fabrication — the other gene
  directory named in the comment (a DR MGI; ... line quoted from Ccne1-uniprot.txt
  inside a [Ccnb1](../../genes/mouse/Ccnb1/Ccnb1-ai-review.html) review), the per-family interpro/panther/PTHR*/*-entries.csv rather
  than just panther.obo, and a PMID cited only in supported_by rather than in the
  top-level references list the extractor read. Widen the corpus before you widen the
  accusation. (iii) Strip prefixes on both sides, or neither. A quote that itself
  carries DR will not match a source you have stripped DR out of; test the raw
  and stripped forms of each.
• origin/main is not a source. Listing "present in this file's own origin/main
  text" among the places a span may be proven against silently exempts every pre-existing
  quote, because that is where pre-existing quotes live. It answers was this introduced by
  this PR and gets read as is this verified - two different questions, and the corpus
  sweep on this branch reported one unmatched span with the clause and ten without it. Keep
  the scope question if you need it, but keep it in a separate column. The exception is a
  span the prose explicitly presents as the file's own superseded text ("the previous shared
  reason offered a disjunction, ..."): there origin/main genuinely is the source.
• A minimum length on the quote regex mis-pairs quotation marks. "([^"]{12,400})"
  skips a short quoted phrase and then matches from its closing mark to the next opening
  one, so the span reported is the prose between two quotations rather than either of them.
  A sentence naming both a "Par complex" and a "Par3" produced exactly that. Match every
  quoted span, then filter by length.
• Sweep on the property, not on the phrasings you have seen. A tell list is built from
  the instances already found, so it cannot find the ones worded differently — and it reports
  a clean pass when it runs out of tells, not when it runs out of instances. This project has
  now shipped that mistake in six places (the abstract-only pass keyed on "The paper concerns"
  and missed "This physiology belongs to"; a MOD-prefix list; the Araport/araport11
  spelling; the PAINT namespace enumeration; a -B30 extraction; and a description sweep
  keyed on Falcon / best curated with / retained as non-core, which missed
  over-extensions, GOA rows, not treated here and conflate — two of them in
  descriptions the same commit had just rewritten). For description the property is one
  sentence: it must be readable by someone who does not know this repository exists, so
  any sentence whose grammatical subject is an annotation rather than the gene belongs in
  review.reason, core_functions or the notes file. Operationalize that grammatically, not
  lexically — split the field into sentences and judge each one's subject, object and
  predicate, flagging any where an annotation, an evidence record or a curation artifact
  appears as subject or object, or where the predicate is a curation act. Read each hit
  and judge; the result is a set you can defend, because what it missed is a judgement you made
  rather than a word you had not thought of.

Do not read any example below as the set to match against — that is this bullet's whole
subject, and the examples fail the test themselves. Every curation-act instance in this
corpus is plural (are treated as, are retained as, are curated as), so the singular
forms that come to mind first — is treated as, is kept as — match zero of the ten
occurrences, spread over eight files ([Calm3](../../genes/mouse/Calm3/Calm3-ai-review.html) alone carries three). One of
the ten is are **best** treated as, which no fixed phrase reaches at all. Say which unit a
count is in: at file scope those same singular forms return over three hundred hits, all of
them outside any description, so the figure inverts if the scope is left implicit. Object
position is the same trap one step further out: [Cdk5r1](../../genes/mouse/Cdk5r1/Cdk5r1-ai-review.html) says the p35/p25 distribution explains
"the mixed cellular localization annotations", where subject and predicate are both clean and
the annotation sits in a trailing participial clause. An outside reader still asks what
annotations?, so it is an instance — which is why the test is a judgement about roles, not a
list of strings.

Expect the screen to produce hits that clear — that is the read each hit and judge step
working, not a sign it is too wide. The Dnajb11 sentence written for this very bullet ("the
observation used an N-terminally tagged construct…") is itself a hit, with an evidence
record as its subject, and it clears: a biologist who has never seen this repository reads it
as ordinary writing about a published experiment. Counting hits as defects is what pulls a
screen back toward being a tell list.

This bullet got that wrong on its first writing, which is the sharpest case in the file: it
diagnosed tell-list sweeping and then specified the remedy as a tell list (annotat,
GOA, GO_REF, an evidence code, curat, term, non-core, over-, treated here, a
provider name). Run that vocabulary over [Dnajb11](../../genes/mouse/Dnajb11/Dnajb11-ai-review.html)'s "The reported APOBEC1/apoB
mRNA-editing interaction is treated as unsupported" and it returns zero hits —
treated as is not treated here, and none of the other nine appears. [Tert](../../genes/mouse/Tert/Tert-ai-review.html) ("are
treated as context-dependent non-canonical activities") and [Syk](../../genes/mouse/Syk/Syk-ai-review.html) ("rather than the
core molecular function") escape identically, on one word each. A vocabulary is always
built from the instances already found, so specifying one as the cure for that very failure
reproduces it at one remove. The grammar is what generalizes: a curation act has a doer, and
in a description the doer is always this review.
- A scare quote is not a quotation. "FB:FBgn0001091 is Gapdh1" in a Gapdh comment is
a proposition the sentence goes on to call "an inference from organism and gene name
rather than a lookup", not a span lifted from a source. Read the surrounding prose
before treating an unmatched span as a citation defect.
- If a block's prose argues its own action may be under-strength, say so in the prose.
root_cause and failure_modes are a fixed vocabulary and cannot carry "the objection
reaches further than the action I am leaving in place". A consumer reading the structured
fields alone gets only the coded story, so the tension belongs in the comment, with the
reason it is being left (e.g. re-typing an inherited action is a separate judgement).

Name the fault with its own action's vocabulary. Calling a defect "the over-annotation"
in a row whose action is REMOVE — beside a sibling row carrying
MARK_AS_OVER_ANNOTATED — reads as an argument for the other action. The disjunctive
reason class this project dismantled (Casp3, Ghr) failed the same way: the limb that made
the text defensible was the limb arguing against the action it sat on.

Project history & methodology

This page records the synthesized findings. The dated project log, the per-pass
verification narrative (what was added, what was retracted and why), and the
lessons learned are kept separately in IBA_REVIEW/HISTORY.md.

A reproducible multiple-sequence-alignment check of the two central pseudo-enzyme
claims (Argonaute catalytic tetrad; CRMP/DPYSL metal-coordinating residues) lives in
IBA_REVIEW/msa/ — uv run python catalytic_residue_msa.py.