Re-examining the RAR misannotation study (Wei et al. 2020)

IN_PROGRESS EVALUATION

Collection: Propagation by homology

Species: ARATH, worm, DROME, mouse, human, CHICK

Genes: At3g08840 KDTA cec-10 npp-20 CG3226 Gprc5c BTBD3 LPXA LPXB LPXK LPXD1 LPXD2 LPXC1 LPXC2 LPXC3 LPXC4 LPXC5

Warnings (2)

Re-examining the RAR misannotation study (Wei et al. 2020)

Bottom line: Wei et al. (2020) flagged 190 GO annotations in UniProt-GOA
2019-06-03 and 250 in 2018-11-06 as probable misannotations, using taxon-specific
annotation-rate ratios (RAR) followed by manual review
[PMID:32470107 "Detecting Gene Ontology misannotations using taxon-specific rate ratio comparisons."].
We rebuilt their complete case list from the supplement and checked every
protein–term pair against current GOA. Of the 290 distinct pairs, 216 are gone
from GOA, 36 more vanished with their UniProt entry, and 38 persist. Most of
the paper's calls therefore hold up and have been acted on. The survivors contain
one clear error by the paper: plant lipid A biosynthesis (25 of the 38 pairs),
which experimental work shows Arabidopsis really has
[PMID:21709257 "our work demonstrates that plants synthesize lipid A precursors using the same enzymatic pathway present in E. coli"].
They also contain one moot case (a GO term since renamed to a general meaning)
and a handful of animal annotations that still look wrong. The paper's two
best-known IBA case studies were right on the facts, but its reasoning for one of
them (plant DDL) was wrong.

This project is a companion to IBA_REVIEW: the paper's headline
claim is that IBA had become the main source of non-electronic GO errors, and
its main mechanism (terms placed at a PAINT node deeper than the evidence
supports) is the same as the LINEAGE_OR_TAXON_MISMATCH pattern catalogued
there and the pathway-level problem in TAXON_PATHWAY_VARIANCE.

The paper's method and claims

For a GO term q, the annotation rate in taxon t is the fraction of that
taxon's annotated proteins that carry q; RAR is the minimum rate over taxa
divided by the maximum. The authors computed RAR for five kingdoms (animals,
bacteria, archaea, fungi, plants; 20 reference proteomes) and for vertebrates
versus invertebrates, then manually inspected every term with RAR < 0.1.

# Claim Our assessment
C1 190 (2019-06-03) and 250 (2018-11-06) manually confirmed misannotations, over 43 and 50 terms Reproduced. Our reconstruction from the supplement matches Table 1 on every line (terms, proteins, annotations).
C2 Most flagged annotations are wrong Largely holds. 216 of 254 pairs on still-active entries have since been removed from GOA. One family of flags (plant lipid A, 25 pairs) is refuted by experiment.
C3 56% of flagged errors are IBA, overtaking ISS as the main non-IEA error source Plausible for 2018–19. The supplement only records each term's most frequent evidence code, so this cannot be re-counted per pair.
C4 Only 25% were fixed six months later Superseded. About 85% of pairs on still-active entries are now gone (see Results).
C5 B. subtilis Udg (P39615) wrongly gets IBA nucleus and mitochondrion from a root-placed PAINT node (PTN000137400) Holds; fixed. These rows are gone; P39615, P56397 and P12295 keep only the uracil DNA N-glycosylase activity and base-excision repair, AP site formation via deaminated base removal IBAs. QuickGO now returns 0 bacterial annotations to GO:0005634 or GO:0005739.
C6 Arabidopsis DDL (A0A1I9LPE3) wrongly gets IBA peptidoglycan biosynthetic process, because plant walls are cellulose Conclusion holds; reasoning wrong. See DDL.
C7 Human JUP wrongly gets IC structural constituent of cell wall Holds; fixed. GOA keeps only NAS GO:0005198 structural molecule activity.
C8 A low RAR can be biologically correct (Salmonella SspH2 E3 ligase) Agreed. This is why the screen needs manual review: only ~9% of RAR < 0.1 terms were confirmed (Supplementary Table S2).
C9 GOA error rate ≈ 7%, IBA ≈ 31%, estimated from later NOT annotations vs later confirmations Biased upward for IBA. NOT annotations are made to counter propagated annotations, so IBA scores high by construction.
C10 12 of 43 flagged terms violated existing GO taxon constraints Not re-checked; GO's constraints have changed since.
C11 neXtProt would downgrade IBA from "Gold" to "Silver" Unverified. See neXtProt.

Results

Scripts and data are in RAR_MISANNOTATION/.
The current-status check was run on 2026-10-18 against QuickGO and UniProt.

Release flagged Pairs Removed from GOA Entry deleted/merged Still annotated
2019-06-03 190 144 33 13
2018-11-06 250 191 26 33
Distinct pairs (both releases) 290 216 36 38

Removal is not proof that a pair was wrong: PAINT re-runs, InterPro2GO changes
and accession changes all remove rows. But it agrees with the paper's call in
almost every case we looked at, and none of the removed terms needs a second
look (bacterial nucleus/mitochondrion/nucleolus, animal fungal-type vacuole membrane and cell wall mannoprotein biosynthetic process, fly
inflammatory response, and so on).

The 38 surviving pairs

Term Proteins (taxon, evidence) Verdict on the paper's flag
GO:0009245 lipid A biosynthetic process 25 Arabidopsis and Selaginella proteins (IBA from E. coli LpxA/LpxB/LpxD nodes; InterPro2GO IEA) Paper wrong. Arabidopsis has LpxA, LpxC, LpxD, LpxB, LpxK and KdtA orthologs. Knocking them out shifts the expected lipid A precursors, and the enzymes are targeted to mitochondria [PMID:21709257 "the Lpx and KdtA proteins are expressed and targeted to mitochondria"]. AtLpxA complements an E. coli lpxA mutant [PMID:22545860 "A. thaliana lpxA complements an E. coli mutant lacking the chromosomal lpxA"]. A low RAR reflected curation density, not absence.
GO:0001216 5 (human ZNF341 IDA, mouse Znf341 IEA, 3 Arabidopsis IDA/IMP) Moot. In 2019 the term was "bacterial-type RNA polymerase transcriptional activator activity, sequence-specific DNA binding"; it is now the general "DNA-binding transcription activator activity". The flag was right against the old label; the current rows are fine.
GO:0021987 cerebral cortex development 4 chicken (BTBD3 IBA, NTRK2 ISS, two Ensembl IEA) Open. Whether "cerebral cortex" ("the outer layered region of the telencephalon") applies to the avian pallium is a homology question, not a clear error.
GO:0030907 MBF transcription complex worm cec-10 (Q9N4L7), IBA from PTN000917496 seeded by fission yeast Res2 Flag holds (provisionally). MBF is the fungal G1/S complex; cec-10 is a chromodomain protein. Needs a family-level check of the node.
GO:0010973 positive regulation of division septum assembly worm npp-20/SEC13 (Q9N4A7), IMP, PMID:12937276 Flag likely holds. The cited paper is an RNAi survey of nucleoporins; division septum assembly is a fungal/bacterial process. This is an experimental row, so per project rules it needs the full text, not a REMOVE from the abstract.
GO:0044548 S100 protein binding fly CacyBP (Q9W3Y3), InterPro2GO IPR037893 Flag holds. S100 proteins are a vertebrate family; the mapping transfers a vertebrate partner interaction to flies.
GO:0005118 sevenless binding mouse Gprc5c (Q8K3J9), InterPro2GO IPR002956 Open. Vertebrates have a sevenless ortholog (ROS1), so this is a question about the InterPro2GO mapping, not the taxon.

DDL in Arabidopsis

The paper said At3g08840 cannot do peptidoglycan biosynthesis because plant cell
walls are cellulose. That misses the point: the question is the plastid, not the
cell wall. Moss chloroplasts do have a peptidoglycan wall, and moss ddl
knockouts give giant chloroplasts that D-Ala-D-Ala rescues
[PMID:27325639 "moss chloroplasts are completely surrounded by peptidoglycan"].
So the PAINT node uniting bacterial and plant Ddl is defensible for land plants
as a group. What makes the Arabidopsis annotation doubtful is lineage-specific
loss: Arabidopsis has no Pbp gene [PMID:16618924 "no Pbp gene was found in A. thaliana"],
and its ddl mutants have no plastid phenotype
[PMID:27325639 "no plastid phenotypes were observed in the T-DNA tagged ddl mutant lines of Arabidopsis thaliana"].
Current GOA has dropped the BP IBA but kept the MF IBA D-alanine-D-alanine ligase activity, which is untested for the Arabidopsis protein. This is the
same "inside the clade, lost the function" question IBA_REVIEW asks, and is a
candidate for a gene review.

Gene reviews of the surviving cases

One full review per surviving case (2026-10-19). Each one checks the flagged
annotation against the literature and, for IBA rows, the PAINT node.

Gene Flagged term Review action Paper's flag
ARATH/KDTA lipid A biosynthetic process (IBA) ACCEPT, core Wrong. Lipid IVA, KdtA's substrate, accumulates only in atkdta mutants; the enzyme is mitochondrial.
ARATH/At3g08840 peptidoglycan biosynthetic process (IBA, now gone); D-Ala-D-Ala ligase activity (IBA) MF ACCEPT; no BP proposed Holds for the process, for plastid reasons. Catalytic residues are conserved, matching the moss enzyme, so the MF stays (untested).
worm/cec-10 MBF transcription complex (IBA) REMOVE (with four other MBF-derived IBAs) Holds. cec-10 lacks the KilA-N DNA-binding domain; PTN000917496 is placed too deep over metazoan Swi6-family members.
worm/npp-20 positive regulation of division septum assembly (IMP) MODIFY → mitotic cytokinesis, non-core Holds for the term, not the experiment. The full text reports that RNAi "blocked cytokinesis"; the curator chose a fungal septum term.
DROME/CG3226 S100 protein binding (InterPro2GO) REMOVE Holds. No S100 proteins outside chordates; the IPR037893 mapping carries a vertebrate partner to flies.
mouse/Gprc5c sevenless binding (InterPro2GO) REMOVE (also visual perception from the same entry) Holds, for a different reason. IPR002956 is a BOSS PRINTS signature matching two transmembrane segments; no GPRC5–ROS1 link is known.
CHICK/BTBD3 cerebral cortex development (IBA) MODIFY → pallium development Holds, weakly. The node is sound; the anatomy term doesn't fit birds. About 1,760 bird annotations use it, constrained only to Bilateria.

The Arabidopsis lipid A pathway

Since the lipid A flags were the paper's one clear error, we reviewed the whole
Arabidopsis pathway rather than relying on KDTA alone (2026-10-26). All eleven
enzymes are reviewed here, with the experimental evidence from Li et al. 2011
and Joo et al. 2012 recorded on their rows.

Gene Key evidence Outcome
ARATH/LPXA Purified enzyme active in vitro; complements E. coli lpxA; mutant has ~21% of wild-type lipid X Lipid A biosynthetic process added as NEW (IDA), the only term the experiments support that GOA lacked
ARATH/LPXC1 to ARATH/LPXC5 Five tandem copies encoding identical proteins; RNAi of the cluster lowers lipid X; LpxC1.1-GFP is mitochondrial Same calls on all five; the experiments cannot tell the copies apart
ARATH/LPXD2 Lipid X undetectable in the null Main LpxD; activity and process accepted as core
ARATH/LPXD1 Lipid X not reduced in the null; R293K and missing acceptor-binding residues Activity and process left UNDECIDED; only acyltransferase activity accepted
ARATH/LPXB Two nulls accumulate lipid X (42-fold) and UDP-2,3-diacyl-GlcN Accepted
ARATH/LPXK Three nulls accumulate the disaccharide 1-phosphate Accepted
ARATH/KDTA Null accumulates lipid IVA Accepted (see above)

Across the set, the predicted chloroplast (ISM) rows were removed, because
GFP fusions in Li et al. 2011 did not colocalise with chloroplasts. Rows using
the obsolete GO:2001289 lipid X metabolic process were moved to lipid A biosynthetic process, except on LPXD1, which has no phenotype.

Taken together, the paper was wrong about one family (plant lipid A). It was right
about five, but for three of those the cause was not the one it suggested (plastid
loss for DDL, a mis-scoped experimental term for npp-20, a bad InterPro signature
for Gprc5c). Only one of the five (cec-10) is the IBA node-depth problem the paper
emphasised. The other is too broad an anatomy term (BTBD3).

The neXtProt claim

The paper says, citing personal communication with Lydie Lane, that neXtProt
"decided to downgrade the status of IBA annotations in the future release of
neXtProt from 'Gold' to 'Silver'". We could not confirm this:

So it is a plausible second-hand report, with no public record we could find,
about a resource that no longer exists. It should not be cited as evidence that
IBA is low quality.

What this means for IBA review

RAR rerun on current GOA (2026-11-11)

rar_rerun.py reruns the screen on current data.
It uses the paper's 20 species and the authors' exact rate-ratio test with Benjamini-Hochberg
correction. The inputs are the GOA proteome files of 2026-11-04 and GO release 2026-07-26.
It counts asserted annotations only (NOT rows excluded).

Kingdom level Vertebrates vs invertebrates
Terms with a rate ratio 8,520 8,613
Rate ratio < 0.1 2,862 (paper, 2019: 2,369) 577 (paper: 618)
Of those, q < 0.05 2,615 550
Review candidates 1,241 280

The flag counts are close to the paper's, so the method reproduces on current data. Two
artefacts inflate the raw counts:

So the committed candidates keep only
stray annotations: q < 0.05, not a root term, and at most three proteins in the flagged taxon.
Only 4 candidate terms were also flagged in 2018 or 2019, so nearly all of today's flags are new:
MBF transcription complex, positive regulation of division septum assembly, periplasmic space,
and (vertebrate vs invertebrate) sevenless binding. Three of these already have in-repo answers:
TPV-004, the npp-20 MODIFY, and TPV-008 with CR-0006.

What is new since 2019

The rest are curated rows from TAIR, MGI, CGD, FlyBase and UniProt. Of the 21 terms
involved, the first pass finds 14 annotation errors, 5 constraints that look too narrow and
2 open cases. The too-narrow constraints are:
- glutamate-gated calcium channel activity, plus the signalling term inferred from it, and
plasma membrane repair: all Metazoa-only, but plant GLR3.3 and SYT1 do these;
- cholesterol binding: never in Fungi, but CAP/PRY proteins bind sterols;
- sperm axoneme assembly: Metazoa-only, but Selaginella makes flagellate sperm.
- TreeGrafter IEA in prokaryotes. 47 prokaryote-flagged terms on 44 proteins come from
TreeGrafter, PANTHER's automatic transfer of PAINT annotations to unreviewed proteins. Most are
eukaryote-only structures and processes on P. aeruginosa PA14, C. difficile and archaeal
proteins: microtubule binding, mitotic spindle, nuclear pore, Golgi organization,
telomerase-dependent telomere maintenance, extracellular matrix. TreeGrafter postdates the
paper, which saw this pattern only as IBA. Grafting a TrEMBL protein onto a PAINT node
without a taxon check turns every node placed too deep into many IEA rows.
- Prokaryote IBA, the paper's Udg pattern. Eighteen PAINT nodes still give bacteria IBAs
flagged at the kingdom level. The first pass finds:
- 6 errors: ubiquitin/proteasome terms on two B. subtilis proteins; sterol metabolism on
YjiC; carotenoid biosynthesis on YisP; a eukaryotic multi-synthetase complex on
Mycoplasma ProS; the AcuC and RppH constraint breaks;
- 5 that are rare but right: ecotin, PaaK, LigD NHEJ, NhaP2, EnvC;
- 1 that is right but too general: PdxS, a MODIFY candidate;
- 6 open cases.

The false flags make the paper's point again: a low rate shows curation density as often as
error.

Verdicts are in rar_rerun_triage.yaml. They are
first-pass: they rest on curated UniProt function text, term definitions and constraints, not on
reading each paper. Where a verdict rests on a cached paper or UniProt record, the file quotes it
verbatim, so just validate-quotes checks it. Plant- and animal-flagged candidates are
screened in the next section; the fungus-flagged ones are not yet.

Plant and animal candidates (2026-11-11)

All 958 kingdom-level candidates flagged in plants (604) or animals (354) have a first-pass
verdict in
rar_rerun_triage_plant_animal.yaml.
They were screened in five groups, by lineage and by evidence type: plant IBA, plant inferred,
plant experimental, animal inferred and animal experimental. A merge step checked that every
term is covered and that every proposed replacement is a live GO term (109 ids); quote_check
verifies all 39 quotes.

Verdict Entries
Annotation wrong 139
Rare but right (flag fails) 363
Experimental, deferred to the curator 379
Right but too general 36
Constraint looks too narrow 4
Uncertain 66

Experimental rows are not overruled on first-pass evidence. They are called wrong only when
the term's definition or GO taxon constraint excludes the lineage, which makes the error a term
choice. Otherwise they are deferred to the curator, who read the full text.

Most flags fail. The flag measures curation density as often as error, as in the first round:
- plant metabolic enzymes look rare because bacteria dominate the denominator;
- conserved eukaryotic machinery looks rare in plants because plants are under-curated;
- many animal flags are real enzymology that is rare in animals, such as the nematode
glyoxylate cycle, phytochelatin synthase and insect trehalose metabolism.

The 139 errors fall into a few kinds:

The four too-narrow constraints repeat the first round's cases: glutamate-gated calcium channel
and plasma membrane repair (both Metazoa-only, but plant GLR3.3 and SYT1 do these), sperm
axoneme assembly (Selaginella), and ionotropic glutamate receptor signalling. One verdict
disagreed with an existing review. Mouse Acadl's carnitine catabolic process row was kept as
non-core, but the term means breaking down carnitine itself, which mammals do not do; the
knockout paper shows only acylcarnitine build-up from blocked beta-oxidation. The Acadl
review now removes the row, and the two CoA-linked carnitine metabolic process rows that
rest on the same paper.

Next steps

Sources