IBA Annotation Quality — History, Methodology & Lessons

← back to the findings page

This is the working log for the IBA quality project. The main page holds the
findings; this page holds the process: the dated log, how each candidate
was verified, what got retracted, and — most importantly — the lessons I want my
future self to apply before ever flagging an IBA as wrong.

---

Lessons learned (read this before flagging any IBA)

These are written to myself. The recurring failure mode is acting on one line of
evidence. Flagging a curated IBA as wrong is a strong claim; treat it like one.

1. No single source is sufficient — synthesize. A review YAML assertion, a
   UniProt keyword, and a PANTHER node label are each individually weak. Before
   REMOVE, I should have at least two independent lines pointing the same way
   (e.g. UniProt CAUTION and direct enzymology; a curated NOT and the term
   definition).

2. Read the GO term definition, not the label. My worst miss: I called
   ATP7B's GO:0015677 "copper ion import" a direction inversion because ATP7B is
   a copper exporter. But the definition is "movement into a cell or
   organelle," and ATP7B pumps Cu into the Golgi lumen — so the term is
   defensible. The label looked opposite; the definition was not. Always fetch
   the definition (QuickGO/OLS) for any "directionally wrong" claim.

3. "By similarity" is weak; direct experiment in the target species beats it.
   My second miss (in the other direction): I retracted the AGK ceramide/sphingosine
   flags because UniProt "curates" them — but the ceramide claim was only a
   propagated "By similarity" tag, UniProt also says "Does not phosphorylate
   sphingosine," and two papers (PMID:15939762, PMID:16269826) report no
   ceramide/sphingosine activity. Deferring to the keyword was exactly backwards.
   Check the evidence code behind the UniProt statement (ECO:0000250 "by
   similarity" vs ECO:0000269 experimental) and read the cited paper.

4. Don't misread fragmented tool output. I mis-paraphrased UniProt's "Does not
   phosphorylate sphingosine" as "phosphorylates sphingosine" from a line-wrapped
   grep. Read the full FUNCTION block (sed -n '/FUNCTION/,/-!-/p'), not a
   keyword grep, before quoting it.

5. Use phylogeny / the PANTHER family composition. Over-propagation almost
   always traces to a family node that lumps members of different
   function/specificity. Inspecting the family pays off: PTHR12358 mixes
   acylglycerol and sphingosine kinases (explains AGK); PTN000135648 mixes pro-
   and anti-apoptotic BCL2 members (explains the BCL2 sign error); PTN000222418
   lumps HMGCS paralogs (explains HMGCS2). Always look at the IBA WITH/FROM and
   what the node contains.

6. Distinguish "absent" from "over-annotated/non-core". HMGCS2 really does
   catalyze the HMG-CoA-synthase step (the mevalonate-pathway entry reaction); what's
   wrong is assigning it the downstream FPP/isoprenoid flux that belongs to the
   cytosolic paralog. That's a paralog over-annotation, not a fabricated activity —
   so the framing must be precise (over-annotation), not "it can't do this."

7. Respect the curator (CLAUDE.md). REMOVE is for IBA/IEA inferences I can argue
   against on biological grounds — not for second-guessing experimental annotations
   whose full text I haven't read. When I genuinely can't adjudicate, UNDECIDED.

8. A label that "looks" like a name collision needs the WITH/FROM to confirm the
   cause. PEX2 is biologically not in the Cdc73/Paf1 complex (solid), but the
   "old synonym PAF1" mechanism is unverified — the WITH/FROM is a PANTHER node,
   not a PAF1 gene. Separate "this is wrong" (provable) from "this is why it's
   wrong" (often a hypothesis).

9. Read the WITH/FROM column before anything else. It names the exact source
   proteins the IBA was transferred from. If they are the wrong family (NTN1's TF
   terms came from POU2F1/POU1F1/POU4F1; NOTCH1's axon guidance from SLIT1/2/3) or
   a single wrong paralog (ABRAXAS1's spindle terms from ABRAXAS2/Q15018; HINT2's
   cytoplasm from HINT1), that is near-mechanical evidence of error. If they are a
   broad, coherent set of true orthologs, the transfer is probably fine. I should
   pull WITH/FROM on every candidate, not just the puzzling ones.

10. Uncertain ≠ wrong — leave the borderline ones (UQCRC1). I nearly added
    UQCRC1 as a "pseudo-peptidase," but UniProt says the bc1 core subunits "seem to
    have preserved their MPP processing properties (By similarity)" and "may be
    involved in" processing UQCRFS1. That is genuinely unresolved, so the right call
    is UNDECIDED, not a confident REMOVE. Excluding it was the disciplined choice.

11. For localization, place both compartments in the GO hierarchy first.
    GO:0005737 cytoplasm is defined as "the contents of a cell excluding the plasma
    membrane and nucleus, but including other subcellular structures" — and
    mitochondrion, ER, Golgi, and lysosome are all part_of cytoplasm. So "cytoplasm"
    is not wrong for a mitochondrial/ER/lysosomal protein (just imprecise); flagging
    it REMOVE is an over-reach. Valid localization REMOVEs need mutually-exclusive
    compartments: nucleus vs cytoplasm, plasma membrane vs internal, one organelle vs
    another. This is the same "read the definition" trap as ATP7B's copper import.

12. Check taxon/lineage appropriateness for process terms. The single biggest BP
    error class is processes transferred across kingdoms to organisms where they don't
    exist: inflammatory response on a mosquito Toll (from human TLR4), JAK-STAT on a
    worm (no JAK), aerobic respiration on chloroplast NDH subunits, smell perception
    on mosquito salivary OBP-fold proteins. The WITH/FROM naming a vertebrate/Drosophila
    source is the tell; GO taxon constraints catch some. Always ask "does this process
    occur in this lineage / this organelle system?"

---

2026-06-20 (e) — Final sweep + first actual MSAs

MSAs (finally — addressing the standing gap). Through passes (a)–(d) I never built
an alignment; "missing catalytic residue" rested on UniProt CAUTION text. Built a
reproducible FAMSA pipeline (msa/, pyfamsa under uv) that
pulls catalytic-residue positions live from UniProt feature tables and inspects them:
- Argonautes: AGO2 tetrad intact (D/D/H); AGO4 has two substitutions (D669G,
H807R) → confirms no slicer activity. Bonus nuance: AGO3 retains the tetrad, so the
blanket "only AGO2 has the residues" framing would have been wrong — the MSA corrected it.
- CRMP/DPYSL: all five paralogs have lost the carbamylated catalytic Lys159 (→L/M/Q)
plus Zn-coordinating His/Asp vs active DPYS — first-hand confirmation of the CAUTION.
This is the lesson-5 / lesson-9 method applied properly; cited from Pattern 7.

Final sweep of remaining ~38 REMOVE stragglers. Added the clearly-verified ones:
- Pseudo-enzyme (Pattern 7): CASP12 (UniProt RecName "Inactive caspase-12") and
Serpinh1/HSP47 (non-inhibitory serpin; collagen ER chaperone).
- Regulator/effector (Pattern 15): sigF/sigG/sigK sigma factors given RNA-polymerase
catalytic activity — a specificity subunit assigned the holoenzyme's catalytic activity
(belongs to RpoB/RpoC).

Excluded for insufficient verification (discipline): PGRPLC/PGRPS1 (the PGRP/amidase-2
family has both catalytic and non-catalytic members; flagging amidase activity needs the
same residue check I did for CRMP — not done). Left in backlog with UQCRC1.

Remaining backlog is now heterogeneous singletons (e.g. lgg-1 GABA-receptor binding,
ubl-5 protein-tag activity, bbs-1 patched/smoothened binding, CFTR basolateral PM,
generic metal/calcium-binding over-annotations) — no further coherent clusters; returns
have diminished.

---

2026-06-20 (d) — Fifth pass: biological-process cluster

Triaged the 53 BP REMOVE candidates (44 genes). Verified each flagship against UniProt
+ WITH/FROM + (where cited) cached papers. Two new patterns, two reinforced.

New Pattern 14 — lineage-inappropriate / cross-kingdom process transfer (largest BP
class): TOLL9 inflammatory response ← human TLR4 (O00206); ndhA/D/K aerobic respiration
(UniProt: chloroplastic NDH, plastid); che-3 cilium motility (dynein-2 IFT; worm cilia
non-motile); D7r2/r4/r5/L1 smell perception (UniProt: blood-feeding salivary proteins);
sta-2 JAK-STAT (no JAK in worms; from fly/mouse STATs); fshr-1 hormone signaling (no
gonadotropins); HEN1 piRNA processing (plant; metazoan over-transfer); DpuGr29 male
courtship (Drosophila-specific term).

New Pattern 15 — regulator/effector & direct/downstream conflation: lys-7 (AMP
effector, not a signal transducer); SIR3 (represses origins ≠ replication initiation);
UBP3 (regulation of protein stability is downstream of its DUB activity).

Reinforced Pattern 8 (sub-activity): worm pseudo-sHSPs — hsp-12.3/hsp-12.6 have no
chaperone activity (PMID:9744800 title verbatim), stronger than CRYAA. Reinforced
Pattern 12 (mis-grouping): opa1/eat-3 peroxisome fission on mitochondrial-fusion OPA1
(DRP1-branch transfer); YAR1 transcription term on an RPS3-binding 40S-biogenesis factor;
ACL4 mito-import by TOM70-family over-transfer; plus BP versions of the PIWI/germ-granule
cluster (pgl-1 mRNA splicing/export, glh-4 SSU-rRNA maturation).

Noted but not separately featured (paralog conflation, already a known pattern):
mid1←mid2 (septin ring), Ang2←angiogenin/AMP paralogs, csr-1 (cell-differentiation IBA
traceable to a nuclear-receptor node, not the Argonaute — a wrong-grouping case I left
out pending tighter confirmation of the source).

---

2026-06-20 (c) — Fourth pass: the generic-localization cluster (two-sided)

Triaged the 32 generic-localization REMOVE candidates (cytoplasm 10, nucleus 7,
plasma membrane 6, mitochondrion 2, membrane 2, cytosol 2, peroxisome/Golgi/nucleoplasm
1 each). First fetched the GO:0005737 definition — it subsumes membrane-bounded
organelles — which split the cluster cleanly:

Valid REMOVEs (mutually-exclusive compartment) → added as Pattern 13 Tier A:
nucleus on cytoplasmic PIWI/Argonaute (PIWIL1, prg-1, wago-1, glh-1; WITH/FROM
includes nuclear-acting Piwi orthologs), nucleus on ER-kinase EIF2AK3 and on BIRC6,
plasma membrane on ribosome-associated chaperones SSB2/SSZ1, nucleoplasm on BAIAP2L2,
peroxisome on PIK3C3, cytoplasm on strictly-nuclear rqh1/HDA1, cytoplasm on secreted
SCGB1A1.

Reviewer over-reaches (cytoplasm subsumes the organelle) → Pattern 13 Tier B, NOT
flagged: "cytoplasm" REMOVE on Aga/GLA (lysosome), DHCR24 (ER), ISCA1/ATP5IF1/gtpbp3
(mito); "membrane" REMOVE on flvcr2a (a multi-pass membrane transporter). These
should be UNDECIDED/KEEP.

Self-correction: HINT2 (added in pass (b) as a wrong-paralog "cytoplasm" case) is
itself a Tier-B over-reach — mitochondrion ⊂ cytoplasm, so "cytoplasm" isn't strictly
wrong. Removed HINT2 from the findings table.

Net lesson: localization is where reviewers (me included) most often over-flag, because
"cytoplasm"/"membrane" are broad subsuming terms. Roughly a third of this cluster were
reviewer over-reaches rather than bad IBAs.

---

2026-06-20 (b) — Third pass: WITH/FROM-driven triage of the backlog

Worked through more of the ~167 un-triaged REMOVE candidates, this time leading
with the WITH/FROM column and confirming each against UniProt + family. Added 11
verified examples and two new patterns (mis-grouping revealed by WITH/FROM;
extended pseudo-enzyme + substrate/neofunctionalization sets).

Added (verified):
- Pseudo-enzyme — AKTIP (UniProt CAUTION: "Lacks the conserved Cys residue
necessary for ubiquitin-[conjugation]"; UEV pseudo-E2) and DPYSL4 (fourth
CRMP-family member on the same non-catalytic basis as DPYSL2/3/CRMP1).
- Substrate / neofunctionalization — SAMD8/SMSr (UniProt: makes ceramide
phosphoethanolamine from PE, not sphingomyelin from PC — "The larger PC prevents
an efficient fit in the enzyme's catalytic pocket") and CPT1C (UniProt RecName
"Palmitoyl thioesterase CPT1C"; lost the ancestral carnitine-transferase activity).
- Mis-grouping (WITH/FROM-confirmed) — Tier A wrong-family: NTN1/NTN3 (secreted
Netrins carrying POU-domain TF terms; WITH/FROM = POU2F1 P14859, POU1F1 P28069,
POU4F1 Q12837, POU4F3 Q15319), NOTCH1 (axon guidance from SLIT1/2/3
O75093/O94813/O75094), IL23R (prolactin-receptor activity from PRLR P16471).
Tier B wrong-paralog: ABRAXAS1 (spindle/MT from ABRAXAS2 Q15018), HINT2
(cytoplasm reflects HINT1).

Examined and deliberately NOT added (insufficiently certain):
- UQCRC1 / GO:0017087 (MPP complex) — UniProt: core subunits "seem to have
preserved their MPP processing properties (By similarity)" and "may be involved
in" UQCRFS1 processing. Residual activity is unresolved → UNDECIDED, not REMOVE.

Still in the backlog (not yet adjudicated): many generic localization
over-propagations (GO:0005737 cytoplasm / GO:0005634 nucleus on DHCR24, PIWIL1,
BIRC6, EIF2AK3, GLA, GK5, SCGB1A1, …) — these look like a high-frequency "default
compartment" category worth a dedicated future pass; CRY1/CRY2 photo-entrainment
(contested); ankzf1 ERAD-vs-RQC; CALCOCO2 PML-body; BAIAP2L2 nucleoplasm.

---

2026-06-20 — Second corpus pass (+ verification)

Method. Mined all 2,732 genes/*/*/*-ai-review.yaml for IBA-evidence
annotations and their review action. Distribution:
ACCEPT 4651 · KEEP_AS_NON_CORE 943 · MODIFY 321 · MARK_AS_OVER_ANNOTATED 189 · REMOVE 190 · UNDECIDED 39 · NEW 33 · PENDING 11. Shortlisted ~12 REMOVE candidates
and verified each against: UniProt *-uniprot.txt (FUNCTION block + evidence codes),
the GO term definition (QuickGO), GOA *-goa.tsv WITH/FROM provenance and PANTHER
family composition, cached publications/, and the gene -notes.md.

Verification first pass — over-retracted three, then corrected on review:

Gene	First call	Why first call was wrong	Final
ATP7B GO:0015677	flag → then retract	n/a — retraction was correct	Stays out. Term covers organelle import; ATP7B loads Cu into Golgi. Not a direction error.
AGK ceramide/sphingosine	flag → retract	I deferred to a UniProt "By similarity" keyword and misread "Does not phosphorylate sphingosine"	Reinstated. Two papers (PMID:15939762, PMID:16269826) + UniProt's own "does not phosphorylate sphingosine" refute the activity; ceramide is only "By similarity".
HMGCS2 GO:0010142	flag → retract	I deferred to UniProt's UniPathway "mevalonate biosynthesis" tag (shared reaction chemistry)	Reinstated as paralog over-annotation. FPP/isoprenoid flux belongs to cytosolic HMGCS1; mito HMGCS2 → ketogenesis.

Final verified additions (on the findings page): pseudo-enzyme cluster
(DPYSL2/CRMP1/DPYSL3, AGO4, UBAC2), partial sub-activity loss (CAPG, CRYAA),
regulatory-sign inversion (BCL2, with caveat), complex/compartment/pathway
over-transfer (EIF4E2, ALDH1L1, HMGCS2, PEX2), substrate over-propagation (AGK).

Evidence anchors per gene:
- DPYSL2/CRMP1/DPYSL3 — UniProt CAUTION: "Lacks most of the conserved residues … essential for binding the metal cofactor".
- AGO4 — UniProt FUNCTION: "Lacks endonuclease activity and does not appear to cleave target mRNAs".
- UBAC2 — rhomboid-like fold (inactive-rhomboid clan) + no curated protease activity (no explicit CAUTION → kept MEDIUM/cautious).
- CAPG — cached PMID:1322908 verbatim "blocks the barbed ends … but does not sever preformed actin".
- CRYAA — GOA already contains a curated NOT|involved_in (ISS) to GO:0042026 (refolding) that the IBA contradicts.
- BCL2 — WITH/FROM node PTN000135648 mixes pro-apoptotic BAX (Q07812)/BAK1 (Q16611) with anti-apoptotic members; caveat: a separate NAS annotation to the same term exists + context-dependent pro-apoptotic roles → "IBA sign unreliable," not "impossible".
- EIF4E2 — UniProt: "is unable to bind eIF4G", "Does not interact with eIF4G".
- ALDH1L1 — UniProt Cytosolic 10-formyltetrahydrofolate dehydrogenase, Cytoplasm, cytosol, cytosol IDA; mito work is ALDH1L2.
- HMGCS2 — IBA from PANTHER node PTN000222418 lumping HMGCS paralogs; notes reason out ketogenesis vs cytosolic mevalonate.
- PEX2 — UniProt: peroxisomal RING E3 for PEX5 retrotranslocation; WITH/FROM is a PANTHER node (name-collision cause unverified).
- AGK — PTHR12358 mixes acylglycerol + sphingosine kinases; direct human enzymology (PMID:15939762, PMID:16269826) shows MAG/DAG only.

Backlog (~175 un-triaged REMOVE-flagged IBAs) — verify each the same way before
adding: e.g. PIWIL1 piRNA cross-aspect, human/CRY1 & human/CRY2 photo-entrainment,
ankzf1 ERAD vs RQC.

---

2026-03-04 — Added ECOLI/arnF (mechanism divergence within SMR superfamily)

PTHR30561 groups the whole SMR/DMT superfamily — EmrE (drug efflux), MdtI/J
(spermidine export), Gdx (guanidinium export), Mmr (multidrug resistance), and
ArnE/ArnF (lipid flipping). Same 4-TM fold, but ArnE/ArnF do intramembrane lipid
translocation rather than transmembrane solute export. The IBA WITH/FROM for
GO:0022857 (transmembrane transporter activity) lists EmrE (P23895), MdtI
(P69210), MdtJ (P69212), Gdx (P69937), Mmr (P9WGF1), ArnE (Q47377) + ArnF — all
genuine exporters, correct for them but misleading for the flippase. Correct term:
GO:0140303 (intramembrane lipid transporter activity). Severity MEDIUM: the IBA
isn't wrong in kind (it is a transporter), just wrong in mechanism. EcoCyc also
contributed two "response to iron(III) ion" over-annotations (IGI PMID:12139617,
IEP PMID:15322361) that conflate BasS-BasR transcriptional regulation of the arn
operon with direct function.

2026-01-26 — Added cds1 (neo-functionalization, opposite reaction) + RIMBP2

cds1 (PTHR10314 SF135) — the most severe error type: GO:0019344 (cysteine
biosynthesis) propagated from the family root (PTN000034104) to a subfamily that
catalyzes cysteine catabolism (EC 4.4.1.1: L-Cys → H2S + pyruvate). Evidence of
neo-functionalization in SF135: longest branch length (0.528), different EC class,
~24% identity with synthases (synthases share 43%), distinct active-site motif
(ASSGST vs PTSGNTG). Experimental: PMID:34439535 (M. tuberculosis), PMID:34283874
(V. cholerae). Detail: interpro/panther/PTHR10314/PTHR10314-notes.md.

RIMBP2 — organism/tissue context transfer: GO:0007274 (neuromuscular synaptic
transmission) transferred from the Drosophila ortholog (FB:FBgn0262483, NMJ is the
fly model), but human RIMBP2 acts at central synapses (hippocampal mossy fiber,
CA3-CA1, auditory ribbon). Family PTHR14234 also contains non-synaptic RIMBP3 (SF21,
testis/spermiogenesis); the family research warns against propagating
"regulation of neurotransmitter release" to RIMBP3 paralogs.

2026-01-22 — Project creation (Epe1)

Documented IBA quality issues found through AI review. First key finding:
pseudo-enzyme propagation — Epe1 has a JmjC fold but an HVD (not HXD) motif and
lacks Fe(II) coordination; mass-spec assays show no demethylation; the H297A
catalytic mutant retains anti-silencing function. So the demethylase IBA
(GO:0032452) should be REMOVE. Recommendation that seeded the project: build a
pipeline to flag IBA enzymatic annotations on proteins with degenerate active-site
motifs.