SwissProt Keywords (SPKW) Unique Terms Project

Overview

This project reviews genes that have GO annotations derived solely from UniProt Keywords (SPKW) via GO_REF:0000043, with no corroborating evidence from experimental, computational, or curator sources. The goal is to identify systematic over-annotation patterns and distinguish legitimate SPKW contributions from problematic mappings.

Key Findings

• Over-annotation rates vary dramatically by organism and term
• Eukaryotic BP terms (apoptosis, meiosis, autophagy, rhythm) show 80-100% over-annotation
• Bacterial annotations are mostly accurate (~5% issues for P. putida)
• Viral annotations are clade-sensitive: phage immune/defense terms are often semantic mismatches, while eukaryotic viral immune-evasion terms may be legitimate but too broad
• Common patterns: process conflation, regulatory vs participatory confusion, caspase substrates
• GOA has retired the SPKW pipeline (≈April 2026): GO_REF:0000043 keyword-to-GO
  annotations have been removed from live GOA for all cellular organisms (verified zero for
  human, mouse, fly, worm, S. pombe, plants; only viruses retain them). The problem this
  project documented is now resolved at the source. Retrospective review of 38 non-Arabidopsis
  plant genes (see PLANTS) shows only ~15% of plant SPKW-unique terms carry
  real over-annotation risk; removal was justified for those, but blanket retirement also
  dropped correct annotations when the keyword was the only carrier of a fact.

Cumulative Results

Subproject	Organism	Total Genes	Reviewed	Issue Rate	Main Pattern
Apoptosis	Human	280	23	87%	Regulatory conflation
Rhythmic Process	Human	146	5	100%	Expression ≠ function
Autophagy	Human	123	14	79%	Signaling over-extension
ANOGA	A. gambiae	5,812	22	Mixed	D7 toxin=100%, immune=17%
SCHPO	S. pombe	1,963	7	100%	ATG-meiosis conflation
DROME	D. melanogaster	2,753	4	50%	Mixed patterns
PSEPK	P. putida	1,098	4	25%	RT defense keyword
ARATH	A. thaliana	8,433	4	75%	Subclade divergence
Virus clades	Viral taxa	54,131	11	55%	Host-context mismatch, specificity
PLANTS	Non-ARATH plants	4,117	38	15% Tier-A	Term-tiering; GOA retired SPKW
BPT4	Phage T4	~300	3	100%	Eukaryote-centric terms
ECO57	E. coli O157	~74,000	2	50%	Toxin vs effector

Methods

See SPKW-METHODOLOGY.md for detailed SQL queries and explanation of closure-based filtering (which reduces false positives by 70%+ in well-curated organisms).

Over-Annotation Patterns Identified

Pattern	Description	Examples	Action
Process conflation	Gene active during process X gets annotated to X	ATG genes → meiosis (S. pombe)	REMOVE
Regulatory conflation	Gene regulates X, annotated to X	AIMP2 → apoptotic process	MODIFY to regulatory term
Caspase substrate	Cleaved by caspases, annotated to apoptosis	AIMP1, BCAP31	REMOVE
Signaling over-extension	4+ steps from direct function	Sin1 → apoptosis	REMOVE
Eukaryote-centric terms	Immune/defense terms for phage-bacteria	T4 DAM → innate immune	REMOVE
Viral host-context mismatch	Same host-pathogen term valid in one viral clade but wrong in another	Phage AcrF8 → innate immune	MODIFY
Toxin vs effector	Effectors incorrectly called toxins	NleB1 (E. coli)	REMOVE
Subclade divergence	Family keyword ignores subfunctionalization	LCR1 (Arabidopsis DEFL)	REMOVE
Kratagonist ≠ toxin	Sequestration ≠ toxin activity	D7 proteins (mosquito)	MODIFY
Enzyme-class keyword → bare process	An activity keyword maps to a generic, substrate-less process term; substrate specificity lives on the MF branch	Methyltransferase → methylation (plant MTases: MET1A, EZ1, CCOAOMT, COQ5)	MARK_OVER / MODIFY

Legitimate SPKW Contributions

Not all SPKW-unique annotations are over-annotations:

• Antimicrobial peptides/lysozymes (D. melanogaster) - "killing of cells" is correct
• Arsenic/antibiotic resistance genes (P. putida) - direct functional annotations
• Conserved functions (Ced-12/ELMO in D. mel) - SPKW captures known biology missing from experimental annotations
• Core circadian genes (ELF4 in Arabidopsis) - accurate but redundant with specific terms
• DELLA growth repressors (RHT1/"Reduced height" in wheat) - "GA signaling pathway" is the core function of a DELLA; retirement was collateral damage (left only "nucleus")
• Storage proteins (patatin PATB1 in potato) - "nutrient reservoir activity" is patatin's defining role; current GOA has no storage term after retirement
• Viral life-cycle terms (phage DGR reverse transcriptase, phage quorum-sensing peptide, influenza M2) - often accurate but may need more specific viral or molecular-function terms

Project Status

• Started: 2025-12-23
• Last updated: 2026-05-30
• Total genes reviewed: 133 across 11 subprojects
• Compiled data: spkw_reviewed_genes.csv

Phase 1 (Original)

• [x] MAP3K5 - COMPLETE
• [x] PHF23 - COMPLETE
• [x] SIRT2 - COMPLETE

Subprojects

• [x] Apoptosis - 23/280 reviewed
• [x] Rhythmic Process - 5/146 reviewed
• [x] Autophagy - 14/123 reviewed
• [x] ANOGA - D7 + immune genes
• [x] SCHPO - ATG-meiosis pattern
• [x] DROME - Case studies
• [x] PSEPK - Bacterial control
• [x] ARATH - Plant patterns
• [x] Virus clades - Virus-wide and clade-specific patterns
• [x] PLANTS - Non-Arabidopsis crops (38 genes, 14 species); term-tier classification + retrospective validation + full Tier-A keyword-watch-list sweep (methylation, developmental, defense, nodulation, hormone-signaling x6, long-tail) + verbatim-quote integrity audit
• [x] BPT4 - Phage semantics
• [x] ECO57 - Toxin/effector

Curation Recommendations

1. Check regulatory vs participatory - many genes regulate processes but don't participate IN them
2. Consider organism biology - same GO term can have different validity across taxa
3. Distinguish toxins from effectors - direct cytotoxicity vs signaling modulation
4. Validate family-level keywords - subfunctionalization can invalidate family annotations
5. Expression ≠ function - upregulation during a process doesn't mean functional involvement
6. Check viral host context - phage-bacterium interactions need different terms from eukaryotic viral immune evasion

Swiss-Prot vs TrEMBL Analysis

Key finding: Keywords on Swiss-Prot entries are manually assigned by curators, not by ARBA/UniRule automatic systems. This means:

Organism	Swiss-Prot %	Implication
Human	99.6%	Over-annotations reflect manual curator keyword choices
T4 Phage	99.6%	Same - curators chose these keywords
E. coli O157	88.6%	Mostly manual
P. putida	32.2%	Mixed manual/automatic
Virus (all)	13.1%	Mostly TrEMBL; errors may reflect mapping or automatic keyword assignment
A. gambiae	3.8%	Mostly automatic keyword assignment

Of 71 genes with over-annotation issues: 70 are Swiss-Prot (99%)

For reviewed high-confidence organism batches, this confirms the problem is usually in the KW→GO mapping layer, not keyword assignment. Virus-wide analysis is different because most candidates are TrEMBL. See SPKW-METHODOLOGY.md and SPKW-VIRUS.md for stratification queries.

---

Session Notes

2026-05-30

• Added a keyword-level view to SPKW-METHODOLOGY.md: reverse-map
  SPKW-unique GO terms to their source UniProt keywords via the public keyword2go
  (external2go) mapping (the GAF stores only the GO term). Produced a tier-annotated
  watch-list of the ~30 process/role keywords that drive plant over-annotation; it is
  organism-independent (same keywords drove the human/pombe/Arabidopsis subprojects).
• Reviewed the largest unreviewed Tier A keyword, Methyltransferase → methylation
  (GO:0032259) (92 plant genes), across 4 substrate classes: MET1A (rice, DNA),
  EZ1 (maize, histone H3K27), CCOAOMT (potato, lignin caffeoyl-CoA), COQ5 (rice, ubiquinone).
• Finding: Methyltransferase → methylation is a reliable, mechanistically-explained
  over-annotation (4/4 flagged; 3 MARK_OVER, 1 MODIFY). GO keeps methylation
  substrate-specificity on the molecular-function branch and has obsoleted the specific
  process terms (DNA methylation GO:0006306, histone methylation GO:0016571, H3K27
  methylation GO:0070734), so the bare methylation (GO:0032259) is structurally doomed to be
  redundant with the specific …-methyltransferase activity MF the gene already carries.
• Bonus: the review pass also caught two unrelated over-predictions — CCOAOMT circadian
  rhythm (ARBA) and EZ1 single-stranded RNA binding (ortholog transfer), both REMOVEd.
• Keyword-watch-list sweep (12 more genes across the next three Tier A keyword classes):
• Developmental (Differentiation/Flowering): GI, HD3A (REMOVE cell differentiation;
  MODIFY flower development on flowering-TIME genes — the ELF4 pattern), MADS3
  (MODIFY→specification of stamen identity), FEA2 (MODIFY→meristem maintenance). "cell
  differentiation" is a catch-all that's wrong or too coarse.
• Defense/killing (Plant defense/toxin): XA21 R-gene (MODIFY→defense response to
  bacterium), CPS4 phytoalexin synthase (MARK_OVER, STS3 enzyme-vs-product), O6/b-32 RIP
  (toxin activity legitimate, kept→rRNA N-glycosylase), CHIB chitinase (mixed).
• Nodulation: NSP1 (ACCEPT — legitimate core Nod-signaling TF, removal=collateral damage),
  CCAMK (MODIFY→arbuscular mycorrhizal association — rice doesn't nodulate; PPC16-style
  organism-context error), LBA leghemoglobin (O2-carrier ACCEPT, nodulation MARK_OVER),
  ENOD2A (MARK_OVER, expression marker).
• Meta-rule confirmed: a process/role keyword is only as good as the gene's position
  relative to that process → (1) core component → keep; (2) wrong organism/pathway → MODIFY;
  (3) right area wrong altitude → MODIFY-to-specific; (4) expression/component ≠ function →
  MARK_OVER/REMOVE. PLANTS now 30 genes / 14 species (added common bean).
• Hormone-signaling-subtype batch (4 genes; completes all six major plant hormones):
  VP1 (maize, ABA B3 TF) MODIFY→regulation of/response to ABA (responsive effector); CKX2
  (rice cytokinin dehydrogenase) MARK_OVER — a catabolic enzyme that degrades the hormone,
  not a signaling component (cleanest case); TUD1 (rice BR U-box E3 ligase) ACCEPT (legitimate
  component); EIL2 (rice ethylene EIN3-like TF) ACCEPT (legitimate; collateral damage). With
  auxin (ABP1) and GA (RHT1/DELLA), one keyword ("X signaling pathway") yields keep / MODIFY /
  remove purely by the gene's pathway position (transduction component vs responsive effector vs
  hormone-metabolism enzyme). PLANTS now 34 genes / 14 species; the full high/medium-value
  Tier A watch-list is covered.
• Long-tail keyword batch (4 genes): PG2 (tomato polygalacturonase) fruit ripening MARK_OVER
  (enzyme-vs-process); JOKA2 (potato NBR1-like) autophagy MODIFY->macroautophagy (legitimate
  autophagy gene — keyword reputation ≠ gene verdict); LSD1 (rice) HR MODIFY->negative
  regulation of HR; DET1 (tomato) light signaling MODIFY->negative regulation of
  photomorphogenesis. New variant: regulatory-direction inversion — LSD1/DET1 are negative
  regulators of the very process their keyword names. PLANTS now 38 genes / 14 species.
• Data-integrity audit (important): the reference-validator
  (conf/reference_validator_config.yaml) lists file in skip_prefixes, so file:
  supporting-text is NEVER verbatim-checked by just validate — only PMID: quotes are. An
  audit found 265/550 (~48%) of file: quotes across all SPKW-PLANTS reviews (incl. the
  original 9 and the EME1 exemplar) were non-verbatim — fabricated/paraphrased/markdown-mangled —
  yet had passed validation. Fixed every one (replaced with exact substrings or verbatim
  primary-PMID quotes; removed a few genuinely fabricated claims) with zero annotation-action
  changes (curation was sound; provenance was not). Re-verified with the validator's own
  matching logic: 0/536 file: quotes non-verbatim across all 38 genes. Lesson recorded:
  prefer PMID quotes; grep-verify any file: quote.

2026-05-29

• Extended PLANTS with five more genes across four new species
  (maize, wheat, the moss Physcomitrium patens, potato), bringing PLANTS to 14 genes /
  13 species. Selected from the plant.ddb closure-filtered TRUE SPKW-unique set
  (spkw_plants_unique table) to stress-test the highest-risk Tier A keyword classes.
• Reviewed (falcon deep research + full ai-review.yaml, all just validate-clean):
• RHT1 (wheat DELLA, GA signaling) → MODIFY→neg. regulation of GA signaling; removal
  NOT justified (core repressor function lost — strongest collateral-damage case yet).
• ABP1 (maize, auxin signaling) → MARK_AS_OVER_ANNOTATED→response to auxin; removal
  justified (contested receptor), but auxin binding kept as core.
• MPK4a (moss, innate immunity/defense) → MARK_AS_OVER_ANNOTATED ×2; removal justified
  (correct but redundant with experimental PRR-signaling-pathway).
• AM1 (maize AMEIOTIC1, chromosome segregation/cell division) → REMOVE / MARK_OVER;
  removal justified (over-broad/mis-placed; specific meiotic terms carry the biology).
• PATB1 (potato patatin, nutrient reservoir/lipid catabolic/defense) → ACCEPT/ACCEPT/
  MARK_OVER; mixed — storage role is collateral damage, defense over-annotated.
• New finding — hormone-signaling keywords cut both ways. Tier alone does not decide a
  hormone-signaling term; the gene's pathway position does. RHT1 (DELLA, a transduction
  component) refines the earlier ARF19/PARA rule (those were merely hormone-responsive).
  Running collateral-damage tally now five: CASP1, EME1, PR1B1, RHT1, PATB1.

2026-05-21

• Added PLANTS subproject: SPKW over-annotation in non-Arabidopsis
  Viridiplantae.
• Built plant.ddb (go-db, make plant) from the Sept 2025 goa_uniprot_gcrp snapshot;
  40.4M annotations, 26,493 SwissProt accessions.
• Closure-filtered TRUE SPKW-unique query (non-ARATH, SwissProt-only): 6,678 annotations,
  4,117 genes, 214 terms — 74% closure reduction; aspect F 54% / P 37% / C 8%. Rice
  dominates (1,927 genes); 200+ species represented.
• Discovered GOA retired GO_REF:0000043 for all cellular organisms since the snapshot.
  Reframed PLANTS as a retrospective validation study.
• Reviewed 4 genes (rice EME1, soybean PPC16, tobacco PARA, tomato PR1B1) — all 4 SPKW-unique
  annotations were over-annotations; GOA's removal justified for every headline term, but
  EME1 (endonuclease activity) and PR1B1 (defense response to fungus) lost correct biology.
• Classified all 214 SPKW-unique terms into 4 tiers: A over-annotation-risk (15%, 29 terms),
  B broad cofactor/enzyme-class MF (48%), C specific informative (31%), D context-dependent
  (6%). Only ~15% carry real over-annotation risk; ~79% (B+C) are correct — "SPKW-unique"
  is not a synonym for "over-annotation".
• Reviewed 5 more genes across broader taxa (grape STS3, Medicago NFP, poplar METK1,
  Chlamydomonas psaC, sorghum CASP1) — 9 genes / 9 species total, sampling all 4 tiers.
  The tier predicts the verdict: every Tier A removal was justified; the Tier C removal
  (CASP1, cell wall organization) discarded correct plant-specific biology.

2026-02-04

• Researched UniProt keyword assignment process (confirmed: Swiss-Prot = manual)
• Created spkw_reviewed_genes.csv compiling 95 reviewed genes
• Added Swiss-Prot vs TrEMBL stratification to methodology
• Cross-species analysis: issue rates 10-40%, all Swiss-Prot dominated

2026-05-21

• Added SPKW-VIRUS.md as the virus-wide and clade-specific counterpart to the organism SPKW subprojects
• Quantified virus.ddb: 180,680 SPKW annotations, 135,117 naive SPKW-unique annotations, and 80,218 closure-filtered SPKW-unique annotations
• Summarized 11 existing viral gene reviews across phage, anti-CRISPR, influenza, phage quorum-sensing, and DGR cases