SPKW Non-Arabidopsis Plants (PLANTS) Subproject

Parent project: SPKW.md

Overview

The ARATH subproject examined SwissProt-keyword (SPKW) over-annotation
in Arabidopsis thaliana. This subproject extends that analysis to the rest of the green
plant lineage (Viridiplantae, NCBITaxon:33090) excluding Arabidopsis — rice, maize,
soybean, tomato, potato, tobacco, wheat, grape, poplar, the green alga Chlamydomonas,
and ~200 other species — to test whether the patterns seen in the premier model plant
generalize across crops and non-flowering plants.

The analysis was run against plant.ddb, a DuckDB GO-annotation database built in the
go-db repository (make plant) from the September 2025 goa_uniprot_gcrp snapshot
(Gene-Centric Reference Proteomes), loaded with a swissprot table of 26,493 reviewed
accessions. It contains 40.4M annotations across the plant kingdom.

The central new finding is temporal. Between the September 2025 snapshot and the
current (≈April 2026) GOA release, GOA retired the keyword-to-GO pipeline
(GO_REF:0000043) for all cellular organisms. Live QuickGO now returns zero
GO_REF:0000043 annotations for human, mouse, fly, worm, S. pombe, Chlamydomonas,
Arabidopsis, rice, soybean, tobacco, and every other model organism checked (only viruses
retain them). plant.ddb is effectively the last high-coverage record of the plant SPKW
landscape, so this subproject is a retrospective validation study: for thirty-eight
genes across fourteen species (nine original, five in the 2026-05-29 species extension,
four methylation, twelve in the developmental/defense/nodulation sweep, four hormone-
signaling subtypes, and four long-tail keywords — the latter five batches a full sweep of
the high/medium-value Tier A keyword watch-list) we ask whether GOA's wholesale removal
was justified, and we classify the 214 SPKW-unique terms by their over-annotation risk.
All 38 reviews have had their file: supporting-text quotes verbatim-audited (the
reference-validator does not check file: prefixes; see SPKW-METHODOLOGY.md).

Key Statistics (2026-05-21)

Scope: non-Arabidopsis Viridiplantae, restricted to SwissProt (reviewed) entries via the
swissprot table; closure-filtered per SPKW-METHODOLOGY.md.

The 74% closure reduction is in line with Arabidopsis (72%) and confirms that
closure-based filtering is essential — a naive query would overstate the problem ~4×.

Distribution by Aspect

Per-Species Breakdown (TRUE SPKW-unique)

Rice dominates (japonica + indica = 1,927 genes, 47% of SPKW-unique genes) simply because
it is the most deeply SwissProt-curated non-Arabidopsis plant. Over 200 species contribute,
including non-flowering lineages (the moss Physcomitrium, the alga Chlamydomonas) and
the basal angiosperm Amborella.

Term-Level Classification: Which SPKW Terms Are Prone to Over-Annotation

A core question for this subproject is whether "SPKW-unique" should be read as a synonym
for "over-annotation". It should not. Classifying all 214 TRUE SPKW-unique terms by
GO-curation principles shows the over-annotation risk is concentrated in a small,
identifiable set of process/pathway terms, while the bulk of SPKW-unique annotations are
correct — either broad-but-true or specific-and-informative.

(Tier boundaries are curatorial judgement; the counts use an explicit GO-ID assignment
and sum to the 6,678 / 214 totals.)

Tier A — Prone to over-annotation (15% of annotations, 29 terms)

The keyword names a process or pathway the gene is associated with, not a function it
performs. This is where keyword→GO mapping systematically fails.

• Hormone signaling — auxin-activated signaling pathway (120 genes), abscisic
  acid-activated signaling (25), gibberellic acid-mediated signaling (21),
  brassinosteroid-mediated signaling (20), cytokinin-activated signaling (16),
  ethylene-activated signaling (9), red/far-red light signaling (7). Applied to
  hormone-responsive genes, not signal transducers.
• Defense / immunity processes — defense response (312), killing of cells of another
  organism (42), defense response to fungus (33), defense response to bacterium (12),
  plant-type hypersensitive response (8), innate immune response (7), defense response to
  virus (5), immune system process (3). Applied to any pathogen-induced or PR protein.
• Pleiotropic development — cell differentiation (62), flower development (46),
  nodulation (40), fruit ripening (26). Expression during development ≠ function.
• Cell cycle / meiosis keywords — cell division (81), meiotic cell cycle (6),
  chromosome segregation (2). Applied to general genome-maintenance genes.
• Vague / previously-flagged — methylation (92), rhythmic process (10), autophagy (4),
  signal transduction (2). Process terms too broad to be informative or
  over-annotated in earlier SPKW subprojects.

Tier B — Broad cofactor / enzyme-class MF (48% of annotations, 37 terms)

The keyword reflects a real bound metal, cofactor, ligand, or broad catalytic class.
These are almost never wrong; they are uninformative. Removing them loses generic-but-true
information.

• Binding — metal ion binding (1,149 genes), nucleotide binding (342), ATP binding
  (319), zinc ion binding (286), quinone binding (178), DNA binding (152), RNA binding
  (110), 4Fe-4S / iron-sulfur cluster binding (~150), GTP binding, lipid binding,
  L-ascorbic acid binding, calmodulin binding, carbohydrate binding.
• Broad enzyme-class MF — oxidoreductase activity (129), hydrolase activity (105),
  kinase activity (27), transferase activity, lyase activity, peptidase activity,
  isomerase activity, nuclease activity, methyltransferase activity.

Tier B is the single largest category (48%) — the dominant content of the SPKW-unique
set is broad, correct, low-information binding/enzyme-class terms.

Tier C — Specific informative function (31% of annotations, 142 terms)

The keyword reflects the protein's genuine catalytic, structural, biosynthetic, or
transport role. These are correct and informative; removing them is a real loss.

• Specific enzymes — monooxygenase activity, endonuclease activity, helicase activity,
  trihydroxystilbene synthase activity, dioxygenase activity, methionine adenosyltransferase
  activity.
• Biosynthetic / metabolic pathways — amino acid biosynthetic process (71), fatty acid
  biosynthetic process (54), polysaccharide catabolic process (66), lipid catabolic
  process (62), photorespiration (53), chlorophyll biosynthetic process, lignin
  biosynthesis, alkaloid metabolic process.
• Inhibitors — serine-type endopeptidase inhibitor activity (38), alpha-amylase
  inhibitor activity (10), peptidase inhibitor activity.
• Structural / transport / translation — cell wall organization (140), ribonucleoprotein
  complex (235), DNA-directed RNA polymerase complex (58), nutrient reservoir activity
  (54), oxygen carrier activity (23), symporter activity (32), porin activity, translation
  and tRNA/rRNA processing.

The 142 Tier-C terms are mostly small (1–20 genes each) — a long tail of specific,
correct functions.

Tier D — Context-dependent (6% of annotations, 6 terms)

Correct for canonical pathway members, over-annotation for accessory or divergent ones:
photosynthesis (110), photosystem I (66), photosystem II (65), chlorophyll binding
(87), toxin activity (22), response to herbicide (25). A core PSI subunit legitimately
gets "photosynthesis"; a non-photosynthetic PEPC paralog does not (see PPC16 below).

Bottom line

Only ~15% of plant SPKW-unique annotations (Tier A) carry real over-annotation risk.
~79% (Tiers B + C) are correct — broad-but-true or specific-and-informative. This reframes
the SPKW problem: it is not that keyword annotations are generally bad, but that a
specific, enumerable set of ~29 process/pathway terms is systematically mis-applied. The
nine gene reviews below were chosen to sample each tier and test this classification.

Full Gene Reviews Completed (2026-05-21)

Nine genes, one per species, sampling all four tiers.

Verdict across ~16 reviewed SPKW annotations: GOA's removal was justified for every
Tier A and Tier B annotation; it was a genuine or partial loss of correct biology for
the Tier C cases (CASP1 cell wall organization; EME1 endonuclease activity; PR1B1 defense
response to fungus). The tier of a term predicts whether its removal was safe.

Case 1: EME1 — Meiosis Keyword on a General Repair Endonuclease (Tier A→C)

Gene: EME1 (Q0J9J6) — Crossover junction endonuclease EME1, Oryza sativa Japonica.
Review file: genes/ORYSJ/EME1/EME1-ai-review.yaml

OsEME1 is the EME1/MMS4-family subunit of the MUS81–EME1 structure-specific endonuclease,
cleaving branched DNA in both mitotic repair and meiotic recombination. The keyword
"Meiosis" (→ meiotic cell cycle, GO:0051321) is over-specific — oseme1 mutants are
MMS/Zeocin-hypersensitive in vegetative tissue [PMID:40333587] and rice mus81 has normal
crossover designation [PMID:36495065] — so removal was justified (MARK_AS_OVER_ANNOTATED).

The instructive twist: of six SPKW-unique terms, five were over-broad and correctly
dropped, but endonuclease activity (GO:0004519) was genuinely correct — EME1 is an
endonuclease (EC 3.1.22.-). Current GOA, having dropped the keyword annotation, now has
no catalytic molecular-function term at all for OsEME1. Action MODIFY → crossover
junction DNA endonuclease activity (GO:0008821), restored as a curated annotation. Three
expression-based IEP annotations (citing a cDNA-collection paper) were also REMOVEd.

Case 2: PPC16 — Photosynthesis Keyword on a C3 Housekeeping Enzyme (Tier D)

Gene: PPC16 (Q02909) — Phosphoenolpyruvate carboxylase, housekeeping isozyme, Glycine max.
Review file: genes/SOYBN/PPC16/PPC16-ai-review.yaml

PEPC is famous as the CO₂-fixing enzyme of C4/CAM photosynthesis, so "Photosynthesis"
is a family-wide keyword. But PPC16 is the anaplerotic housekeeping isozyme of soybean,
a C3 plant — C3 plants have no photosynthetic PEPC. Decision REMOVE; removal fully
justified. The same family-level error in two non-SPKW annotations ("carbon fixation",
"tricarboxylic acid cycle") was MODIFYed to oxaloacetate metabolic process. This is the
Tier D context-dependent term in action: "photosynthesis" is right for a PSI subunit,
wrong for a C3 anaplerotic enzyme.

Case 3: PARA — Auxin-Signaling Keyword on an Auxin-Induced Enzyme (Tier A)

Gene: PARA (P25317) — Probable glutathione S-transferase parA / STR246C, Nicotiana tabacum.
Review file: genes/TOBAC/PARA/PARA-ai-review.yaml

parA is a tau-class glutathione S-transferase. It carries the name "Auxin-regulated
protein" because its transcription is auxin-induced — not because it transduces the
auxin signal. "Auxin-activated signaling pathway" (GO:0009734) belongs to the TIR1/AFB–
Aux/IAA–ARF cascade. Decision REMOVE; removal fully justified. The correct term,
"response to auxin" (GO:0009733), was added. Classic "expression ≠ function" pattern.

Case 4: PR1B1 — "Killing" Keyword on the Canonical SAR Marker (Tier A→D)

Gene: PR1B1 (P04284) — Pathogenesis-related leaf protein 6 (PR-1b), Solanum lycopersicum.
Review file: genes/SOLLC/PR1B1/PR1B1-ai-review.yaml

The genuinely mixed case. "Killing of cells of another organism" (GO:0031640)
over-states the evidence — tomato P14a inhibits Phytophthora growth modestly via sterol
sequestration [PMID:7784503], not overt cytotoxicity → MARK_AS_OVER_ANNOTATED. But
"defense response to fungus" was substantially correct biology; its only flaw is
taxonomic (the characterized target is an oomycete) → MODIFY to defense response to
oomycetes (GO:0002229). Removal only partly justified.

Case 5: STS3 — Defense Keyword on a Phytoalexin Biosynthetic Enzyme (Tier A)

Gene: STS3 (P51071) — Stilbene synthase 3, Vitis vinifera.
Review file: genes/VITVI/STS3/STS3-ai-review.yaml

STS3 is a type III polyketide synthase that makes trans-resveratrol, grapevine's major
phytoalexin. Because resveratrol is antimicrobial, the keyword "Plant defense" was
mapped to "defense response" (GO:0006952). But STS3 is a biosynthetic synthase, not a
defense effector or sensor — the annotation conflates the protective role of the product
with the function of the enzyme. Decision MARK_AS_OVER_ANNOTATED; removal justified.
Two NEW biosynthetic terms — stilbene biosynthetic process (GO:0009811) and phytoalexin
biosynthetic process (GO:0052315) — capture the defense context correctly. This is the
"biosynthetic enzyme labelled with its product's downstream process" pattern.

Case 6: NFP — Kinase Activity on a Pseudokinase (Tier A)

Gene: NFP (Q0GXS4) — Nod Factor Perception LysM receptor-like kinase, Medicago truncatula.
Review file: genes/MEDTR/NFP/NFP-ai-review.yaml

NFP is the LysM receptor that perceives rhizobial Nod factors to initiate the
nitrogen-fixing symbiosis. Its intracellular kinase domain is a catalytically dead
pseudokinase — it lacks conserved catalytic residues and has no phosphotransfer activity
[PMID:16844829]; signalling proceeds via the active co-receptor LYK3. "Protein tyrosine
kinase activity" (GO:0004713) is therefore doubly wrong (pseudokinase; and plant RLKs are
Ser/Thr, not Tyr) → REMOVE, removal justified. The keyword-derived "defense response"
was also REMOVEd as redundant with specific curated defense terms. Notably the current
InterPro-derived "protein kinase activity" (GO:0004672) has the same pseudokinase
problem — sequence-motif pipelines (SPKW and InterPro2GO alike) cannot tell an active
kinase from a pseudokinase. Even a Tier-C-looking specific MF can be wrong.

Case 7: METK1 — Broad Cofactor Keyword on a SAM Synthase (Tier B)

Gene: METK1 (A9P822) — S-adenosylmethionine synthase 1, Populus trichocarpa.
Review file: genes/POPTR/METK1/METK1-ai-review.yaml

METK1 makes S-adenosylmethionine, the universal methyl donor. Its two SPKW-unique terms
are both correct-but-broad: "metal ion binding" (GO:0046872) is true — the enzyme needs
Mg²⁺ and K⁺ — but uninformative → MARK_AS_OVER_ANNOTATED (proposed: magnesium/potassium
ion binding); "one-carbon metabolic process" (GO:0006730) is a defensible broad parent of
the precise, still-current "S-adenosylmethionine biosynthetic process" → KEEP_AS_NON_CORE.
A clean Tier B exemplar: removal pruned coarse, redundant terms without losing real
biology.

Case 8: psaC — Broad Keywords Superseded by Specific Terms (Tier B)

Gene: psaC (Q00914) — Photosystem I iron-sulfur center, Chlamydomonas reinhardtii.
Review file: genes/CHLRE/psaC/psaC-ai-review.yaml

PsaC is the PSI subunit carrying the two terminal [4Fe-4S] clusters that pass electrons to
ferredoxin. Its two SPKW-unique terms are correct but imprecise: "metal ion binding"
gestures at what the still-current "4 iron, 4 sulfur cluster binding" (GO:0051539) states
precisely; "oxidoreductase activity" gestures at what the still-current "electron transfer
activity" (GO:0009055) states precisely. Both REMOVE; removal justified — no
information lost because more specific current annotations cover the same biology. (A
nuance: the relationship is "imprecise generalization superseded by a better term", not
strict GO parent→child subsumption.)

Case 9: CASP1 — Cell-Wall Keyword That Carried Real Biology (Tier C)

Gene: CASP1 (C5YAP3) — Casparian strip membrane protein 1, Sorghum bicolor.
Review file: genes/SORBI/CASP1/CASP1-ai-review.yaml

CASP proteins are MARVEL-domain membrane proteins that self-assemble into the Casparian
Strip Membrane Domain and recruit the lignin-polymerizing machinery that builds the
endodermal Casparian strip — a localized cell wall modification forming the root
diffusion barrier. "Cell wall organization" (GO:0071555) is therefore genuinely correct
and informative. Decision MODIFY to the precise Casparian strip assembly
(GO:0160073) — but GOA's outright removal of this SPKW annotation discarded correct,
plant-specific biology. This is the key counter-example: a Tier C term where blanket
retirement was not the right call.

Extension: Five More Species (2026-05-29)

A second round added five genes across four further species — maize (Zea mays),
bread wheat (Triticum aestivum), the moss Physcomitrium patens (a non-vascular
plant ~450 My diverged from the crops), and potato (Solanum tuberosum) — selected to
stress-test the tier→verdict rule on the highest-risk Tier A keyword classes
(hormone signaling, immunity, meiosis/cell-cycle) plus a storage-protein keyword.

Verdict: every bona fide over-annotation (ABP1 signaling, MPK4a broad immunity,
AM1 cell-cycle keywords, PATB1 defense) was correctly removed; but two more cases —
RHT1 and PATB1's storage role — are collateral damage where the keyword was the
sole carrier of correct biology. The extension brings the running collateral-damage
tally to five (CASP1, EME1, PR1B1, RHT1, PATB1).

Case 10: RHT1 — A Hormone-Signaling Keyword That Was Correct (Tier A, but legitimate)

Gene: RHT1 (Q9ST59) — DELLA protein RHT-1 / "Reduced height-1", Triticum aestivum.
Review file: genes/WHEAT/RHT1/RHT1-ai-review.yaml

RHT1 is the wheat DELLA protein — one of the Green Revolution semi-dwarfing genes — and
DELLA proteins are the central repressors of gibberellin signaling: bioactive GA binds
the GID1 receptor, the GA–GID1–DELLA complex recruits an SCF E3 ligase, and DELLA is
degraded by the proteasome, de-repressing growth. "Gibberellic acid mediated signaling
pathway" (GO:0009740) is therefore not an over-annotation at all — it is RHT1's core
process. Yet the keyword2GO retirement stripped it, leaving RHT1 with essentially only
"nucleus". Decision MODIFY to the more precise "negative regulation of gibberellic
acid mediated signaling pathway" (GO:0009938), since DELLA is specifically a repressor,
plus a NEW transcription-corepressor MF (GO:0003714 — DELLA acts via protein–protein
interactions with TFs, not sequence-specific DNA binding). GOA's removal was NOT
justified. RHT1 is the mirror image of Arabidopsis ARF19 and tobacco PARA: those were
hormone-responsive genes wrongly called signaling components; RHT1 is a signaling
component wrongly stripped of the label.

Case 11: ABP1 — Auxin-Signaling Keyword on the Contested Auxin Receptor (Tier A)

Gene: ABP1 (P13689) — Auxin-binding protein 1, Zea mays.
Review file: genes/MAIZE/ABP1/ABP1-ai-review.yaml

ABP1 is the textbook auxin-binding protein: maize ABP1 binds auxin in an ER-luminal cupin
pocket with a coordinated Zn²⁺ (crystal structure), so "auxin binding" (GO:0010011, IDA)
and "endoplasmic reticulum lumen" (EXP) are retained as correct core terms. Unlike
PARA/ARF19, ABP1 genuinely binds the hormone. But "auxin-activated signaling pathway"
(GO:0009734) is defined around auxin binding to a receptor that drives a downstream,
largely transcriptional response — a role now firmly attributed to the nuclear
TIR1/AFB–Aux/IAA–ARF module. ABP1's status as a functional auxin receptor is famously
contested (CRISPR abp1 nulls are phenotypically normal; the original knockout carried a
confounding second-site mutation). Decision MARK_AS_OVER_ANNOTATED, replacement
"response to auxin" (GO:0009733). The lesson: binding a hormone is not the same as being
a component of its signaling pathway — removal justified, but the binding MF must be kept.

Case 12: MPK4a — Correct-but-Redundant Immunity Terms in a Moss (Tier A)

Gene: MPK4a (A9T142) — Mitogen-activated protein kinase 4a, Physcomitrium patens.
Review file: genes/PHYPA/MPK4a/MPK4a-ai-review.yaml

PpMPK4a is the moss orthologue of the immunity MAP kinase MPK4, experimentally required
for pathogen defense: the CERK1→MEKK1→MKK1→MPK4a cascade transduces chitin/chitosan
perception to defense-gene activation, and mpk4 mutants are susceptible to Botrytis
cinerea [PMID:27268428]. So the retired keywords "innate immune response" (GO:0045087)
and "defense response" (GO:0006952) are correct. But current GOA already carries the
more specific, experimentally backed "pattern recognition receptor signaling pathway"
(GO:0002221, EXP) — of which innate immune response is a broad parent — so the SPKW terms
add little. Decision MARK_AS_OVER_ANNOTATED for both, with NEW specific terms (defense
response to fungus; cellular response to chitin). Removal justified (low information
loss). The moss case shows the over-annotation pattern is identical in a non-vascular
plant deeply diverged from the crops — the same ELF4/psaC "correct-but-redundant" shape.

Case 13: AM1 — Cell-Cycle Keywords on a Meiotic-Entry Switch (Tier A)

Gene: AM1 (C0RWW9) — Protein AMEIOTIC 1, Zea mays.
Review file: genes/MAIZE/AM1/AM1-ai-review.yaml

AMEIOTIC1 is the master switch for meiotic entry in maize: am1 mutants fail the
mitosis-to-meiosis transition and arrest at leptotene, and AM1 is required for homologous
pairing and initiation of meiotic recombination [PMID:19204280]. Its biology is
meiosis-specific and is already captured by current GOA's "meiotic cell cycle" (IDA),
"reciprocal meiotic recombination" and "meiotic sister chromatid cohesion". The retired
SPKW terms are over-broad or mis-placed: "cell division" (GO:0051301) is uninformative
(→ MARK_AS_OVER_ANNOTATED), and "chromosome segregation" (GO:0007059) is mis-placed —
am1 arrests at leptotene, so chromosomes never reach the meiotic divisions where
segregation occurs (→ REMOVE). Removal justified. This parallels rice EME1 and
Arabidopsis BUB3.1: generic cell-cycle keywords pasted onto genes with a specific, earlier
role.

Case 14: PATB1 — A Storage Protein That Lost Its Defining Annotation (Tier C/A, mixed)

Gene: PATB1 (P15476) — Patatin-B1, Solanum tuberosum.
Review file: genes/SOLTU/PATB1/PATB1-ai-review.yaml

Patatin is the major potato tuber storage glycoprotein (~40% of soluble tuber protein) and
also a lipid acyl hydrolase. Its three SPKW terms split by tier: "nutrient reservoir
activity" (GO:0045735) is patatin's defining role — current GOA has no storage term, so
its removal is collateral damage (→ ACCEPT, re-assert as core); "lipid catabolic
process" (GO:0016042) is correct and more specific than the retained generic "lipid
metabolic process" (→ ACCEPT); but "defense response" (GO:0006952) over-states an
indirect role (defense is a downstream consequence of the lipid-hydrolase activity
releasing antifungal fatty acids; UniProt itself hedges "thought to be involved")
(→ MARK_AS_OVER_ANNOTATED). The current generic "lipase activity" was MODIFYed to the
precise "glycerophospholipase activity". Removal verdict mixed — storage and
lipid-catabolism correct, defense over-annotated. Like CASP1, the Tier C term carried real
biology.

What the extension adds: hormone-signaling keywords cut both ways

The original PLANTS finding was that hormone-signaling keywords land on hormone-responsive
genes (ARF19, PARA) that are not signal transducers. RHT1 (DELLA) is the mirror image: a
hormone-signaling keyword on the gene that is the central signaling component — here
removal was wrong. ABP1 sits between them: a genuine auxin binder whose role as a pathway
component is contested. Refined rule: for a hormone-signaling keyword, ask whether the
gene is (i) a component of the transduction machinery (keep — DELLA), (ii) a ligand-binder
of contested pathway role (case-by-case — ABP1), or (iii) merely hormone-responsive
(over-annotation — PARA/ARF19). The tier alone does not decide a hormone-signaling keyword;
the gene's position in the pathway does.

Methylation-Keyword Batch (2026-05-30): One Keyword, Four Substrates

Following the keyword-level watch-list in SPKW-METHODOLOGY.md,
Methyltransferase → methylation (GO:0032259) was the largest unreviewed Tier A
keyword (92 plant genes). Four methyltransferases spanning four substrate classes were
reviewed to test whether the bare process term is a reliable over-annotation.

Result: in all four, the bare methylation (GO:0032259) is an over-annotation. It drops
the one thing that matters — the substrate — and in every case the substrate-specific
molecular function (DNA (cytosine-5)-methyltransferase, histone H3K27 methyltransferase,
caffeoyl-CoA O-methyltransferase, benzoquinol methyltransferase activity) is already present,
making the bare process term redundant.

Why methylation is almost always redundant: GO puts the substrate on the MF branch

A structural insight made this verdict near-deterministic. GO encodes methylation
substrate-specificity on the molecular-function branch (the specific
…-methyltransferase activity terms), not the process branch. The substrate-specific
methylation process terms one would want to MODIFY to are now obsolete — DNA
methylation (GO:0006306), histone methylation (GO:0016571), and histone H3-K27
methylation (GO:0070734) are all obsoleted in the current ontology. So for a
Methyltransferase-keyword gene there is usually no specific process term to MODIFY to,
and the right action is to drop the bare methylation in favour of the specific MF the gene
already carries. The one exception in this batch was MET1A, where a maintenance-specific
process term survives (GO:0141119, chromosomal DNA methylation maintenance following DNA
replication, whose GO definition explicitly names plant MET1).

Case 15: MET1A — DNA methylation maintenance (DNA substrate)

Gene: MET1A (Q7Y1I7) — DNA (cytosine-5)-methyltransferase 1A, Oryza sativa.
Review file: genes/ORYSJ/MET1A/MET1A-ai-review.yaml

OsMET1a is the rice DNMT1-orthologue that maintains CG methylation after replication. The
keyword-derived methylation drops both substrate (DNA) and mechanism (maintenance); decision
MODIFY → GO:0141119 (the surviving substrate-specific process term, whose definition names
MET1). The redundant bare methyltransferase activity (GO:0008168) was also
MARK_AS_OVER_ANNOTATED, leaving the specific DNA (cytosine-5-)-methyltransferase activity
(GO:0003886) as the core MF. Tier A.

Case 16: EZ1 — histone H3K27 methylation (protein substrate)

Gene: EZ1 (Q8S4P6) — Histone-lysine N-methyltransferase EZ1 (maize Enhancer-of-zeste, "Mez1").
Review file: genes/MAIZE/EZ1/EZ1-ai-review.yaml

EZ1/Mez1 is the catalytic SET-domain subunit of Polycomb Repressive Complex 2; it
trimethylates histone H3 Lys27 to silence chromatin. methylation → MARK_AS_OVER_ANNOTATED:
the substrate-specific histone methylation process terms are obsolete, and the specific MF
histone H3K27 methyltransferase activity (GO:0046976) is already present, so the bare term is
redundant with no MODIFY target. The review also tightened the generic histone
methyltransferase activity → H3K27-specific, and REMOVEd a spurious ortholog-transferred
single-stranded RNA binding. Tier A.

Case 17: CCOAOMT — lignin O-methylation, with collateral damage (small-molecule substrate)

Gene: CCOAOMT (Q8H9B6) — Caffeoyl-CoA O-methyltransferase, Solanum tuberosum.
Review file: genes/SOLTU/CCOAOMT/CCOAOMT-ai-review.yaml

The Mg²⁺-dependent O-methyltransferase that converts caffeoyl-CoA to feruloyl-CoA in
phenylpropanoid/lignin biosynthesis — and the cleanest redundancy case: methylation is
fully covered by the specific MF caffeoyl-CoA O-methyltransferase activity (GO:0042409) plus
the pathway process, so → MARK_AS_OVER_ANNOTATED (Tier A). But this gene is also multi-tier
and partly collateral damage: its retired lignin biosynthetic process (GO:0009809) is
correct and informative → ACCEPT (Tier C; the retained phenylpropanoid biosynthetic
process is only the broader parent, not a substitute), and metal ion binding (GO:0046872) is
broad-but-correct → MARK_OVER (Tier B). A spurious ARBA-predicted circadian rhythm
(GO:0007623) was also REMOVEd. Tier A/C/B.

Case 18: COQ5 — ubiquinone C-methylation (cofactor substrate)

Gene: COQ5 (Q5JNC0) — 2-methoxy-6-polyprenyl-1,4-benzoquinol methylase, mitochondrial, Oryza sativa.
Review file: genes/ORYSJ/COQ5/COQ5-ai-review.yaml

The C-methyltransferase of the coenzyme-Q (ubiquinone) biosynthesis pathway. methylation →
MARK_AS_OVER_ANNOTATED: both the specific MF (GO:0008425, benzoquinol methyltransferase
activity) and the specific process (GO:0006744, ubiquinone biosynthetic process) are already
present, so the bare term is purely redundant. Tier A.

What the methylation batch adds

Methyltransferase → methylation is a reliable, mechanistically-explained over-annotation:
4/4 genes, across DNA / histone / secondary-metabolite / cofactor substrates, had the bare
process term flagged (3 MARK_OVER, 1 MODIFY). Because GO deliberately keeps methylation
substrate-specificity on the MF branch and obsoleted the specific process terms, the keyword's
target term is structurally doomed to be uninformative or redundant. As a bonus, reviewing the
keyword-derived terms surfaced two unrelated over-predictions (CCOAOMT circadian rhythm
from ARBA; EZ1 ssRNA binding from ortholog transfer) — a reminder that the SPKW review pass
is also a general annotation-quality pass.

Keyword-Sweep: Developmental, Defense & Nodulation Batches (2026-05-30)

Continuing down the keyword watch-list (SPKW-METHODOLOGY.md), twelve
more genes were reviewed across the next three high-count Tier A keyword classes:
developmental (Differentiation→cell differentiation, Flowering→flower development),
defense/killing (Plant defense→defense response, toxin/killing), and
Nodulation→nodulation. Genes were chosen to span each keyword's verdict range;
per-gene detail is in the individual ai-review.yaml files. All 12 pass just validate.

Developmental keywords (cell differentiation GO:0030154; flower development GO:0009908)

"cell differentiation" is a generic catch-all that is essentially never the right term. On
flowering-time signals/regulators (GI, HD3A) it is simply wrong → REMOVE; on genuine
developmental genes (MADS3 organ identity, FEA2 meristem homeostasis) it is correct-but-far-too-
coarse → MODIFY to the precise process. "flower development" on flowering-time genes (GI, HD3A)
is the ELF4 pattern — they control the timing/transition, not organ morphogenesis → MODIFY to a
"regulation of flower development" term. GI's "rhythmic process" repeats the 100%-over-annotated
human/ELF4 result.

Defense / killing keywords (defense response GO:0006952; toxin/killing)

"defense response" spans the full verdict range, set by the gene's mechanism. A genuine
immune receptor (XA21) → correct-but-broad, MODIFY to the specific defense-to-bacterium (it
was the gene's only process term, so collateral-damage-adjacent). A biosynthetic enzyme for
a defense metabolite (CPS4) → over-annotation, the STS3 enzyme-vs-product conflation in a crop.
A genuine cytotoxin (b-32 RIP) → "toxin activity" is legitimate (kept, refined to the
precise rRNA N-glycosylase MF), while the broad "defense response" process is trimmed. An
antifungal enzyme (CHIB chitinase) → the substrate-catabolism term is correct (MODIFY to
chitin-specific) but "defense response" is over-broad. In every case the informative annotation
is the molecular function; "defense response" is either redundant with it or wrong.

Nodulation keyword (nodulation GO:0009877)

"nodulation" splits cleanly by what the gene does. For a core signaling component (NSP1 —
literally NODULATION SIGNALING PATHWAY 1) it is correct and central; removal was collateral
damage (the RHT1/DELLA pattern — current GOA left it with no symbiosis term). For CCAMK it is
a pathway/organism-context error: rice does not nodulate, so the legume-inherited keyword is
wrong — but the gene's genuine arbuscular mycorrhizal symbiosis role must be preserved
(MODIFY), exactly the PPC16 photosynthesis-in-a-C3-plant pattern. For a component that merely
operates in the nodule (leghemoglobin's O₂ transport) or a nodule-specific expression marker
(ENOD2), "nodulation" is over-broad — the real biology is the molecular function / structural
role, which is kept.

Hormone-signaling subtypes (ABA, cytokinin, BR, ethylene) — completing the picture

With auxin (ABP1) and GA (RHT1/DELLA) from the species extension, all six major plant
hormones are now covered, and the "hormone-signaling cuts both ways" rule resolves cleanly by
the gene's position in the pathway:

The keyword is right for genuine transduction components (DELLA, TUD1, EIL2 — keep), too
coarse/misframed for downstream responsive effectors (VP1, ABP1 — MODIFY to "response
to"/"regulation of"), and simply wrong for a hormone-metabolism enzyme (CKX2, which
degrades cytokinin and has nothing to do with signal transduction). One keyword
("X signaling pathway") thus yields keep / MODIFY / remove verdicts depending entirely on
whether the gene perceives-and-transduces, merely responds to, or metabolizes the hormone.

Long-tail keywords (fruit ripening, autophagy, hypersensitive response, light signaling)

A final pass covered the remaining smaller Tier A keywords (each ≤30 genes):

These reconfirm the four outcomes and add two sharp variants:
- Regulatory-direction inversion (LSD1, DET1): both are negative regulators of the process
their keyword names — LSD1 restrains the hypersensitive response (lsd1 = runaway cell death),
DET1 represses photomorphogenesis. The keyword maps them to the bare process as if they
execute/promote it; the correct term is "negative regulation of …". This is a sharper form of
the regulator-vs-participant conflation — the keyword gets the direction backwards.
- Keyword reputation ≠ gene verdict (JOKA2): the Autophagy keyword was ~80% over-annotated in
the human subproject, but potato JOKA2 is a bona fide NBR1/p62-family selective-autophagy cargo
receptor — so "autophagy" is correct here (MODIFY to the more specific macroautophagy/
autophagosome). Gene-specific biology overrides keyword priors.
- PG2 (fruit ripening on the softening polygalacturonase) is the STS3/CPS4 enzyme-vs-process
pattern in developmental guise; the pass also removed spurious anther/fruit-dehiscence
cross-paralog transfers.

What the keyword-sweep confirms

Across all three batches the same meta-rule holds — a process/role keyword is only as good as
the gene's actual position relative to that process — and resolves into four recurring
outcomes that recur regardless of which keyword is involved:

1. Core component → keep (NSP1 nodulation, XA21 receptor; cf. RHT1/DELLA) — removal is collateral damage.
2. Wrong organism / pathway context → MODIFY (CCAMK nodulation→mycorrhiza; cf. PPC16 photosynthesis).
3. Right area, wrong altitude → MODIFY to specific (MADS3, FEA2, HD3A/GI flower development, CHIB, and the CPS4 product-process).
4. Expression / component ≠ function → MARK_OVER or REMOVE (GI & HD3A cell differentiation, ENOD2, leghemoglobin nodulation, GI rhythmic process).

The bare process term almost never adds information beyond the gene's molecular function; where
it does (NSP1, XA21), it is the sole carrier of a correct fact and must not be silently
dropped. This is the same tier→verdict logic established for the first 18 genes, now stress-
tested across four more keyword classes and a 14th species (common bean).

Over-Annotation Patterns (Non-Arabidopsis Plants)

Comparison: Non-Arabidopsis Plants vs Other Organisms

The headline number is different in kind: rather than an organism-level "issue rate", the
plant analysis localizes the risk to Tier A (15% of annotations, 29 terms). Of the
nine genes reviewed, every Tier A SPKW term was a genuine over-annotation, while Tier B/C
terms were correct.

Key Lessons from Non-Arabidopsis Plants

1. "SPKW-unique" is not "over-annotation". Only ~15% of plant SPKW-unique annotations
   (Tier A) are over-annotations. ~48% (Tier B) are broad-but-correct cofactor/enzyme-class
   terms and ~31% (Tier C) are specific, correct, informative functions. The problem is a
   small, enumerable set of ~29 process/pathway terms.

2. The patterns generalize, but the tier predicts the verdict. Crops, the moss
   Physcomitrium, the alga Chlamydomonas, and the basal angiosperm Amborella all show
   the same categories. The tier of an SPKW term reliably predicts whether GOA's removal
   was safe: Tier A removals were always justified; Tier B removals lost only redundant
   coarse terms; Tier C removals risked — and in CASP1, EME1, and PR1B1 actually caused —
   loss of correct biology.

3. GOA's retirement of the SPKW pipeline was right for the over-annotations — but
   collateral damage is real. EME1 lost its only catalytic molecular-function term;
   CASP1 lost a correct cell-wall-organization annotation; PR1B1 lost correct (if
   imprecise) defense biology. A keyword annotation that is the sole carrier of a fact
   should be re-asserted with a proper evidence code, not silently dropped.

4. Sequence-motif pipelines cannot detect catalytic death. NFP shows that both SPKW
   keyword2GO and InterPro2GO assign "kinase activity" to a known pseudokinase. This
   failure mode affects Tier-C-looking specific MF terms, not just broad ones.

5. A hormone-signaling keyword cuts both ways (2026-05-29 extension). The tier alone
   does not decide a hormone-signaling term — the gene's position in the pathway does. RHT1
   (wheat DELLA) is a core GA-signaling repressor whose keyword was correct and wrongly
   retired (collateral damage); PARA and ARF19 are merely hormone-responsive and were
   correctly flagged; ABP1 (a genuine auxin binder of contested receptor status) is
   case-by-case. Ask: transduction component → keep; ligand-binder of contested role →
   case-by-case; merely responsive → over-annotation.

Recommendations

1. Triage retired SPKW annotations by tier. Tier A removals can be accepted wholesale.
   Tier C removals must be checked individually — if the keyword was the only source of a
   correct fact, re-assert it (ideally with a more specific term).

2. The 29 Tier A terms are a reusable watch-list. Any future keyword→GO mapping, in
   any organism, should treat hormone-signaling, defense/immunity-process,
   pleiotropic-development, and cell-cycle/meiosis keywords as high-risk.

3. Pathway-context check for enzyme families (Tier D): confirm the specific
   isozyme/organism performs the family's signature role before applying the family
   keyword — C3 vs C4 status, paralog subfamily, and tissue expression all matter.

4. Distinguish hormone-responsive from hormone-signaling genes, and
   biosynthetic enzymes from the defense processes their products serve.

Review Files Location

genes/ORYSJ/EME1/EME1-ai-review.yaml      (A→C: meiosis over-annotated; endonuclease MF restored)
genes/SOYBN/PPC16/PPC16-ai-review.yaml    (D: photosynthesis on a C3 housekeeping PEPC)
genes/TOBAC/PARA/PARA-ai-review.yaml      (A: auxin signaling on an auxin-induced GST)
genes/SOLLC/PR1B1/PR1B1-ai-review.yaml    (A/D: PR-1 defense keywords)
genes/VITVI/STS3/STS3-ai-review.yaml      (A: defense response on a phytoalexin synthase)
genes/MEDTR/NFP/NFP-ai-review.yaml        (A: kinase activity on a pseudokinase)
genes/POPTR/METK1/METK1-ai-review.yaml    (B: broad cofactor keyword on a SAM synthase)
genes/CHLRE/psaC/psaC-ai-review.yaml      (B: broad keywords superseded by specific terms)
genes/SORBI/CASP1/CASP1-ai-review.yaml    (C: correct cell-wall biology, wrongly retired)
# 2026-05-29 extension:
genes/WHEAT/RHT1/RHT1-ai-review.yaml      (A-legit: GA signaling on a DELLA; removal = collateral damage)
genes/MAIZE/ABP1/ABP1-ai-review.yaml      (A: auxin signaling on the contested auxin receptor; binding kept)
genes/PHYPA/MPK4a/MPK4a-ai-review.yaml    (A: broad immunity terms redundant with PRR-signaling, in a moss)
genes/MAIZE/AM1/AM1-ai-review.yaml        (A: cell-division/segregation keywords on a meiotic-entry switch)
genes/SOLTU/PATB1/PATB1-ai-review.yaml    (C/A: storage role kept; defense over-annotated)
# 2026-05-30 methylation-keyword batch (Methyltransferase -> methylation, GO:0032259):
genes/ORYSJ/MET1A/MET1A-ai-review.yaml    (A: DNA MTase; methylation MODIFY -> DNA-maintenance GO:0141119)
genes/MAIZE/EZ1/EZ1-ai-review.yaml        (A: histone H3K27 MTase; methylation redundant w/ specific MF)
genes/SOLTU/CCOAOMT/CCOAOMT-ai-review.yaml (A/C/B: lignin OMT; methylation redundant, lignin process correct)
genes/ORYSJ/COQ5/COQ5-ai-review.yaml      (A: ubiquinone MTase; methylation redundant w/ specific MF+process)
# 2026-05-30 keyword sweep — developmental (Differentiation/Flowering):
genes/ORYSJ/GI/GI-ai-review.yaml          (A: clock/photoperiod; cell diff REMOVE, flower dev MODIFY, rhythmic MARK_OVER)
genes/ORYSJ/HD3A/HD3A-ai-review.yaml      (A: florigen; cell diff REMOVE, flower dev MODIFY->positive regulation)
genes/ORYSJ/MADS3/MADS3-ai-review.yaml    (A: C-class floral TF; cell diff MODIFY->specification of stamen identity)
genes/MAIZE/FEA2/FEA2-ai-review.yaml      (A: CLV meristem receptor; cell diff MODIFY->meristem maintenance)
# 2026-05-30 keyword sweep — defense/killing:
genes/ORYSJ/XA21/XA21-ai-review.yaml      (A: R-gene/PRR; defense response MODIFY->defense response to bacterium)
genes/ORYSJ/CPS4/CPS4-ai-review.yaml      (A: phytoalexin synthase; defense response MARK_OVER, enzyme-vs-product)
genes/MAIZE/O6/O6-ai-review.yaml          (legit toxin: RIP; toxin activity MODIFY->rRNA N-glycosylase, defense MARK_OVER)
genes/MAIZE/CHIB/CHIB-ai-review.yaml      (A/C: chitinase; polysacc catab MODIFY->chitin catab, defense MARK_OVER)
# 2026-05-30 keyword sweep — nodulation:
genes/MEDTR/NSP1/NSP1-ai-review.yaml      (legit core: Nod-signaling TF; nodulation ACCEPT, removal=collateral damage)
genes/ORYSJ/CCAMK/CCAMK-ai-review.yaml    (A: rice common-symbiosis kinase; nodulation MODIFY->arbuscular mycorrhizal association)
genes/PHAVU/LBA/LBA-ai-review.yaml        (C/A: leghemoglobin; O2-carrier ACCEPT, nodulation MARK_OVER)
genes/SOYBN/ENOD2A/ENOD2A-ai-review.yaml  (A: early nodulin; nodulation MARK_OVER, expression marker)
# 2026-05-30 hormone-signaling-subtype batch (ABA/cytokinin/BR/ethylene):
genes/MAIZE/VP1/VP1-ai-review.yaml        (A: ABA B3 TF; ABA signaling MODIFY->regulation/response, responsive effector)
genes/ORYSJ/CKX2/CKX2-ai-review.yaml      (A: cytokinin dehydrogenase; cytokinin signaling MARK_OVER, catabolic enzyme)
genes/ORYSJ/TUD1/TUD1-ai-review.yaml      (A: BR E3 ligase; BR signaling ACCEPT, legitimate component)
genes/ORYSJ/EIL2/EIL2-ai-review.yaml      (A: ethylene EIN3-like TF; ethylene signaling ACCEPT, legitimate component)
# 2026-05-30 long-tail keyword batch (fruit ripening, autophagy, HR, light signaling):
genes/SOLLC/PG2/PG2-ai-review.yaml        (A: polygalacturonase; fruit ripening MARK_OVER, enzyme-vs-process)
genes/SOLTU/JOKA2/JOKA2-ai-review.yaml    (A-legit: NBR1-like autophagy receptor; autophagy MODIFY->macroautophagy)
genes/ORYSJ/LSD1/LSD1-ai-review.yaml      (A: neg. regulator of cell death; HR MODIFY->negative regulation of HR)
genes/SOLLC/DET1/DET1-ai-review.yaml      (A: photomorphogenesis repressor; light signaling MODIFY->neg. reg. photomorphogenesis)

Methods Note

The nine genes were selected from the plant.ddb closure-filtered TRUE SPKW-unique set to
sample all four term tiers across nine species. Because GOA had already retired
GO_REF:0000043, just fetch-gene seeds existing_annotations from live QuickGO and no
longer includes the SPKW annotations; the historical SPKW annotations were re-added from
the September 2025 plant.ddb snapshot and flagged retired: true so they are reviewed
but excluded from the live-GOA consistency check. The term-tier classification uses an
explicit GO-ID assignment over all 214 terms (the spkw_plants_unique analysis table in
plant.ddb). See SPKW-METHODOLOGY.md for the closure-filtered SQL.