Top-Nots: Candidate NOT Annotations from Existing Reviews

Bottom line: a NOT annotation states that a gene product lacks a function, and it is
the only thing that stops automated pipelines from re-asserting a wrong activity by IEA
or IBA. Many of the repo's ~6,319 REMOVE and MARK_AS_OVER_ANNOTATED decisions are really
negative findings, so we mined them for the strongest NOT candidates with a keyword score
over the review summaries ("lacks catalytic", "pseudoenzyme", "no detectable activity"
and similar). We did this because a REMOVE only cleans one review, while a NOT, filed
upstream, prevents the same error from propagating again. The scan (2026-03-06) found
250 candidates across 19 species, 115 strong (score 4 or more) and 22 very strong,
dominated by pseudo-enzymes, "is phosphorylated" misread as "does phosphorylation", and
assembly factors given the activity of their complex; 21 of the 22 Tier 1 rows are
REMOVE in their reviews and one is MARK_AS_OVER_ANNOTATED. A later BGC addendum adds
five more high-score candidates from a heme-less P450 homologue and active-site-less
condensing-enzyme folds. No candidate has yet been
literature-verified or proposed as a formal NOT, so the list is a worklist, not a
submission.

Overview

NOT annotations (negated annotations) in GO explicitly state that a gene product does NOT have a particular function, localization, or involvement in a process. They are valuable because they:

Prevent propagation of incorrect IBA/IEA annotations to other species
Document experimentally-tested negative results
Correct systematic errors from domain-based prediction pipelines
Provide essential context for computational analyses

This project mines existing AI gene reviews to find the strongest candidates for formal NOT annotations - cases where the reviewer found strong evidence that the gene definitively does NOT have the annotated function, not merely that evidence is insufficient.

Key distinction: A REMOVE action means "this annotation should be removed" - but only a subset of these warrant a NOT annotation. The best NOT candidates are cases where:

There is positive experimental evidence the gene lacks the function
The gene has a domain that predicts an activity it demonstrably lacks (pseudo-enzyme pattern)
The misannotation is likely to recur via automated pipelines without an explicit NOT

Methodology

Candidates were identified by scanning all *-ai-review.yaml files for action: REMOVE or action: MARK_AS_OVER_ANNOTATED annotations whose review summaries contain multiple strong negative-evidence keywords (e.g., "does not have", "lacks catalytic", "pseudoenzyme", "no detectable activity", "domain lacks", "misannotation", "does not catalyze", "no intrinsic").

A keyword-based scoring system ranks candidates by confidence of negative evidence. Score >= 6 indicates very high confidence; score 4-5 indicates strong candidates; score 3 indicates moderate candidates worth manual review.

Statistics

Total REMOVE/MARK_AS_OVER_ANNOTATED annotations across all reviews: ~6,319
NOT candidates (score >= 3): 250 across 19 species
Strong NOT candidates (score >= 4): 115
Very strong NOT candidates (score >= 6): 22
Already marked as negated: true: 295 rows in 187 reviews (2026-10-08 recount; the earlier ~15 was an undercount). See NOT_NOTs

Tier 1: Highest Confidence NOT Candidates (score >= 6)

These are the strongest cases where a domain predicts an activity that is demonstrably absent.

Species	Gene	UniProt	Incorrect Term	Evidence	Score	Pattern
DROME	CG6051	Q9VB70	`GO:0004438` phosphatidylinositol-3-phosphate phosphatase activity	IEA	10	FYVE domain binds PI3P but has no phosphatase activity
DROME	CG6051	Q9VB70	`GO:0052629` phosphatidylinositol-3,5-bisphosphate 3-phosphatase activity	IEA	9	Same gene, second phosphatase term
human	MEX3B	Q6ZN04	`GO:0006468` protein phosphorylation	ISS	9	RNA-binding/E3 ligase, NOT a kinase
ANOGA	PGRPS1	Q7QFK2	`GO:0008745` N-acetylmuramoyl-L-alanine amidase activity	IBA	7	Non-catalytic PGRP, lacks zinc-binding residues
BACSU	spoVAD	P40869	`GO:0016746` acyltransferase activity	IEA	7	Thiolase-like fold but no enzymatic activity
SCHPO	Epe1	O94603	`GO:0016491` oxidoreductase activity	IEA	7	JmjC pseudo-enzyme, degenerate active site
human	AKTIP	Q9H8T0	`GO:0061631` ubiquitin conjugating enzyme activity	IBA	7	UEV domain lacks catalytic cysteine
human	CHRAC1	Q9NRG0	`GO:0003887` DNA-directed DNA polymerase activity	IEA	7	Histone-fold structural protein, no catalytic activity
human	PLD5	Q8N7P1	`GO:0003824` catalytic activity	IEA	7	Pseudoenzyme, lacks conserved HKD motifs
yeast	YDJ1	P25491	`GO:0005524` ATP binding	IEA	7	DnaJ/HSP40 stimulates Hsp70 ATPase but does not bind ATP itself
ACET2	cipB	Q01866	`GO:0004553` hydrolase activity, hydrolyzing O-glycosyl compounds	IEA	6	Dockerin domain =/= hydrolase
SCHPO	Epe1	O94603	`GO:0051213` dioxygenase activity	IEA	6	JmjC pseudo-enzyme
human	CREB1	P16220	`GO:0006468` protein phosphorylation	IDA	6	Is phosphorylated BY kinases, not itself a kinase
human	DNAJA4	Q8WW22	`GO:0005524` ATP binding	IEA	6	DnaJ/HSP40, same pattern as YDJ1
human	GMFB	P60983	`GO:0006468` protein phosphorylation	TAS	6	Glia maturation factor, not a kinase
human	ILF3	Q12906	`GO:0006468` protein phosphorylation	IDA	6	dsRNA-binding protein, not a kinase
human	MEX3B	Q6ZN04	`GO:0046777` protein autophosphorylation	ISS	6	Same gene, second phosphorylation term
human	PRRT1	Q99946	`GO:0006468` protein phosphorylation	ISS	6	SynDIG4, regulates phosphorylation of GRIA1 but is not a kinase
human	RASA1	P20936	`GO:0003924` GTPase activity	TAS	6	Stimulates Ras GTPase, has no intrinsic GTPase activity
human	SURF1	Q15526	`GO:1902600` proton transmembrane transport	IEA	6	Assembly factor, not itself a proton transporter
human	SURF1	Q15526	`GO:0004129` cytochrome-c oxidase activity	IEA	6	Assembly factor for CIV, does not have CcO activity itself
yeast	ATP10	P18496	`GO:0051082` unfolded protein binding	IPI	6	Assembly factor for ATP synthase, not a chaperone

BGC project: catalytic/cofactor terms on non-catalytic subunits & pseudoenzymes

NOT candidates from the biosynthetic-gene-cluster reviews (BGC.md). EryCII is a heme-less
cytochrome-P450 homologue; PqsB and the act chain-length factor are condensing-enzyme folds
with no active site (the catalytic residues are in their partner subunit).

Species	Gene	UniProt	Incorrect Term	Evidence	Score	Pattern
SACEN	eryCII	A4F7P2	`GO:0020037` heme binding	IEA	9	P450 homologue lacks the conserved heme-ligating Cys; apo structure (PDB 2YJN); UniProt "lacks the heme-binding sites"
SACEN	eryCII	A4F7P2	`GO:0004497` monooxygenase activity	IEA	9	No heme → no P450 monooxygenase chemistry; functions as a GT activator
SACEN	eryCII	A4F7P2	`GO:0005506` iron ion binding	IEA	8	No heme iron; depends on the absent heme cofactor
PSEAE	pqsB	Q9I4X2	`GO:0016746` acyltransferase activity	IEA	7	FabH/KAS-III fold but no active site (Cys-129/His-269 are in PqsC); contributes_to only
STRCO	actI-ORF2	Q02062	`GO:0016746` acyltransferase activity	IEA	7	Type II PKS chain-length factor; "does not have an active site"

Tier 2: Strong NOT Candidates (score 4-5)

The March 2026 scan returned 93 score-4 and score-5 rows. The selected highlights
below have been pruned against the current reviews; rows since accepted, moved to
UNDECIDED, or already represented by a curated NOT row are no longer listed.

Pseudo-enzymes and Wrong Catalytic Activity

Species	Gene	UniProt	Incorrect Term	Evidence	Score
ACET2	cipA	Q06851	`GO:0004553` hydrolase activity, hydrolyzing O-glycosyl compounds	IEA	5
ANOGA	PGRPLD	A7UTR2	`GO:0008745` N-acetylmuramoyl-L-alanine amidase activity	IEA	5
BACSU	fliY	P24073	`GO:0003774` cytoskeletal motor activity	IEA	5
DESVH	DVU_3336	Q725T9	`GO:0004673` protein histidine kinase activity	IEA	5
DESVH	DVU_3336	Q725T9	`GO:0016301` kinase activity	IEA	5
DESVH	rpoN	Q72BK7	`GO:0016779` nucleotidyltransferase activity	IEA	5
DESVH	rpoN	Q72BK7	`GO:0016740` transferase activity	IEA	5
DROME	CG6051	Q9VB70	`GO:0004721` phosphoprotein phosphatase activity	IEA	5
DROME	CG6051	Q9VB70	`GO:0046856` phosphatidylinositol dephosphorylation	IEA	5
DROME	Ccs	A1Z850	`GO:0016491` oxidoreductase activity	IEA	5
SALTY	slrP	Q8ZQQ2	`GO:0051082` unfolded protein binding	IPI	5
human	ADM2	Q7Z4H4	`GO:0006468` protein phosphorylation	IDA	5
human	AIP	O00170	`GO:0003755` peptidyl-prolyl cis-trans isomerase activity	IEA	5
human	CHRAC1	Q9NRG0	`GO:0016740` transferase activity	IEA	5
human	CHRAC1	Q9NRG0	`GO:0003887` DNA-directed DNA polymerase activity	NAS	5
human	CTBP1	Q13363	`GO:0006468` protein phosphorylation	TAS	5
human	DNAJA2	O60884	`GO:0005524` ATP binding	IEA	5
human	EGFR	P00533	`GO:0004709` MAP kinase kinase kinase activity	NAS	5
human	IDH3B	O43837	`GO:0016616` oxidoreductase activity, acting on CH-OH donors	IEA	5
human	LIPE	Q05469	`GO:0006468` protein phosphorylation	TAS	5
human	MORC3	Q14149	`GO:0018105` peptidyl-serine phosphorylation	IDA	5
human	PICK1	Q9NRD5	`GO:0006468` protein phosphorylation	ISS	5
human	RARA	P10276	`GO:0006468` protein phosphorylation	IMP	5
human	RUNX3	Q13761	`GO:0006468` protein phosphorylation	IDA	5
human	SLC3A2	P08195	`GO:0005975` carbohydrate metabolic process	IEA	5
mouse	Dnaja3	Q99M87	`GO:0005524` ATP binding	IEA	5

Wrong Binding, Localization, and Process (score 4-5)

Species	Gene	UniProt	Incorrect Term	Evidence	Score
ANOGA	PGRPS1	Q7QFK2	`GO:0008270` zinc ion binding	IEA	4
ANOGA	PGRPS1	Q7QFK2	`GO:0009253` peptidoglycan catabolic process	IEA	4
CLOCL	cbpA	P38058	`GO:0030245` cellulose catabolic process	IEA	4
CLOCL	hbpA	Q9RGE7	`GO:0000272` polysaccharide catabolic process	IEA	4
DESVH	DVU_3335	Q725U0	`GO:0006355` regulation of DNA-templated transcription	IEA	4
DESVH	DVU_3336	Q725T9	`GO:0000155` phosphorelay sensor kinase activity	IEA	4
DESVH	DVU_3336	Q725T9	`GO:0016740` transferase activity	IEA	4
DESVH	DVU_3336	Q725T9	`GO:0034220` monoatomic ion transmembrane transport	IEA	4
DESVH	kdpC	Q725T8	`GO:0005524` ATP binding	IEA	4
DESVH	kdpC	Q725T8	`GO:0016787` hydrolase activity	IEA	4
DROME	Ccs	A1Z850	`GO:0019430` removal of superoxide radicals	IEA	4
DROME	Ccs	A1Z850	`GO:0016209` antioxidant activity	IEA	4
ECOLI	DnaJ	P08622	`GO:0005524` ATP binding	IEA	4
SCHPO	pmp20	O14313	`GO:0008379` thioredoxin peroxidase activity	IBA	4
human	AIP	O00170	`GO:0003755` peptidyl-prolyl cis-trans isomerase activity	IDA	4
human	APBB1	O00213	`GO:0006915` apoptotic process	IEA	4
human	BCL2	P10415	`GO:0000209` protein polyubiquitination	IDA	4
human	BECN1	Q14457	`GO:0006915` apoptotic process	IEA	4
human	CDC25B	P30305	`GO:0006468` protein phosphorylation	IDA	4
human	CDK1	P06493	`GO:0006915` apoptotic process	IEA	4
human	CDK1	P06493	`GO:0016579` protein deubiquitination	TAS	4
human	CFTR	P13569	`GO:0016853` isomerase activity	IEA	4
human	CHMP3	Q9Y3E7	`GO:0140678` molecular function inhibitor activity	EXP	4
human	CHRAC1	Q9NRG0	`GO:0071897` DNA biosynthetic process	IEA	4
human	CPT1C	Q8TCG5	`GO:0016740` transferase activity	IEA	4
human	GMFG	O60234	`GO:0006468` protein phosphorylation	TAS	4
human	GRPEL1	Q9HAV7	`GO:0051082` unfolded protein binding	IBA	4
human	HSPB6	O14558	`GO:0005212` structural constituent of eye lens	IEA	4
human	HSPG2	P98160	`GO:0005509` calcium ion binding	IEA	4
human	IL7R	P16871	`GO:0003823` antigen binding	TAS	4
human	KCTD11	Q693B1	`GO:0016740` transferase activity	IEA	4
human	MORC3	Q14149	`GO:0006468` protein phosphorylation	IDA	4
human	PDGFA	P04085	`GO:0038083` peptidyl-tyrosine autophosphorylation	NAS	4
human	PEX14	O75381	`GO:0034614` cellular response to reactive oxygen species	IDA	4
human	PPP3CB	P16298	`GO:0006468` protein phosphorylation	ISS	4
human	SCG5	P05408	`GO:0005634` nucleus	IEA	4
human	SIRT1	Q96EB6	`GO:0004857` enzyme inhibitor activity	IEA	4
human	SLC14A1	Q13336	`GO:0005372` water transmembrane transporter activity	IEA	4
human	SPR	P35270	`GO:0008106` alcohol dehydrogenase (NADP+) activity	TAS	4
human	TMEM67	Q5HYA8	`GO:0051082` unfolded protein binding	IPI	4
mouse	Pld4	Q8BG07	`GO:0004630` phospholipase D activity	TAS	4

Tier 3: Moderate Candidates (score 3, selected highlights)

These need manual review but include notable patterns. Many of these are better treated as simple removals rather than formal NOT annotations — a NOT annotation requires positive evidence that the function is absent, not merely insufficient evidence that it is present.

Species	Gene	UniProt	Incorrect Term	Evidence	Pattern
ACET2	celC	A3DJ77	`GO:0008422` beta-glucosidase activity	IEA	Xylanase, not a glucosidase
ACET2	xynZ	P10478	`GO:0033905` xylan endo-1,3-beta-xylosidase activity	IDA	Wrong xylanase specificity
DROME	Ccs	A1Z850	`GO:0004784` superoxide dismutase activity	IEA	Copper chaperone for SOD1, not SOD itself
DESVH	fliA	Q726C4	`GO:0003899` DNA-directed RNA polymerase activity	IEA	Sigma factor, not the catalytic subunit
METEA	mxaI	P14775	`GO:0004022` alcohol dehydrogenase (NAD+) activity	IEA	Beta subunit, catalysis is in alpha
SCHPO	Epe1	O94603	`GO:0032452` histone demethylase activity	IBA	Core pseudo-enzyme case
SCHPO	sou1	Q9Y6Z9	`GO:0050085` mannitol 2-dehydrogenase (NADP+) activity	IEA	Different substrate specificity
human	APEX1	P27695	`GO:0033892` deoxyribonuclease (pyrimidine dimer) activity	IDA	AP endonuclease, not pyrimidine dimer nuclease
human	APEX1	P27695	`GO:0004844` uracil DNA N-glycosylase activity	TAS	Misattributed activity
human	ATF2	P15336	`GO:0018107` peptidyl-threonine phosphorylation	IDA	Is phosphorylated, not a kinase
human	ATP6V0C	P27449	`GO:0046933` proton-transporting ATP synthase activity, rotational	TAS	V-ATPase, not F-ATP synthase
human	CAPG	P40121	`GO:0051014` actin filament severing	IBA	Caps filaments, does not sever
human	CFTR	P13569	`GO:0006695` cholesterol biosynthetic process	IEA	Chloride channel, not cholesterol synthesis
human	CTLA4	P16410	`GO:0050853` B cell receptor signaling pathway	IBA	T cell inhibitory receptor
human	DCN	P07585	`GO:0003723` RNA binding	HDA	HTP artifact; decorin is extracellular
human	IL7R	P16871	`GO:0003823` antigen binding	TAS	Cytokine receptor, not antigen-binding

Emerging Patterns

1. Pseudo-enzyme Pattern (Highest value for NOT annotations)

Proteins with enzyme-family domains that have lost catalytic activity. These are the strongest NOT candidates because:

Automated pipelines will repeatedly re-annotate them
The negative evidence is biochemically definitive
NOT annotations prevent IBA propagation

Examples: Epe1 (JmjC pseudo-demethylase), PLD5 (pseudo-phospholipase D), AKTIP (pseudo-E2), AIP (pseudo-PPIase), CG6051 (pseudo-phosphatase), CPT1C (pseudo-transferase), EryCII (heme-less P450)

2. Domain =/= Function Pattern

Domains that serve structural/binding roles but are annotated with the catalytic activity of the domain family.

Dockerin domains annotated as hydrolases (cipB, cipA, celX)
Sensor domains annotated as kinases (DVU_3336)
DnaJ domains annotated as ATP-binding (YDJ1, DNAJA2, DNAJA4, Dnaja3, DnaJ)
PGRP domains annotated as amidases (PGRPS1, PGRPLD)

3. "Is Phosphorylated" =/= "Does Phosphorylation" Pattern (Over-annotation — removals, not NOTs)

A systematic class of over-annotation where proteins annotated to GO:0006468 (protein phosphorylation) are actually kinase SUBSTRATES, and the cited paper only demonstrates their phosphorylation BY kinases, not that they are kinases. This is especially common with TAS and IDA evidence codes.

Important caveat: Being a phosphorylation substrate does NOT definitively mean the protein lacks kinase activity — autophosphorylation exists, and some proteins are both substrates and kinases. These are therefore primarily over-annotation candidates (the evidence cited doesn't support the annotation), not necessarily true NOT candidates (where there is positive evidence the activity is absent).

However, none of these 19 genes have a kinase domain in InterPro, which strengthens the case. The nuance is:

GO:0006468 (protein phosphorylation) is a process term — a gene product can be "involved in" phosphorylation via regulation without being a kinase. So a NOT for process involvement is a higher bar than a NOT for kinase MF.
For a formal NOT on the MF (e.g. NOT GO:0004672 protein kinase activity), the absence of a kinase domain IS strong structural evidence and could justify NOTs for some of these.
For the BP annotation GO:0006468, these are better treated as evidence-insufficient removals unless the gene has no plausible regulatory role in phosphorylation either.

Examples: CREB1, MEX3B, ATF2, RARA, RUNX3, ILF3, GMFB, ADM2, PICK1, PRRT1, CTBP1, LIPE, CDC25B, PPP3CB, GMFG, MORC3, PDGFA

This is the single largest category by count. Most should be simple removals, not formal NOTs. The exception: proteins with no kinase domain AND no plausible regulatory role in phosphorylation could warrant NOT annotations specifically on the MF term GO:0004672 (protein kinase activity), where structural absence of the kinase domain is positive evidence.

4. Assembly Factor =/= Complex Activity Pattern

Proteins required for assembly of multi-subunit complexes but annotated with the activity of the complex itself.

SURF1: Assembly factor for Complex IV, annotated with cytochrome-c oxidase activity and proton transport
ATP10: Assembly factor for ATP synthase, annotated with unfolded protein binding
IDH3B: Regulatory subunit annotated with catalytic activity of the alpha subunit
Ccs: Copper chaperone for SOD1, annotated with SOD activity

5. Upstream Regulator =/= Direct Activity

Genes annotated with the activity of their downstream targets.

EGFR annotated as MAP3K (activates RAF, the actual MAP3K)
RASA1 annotated with GTPase activity (stimulates Ras GTPase)

6. Non-catalytic Family Members

Family members that have lost the signature catalytic activity.

PGRPS1, PGRPLD: PGRP family, lost amidase activity, retain binding
Epe1: JmjC family, lost demethylase activity
CG6051: Myotubularin family, lost phosphatase activity

NOT_NOTs: existing NOT annotations that do not hold up

The rest of this project mines REMOVE decisions for NOT annotations that
should exist. This section goes the other way: existing NOT annotations
that reviewers overturned. One test settles both directions. A NOT is worth
having only when the negative result rules out the function: the activity is
absent, or the process or location is excluded. A failure to detect something
is not enough.

Scan (2026-10-08). scripts/not_nots.py
writes not_nots.tsv, one row per negated annotation
in any review.

Review action on the NOT	Rows
ACCEPT	234
KEEP_AS_NON_CORE	25
UNDECIDED	25
REMOVE	8
MARK_AS_OVER_ANNOTATED	3
Total	295 (187 reviews)

Most NOTs hold up. The 11 overturned ones are below, grouped by why the
negative result failed to support a NOT. Each was checked against the
reviewer's reason.

Worked example: ARL3 NOT located in cilium

ARL3 carried NOT|located_in cilium (IDA, PMID:17646400). The full text
shows the basis was a screen of EGFP-tagged Rab and Arl GTPases
overexpressed in serum-starved RPE1 cells. Arl family members other than
ARL13B "were also absent from primary cilia (Fig. S2 A)". That shows only
that tagged ARL3 is not selectively enriched in cilia under one condition.
It says nothing about where ARL3 acts, and ARL3 acts inside the cilium:
ARL13B, its GEF, is confined to the ciliary membrane, so ARL3-GTP is made
there and releases lipidated cargo from PDE6D and UNC119 (PMID:26551564).
Endogenous ARL3 also stains cilia. The NOT was removed; see
genes/human/ARL3/ARL3-ai-review.yaml.

Failure types among the 11 overturned NOTs

Type	Gene: NOT term (evidence, reference)	Why the NOT fails
Non-detection of an overexpressed tagged protein	human ARL3: cilium (IDA, PMID:17646400)	GFP-ARL3 was not enriched in cilia. Native ARL3 is in cilia and functions there.
Absence from a fractionation proteome	ARATH OST1: cytosol (RCA, PMID:21166475); ARATH AT5G02500 (HSC70-1): cytosol (RCA, PMID:21166475)	Not recovered in one cytosolic-proteome dataset. Direct IDA shows both are cytosolic. Missing from a proteome is not exclusion from a compartment.
One allele or condition generalized to the gene	human DCTN1: axonal transport (IMP, PMID:18364389); human AGR2: response to ER stress (IMP, PMID:25666625)	G59S p150Glued mice showed no bulk transport defect, but dynactin is required for dynein-driven axonal transport. AGR2 was not induced in one context, but later work shows it alleviates ER stress.
Early negative assay superseded	human AGO3: miRNA-mediated silencing by mRNA destabilization (IDA, PMID:15260970)	The 2004 assay detected no AGO3 slicing. PMID:29040713 later showed guide- and target-dependent AGO3 slicer activity.
NOT on a parent of an asserted child	ARATH RGA: regulation of developmental vegetative growth (IMP, PMID:11606552)	RGA also carries the positive annotation negative regulation of vegetative growth. NOT on the unsigned parent logically contradicts the child. This is a signed/unsigned curation slip, and could be caught automatically.
Probable identity confusion	human TOMM40: mitochondrion (IMP, PMID:11745481)	The paper studies TOMM40 under the name p38.5/Haymaker. TOMM40 is the core outer-membrane translocase pore.
Negative is real but uninformative (marked over-annotated, not removed)	human GBA1: mitochondrion organization, neuron projection development (IMP, PMID:25456120); human CDK5: microtubule binding (TAS, PMID:17491008)	GBA1 iPSC neurons had normal mitochondrial morphology. True, but these are not processes GBA1 would be expected to act in, so the NOT prevents no plausible error. For CDK5, microtubule binding belongs to its activator p35.

Lessons for making and trusting NOTs

Localization NOTs need positive evidence of exclusion. Examples are a
validated antibody against the native protein, or a compartment marker plus
a functional test. A tagged construct or a proteome list that misses the
protein is not enough.
Process NOTs from loss-of-function need the whole gene tested. One
missense allele, one cell line or one condition can show the allele is
dispensable. It cannot show the gene is uninvolved.
Check NOTs against the ontology. A NOT on a term whose descendant is
asserted positively for the same gene is a contradiction, and a query
could find these systematically (the RGA case).
Old negative assays age badly. AGO3's 2004 negative was overturned by
later work. The same 2004 paper (PMID:15260970) supplies NOTs for AGO1
and AGO4 that are still UNDECIDED (below), so they should be reviewed
together with AGO3.

Undecided NOTs (follow-up list)

25 negated rows are UNDECIDED, usually because the reviewer could not see
the full text. They cluster:

AGO1, AGO4 (×4): NOT siRNA/miRNA silencing by mRNA destabilization,
PMID:15260970. These come from the same paper as the overturned AGO3 NOT.
AGO1 and AGO4 are generally considered slicer-deficient, so these NOTs may
be right where AGO3's was not.
SIRT1, SIRT5 (×5): NOT ADP-ribosyltransferase activities (TAS,
PMID:17456799) and SIRT5 NOT protein deacetylation (IDA, PMID:22076378).
These reflect the debated sirtuin activities.
BMPR2, BMPR1A (×5): NOT cardiac and neural-crest developmental
processes, ISS (GO_REF:0000024). These are inferred negatives, which are
unusual.
ANO5, ANO10: NOT calcium-gated chloride channel activity (IDA).
Singletons: Epe1 heterochromatin formation; A3GALT2
alpha-1,3-galactosyltransferase; ABCB4 ceramide floppase and ceramide
translocation; RAB1B; TMEM65; VAPA; YTHDF3; rat Slc5a1 (ISO).

Relationship to Other Projects

CONTESTED_FUNCTION.md: Overlaps with pseudo-enzyme cases (Epe1)
OVER_ANNOTATION_PATTERNS.md: Overlaps with domain-based prediction errors
PAINT.md: NOT annotations would prevent IBA propagation of these errors
PHOSPHORYLATION_REFACTOR.md: The "is phosphorylated" pattern overlaps with phosphorylation curation

Next Steps

Tracked in ai-gene-review#4019.

Todo: prioritize Tier 1 candidates for formal NOT annotation submission
Todo: for each Tier 1 candidate, verify literature reference supporting the negative claim
Todo: add negated: true to confirmed candidates in review YAML files
Todo: identify which NOT annotations would have the highest impact on preventing IBA propagation
Todo: develop a systematic screen for additional pseudo-enzymes across all reviewed genes
Todo: consider building automated detection of pseudo-enzyme motifs (degenerate active sites)
Todo: file GO tracker issues for the "protein phosphorylation" misannotation class
Todo: investigate whether DnaJ ATP-binding can be fixed at the InterPro2GO mapping level
Todo: NOT_NOTs: review the AGO1/AGO4 NOTs (PMID:15260970) together with the overturned AGO3 NOT
Todo: NOT_NOTs: query for NOTs on a parent term where a descendant is asserted positively for the same gene (RGA pattern)
Todo: NOT_NOTs: work through the 25 UNDECIDED negated rows

STATUS

Completed

Done: initial scan of all reviews for NOT annotation candidates
Done: keyword-based scoring and ranking
Done: pattern categorization
Done: expanded keyword set and rescoring (250 candidates at score >= 3)
Done: NOT_NOTs scan of existing negated annotations and typing of the 11 overturned ones (2026-10-08)

In Progress

Todo: literature verification of Tier 1 candidates
Todo: formal NOT annotation proposals

Last updated: 2026-10-08

NOTES

2026-10-08

NOT_NOTs section added. This started from the ARL3 review in the
HUMAN_PROTEIN_ATLAS cilium project, where NOT|located_in cilium turned out
to rest on a tagged-construct screen (see the worked example). Recounted all
negated annotations: there are 295, not ~15. Of these, 11 were overturned and
25 are undecided. Recorded the overturned ones by failure type, plus the
lessons and follow-ups, as a section here rather than a separate project.
The numbers are small, and the test for a good NOT is the same in both
directions.

2026-03-06

Project Creation

Scanned 6,319 REMOVE/MARK_AS_OVER_ANNOTATED annotations across all gene reviews. Identified 22 very high-confidence (score >= 6) and 115 strong (score >= 4) candidates for formal NOT annotations across 19 species.

The most impactful pattern is pseudo-enzymes: proteins with enzyme-family domains that have demonstrably lost catalytic activity. These are particularly valuable NOT annotations because automated pipelines (InterPro2GO, IBA) will repeatedly re-annotate them without explicit NOT annotations to prevent propagation.

Key insight: The DnaJ/HSP40 ATP binding pattern (YDJ1, DNAJA2, DNAJA4, Dnaja3, DnaJ) represents a systematic class error worth addressing at the pipeline level - DnaJ proteins stimulate HSP70 ATPase but do not themselves bind ATP.

Expanded Search

Added keywords for "misannotation", "does not catalyze", "no intrinsic", "erroneous", "does not bind", "not itself", "false positive". This expanded the candidate pool from 53 to 250 at score >= 3.

Major new finding: the "is phosphorylated" =/= "does phosphorylation" pattern is the single largest category by count, with ~19 human genes incorrectly annotated to GO:0006468 protein phosphorylation because a paper described their phosphorylation by kinases. This is a systematic issue with TAS/IDA evidence codes.

Also identified the assembly factor pattern (SURF1, ATP10, IDH3B, Ccs) where proteins required for complex assembly get annotated with the complex's activity.