UniProt CAUTION Note Project

Overview

UniProtKB records can carry one or more free-text CAUTION comments in the
flat file:

CC   -!- CAUTION: In contrast to other JHDM1 histone demethylases, it lacks the
CC       iron catalytic His in position 370 which is replaced by a Tyr residue
CC       and has no histone demethylase activity in vitro (PubMed:16362057). It
CC       therefore may not be functional in vivo. {ECO:0000305|PubMed:16362057}.

A CAUTION comment is a curator's explicit warning to the reader: a function is
contested, an activity was mis-attributed, a supporting paper was
retracted, a domain is degenerate (pseudo-enzyme), or a feature is a
likely artifact. These are exactly the situations where automated GO
annotation (IEA/IBA, domain-to-function mappings) is most likely to be wrong,
and where AI-assisted review adds the most value.

This project systematically harvests CAUTION comments, categorizes them, and
uses them as a prioritized worklist for annotation review — a CAUTION note is
a strong signal that one or more existing GO annotations on that gene deserve
scrutiny.

Yes, this is a real UniProt feature. CAUTION is a standard UniProtKB
"General annotation (Comments)" topic
(https://www.uniprot.org/help/caution), distinct from the structured
SEQUENCE CAUTION block. We extract CC -!- CAUTION: free-text comments
and exclude the separate structured CC -!- SEQUENCE CAUTION: block
(erroneous initiation / frameshift / predicted-gene-model issues), which is
about sequence construction rather than function.

Headline numbers

Database-wide (Swiss-Prot / reviewed, 2026-06-16):

• 14,513 reviewed entries carry a CAUTION comment (vs 45,468 with the
  unrelated SEQUENCE CAUTION block — confirming cc_caution is the right field).
• 14,830 individual CAUTION notes; ~7,200 fall in high-signal categories.

Distribution by organism (top): Human (2,298 entries), S. cerevisiae (895),
Mouse (844), A. thaliana (714), E. coli K12 (334), Rat (329), Bovine (206),
Dictyostelium (192), Drosophila (160), Rice (150), C. elegans (122),
Zebrafish (109).

Local corpus (genes already fetched in this repo) — 209 CAUTION notes
across 201 cached *-uniprot.txt records:

The WGS-preliminary boilerplate that dominated the local corpus essentially
vanishes in reviewed entries (0–6), confirming it was a TrEMBL/unreviewed
artifact rather than a curatorial signal.

Why this matters for GO review

The high-value categories map directly onto well-known GO over-annotation
failure modes:

• degenerate-domain / pseudo-enzyme → domain-based IEA/IBA assigns a
  catalytic MF the protein cannot perform. Overlaps strongly with the
  Contested Function project and
  Pseudoenzymes project.
• reclassified-function → the gene was historically placed in the wrong
  complex/pathway; legacy GO terms may persist (e.g. NDUFA4 "was initially
  believed to be a subunit of complex I").
• retracted-reference → annotations whose original_reference_id points to
  a retracted paper should be flagged is_invalid and the function re-examined.
• contested-function / possible-artifact → candidates for the schema's
  finding_review (DISPUTED) and reference_review machinery, or
  KEEP_AS_NON_CORE / MARK_AS_OVER_ANNOTATED actions.

Cross-referencing the DB-wide survey against the accessions we have already
fetched (148 of our local genes overlap the reviewed-CAUTION set) yields
4,046 high-value, not-yet-reviewed candidates: reclassified-function (2,401),
degenerate-domain (1,342), retracted-reference (255), possible-artifact (48).
Human alone contributes 95 pseudo-enzyme (degenerate-domain) and 111
retracted-reference candidates, e.g.:

• RHBDF1 (Q96CC6) — "Lacks serine protease activity ... lacks the catalytic
  Ser residue at position 720" (iRhom pseudo-protease).
• SUMF2 (Q8NBJ7) — "strongly similar to formylglycine-generating enzyme,
  lacks the catalytic Cys".
• PANK4 (Q9NVE7) — pantothenate kinase domain is degenerate.
• CHI3L1 (P36222), SPACA3 (Q8IXA5) — glycosyl hydrolase folds with
  substituted catalytic residues.
• ERCC8 (Q13216), H3-3A/B (P84243), CBX1/CBX5 — annotations resting
  on retracted papers.

Findings

Deep dive batch 1 (degenerate-domain)

Fetched 5 human degenerate-domain candidates (RHBDF1, SUMF2, PANK4, DPYSL5,
NAALADL2) and checked their curated GO molecular functions against the CAUTION
note:

• 4 of 5 are already correctly handled by GO, usually via an explicit
  NOT|enables on the lost catalytic term — frequently citing the same paper
  named in the CAUTION (RHBDF1↔PMID:21439629, PANK4↔PMID:30927326). The GO NOT
  qualifier is effectively the curated counterpart of the UniProt CAUTION.
• One suspect over-annotation — DPYSL5 (CRMP5): GO:0004157
  dihydropyrimidinase is correctly NOT-ed, while two positive InterPro IEA
  parent terms (GO:0016787, GO:0016810 hydrolase activity) survive. This is
  not a logical contradiction (GO NOT propagates down, not up — so a
  negated child does not negate the parent), but it is an evidentially weak
  fold/domain-based propagation: the metal-site loss that justifies the NOT
  also removes the mechanistic basis for any metallo-hydrolase activity, with
  no positive evidence for an alternative → candidate MARK_AS_OVER_ANNOTATED.

Systematic over-annotation queries (local corpus, 2,675 genes)

Two CAUTION-driven detectors, run over the locally-fetched corpus (GO ancestry
via oaklib; no network).

Query A — negated-child / positive-parent conjunction. Within a gene, a
molecular-function term X is annotated positively while a more specific
descendant Y is annotated NOT. 33 hits / 22 genes (24 with an electronic
IEA/IBA parent). A triage column marks a hit STRONG when the parent's
specificity has no experimental positive support in GOA (the DPYSL5 pattern):

Query A also surfaces 9 DIRECT same-term conflicts — a term annotated both
positively and NOT on the same gene (genuine GOA inconsistencies worth
reporting upstream), e.g. EDEM1/EDEM2 mannosyl-oligosaccharide 1,2-α-mannosidase
(TAS vs NOT IDA), ENDOU serine-type peptidase (IDA vs NOT IDA), PARK7
protein deglycase (IDA vs NOT IDA), CYB5R4 NAD(P)H oxidase (IDA vs NOT IDA).

Query B — CAUTION PMID cited positively, never negated. Flags genes where a
UniProt CAUTION cites a PMID, a GO annotation is made to that same PMID, and
there is no NOT annotation citing it. 69 flags / 39 genes. Highest-value
molecular-function hits:

• CHMP1A ↔ PMID:8863740 — metallopeptidase activity + zinc ion binding
  (TAS) from a paper whose "metalloprotease (PRSM1)" reading came from a wrong
  ORF translation → strong REMOVE candidates.
• ENDOU ↔ PMID:2350438 — serine-type peptidase activity (IDA) from the paper
  later refuted (it is an endoribonuclease, not a protease).
• HDAC6 ↔ PMID:10220385 — histone deacetylase activity (IDA) from the
  "originally thought to be a histone deacetylase" paper (it is predominantly a
  cytoplasmic tubulin deacetylase).
• TCF25 ↔ PMID:16574069 — DNA-binding shown only via a GAL4-DBD fusion (CAUTION:
  "uncertain"); CSNK1D ↔ PMID:20637175 — DCK phosphorylation "probably not in
  vivo"; NME2 ↔ PMID:11121025 — contested intrinsic nuclease activity.

Many other Query B hits are localization debates (AGK, AIFM2, ASAH2, C1QBP,
ATP13A1, Dnajb11) where the positive annotation may have been deliberately
retained — these are flags for a curator's eye, not automatic removals.

Validation against existing curated reviews

Because the local corpus is already curated, the query hits double as a test of
the method — each flag is joined to the action the existing review assigned:

• Query A STRONG: 8/11 CONFIRMED — the curator had already set
  REMOVE/MARK_AS_OVER_ANNOTATED/MODIFY on the flagged parent (DPYSL5, CPT1C
  ×2, pmp20 ×2…). The 3 "kept" are correctly-retained parents on a different
  branch (ENDOU is a ribonuclease → hydrolase is right; MAP1S really does bind
  actin; PEX12 is a ubiquitin ligase). The STRONG triage tracks expert decisions.
• Query B: 17 CONFIRMED real catches (CHMP1A metalloprotease/zinc from a
  mistranslated ORF; ENDOU serine protease; HDAC6 histone-deacetylase; NME2
  nuclease; the retracted TP53–BANP interactions…) — but ~47 "kept" are
  expected false-positives (see Lessons learned).

Net: the two genes spot-checked — CPT1C and CHMP1A — were both already
correctly curated (CPT1C REMOVEs the unsupported transferase parents and keeps
catalytic activity because the real palmitoyl-hydrolase activity sits under it,
exactly matching the STRONG-vs-supported split; CHMP1A REMOVEs the mistranslated-ORF
metalloprotease/zinc terms). No edits were needed — the queries reproduced the
experts' decisions, the validation we wanted before scaling UniProt-wide.

UniProt-wide scaling (QuickGO) — net-new pseudoenzyme families

Pulling molecular-function GOA from QuickGO for all 14,513 reviewed CAUTION
accessions (MF annotations found for 12,194) and running both queries
database-wide, flagging genes not yet in this repo (net_new):

• Query A: 247 conjunctions → 105 STRONG (100 in net-new genes), plus 97
  DIRECT same-term conflicts.
• Query B: 1,140 flags (1,083 net-new).

The STRONG net-new hits land squarely on known pseudoenzyme families, recovered
automatically and extended across orthologs — strong external validation:

That the method independently rediscovers ILK, ROR1, CASP12, AZIN2 and the CRMP
family — canonical pseudoenzymes — from nothing but "CAUTION text + GO NOT +
GO ancestry" is the validation that matters. Immediate human review targets (not
yet in this repo): DPYSL2 (Q16555), DPYSL3 (Q14195), DPYSL4 (O14531), CRMP1
(Q14194), ILK (Q13418), ROR1 (Q01973), CASP12 (Q6UXS9), AZIN2.

Featured examples (from the cached corpus)

• SCHPO/Epe1 (O94603) — "lacks the iron catalytic His ... no histone
  demethylase activity in vitro ... may not be functional in vivo." Textbook
  pseudo-enzyme; already a flagship of CONTESTED_FUNCTION.
• human/NDUFA4 (O00483) — "Was initially believed to be a subunit of
  complex I" (it is a subunit of complex IV). Legacy complex-I GO terms are the
  risk.
• human/AKT1 (P31749) — two CAUTION notes flag retracted papers
  (PUBMED:20231902, PUBMED:19940129) plus an isoform-redundancy caveat.
• human/PARK7 / DJ-1 (Q99497) — glyoxalase vs deglycase activity is
  explicitly called controversial; a clean test case for contested MF.
• human/IL4I1 (Q96RQ9) — an enzymatic-independence claim is flagged as
  needing confirmation (only Phe, not Trp, deprivation was tested).
• human/UCHL1 (P09936) — disease association and ubiquitin-ligase activity
  both flagged as contested across conflicting papers.
• SACEN/eryCII (Q939Z0) — "related to the cytochrome P450 family, lacks the
  heme-binding sites"; pseudo-enzyme that is actually a glycosyltransferase
  activator (already reviewed under CONTESTED_FUNCTION).
• ARATH/CRY1, CRY2 — "Was originally thought to be a DNA photolyase"
  (cryptochromes are photoreceptors, not repair enzymes).

Lessons learned

• A flag is a prioritization signal, not a verdict. Query A STRONG means "no
  experimental backing for the parent's specificity in GOA" — but some flagged
  parents are real activities that simply lack an experimental annotation (e.g.
  pmp20 is a genuine peroxiredoxin; the peroxidase parent is only IEA-annotated).
  Adjudication stays evidential.
• GO NOT propagates down, not up. A negated child does not negate the
  parent, so a positive-parent / negated-child pair is logically admissible. The
  generalizable review flag is a positive IEA parent whose curated,
  more-specific child is NOT-ed — admissible but evidentially weak when the
  same residue loss that justifies the NOT removes the basis for the parent too.
• Genericity ≠ wrongness. The maximally generic MF term can be the correct
  family-level annotation (cf. the InterPro Radical-SAM case): the issue is
  whether the specificity is supported, not how general the term is.
• PMID-level matching over-flags (Query B). A contested CAUTION paper usually
  makes several claims and only one is doubted, so matching on PMID alone catches
  the real refutation plus collateral (e.g. CPT1C↔PMID:30135643 is the real
  palmitoyl-hydrolase paper; UFSP1 is now known active; CSNK1D's kinase activity is
  bona fide — only its DCK substrate is doubted). Precise use needs matching the
  contested activity to the GO term, not just the citation.
• The GO NOT qualifier is the curated counterpart of a UniProt CAUTION —
  when both exist they usually cite the same paper, so the highest-value targets
  are CAUTIONs with no corresponding GO NOT.

Reproducibility

All numbers above are produced by reproducible scripts under
UNIPROT_CAUTION_NOTE/; the multi-MB GOA dumps are
gitignored but regenerable. Rerun after fetching new genes.

The deep-dive write-up is in
deep_dive_batch1.md. To run UniProt-wide
the queries need GOA (evidence codes + NOT qualifiers) per accession, which the
cc_caution API survey does not carry — supplied here by the QuickGO pull.

Workflow

1. (Re)generate the worklist: run extract_caution_notes.py.
2. Prioritize the high-value categories (contested / reclassified /
   degenerate-domain / retracted / possible-artifact).
3. For each gene that already has a *-ai-review.yaml, check whether its
   existing annotations are consistent with the CAUTION note; if a retracted
   PMID is an original_reference_id, flag it.
4. For genes without a review yet, treat the CAUTION note as a strong reason to
   prioritize a full review (just fetch-gene <org> <gene>).
5. Record findings using reference_review (retracted → is_invalid),
   finding_review (DISPUTED), and appropriate annotation actions.

Related projects

• CONTESTED_FUNCTION.md
• PSEUDOENZYMES.md
• OVER_ANNOTATION_PATTERNS.md

STATUS

Done

• [x] Confirmed UniProt CAUTION comments exist and are present in the cached
  corpus (209 notes / 201 records).
• [x] Wrote reproducible local extractor extract_caution_notes.py.
• [x] Surveyed the whole reviewed UniProt via the REST API
  (uniprot_api_survey.py): 14,513 entries / 14,830 notes, with category and
  per-organism distributions — no fetch-gene required.
• [x] Built shortlist_candidates.py → 4,046 high-value, not-yet-reviewed
  candidates (candidates_high_value.tsv, candidates.md).

• [x] Deep dive batch 1 (deep_dive_batch1.md): fetched RHBDF1, SUMF2,
  PANK4, DPYSL5, NAALADL2; found 4/5 already correctly negated by GO and 1
  suspect domain-based over-annotation (DPYSL5 hydrolase IEAs — evidentially
  weak, not a logical contradiction).

• [x] DPYSL5 (CRMP5) full review written and validated
  (genes/human/DPYSL5/DPYSL5-ai-review.yaml, DRAFT): REMOVE on the two IEA
  hydrolase over-annotations (GO:0016787, GO:0016810), ACCEPT on the curated
  NOTs, core function = negative regulation of dendrite morphogenesis.

• [x] All 5 batch-1 reviews completed and validated (RHBDF1, SUMF2, PANK4,
  NAALADL2 + DPYSL5). DPYSL5 was the only one needing a REMOVE of a positively
  asserted catalytic term; PANK4 is a domain-swap pseudoenzyme (dead PanK domain

• real, correctly annotated 4'-phosphopantetheine phosphatase).

• [x] Two systematic over-annotation queries implemented
  (caution_conjunction_queries.py): Query A (negated-child/positive-parent
  conjunction, 33 hits + 9 direct conflicts, with a STRONG/experimental-support
  triage) and Query B (CAUTION-PMID-cited-positively-never-negated, 69 flags).

• [x] Validated the queries against existing reviews
  (audit_queries_vs_reviews.py): Query A STRONG 8/11 CONFIRMED; Query B 17
  CONFIRMED catches (rest are expected PMID-level false-positives). CPT1C and
  CHMP1A were already correctly curated — the queries reproduced the experts'
  REMOVE/MODIFY decisions, so no edits needed.
• [x] Scaled both queries UniProt-wide via QuickGO (uniprot_wide_queries.py):
  14,513 CAUTION accessions → Query A 247 conjunctions / 105 STRONG (100 net-new) +
  97 direct conflicts; Query B 1,140 flags. STRONG net-new hits recover known
  pseudoenzyme families (CRMP/DPYSL, ILK, ROR1, CASP12, AZIN2) across orthologs.

• [x] Reviewed the human CRMP family (DPYSL2, DPYSL3, DPYSL4, CRMP1) — full
  validated DRAFT reviews applying the DPYSL5 fix: REMOVE the electronic
  hydrolase activity parents (GO:0016787/0016810/0016812), ACCEPT the curated
  NOT|dihydropyrimidinase; DPYSL2 additionally REMOVEs a legacy positive
  dihydropyrimidinase (TAS) + its dependent nucleobase-metabolism term. Core
  function recorded as catalytically dead cytoskeletal regulator in semaphorin signaling.

• [x] Reviewed net-new pseudoenzymes ILK, ROR1, CASP12, AZIN2 — full validated
  DRAFT reviews. Each REMOVEs the electronic catalytic over-annotation flagged by
  the query (ILK protein kinase activity; ROR1 protein kinase activity + legacy
  TAS RTK/PTK-signaling; CASP12 cysteine peptidase/endopeptidase/proteolysis;
  AZIN2 catalytic activity/carboxy-lyase/legacy arginine-decarboxylase), ACCEPTs
  the curated NOTs, and records the real non-catalytic core (ILK = IPP-complex
  scaffold; ROR1 = Wnt coreceptor; CASP12 = dominant-negative inflammasome modulator;
  AZIN2 = ODC-activator/antizyme inhibitor).

Pending

• [ ] Report the 9 local + 97 UniProt-wide DIRECT same-term GOA conflicts upstream
  to GO.
• [ ] Scale Queries A/B UniProt-wide via a QuickGO GOA pull keyed on the 14k
  CAUTION accessions.
• [ ] Audit the 255 retracted-reference candidates for retracted
  original_reference_ids in existing reviews.
• [ ] Scale the "positive-IEA-catalytic + curated-NOT" conjunction query across
  the 1,342 degenerate-domain candidates to auto-surface over-annotations.
• [ ] Audit retracted-reference candidates for retracted original_reference_ids
  in any existing review.
• [ ] Refine the keyword classifier to shrink the large "other" bucket (7,605
  DB-wide notes) — many are genuine but unmatched.
• [ ] Add a just target to regenerate the survey + shortlist on demand.

Last updated: 2026-06-16

NOTES

2026-06-16

Project creation. Verified the premise before building: UniProt CAUTION is a
real, documented comment topic, and grep -c '\-!- CAUTION:' across the cached
*-uniprot.txt files returns 209 matches in 201 records. Built a parser that
correctly handles multi-line CAUTION comments and excludes the structured
SEQUENCE CAUTION block. Keyword categorizer separates curation-meaningful
notes (contested / reclassified / degenerate-domain / retracted / artifact) from
automatic boilerplate (WGS-preliminary, lacks-conserved-residue). The
degenerate-domain bucket overlaps heavily with existing CONTESTED_FUNCTION and
PSEUDOENZYMES work, confirming CAUTION notes are a good signal for finding
over-annotated catalytic functions.

API survey added (same day). Per request, queried the live UniProt REST API
on the cc_caution field (distinct from cc_sequence_caution) to get a
database-wide distribution without fetching genes: 14,513 reviewed entries,
14,830 notes, ~7,200 high-signal. Human dominates (2,298). Cross-referencing
against the 148 reviewed-CAUTION genes we already have yields a 4,046-entry
high-value worklist — the basis for a prioritized deep dive (starting with human
pseudo-enzymes flagged by degenerate-domain cautions).