Deep Dive Batch 1 — Human pseudo-enzymes flagged by UniProt CAUTION

← back to UNIPROT_CAUTION_NOTE

Five human entries from the degenerate-domain candidate list
(candidates.md) were fetched (fetch-gene human <GENE>) and
their curated GO molecular-function annotations were checked against the
UniProt CAUTION note. The question for each: does GOA still assert the
catalytic activity that the CAUTION note says the protein has lost?

Result summary

The one actionable finding: DPYSL5 (CRMP5)

DPYSL5 has the specific catalytic term correctly negated:

• GO:0004157 dihydropyrimidinase activity — NOT|enables (IBA, GO_REF:0000033)

…and still carries two positive IEA parent terms derived from
metallo-hydrolase fold signatures:

• GO:0016787 hydrolase activity — IEA, InterPro:IPR006680 (Amidohydrolase-related)
• GO:0016810 hydrolase activity, acting on carbon-nitrogen (but not peptide) bonds — IEA, InterPro:IPR011059 (Metal-dependent hydrolase composite)

This is not a logical contradiction. Under GO's true-path rule a NOT
annotation propagates downward (to subtypes of dihydropyrimidinase), not
upward — so NOT enables dihydropyrimidinase does not entail NOT enables hydrolase. A protein can consistently be "some hydrolase" while "not a
dihydropyrimidinase," so the positive-parent + negated-child pair is logically
admissible.

It is, however, an evidentially suspect domain-based over-annotation. Both
positive terms rest only on fold/superfamily signatures (IPR006680, IPR011059;
UniProt SIMILARITY: Belongs to the metallo-dependent hydrolases superfamily) —
pure "has-the-fold ⇒ has-the-activity" propagation. The CAUTION states DPYSL5
"Lacks most of the conserved residues that are essential for binding the metal
cofactor and hence for dihydropyrimidinase activity," and catalysis across the
entire metallo-dependent hydrolase superfamily depends on that metal site. So
the very fact that justifies the curated NOT (no metal-binding residues) also
undermines the evidential basis for the generic hydrolase parents: there is no
positive evidence DPYSL5 performs any hydrolysis, and its curated function is
non-catalytic (FUNCTION: negative regulation of dendrite outgrowth; a CRMP5
cytoskeletal adapter).

Proposed action: MARK_AS_OVER_ANNOTATED for GO:0016787 and GO:0016810
on DPYSL5 — domain-based IEA propagations that are unsupported and biologically
improbable given metal-site loss. This is a deliberately weaker claim than
REMOVE-as-contradiction, and weaker than asserting NOT hydrolase (which would
overclaim absence of activity without positive evidence). The CRMP5
protein-binding / cytoskeleton annotations are the real function and are
retained. Same pattern as
OVER_ANNOTATION_PATTERNS /
CONTESTED_FUNCTION.

Action taken. genes/human/DPYSL5/DPYSL5-ai-review.yaml is now a full
DRAFT review (schema- and term-validated). The two IEA hydrolase terms
(GO:0016787, GO:0016810) are set to REMOVE with the reasoning above
(electronic over-propagation refutable on biological grounds — not
MARK_AS_OVER_ANNOTATED, which is reserved for terms the gene is genuinely but
peripherally related to). The two curated NOT annotations are ACCEPT-ed as
correct, and the IMP negative regulation of dendrite morphogenesis
(PMID:33894126) is recorded as the core function.

All five reviews completed

Full, schema- and term-validated DRAFT reviews are now written for every gene in
the batch:

Only DPYSL5 required a REMOVE of a positively-asserted catalytic term; the
others were already self-consistent (curators had either negated the lost
activity or never asserted it). PANK4 is the most instructive: the CAUTION
negates one domain's activity while a different, real activity is correctly
annotated — a reminder that "pseudoenzyme" can mean domain-specific, not
whole-protein, inactivation.

Cross-resource insight (worth generalizing)

For 4 of 5 genes the GO NOT qualifier is the curated counterpart of the
UniProt CAUTION note — curators have already encoded "this family member lost
the activity" as an explicit negative annotation, frequently citing the same
primary paper named in the CAUTION (e.g. RHBDF1 ↔ PMID:21439629; PANK4 ↔
PMID:30927326). This is a strong cross-resource consistency signal and a good
sanity check on the CAUTION→over-annotation hypothesis.

The remaining leak is almost always IEA from InterPro/UniRule domain rules
(DPYSL5's two hydrolase terms), applied before/independently of the
member-specific NOT. So a useful review flag (not a contradiction detector)
is: "a positive IEA parent term whose curated, more-specific child is
NOT-ed (or whose activity a CAUTION reports as lost)." Such a pair is
logically consistent — NOT does not propagate up — but it marks a
fold/domain-based parent annotation that may have lost its evidential basis,
and is worth a curator's eye. The adjudication is then evidential, not logical:
is there any positive support for the parent activity, or does it rest solely on
the fold signature that the CAUTION/NOT has specifically discredited?

Reproducibility

for g in RHBDF1 SUMF2 PANK4 DPYSL5 NAALADL2; do
  uv run ai-gene-review fetch-gene human "$g" --output-dir .
done
# then inspect molecular_function rows in genes/human/<G>/<G>-goa.tsv

Last updated: 2026-06-16