Hard cases — pseudoenzymes & disputed-function human genes (batch 4)

Hard cases — pseudoenzymes & disputed-function human genes (batch 4)

The earlier cohorts used mostly well-behaved genes. This cohort (batch4-genes.txt;
results) deliberately picks human genes flagged as
hard-to-curate in our own project docs (HUMAN_GENES_RE_REVIEW.md,
UNIPROT_CAUTION_NOTE/, PSEUDOENZYMES.md, TOP_NOTS.md, TRANSCRIPTION_FACTORS/):
pseudoenzymes, literature-contested activities, "was-X-actually-Y" reclassifications,
and family-label misdirection. It asks two questions the easy cohorts couldn't:
(a) does the GO layer assign the ancestral/dead/disputed activity? and (b) does
the narrative acknowledge the pseudoenzyme status or the controversy?

Exact GO capture stays low (1/12). The one hit is GAPDH RNA binding, a secondary core
term, so specific primary-function capture is 0/12, as on the project page. Because Affinage
only emits goslim_generic bins, the informative scores are slim-level. The core-MF bin is
emitted for 10/12. The top-supported MF is a core bin for only 6/11 (CASP12 has no MF), the
lowest of any cohort. See batch4/summary.md. But the two-layer behaviour is
the real result.

The narrative is genuinely pseudoenzyme- and reclassification-aware

Because Affinage reasons from the literature (not from domain/family labels), its
narrative repeatedly gets the hard call right — verbatim:

Gene Curation challenge Narrative verdict (verbatim)
ILK pseudokinase (scaffold) "Despite its kinase-like domain, ILK is a bona fide pseudokinase: recombinant ILK has no detectable activity toward GSK-3beta"
ROR1 pseudokinase "Although it adopts a kinase fold, ROR1 is a pseudokinase devoid of intrinsic catalytic activity and is instead transphosphorylated by partner kinases"
CPT1C pseudo-transferase "it retains weak carnitine palmitoyltransferase activity… catalytic efficiency 20–300 times lower than CPT1A, and it localizes to the ER"
HDAC6 not a histone deacetylase "cytoplasmic class IIb deacetylase… deacetylating a broad set of non-histone substrates… alpha-tubulin"
PARK7 glyoxalase vs deglycase controversy "A reported nucleotide/protein deglycase activity is reframed by rigorous kinetics as glyoxalase-like with only a minor role in neuronal methylglyoxal defense"

On KEAP1, InterPro2GO→BioReason assigned actin binding (from the BTB-Kelch fold). Affinage's
literature grounding avoids that error. Its terms by support are molecular sensor activity
(4), catalytic activity, acting on a protein (3), ligase activity (2), molecular function regulator activity (2) and molecular adaptor activity (2). The two bins of our curated core
(ubiquitin-like ligase-substrate adaptor → adaptor; transcription regulator inhibitor →
regulator) are present but are not the top terms. The acting on a protein and ligase bins
are wrong for a non-catalytic substrate adaptor. So KEAP1 is "not actin", not a clean win.

…but the GO layer is a lossy down-cast that can contradict its own narrative

The mechanism_profile does not inherit the narrative's nuance. The sharpest case:

An earlier version of this table picked the favourable term for each gene rather than the
top-supported one, and so overstated how often the trap is avoided. Read by support count
(from batch4/per-gene.json):

Gene Trap (ancestral/wrong activity) Affinage MF terms, most-supported first (support) Top term avoids trap?
ILK protein kinase activity molecular adaptor (4); cytoskeletal protein binding, ATP-dependent (2); regulator (1) ✅
RASA1 GTPase activity catalytic activity, acting on a protein (5); regulator, transducer (3); lipid binding, adaptor (2) ❌ (the correct GAP bin, regulator, is second)
PLD3 phospholipase D activity catalytic activity acting on RNA, hydrolase (5); acting on DNA, DNA binding (3); RNA binding (2); transferase (1) partial (hydrolase/nuclease zone; DNA exonuclease bin is third)
KEAP1 actin binding molecular sensor (4); acting on a protein (3); ligase, regulator, adaptor (2) partial (no actin, but catalytic/ligase bins for a non-enzyme)
CASP12 cysteine endopeptidase activity (empty — no MF grounded) ✅ by omission
ROR1 (pseudo)kinase catalysis transducer (3); acting on a protein, adaptor, virus receptor (2) ✅ at top, ❌ in the list
CPT1C carnitine transferase molecular sensor (2); transferase, acting on a protein, regulator (1) ✅ at top, ❌ in the list

So literature grounding keeps the pseudo-activity off the top of the list more often than
not (ILK, ROR1, CPT1C, CASP12). It rarely keeps it off the list entirely: only ILK and CASP12
emit no catalytic bin at all.

Controversy handling: mixed

So the "current model" format can represent a controversy when the literature has a
clear resolution (PARK7), but tends to flatten genuinely unresolved ones (UCHL1) —
it has no slot for "annotated both positively and NOT," which is exactly how GOA/our
reviews encode contested functions.

Bottom line

On the hardest human genes, Affinage's narrative is impressively literature-faithful
— it names pseudokinases as catalytically dead, reclassifies HDAC6, and adjudicates
PARK7 — and beats domain-based tools on family-label misdirection (KEAP1). Its GO
layer, by contrast, is coarse. Exact capture is 1/12, and that one hit is a secondary term,
so specific capture is 0/12. At slim level the top-supported bin is right for 6/11. It often
lists the dead ancestral activity (ROR1, CPT1C), and in ROR1's case it flatly contradicts the
very narrative it is derived from. The gap between the two layers is widest exactly where
curation is hardest — reinforcing the project's core result: judge Affinage by its
narrative, not its GO grounding, and even the narrative cannot encode a live dispute.