Taxon-absent-component detector (genome-content satisfiability for IBA QC)

Taxon-absent-component detector

A small, runnable prototype that turns one recurring IBA over-annotation pattern
into an automatable screen. It is the signaling-domain, genome-content
special case of the Pathway satisfiability
engine (which resolves metabolic modules against an expression oracle).

parent project ← Pathway satisfiability ·
origin ← IBA Annotation Quality

The pattern

Some process/pathway GO terms entail an obligate component. GO:0007259
"signaling via JAK-STAT" entails a Janus kinase. Read as a boolean formula
over required components under a genome-content oracle, the annotation is
unsatisfiable in a genome that encodes no such component — so an IBA that
transferred it (from metazoan seeds, say) is a candidate
LINEAGE_OR_TAXON_MISMATCH over-propagation.

This came out of the Dictyostelium development review: Dd-STATa/c were
annotated JAK-STAT, but Dictyostelium has no JAK (the STATs are activated
by the TKL kinases Pyk2/Pyk3), so the term was rescoped to the JAK-independent
parent GO:0097696.

Two oracles — and why PANTHER is primary

The naïve version of this check asks whether the target genome has an InterPro
domain signature for the component. That has a serious failure mode:
signature-absence ≠ genome-absence for divergent lineages — a real ortholog
that has diverged past its metazoan domain signature scores zero.

The fix is the PANTHER family (PTHR…) membership oracle. IBA is itself
propagated along the PANTHER tree, so the very orthologs an IBA touches are, by
construction, placed in the PANTHER family — even when their domains have
diverged. detect.py therefore queries both and lets PANTHER decide:

Status Rule
COMPONENT_PRESENT PANTHER family has a target member → satisfiable; not flagged
ABSENT PANTHER target 0, control >0 → candidate unsatisfiable (confidence HIGH when InterPro agrees, MEDIUM otherwise)
INCONCLUSIVE PANTHER control 0 (bad family id) or an API error
divergence flag present by PANTHER but InterPro domain screen scores 0 — a domain-only screen would have wrongly called it absent

Counts are fetched live from UniProt REST and never hard-coded; a failed
fetch yields INCONCLUSIVE, not a made-up number.

Live results (this is the whole point)

Current RESULTS.md, D. discoideum (44689) vs human (9606):

Case Component PANTHER ctrl/tgt InterPro ctrl/tgt Verdict
jak_stat JAK 56 / 0 56 / 0 ABSENT (HIGH) — true positive; both oracles agree
stat_control STAT 128 / 4 100 / 0 PRESENT + ⚠divergence flag
gpcr_purinergic P2Y (GPCR) 21 / 0 3 / 0 ABSENT (HIGH) — P2Y GPCR truly absent
p2x_control P2X 47 / 5 26 / 0 PRESENT + ⚠divergence flag

The two divergence controls are the payoff: an InterPro-only screen calls
STAT and P2X absent in Dictyostelium (both score 0) — but they are present
(the 4 Dd-STATs; ~5 divergent P2X receptors, matching Fountain et al. 2007,
Nature 448:200). PANTHER recovers both, so the dual-oracle detector does
not flag them, while still correctly flagging the genuine JAK and P2Y-GPCR
absences. The PTHR node is the divergence-robust handle the domain screen
lacked.

Run it

cd projects/PATHWAY_SATISFIABILITY/taxon_absent_component
uv run python detect.py --write RESULTS.md          # DICDI vs human
uv run python detect.py --target 44689 --control 9606

Still a triage screen, not an oracle

PANTHER sharply cuts the false-positive rate but does not eliminate uncertainty:
a genome could carry a member PANTHER has not classified, and family membership
alone does not prove the specific sub-activity is retained. So an ABSENT
verdict — even at HIGH confidence — is a strong lead for review, not an
automatic REMOVE. The deciding evidence stays in corroboration (for JAK: the
documented JAK-independence of Dictyostelium STAT activation).

Possible extensions