NCBIFAM → GO Annotation Gain

Parent project: NCBIFam.md

How many GO annotations would be gained if NCBIFAM's own NCBI-curated
go_terms (from hmm_PGAP.tsv) were ingested as an ncbifam2go mapping and
propagated to every UniProtKB entry that already carries the family signature?
This is the NCBIFAM analog of RHEA-ANNOTATION-GAIN.md,
computed live by ncbifam_go_gain.py.

Method

For each NCBIFAM model with a GO term:

gain = N(xref:ncbifam-<id>) − N(xref:ncbifam-<id> AND go:<id>)

against the UniProtKB REST API (x-total-results, no records fetched). UniProt's
go: query is closure-aware (matches the term and its descendants), so a
family whose GO term is already covered by a more specific child counts as no
gain — the gain is closure-filtered, not naive exact-match. Gain is reported for
Swiss-Prot (reviewed) — the curation-relevant, trustworthy number — and for
all of UniProtKB (mostly automated TrEMBL) as context.

Headline: the gain is real but lives in TrEMBL, not Swiss-Prot

Two deterministic samples of equivalog models with a single GO term:

The curation-relevant (reviewed) gain is small; the all-UniProtKB gain is
large. This is the opposite emphasis from RHEA (which gained ~42 reviewed
annotations). The reason is structural: NCBIFAM is prokaryote-heavy, and
Swiss-Prot's curated prokaryotic enzymes already carry the GO term (often added by
hand or via the InterPro route), whereas the millions of unreviewed TrEMBL
entries carrying the same signature do not. So ingesting ncbifam2go would mostly
improve automated annotation depth, with a modest but non-zero curated-entry gain
concentrated in newer biology.

⚠️ The all-UniProtKB numbers inherit the reliability of automated annotation and
should not be extrapolated naively to "true gaps" — they are an upper bound on
propagation volume, not a count of curation-ready additions.

Where the gap concentrates — mobile elements, defense, secretion, division

Top propagation gaps in the 60-model equivalog sample (gain_all = entries
carrying the signature that lack the GO term or any descendant):

The pattern is consistent: transposases / IS elements, conjugation and type
III/IV/VI secretion, CRISPR and anti-phage defense, sporulation, encapsulins, cell
division — the mobile-genetic-element and microbial-defense biology where
InterPro integration of the NCBIFAM signature lags, so NCBIFAM's own GO assignment
never propagates. Several even show a reviewed gap (FtsX 5, VirB5 4), i.e.
Swiss-Prot entries that are missing the term too — the cleanest gap-fill targets.

The verification caveat is mandatory

NCBIFAM's go_terms are NCBI-assigned and not automatically current or correct.
Three distinct failure modes turned up while building the 28-row seed, each of which
a naive bulk ingest would have propagated:

• Obsolete — GO:0009448 on a GABA transaminase is an obsolete id.
• Outright wrong — the diacylglycerol-kinase model NF009874 (EC 2.7.1.107)
  is tagged GO:0003951 NAD+ kinase activity; the correct term GO:0004143 ATP-dependent diacylglycerol kinase activity exists. A wrong-MF assignment, not
  just an altitude issue — excluded.
• Altitude-broad — several sit far above the family's real function, up to the
  ontology near-root GO:0003824 catalytic activity (enoyl-CoA hydratase
  NF005804, spermidine synthase TIGR00417), or a class parent (dGTPase NF002326 →
  GO:0016793 when GO:0008832 dGTPase activity exists). For these we propose our
  own specific term as the exactMatch mapping (the EC-bridged child) instead of
  recording NCBI's broad value.

Proposing the specific term unmasks the gain

The altitude fix is not cosmetic. Because UniProt's go: query is closure-aware,
the gain measured against a broad parent is near-zero (the parent is already
near-universal), so the broad NCBI term makes a real gap look like no gap.
Re-measuring against the specific child we propose flips this:

So suggesting our own term is what converts these from invisible to actionable — and
surfaces genuine reviewed/Swiss-Prot gaps (dGTPase 13, LL-DAP 2) that the broad
NCBI assignment hid entirely.

…but high gain can also be an over-annotation trap (FtsX)

The same machinery warns against over-reach. For TIGR00439 (cell division protein
FtsX) the candidate terms give:

A naive "maximise gain" rule would pick GO:0043093 (3,304!), but the ontology says
cytokinesis is part_of cell division (so the specific terms are narrower, not
parents), and curators apply them to only 2/7 FtsX — FtsX/FtsEX regulates septal
hydrolysis rather than performing constriction. Propagating the specific term to all
3,306 entries would assert more than the evidence supports. The seed therefore maps
FtsX to the consensus GO:0051301 (gain 22, confirmatory) and declines the
high-gain specific term. High gain is a flag for review, not an instruction to
propagate — it can mean a real masked gap (the five rows above) or an
over-annotation waiting to happen (FtsX).

So an ncbifam2go build must, like the curated seed in
ncbifam2go.sssom.yaml:

1. drop obsolete and incorrect GO ids (check against EC where available),
2. propose the specific child where the NCBI term is a broad parent, and measure
   gain against that term, not the parent,
3. EC-bridge-confirm enzyme rows via ec2go.

Reproduce

uv run python ncbifam_go_gain.py --family-type equivalog --single-go --sample 60
uv run python ncbifam_go_gain.py --accessions NF033545.0,TIGR02791.1,TIGR00439.2