49 of 67 components have never been examined for a grounding

[realmarcin]

Found in an open-issue review. Filed because the queue had no issue for the
largest content gap in the corpus.

grounded                      6
REVIEWED_LABEL_ONLY          12   (looked at, deliberately declined)
no grounding, no status      49   (never examined)

Of the 49 never examined, 35 carry gene_symbols and 6 carry
protein_examples, so most have something to match on.

This is not "run the adopted script"

I nearly filed this as a mechanical gap and it is not one.
scripts/interpro_groundings.py is a curator-reviewed allow-list, not a
matcher — its own docstring says it "never searches labels or chooses the first
result". Running it today touches 8 components and writes nothing:

bacterial_type_flagellum#flagellin   InterPro:IPR001492    unchanged
carboxysome#bmc_h                    REVIEWED_LABEL_ONLY   unchanged
carboxysome#bmc_p                    REVIEWED_LABEL_ONLY   unchanged
... 5 more REVIEWED_LABEL_ONLY
nothing to write

Seven of its eight entries resolved to REVIEWED_LABEL_ONLY — the adapter
looked and declined because the family consensus was ambiguous across the
component's scope. That is the adapter working correctly, and it is why only
2 of the 6 groundings are InterPro at all (the other four are CHEBI and SO).

So the gap is not un-run tooling. It is 49 components nobody has reviewed, and
closing it means extending the allow-list one component at a time with a
recorded decision — including recording "declined, and why", which is what
REVIEWED_LABEL_ONLY exists to say.

Suggested shape

Batch by record rather than corpus-wide, largest first — carboxysome (8),
flagellum (7), ATP synthase (5), injectisome (5) — so each batch is one
coherent review against one structure's scope.

bacterial_type_flagellum already has an open component_groundings
CURATION_TODO covering its 7; that discussion and this issue should not drift
apart.

Why it matters

components is where this corpus says what a structure is made of. Without
groundings those components are labels and gene symbols — not resolvable
identifiers — so nothing downstream can join them to protein data, and the
just validate-products id↔label gate has nothing to check.

Related: #7 item 3 asks for exactly this on the flagellum.

[realmarcin]

The 2026-09-02 source-discovery pass found a concrete second family source for
this review backlog: NCBIfam's current PGAP model index. Exact NCBIFAM canaries
include NF053793.1 (CcmN), NF040964.1 (MamK), and NF041155.2 (family-1
encapsulin shell). The strongest first target is
carboxysome#assembly_adaptor -> NF053793.1.

Important boundary: the bulk index mixes NCBIFAM, Pfam, TIGRFAM, and PRK. Any
adapter should filter source == NCBIFAM, start from a reviewed component/protein
example rather than label search, and decline combined component scopes. Full
evidence and official endpoints are in
research/2026-09-02-next-data-sources.md; the source is now a P1 queue
candidate.

[realmarcin]

Correcting this issue's own framing. I wrote that 35 of the 49 carry
gene_symbols "so most have something to match on". That is misleading, and
working the flagellum showed why. The 49 are not one gap — they are three, with
different authorities and a prerequisite.

Broken down by component_type:

type	count	InterPro applicable?
PROTEIN	26	yes
PROTEIN_COMPLEX	16	no — see below
PEPTIDOGLYCAN / LIPID / POLYSACCHARIDE / DNA	7	no — needs CHEBI or GO

interpro_groundings.py refuses a non-PROTEIN component outright:

InterPro grounding import refused: automatic InterPro review accepts only
PROTEIN components; GO:0009288#ms_ring is 'PROTEIN_COMPLEX'

That guard is correct. An InterPro family denotes a protein; the MS ring is ~26
copies of FliF, and tagging the ring with the FliF family would say the ring is
a protein. So 23 of the 49 can never be closed by this adapter, and no
amount of running it will help.

And the 26 that are eligible have a prerequisite: none of them has a
protein_examples accession. I checked all 26 — zero have one. The adapter
matches on accessions, not gene symbols, so every one of them needs an accession
added before InterPro can say anything.

So the real sequence is:

1. Add reviewed UniProt accessions to the 26 PROTEIN components (uniprot_sl.py
   already seeds these where a GO→SL mapping exists).
2. Then extend the InterPro allow-list over them.
3. Separately, decide what grounds a PROTEIN_COMPLEX — Complex Portal is
   already an adopted source and is the obvious candidate.
4. Separately again, ground the 7 material components against CHEBI/GO.

The flagellum work in the PR closing this comment covers exactly one component,
hook, because it is the only ungrounded flagellum component that is a PROTEIN
at all — and its answer is a recorded decline, not a grounding.