SL-0162 Membrane → GO:0016020

The largest under-specified subcellular location in the corpus, and the cleanest example of
the SL project's core finding. 61 SL-unique annotations reviewed; 17 carry a
hard issue (28%) — the highest rate of any location with a substantial sample. Only
SL-0171 Mitochondrion membrane rates higher (31%), on a sample a fifth the size (4/13).

What the reviewers actually objected to

Every single flagged case is the same complaint, stated in different words: the term is true
and useless. A sample, verbatim from review.reason:

gene	verdict	reviewer's words
human PET100	MARK_OVER	"'membrane' is an unspecific location that adds no information beyond the more precise GO:0005743 … It reflects the generic single-pass-membrane-protein SubCell keyword rather than a distinct localization."
yeast DCV1	MARK_OVER	"a bona fide integral membrane protein … so the annotation is not wrong, but 'membrane' is the uninformative parent of the more specific compartment terms already present"
worm daf-2	MODIFY	"too general. The more specific term plasma membrane (GO:0005886) is already annotated"
human ABHD14A	MARK_OVER	"Marked over-annotated for lack of specificity rather than for lack of membrane association"
human GP9	MODIFY	"While 'membrane' is technically correct, this term is too general."

Not one flagged case disputes that the protein is in a membrane. GO:0016020 is being
flagged for what it fails to say.

Where reviewers proposed a replacement, it was always a specific compartment already implied
by the protein's topology or known biology: ER membrane (DANRE cyp51, worm atf-6, yeast
PHO86), plasma membrane (worm daf-2, mouse Egf, human GP9), lysosomal
membrane (human ATP6V0E1), mitochondrial inner membrane (human PET100).

Batch reviewed under this subproject

Three annotations moved ACCEPT → MARK_AS_OVER_ANNOTATED, selected as the clearest cases:
SL-unique, and the gene already carries a strictly more specific CC term from independent
(non-SL) evidence.

gene	already carries	source
human IL2RA	GO:0005893 interleukin-2 receptor complex; GO:0009897 external side of plasma membrane	TAS
human FGFRL1	GO:0005886 plasma membrane	IBA
ANOGA TOLL9	GO:0005886 plasma membrane	IBA

MARK_AS_OVER_ANNOTATED rather than MODIFY: the precise term is already present, so there
is nothing to replace the broad one with.

The batch started at five. Human DNAJC25 and yeast MCH2 were reverted to ACCEPT during
review, because in both cases the "strictly more specific CC term" the selection criterion
relies on is itself the uncertain one — MCH2's GO:0005886 is KEEP_AS_NON_CORE because
plasma-membrane localization is not established for that paralog, and DNAJC25's own
suggested_questions still asks which membrane it resides at. So the criterion as stated —
"the gene already carries a strictly more specific CC term from independent evidence" — is
not sufficient on its own: the more specific term also has to be accepted as established.
Two of the five hand-picked clearest cases failed that second condition.

One candidate was dropped after checking — yeast MNT4 has no independent more-specific CC
term, so the broad annotation is the only thing placing it anywhere.

What this subproject does not show

The obvious mechanical fix — suppress an SL-derived CC term when the gene already carries a
descendant — does not work. See the redundancy test:
across the whole corpus, redundant SL-unique annotations were flagged at a rate barely
distinguishable from non-redundant ones. Within SL-0162 itself, before this batch, redundant
cases ran 26% and non-redundant 19% — a gap, but nothing like the clean separation the rule
would need.

The reason is visible in the reviewer quotes above: ABHD14A and DCV1 were flagged for
vagueness, not duplication. A gene whose only annotation is membrane is just as
uninformatively annotated as one that has membrane plus five specific terms — arguably
worse. Deduplication would fix the second case and miss the first.