SL Methodology

Supporting page for the SL project. All tables are generated:

uv run python projects/SL/scripts/scan_sl_unique.py

Takes ~5 minutes over the full genes/ tree. Rerun and replace the tables wholesale.

What counts as SL-unique

A gene's annotation to a GO term is SL-unique when GO_REF:0000044 is the only
reference supporting that gene-term pair. If the same term also carries an IDA, an IBA, an
ISS, or any other reference, it is excluded — the point is to isolate annotations where the
UniProt subcellular-location mapping is the sole load-bearing evidence.

This mirrors the SPKW project's definition of keyword-unique annotations, with one
simplification. SPKW had to reverse-map GO terms to their originating keywords through the
external2go keyword2go file, because the GAF stores only the GO term. Here the source is in
the annotation: the WITH/FROM column carries UniProtKB-SubCell:SL-xxxx, so per-location
statistics fall out of the same scan.

Two caveats on the definition:

• Closure is not applied. SPKW's methodology filters annotations whose term is an ancestor
  of another term the gene already has, which removed 70%+ of naive hits in well-curated
  organisms. That filter is not applied here, so some SL-unique annotations are redundant
  rather than wrong. Given that the headline finding is precisely about under-specified terms,
  adding closure filtering would remove the signal being measured — but it means the 40%
  "downgraded or worse" figure mixes redundancy with error. The 9% hard-issue rate is the more
  conservative number.
• The corpus is not a sample of GOA. These 986 gene folders were selected for review for
  unrelated reasons. Rates here describe this corpus, not the pipeline at large.

Reading the tables

• The by GO term and by SL location tables are near-duplicates by construction, since
  the mapping is close to one-to-one. Divergences between them are worth a look: they mean one
  GO term is reached from more than one SL location, or vice versa.
• "Issues" counts REMOVE, MARK_AS_OVER_ANNOTATED, and MODIFY. KEEP_AS_NON_CORE is
  excluded — it is a judgment that the location is real but peripheral, which is a different
  claim from the annotation being wrong.
• Rows below 10 reviewed annotations are omitted. SL-0221, the subject of the first
  subproject, falls below that threshold in this corpus (9 assertions) and is documented
  separately.

SL names

Names are resolved live from https://rest.uniprot.org/locations/SL-xxxx. Pass --offline
to skip the lookup when running without network access.

Corpus totals

• SL-unique annotations (sole source GO_REF:0000044): 1300
• distinct gene folders: 986
• reviewed SL-unique annotations: 1297
• aspect: {'cellular_component': 1299, 'C': 1}
• actions: ACCEPT 760, KEEP_AS_NON_CORE 406, MODIFY 58, MARK_AS_OVER_ANNOTATED 43, REMOVE 15, UNDECIDED 9, PENDING 6
• downgraded or worse: 522/1297 (40%)
• issue rate (REMOVE/MARK_AS_OVER_ANNOTATED/MODIFY): 116/1297 (9%)

By GO term (>= 10 reviewed)

By UniProt subcellular location (>= 10 reviewed)

Redundancy test

uv run python projects/SL/scripts/sl_redundancy.py

Asks, for each SL-unique annotation to term T, whether the gene carries any other CC term that
is a proper descendant of T under is_a/part_of, using the local GO SQLite build. Used to
test — and refute — the hypothesis that SL over-annotation is duplication; see
the project page.

The output below is post-intervention and partly circular: the SL-0162 and SL-0090 review
batches deliberately selected redundant annotations, which inflates the redundant group's issue
rate from 10% to 12%. The result quoted on the project page is the pre-batch measurement.

Redundancy of SL-unique annotations

• SL-unique annotations examined: 1300
• with a review action: 1297
• of those, the gene already carries a more specific CC term from another source: 445/1297 (34%)

Issue rate, split by redundancy

By SL location (>= 10 reviewed)