UniProt subcellular locations

Reviewing GO annotations that come only from GO_REF:0000044

AI Gene Review · projects/SL · 2026

Bottom line

  • The UniProt SL → GO pipeline is still running and supplies many CC annotations with no other evidence.
  • Failures track granularity: membrane, cytoskeleton, mitochondrial membrane flagged at 23–36%; precise membranes at 0%.
  • Dropping SL terms made redundant by a more specific term was tested and refuted (10% vs 8%). 27 annotations moved so far.

Why SL, after SPKW

  • SPKW (keywords, GO_REF:0000043) was retired by GOA around April 2026. GO_REF:0000044 is live.
  • The GAF writes UniProtKB-SubCell:SL-xxxx into WITH/FROM, so each row names its source location.
  • Scan: every annotation whose only source is GO_REF:0000044, joined to the reviewer's verdict.
  • Now 1,580 such annotations in 1,179 gene folders; 1,565 reviewed.

Vague locations fail, precise ones do not

A controlled comparison inside one organelle

Four failure patterns

Pattern What goes wrong Examples
A under-specification true, but says nothing membrane on an ER protein
B association ≠ residence binds the structure, not in it SGCA, SGCE → MODIFY to complex terms
C family-rule propagation a family trait attached to all enolase Secreted via HAMAP MF_00318
D transit as destination passes through on the way out SALTY slrP, STAAU lytN

Plus SL-0221: the target term GO:0034045 is itself defective (a membrane bounding a condensate); 18 of 29 annotations moved.

The cheap fix that does not work

Group n Issue rate
more specific CC term already present 445 10%
SL term is the most specific the gene has 852 8%
  • Reviewers flag vagueness, not duplication: "not wrong, but 'membrane' is the uninformative parent" (yeast DCV1).
  • So GOA cannot suppress these by graph logic; it is a per-annotation judgement.

Status and next steps

  • ✅ Scanner and redundancy test committed: projects/SL/scripts/.
  • ✅ 22 genes re-reviewed; 27 annotations moved (18 SL-0221, 9 SL-0162/SL-0090).
  • ⬜ Audit the few family rules that attach Secreted to housekeeping enzymes (pattern C).
  • ⬜ Decide whether "true but uninformative" deserves its own verdict (31% land on KEEP_AS_NON_CORE).
  • ⬜ Systematic check for SL → GO mappings whose axioms fail, like SL-0221.

Read more: projects/SL.md · projects/SL/SL-METHODOLOGY.md