SwissProt keyword (SPKW) annotations

A retrospective audit of GO_REF:0000043, where a UniProt keyword was the only source

AI Gene Review · projects/SPKW · 2026

Bottom line

  • Eukaryotic process keywords were unsafe: 79–100% of reviewed SPKW-only apoptosis, autophagy, rhythm and S. pombe meiosis rows were over-annotations.
  • Bacterial keyword rows were mostly sound; plant SPKW-unique terms carried real risk in only ~15% of cases.
  • GOA retired SPKW for cellular organisms (~April 2026). That removed the bad rows and also correct ones, e.g. DELLA GA signalling (RHT1), patatin storage (PATB1).

What we asked

  • A keyword such as Apoptosis maps to a GO term via GO_REF:0000043. For many genes it was the only evidence.
  • Closure filter: a row counts only if no other source supports that term or a more specific one.
  • 12 subprojects: human apoptosis / rhythm / autophagy; S. pombe, Drosophila, Anopheles, P. putida, Arabidopsis; phage T4, E. coli O157, virus clades; non-Arabidopsis plants.
  • Swiss-Prot keywords are manually assigned, so errors in reviewed organisms sit in the keyword→GO mapping, not keyword choice.

Issue rates by subproject

One keyword, several verdicts

Recurring failure patterns

Pattern Example
Regulatory conflation AIMP2 → apoptotic process (MODIFY)
Caspase substrate ≠ apoptosis AIMP1, BCAP31 (over-annotated)
Process conflation S. pombe ATG genes → meiosis
Eukaryote-centric terms on phage T4 DAM → innate immune response
Enzyme-class keyword → bare process Methyltransferase → methylation (plant MTases)
Expression ≠ function rhythmic-process genes; ENOD2A nodulation

The repository-wide picture

All SPKW rows reviewed anywhere in this repo, not only this project. Nearly half were accepted as they stand.

What retirement cost

  • Plants: 214 SPKW-unique terms tiered. Tier A (over-annotation risk) 15%; Tier B+C (correct) ~79%.
  • Correct facts lost where the keyword was the only carrier: RHT1 GA signalling, PATB1 nutrient reservoir, NSP1 nodulation, EME1 endonuclease, PR1B1 defense.
  • Lesson for any mapping source: tier terms by risk and review by gene position, rather than keep-all or drop-all.

Status

  • Complete as a retrospective: 12 subprojects, 137 genes in the results table.
  • Side finding: 265/550 plant file: quotes were non-verbatim because the validator skips file: refs; all fixed, no action changes.
  • Methods and SQL: projects/SPKW/SPKW-METHODOLOGY.md.

Read more: projects/SPKW.md · projects/SPKW/SPKW-PLANTS.md · projects/SPKW/SPKW-APOPTOSIS.md