NOT Annotation Usage Audit

IN_PROGRESS EVALUATIONPIPELINE

Species: human, rat, ARATH, SCHPO

Genes: FGFR4 AGR2 GCH1 HSPA6 CRY2 Ghr CLV3 GID1A chk1

NOT Annotation Usage Audit

Bottom line: GOA contains 10,622 NOT annotations on 7,408 gene products. The
qualifier is used in very different ways. Its strongest use is to negate a molecular
function, typically a catalytic activity that a pseudoenzyme or a diverged family
member has lost (3,405 rows), backed by an assay or by a PAINT loss-of-function call
(IKR). Its weakest use is to negate a process from an expression or perturbation readout,
above all "NOT response to ". There are 490 such rows, and the most
common evidence for them is IEP, an expression pattern, which by itself cannot show that
a gene product lacks a function. In total 169 NOTs rest on IEP alone, and another 627 are
TAIR computational (RCA) localization calls. UniProt CAUTION notes track the useful
end: 27% of reviewed proteins with an MF NOT carry a CAUTION, 19.5% with loss-of-activity
wording, against 7% and 1.9% for proteins whose NOTs are only to processes or
locations. Our own gene reviews have so far ACCEPTed 83% of the NOTs they met, so the
review guidance needs a rule for judging NOTs, not just positive annotations.

We started this because the FGFR4 review met NOT involved_in response to bile acid
(GO:1903412, IMP, PMID:19085950). The paper's only negative result is that bile acid did
not change FGFR4 protein level, which is an expression observation, not evidence that
FGFR4 lacks a function. FGFR4 acts in the FGF19 signal that bile acid induces, which is
already captured as negative regulation of bile acid biosynthesis.

Question

How is the NOT qualifier used across GO, which uses are sound, and how do NOTs relate
to UniProt CAUTION notes? The working hypothesis is that NOT is meaningful mainly for
molecular functions (a gene product demonstrably lacks an activity its family or domain
predicts), and much weaker for "response to" and other process terms, where the
negative readout is usually an expression or phenotype null.

Methods

All code and data are in NOT_ANNOTATION_USAGE/.

  1. fetch_not_annotations.py downloads every NOT-qualified annotation from QuickGO
    (one download per NOT|<relation>), filling GO labels from the local GO build.
    Output: not_annotations.tsv (10,622 rows, fetched 2026-10-01).
  2. fetch_caution_notes.py fetches the UniProt CAUTION text for each of the 7,369
    distinct UniProt accessions. Output: not_accessions_caution.tsv.
  3. analyze_not_usage.py assigns each negated term to a coarse category by GO
    is_a/part_of ancestry (first match wins: signaling, defense/immune response, other
    response to stimulus, regulation, development, metabolism, localization for BP;
    catalytic, transporter, transducer, transcription regulator, binding for MF). It
    flags CAUTION notes with loss-of-activity wording, and joins the actions our gene
    reviews gave negated rows. It writes not_annotations_classified.tsv,
    review_negated_actions.tsv and RESULTS.md.

Re-run all three scripts in order with uv run python <script>; RESULTS.md is
regenerated each time.

Findings so far

Full tables are in RESULTS.md.

Reviewed proteins Any CAUTION Loss-of-activity CAUTION
with an MF NOT (1,838) 27.0% 19.5%
with only BP/CC NOTs (2,388) 7.0% 1.9%

The reverse direction (CAUTION notes describing lost activity on proteins with no MF
NOT) is covered by Query B of the UniProt CAUTION Note project.
- Our reviews have been deferential.
- 270 negated rows in 174 reviews: 223 ACCEPT, 20 UNDECIDED, 16 KEEP_AS_NON_CORE,
only 8 REMOVE.
- The worklist in RESULTS.md lists the 9 reviewed NOTs to non-defense "response to"
terms or with IEP evidence. FGFR4 and AGR2 are already REMOVE.
- Not every item on that list is bad. CRY2 NOT photoreactive repair records that a
photolyase homolog lacks repair activity, which is a process-shaped statement of an
MF negative and is sound.

Proposed guidance (draft, for discussion)

  1. A NOT is strongest when it negates a molecular function a reader would otherwise
    predict from family, domain or orthology, with assay or key-residue evidence.
  2. A NOT to a process should be judged by asking whether the gene product was shown
    unable to do the process's work. "Expression did not change" never meets this bar,
    so IEP-only process NOTs should be REMOVE in reviews.
  3. "NOT response to X" is acceptable only when the gene product was tested for a direct
    role, for example a receptor shown not to bind and transduce X. It is not acceptable
    for an expression or abundance null.
  4. Where a CAUTION note records lost activity and no MF NOT exists, propose an MF NOT
    (see Top-Nots).

STATUS

Updated 2026-10-01.

NOTES

2026-10-01

Project started from the FGFR4 review (FGFR signaling module work). First full
download and classification run; see Findings. The "response to stimulus" bucket was
initially too broad because GO places signaling pathways under cellular response to
stimulus, and defense responses from antimicrobial assays behave like activity
negatives; both are now separate categories.