Regulation Sign Errors: Activator vs Inhibitor

Bottom line: a GO annotation can name the right process and the right gene and
still say the opposite of what happens: positive regulation where the gene
represses, an adenylate cyclase-activating pathway for a G protein that inhibits
the cyclase. Most over-annotation is a matter of degree (too broad, too indirect);
a sign error is a flat contradiction, and it propagates — through orthology
transfer, IBA and pathway models — with its sign intact. This project collects the
sign corrections already made during gene review, asks where they come from, and
looks for the ones not yet found.

A seed census of existing reviews finds 35 MODIFY rows on 25 genes whose
replacement is the same term with the opposite sign
(census, produced by
sign_flip_census.py). Sign problems resolved by
REMOVE, UNDECIDED or a sign-neutral parent are not in that count; a keyword pass over
review reasons (sign_reason_census.yaml) gives
152 candidate rows that still need screening by hand, since it also catches reasons
that merely discuss direction. A third, structured source already exists: reviewers
have tagged failure_modes: [REGULATORY_SIGN_INVERSION] in the propagation_review of
88 rows across 63 review files (a grep count, not yet analysed).

Principle: sign is judged in vivo

GO annotates what a gene product does in the organism, not what it can be made to do
in an assay. A sign established only in vitro (a purified protein at non-physiological
concentration, an overexpression readout) does not outweigh the in vivo direction, and
is not a reason to hold a contradicted sign as UNDECIDED. When the in vivo role has the
opposite sign, the annotation is corrected (MODIFY to the opposite-sign term) or, if
the row also has no sound source, removed.

Why this deserves its own project

The existing over-annotation work treats regulation terms mostly as a question of
directness — is the gene upstream, pleiotropic, or doing the work? Sign is a
separate axis with a separate failure mode:

• It is a contradiction, not an imprecision. A too-general term is still true; an
  inverted one is false, and a reader using GO for enrichment or causal modelling
  will draw the wrong conclusion.
• It survives every refinement step. IBA, ISS and Ensembl Compara (GO_REF:0000107)
  transfers copy the sign along with the term; GO-CAM causally upstream of, positive effect edges inherit it too.
• It has characteristic origins, which are testable (below).

Working hypotheses for where sign errors come from

1. Perturbation read the wrong way round (IMP). Knockout raises X → the gene
   negatively regulates X. Recording the phenotype direction rather than the gene's
   effect inverts the sign. IMP is the largest evidence class in the census (15/35),
   e.g. NLRP3 negative regulation of inflammatory response (×3) and LRP5 osteoblast /
   fat cell differentiation, both corrected to the opposite sign.
2. Context-dependent direction collapsed to one sign. A gene that promotes X in one
   cell type and inhibits it in another gets whichever sign the first paper showed.
   NTSR1 locomotion (hypolocomotion in knockouts, hyperactivity with stimulants) was
   generalized to the sign-neutral GO:0040012 regulation of locomotion.
3. Orthology transfer of a sign from a different assay or paralog. In this PR's Gi
   reviews both GNAI2 GO:0007189 adenylate cyclase-activating GPCR signaling and two
   NTSR1 rows (positive regulation of apoptotic process, negative regulation of release of sequestered calcium) arrived by GO_REF:0000107. The GNAI2 source paper
   itself reports Gαi2 mediating cyclase inhibition.
4. Wrong donor entirely. GNAI1's GO:0007189 ISS was transferred from an adenylyl
   cyclase (ADCY6), not a Gα ortholog — a sign error manufactured by a mis-matched
   similarity source. The merged #5582 review had held it UNDECIDED because
   high-concentration Gαi can stimulate type I/II cyclases in vitro (PMID:8119955);
   under the in vivo principle above it is now REMOVE, tagged
   PROPAGATION_BAD with failure modes WRONG_ORTHOLOG_OR_PARALOG and
   REGULATORY_SIGN_INVERSION. The episode is a pattern worth looking for elsewhere:
   an in vitro activity of the opposite sign can keep a physiologically wrong sign alive.
5. Family-level sign defaults. Gα proteins split into activating (Gs) and inhibiting
   (Gi/o) branches of the same pathway; a propagation that ignores the split imports
   the wrong half.

Seed cases

The full list is in the census file. The CD47 case is notable: an IBA, so the sign
sits on a PAINT node and every descendant inherits it.

Questions

• Does a sign error on a source row predict the same error on rows propagated from it?
  Trace each census row with IEA/IBA/ISS evidence back to its donor annotation.
• Are IMP sign errors concentrated in particular curation sources or eras?
• For GO-CAM: do the cached models (gocams/) carry any causal edge whose sign
  contradicts a reviewed annotation of the same gene?
• Which corrections should go upstream as GO tickets (the source annotation is wrong)
  versus stay as review-level judgments (the direction is genuinely context-dependent)?

STATUS

2026-10-24

• [x] Seed census of sign-flip MODIFYs across all reviews (35 rows / 25 genes)
• [x] Keyword census of review reasons mentioning sign/direction (152 candidate rows, unscreened)
• [ ] Screen the keyword census by hand; classify each true sign issue by origin hypothesis
• [ ] Harvest the 88 REGULATORY_SIGN_INVERSION propagation_review rows (63 files) into a census
• [ ] Look for other sign calls held UNDECIDED (or ACCEPTed) on in vitro evidence alone
• [ ] Extend the flip census to REMOVE/UNDECIDED rows and to MF activator/inhibitor pairs
• [ ] Trace propagated rows (IEA/IBA/ISS) to their donor annotations
• [ ] Cross-check GO-CAM causal-edge signs against reviews
• [ ] Draft upstream GO tickets for clear source-level errors

NOTES

2026-10-24

Project started from the Gi/opioid signaling reviews (PR #5350), where GNAI2 and
GNAI1 both carried the adenylate cyclase-activating pathway term despite being the
canonical inhibitory Gα. The flip census matches term labels that differ only by a
sign word (positive/negative regulation, activating/inhibiting, activator/inhibitor,
agonist/antagonist, up/down-regulation, stimulation/inhibition), so it misses
corrections to a sign-neutral parent (NTSR1 locomotion) and sign problems resolved by
REMOVE or UNDECIDED (GNAI1); those are listed by hand above.

2026-10-31

Principle added: sign is judged on in vivo evidence. GNAI1 GO:0007189 changed from
UNDECIDED to REMOVE on that basis. Found that PropagationFailureModeEnum already has
REGULATORY_SIGN_INVERSION ("Family contains activators and inhibitors, and a
positive/negative regulatory term leaks across members"), used on 88 rows.