Miscitations

Citations that resolve, match their title, quote verbatim, and are still wrong

AI Gene Review · projects/MISCITATIONS · 2026

Bottom line

  • We aggregated every reference_review in the repo: 14,559 adjudicated references, 559 flagged (3.8% of an enriched sample, not an error rate).
  • Six seed cases re-verified live: all six stand, and all six errors live in GOA, IntAct or UniProt, not in this repo.
  • 4 of 6 are defects in how a correct paper was attached (WITH/FROM, evidence code, missing NOT), and the schema has no field for that yet.

Why the validators cannot see this

  • Check 1: fetched title must match recorded title. A wrong PMID imported with its own title passes.
  • Check 2: every quote must be verbatim in the cached paper. A verbatim quote can support the opposite conclusion.
  • Skipped prefixes: file:, GO_REF:, Reactome: are never snippet-checked; 127 of 559 flags are on them.

So the flag is a manual judgement, recorded in references[].reference_review.correctness.

Where the error lives

NLRP3: a dropped digit

  • Two IDA rows (GO:0060090, GO:0030674) cite PMID:1189953.
  • That PMID is "[Profanities and the profane person]", a 1975 psychiatry abstract.
  • Intended: PMID:31189953, the 2019 NEK7-NLRP3 structure paper, already cited for five other NLRP3 rows.
  • Live QuickGO returns the same two rows: current GOA, not a stale snapshot.

The review records the correct title for 1189953. Internally perfect, externally absurd.

PNPLA3: the paper reports the negative

  • GOA: GO:0003841 LPAAT activity, EXP from PMID:21878620, no NOT.
  • The paper assayed that exact reaction with a positive control:

"Neither the wild-type nor mutant enzyme catalyzed transfer of oleic acid from oleoyl-CoA to glycerophosphate, lysophosphatidic acid, or diacylglycerol"

  • The quote is verbatim, so check 2 passes; the annotation asserts the reverse of what was measured.

What the register holds

Patterns

  • A wrong identifier is rarely wrong once: 11 of 26 WRONG_IDENTIFIER rows come from five PMIDs.
    • paralog spread: ELOVL1/2/3 ← an ELOVL5 paper; NAA10/NAA40 ← a NAA60 paper
    • complex-partner spread: NPLOC4/UFD1, SERP1/SRPRB, UPF1/UPF2
  • Symbol collision: ADPRH ← "ARH1" hypercholesterolaemia; BRIP1 ← transcription factor BACH1.
  • Not only PMIDs: P2RX7 takes a localisation from a Reactome P2X1 event.

Status and next steps

  • ✅ Aggregator, generated register and TSV; six seed cases verified and written up.
  • ✅ Two-kind taxonomy: reference-level vs evidence-attachment miscitation.
  • ⬜ Decide where attachment defects go (per-annotation flag, FindingReviewStatusEnum, or an evidence_review slot).
  • ⬜ Triage 274 MISCITED + 26 WRONG_IDENTIFIER; neighbour sweep on the five recurring PMIDs.
  • ⬜ Decide what to report upstream to GOA, UniProt, IntAct.

Read more: projects/MISCITATIONS.md · MISCITATIONS/miscitation-register.md · sibling projects/MISCITATION_AUDIT.md