Miscitation Review Project

Bottom line: a miscitation is a citation that passes every mechanical check and is
still wrong: the identifier resolves, but to the wrong paper, or to the right paper used
for a claim it does not make. We built an aggregator over every reference_review block
in the gene reviews and re-verified six seed cases (NLRP3, ZBP1, GRID1, PEX39, SULT1B1,
PNPLA3) against live QuickGO, NCBI, EuropePMC and UniProt. We did this because the
validators check only internal consistency, and reviewer judgements about bad citations
were buried one gene at a time. The register covers 14,559 adjudicated references in
2,074 of 4,513 reviews and flags 559 (3.8%, an enriched sample rather than an error
rate); all six seed cases stand. The main finding is structural: four of the six are
defects in how a source database attached a correct paper (a wrong WITH/FROM
interactor, an experimental code on a meeting report, a missing NOT), and the schema
has no field for that yet. Most defects live in GOA, IntAct or UniProt rather than in
this repo, and nothing has been reported upstream.

The sibling Miscitation Audit keys the same flags on the
citation to find bad PMIDs copied across genes. Both count one flag per (gene,
citation) pair; the audit's register is an earlier run over 4,467 reviews, and between
the two runs the counts moved in both directions (WRONG_IDENTIFIER 28 to 26,
MISCITED 257 to 274), so they do not match exactly.

Overview

A miscitation is a citation that passes every mechanical check and is still
wrong. The identifier is well formed. It resolves. A paper comes back. But it is
not the paper that was meant, or it is the right paper and it does not say what
it is being cited for.

The schema already has a place to record this — references[].reference_review.correctness,
a ReferenceCorrectnessEnum — and reviewers have been filling it in for some time.
Nothing aggregated it, so a miscitation found while reviewing one gene stayed buried
in that gene's YAML. This project builds the register, works out what kinds of
miscitation there are, and separates the ones that are defects in this repo's
reviews from the ones that are defects in the source databases we import from.

The register is generated from the YAML, not
hand-maintained.

A citation can also go bad after it was written: see
Retracted Literature Behind Annotations, which checks every cited PMID
against PubMed's retraction metadata and feeds correctness: DISPUTED cases back here.

Why the existing checks cannot see this

The repo already validates citations, hard. Two checks in particular:

1. The reference resolves and the title matches. linkml-reference-validator
   fetches each cited PMID: and compares the fetched title against the title:
   recorded in the review. A transposed or invented PMID whose title no longer
   matches the intended paper fails.
2. Every supporting_text is verbatim. Each quote must be an exact substring
   of the cached publications/PMID_*.md. A paraphrased quote fails; an invented
   quote fails; a quote lifted from a different paper fails.

Both checks are about internal consistency. Neither can catch either half of a
miscitation:

• "Resolves to a different paper than intended." Check 1 compares the fetched
  title to the recorded title. If a wrong identifier was imported together with
  the wrong paper's title — which is exactly what happens when a PMID is taken from
  GOA and its title fetched from that same PMID — the two agree perfectly. NLRP3's
  PMID:1189953 has the correct title for PMID:1189953. It is simply not the
  paper anybody meant to cite.
• "Right paper, wrong claim." Check 2 verifies a quote is present in the
  paper. It cannot verify the quote supports the assertion. A verbatim sentence
  can be quoted in support of a conclusion the paper explicitly rejects — PNPLA3
  below is precisely that.

There is a third blind spot. The validator's skip_prefixes
(conf/reference_validator_config.yaml) exempt every non-literature prefix —
GO_REF, Reactome, file, UniProt, InterPro and others — from snippet
checking entirely. 127 of the 559 currently flagged references carry one of those
prefixes (72 file:, 28 GO_REF:, 27 Reactome:), i.e. they were found by a
human reading them, because no automated check was ever going to look.

So the flag is, and has to be, a manual judgment. reference_review is where it
goes.

Taxonomy

ReferenceCorrectnessEnum distinguishes four failure modes plus two states:

Cutting across that enum, the seed cases below separate into two kinds by where the
error lives, and this turns out to matter more than the enum value:

• Reference-level miscitation. The reference itself is wrong. correctness
  captures it directly and the register picks it up. NLRP3 and SULT1B1 are of this kind.
• Evidence-attachment miscitation. The reference is fine; the link from that
  reference to the assertion is wrong — a wrong interactor in WITH/FROM, an
  experimental evidence code on a non-experimental paper, an annotation asserting the
  positive of a result the cited paper reports as negative. ZBP1, GRID1, PNPLA3 and
  PEX39 are of this kind.

The second kind currently has no structured home. reference_review grades the
reference, and by the reference's own lights these references are all VERIFIED —
which is how they are in fact graded. The defect survives only as free text in
review_notes, where no aggregation can find it. Four of the six seed cases are like
this. That is the main structural finding so far, and the main open question (see
Status).

Source-database defects vs. review defects

Most of what follows is not a mistake in this repo. NLRP3's digit-dropped PMID is
in GOA. ZBP1's and GRID1's wrong interactors are in IntAct. PEX39's evidence code is in
UniProt. PNPLA3's EXP row is in GOA. These reviews are the first place the errors
have been written down; reference_review is the recording mechanism, not the cause.

That distinction should be kept sharp, both because it is true and because it
determines what to do next: a review defect is fixed by editing the YAML, a source
defect is fixed by reporting it upstream.

Seed cases

Six cases surfaced during the contested-functions review. Each was re-verified here
against the local GOA/UniProt files and, where possible, a live query (QuickGO,
NCBI E-utilities, EuropePMC, UniProt REST). All six are confirmed, in the sense
that the factual claim in the last column was checked, not taken on report.

NLRP3 — a dropped digit

genes/human/NLRP3/NLRP3-goa.tsv carries two IDA rows referencing PMID:1189953:

GO:0060090  molecular adaptor activity              ECO:0000314  IDA  PMID:1189953
GO:0030674  protein-macromolecule adaptor activity  ECO:0000314  IDA  PMID:1189953

A live QuickGO query for UniProtKB:Q96P20 + PMID:1189953 returns those same two
annotations, so this is live GOA and not a stale local snapshot. PubMed resolves
1189953 to "[Profanities and the profane person]" (Alva Quinones J, Acta
Psiquiatr Psicol Am Lat, June 1975) — a Spanish-language psychiatry abstract.

The intended reference is almost certainly PMID:31189953, "Structural mechanism for
NEK7-licensed activation of NLRP3 inflammasome" (Nature, 2019). The same GOA file
already cites 31189953 for five other NLRP3 annotations. A leading 3 was dropped.

This is the cleanest possible illustration of why check 1 cannot help: the review
records the title "[Profanities and the profane person]", and that is genuinely the
title of PMID:1189953. The citation is internally perfect and externally absurd.

MYH9, RAB3A, SYTL4, RAB10 — a dropped digit spread across a complex

Found on 2026-09-27 while reviewing MYH9 for the
nucleokinesis module. It is the NLRP3 defect again,
but spread across the members of a complex.

PMID:2732579 resolves to "[Two sisters with pseudoidiopathic hypoparathyroidism
presenting extensive intracranial calcification]" (1989), a Japanese case report.
A live QuickGO query returns 11 annotations citing it, all assigned_by UniProt
on 2019-02-19:

The intended reference is PMID:27325790, "A Rab3a-dependent complex essential for
lysosome positioning and plasma membrane repair" (J Cell Biol, 2016): the cited number
is 27325790 with its last digit dropped. That paper studies exactly these four
proteins (Rab3a, its effector Slp4-a/SYTL4, nonmuscle myosin heavy chain IIA/MYH9, and
Rab10 as a second repair regulator) and exactly these processes. UniProt also cites
27325790 correctly for two MYH9 rows (GO:0005515 with RAB3A, GO:0001778 plasma
membrane repair) entered eight days earlier, on 2019-02-11. So the two numbers were
used side by side for the same paper.

MYH9's review records correctness: WRONG_IDENTIFIER with a replacement to
PMID:27325790 (reason: WRONG_IDENTIFIER). RAB3A, SYTL4 and RAB10 have no reviews
in this repo yet, so their seven rows are not in the register. That is a concrete
case where the register under-counts a complex-partner spread.

ARATH/WIP1 — an IntAct interaction citing a dental paper

Found on 2026-09-27 while reviewing Arabidopsis WIP1 for the
LINC complex module. A live QuickGO query returns
2 annotations citing PMID:20579133: the reciprocal GO:0005515 IPI pair
WIP1 (Q8GXA4) with RANGAP1 (Q9LE82), assigned_by IntAct on 2026-07-25.
PubMed resolves 20579133 to "A laboratory evaluation of the physical and
mechanical properties of selected root canal sealers" (Int Endod J, 2010).

This is not a dropped digit. The IntAct record behind the rows (IMEx
IM-19345) credits "Xu et al. (2007)", which points to PMID:17600715,
"Anchorage of plant RanGAP to the nuclear envelope involves novel
nuclear-pore-associated proteins" (Curr Biol, 2007). That is the paper that
characterized the WIP1-RanGAP1 interaction. But PMID:17600715 carries its own
IMEx id (IM-19776), so the intended reference is not certain. WIP1's review
records correctness: WRONG_IDENTIFIER without a replacement. This is the first
IntAct-sourced WRONG_IDENTIFIER in the project, and the error sits in the
publication identifier of the IMEx record, not in a GO curation step.

ZBP1 and GRID1 — the interactor, not the paper

ZBP1 GO:0005515 IPI from PMID:19590578 lists UniProtKB:Q13601 in WITH/FROM.
The paper is "DAI/ZBP1 recruits RIP1 and RIP3 through RIP homotypic interaction motifs
to activate NF-kappaB" — an entirely appropriate citation. But Q13601 resolves (UniProt
REST) to KRR1_HUMAN, "KRR1 small subunit processome component homolog". RIPK1 is
Q13546. The other partner recorded for the same reference, Q9Y572, is RIPK3 and
does match the paper.

GRID1 has exactly two GO:0005515 IPI rows, from PMID:28514442 and
PMID:33961781, both proteome-scale interactome maps. Both name UniProtKB:P68871 — HBB_HUMAN, haemoglobin
subunit beta. Both source papers are legitimate proteome-scale interactome screens,
correctly cited. Haemoglobin is one of the best-known contaminants of affinity
purification from tissue, and it is not a credible partner for a postsynaptic glutamate
receptor. The whole of GRID1's protein binding evidence rests on it.

In both cases the reference is right and the correctness is (correctly) VERIFIED.
The error is one column over.

PEX39 — an experimental evidence code on a meeting report

UniProt's FUNCTION block for Q5I0X4 (PEX39) reads, in part:

CC   -!- FUNCTION: Cytosolic peroxin that promotes the peroxisomal import of
CC       proteins containing a type-2 peroxisomal targeting signal (PTS2) by
CC       binding to PEX7 ... (PubMed:37160800, PubMed:40739340).
CC       {ECO:0000269|PubMed:37160800, ECO:0000269|PubMed:40739340}.

PMID:40739340 is the primary paper ("PEX39 facilitates the peroxisomal import of
PTS2-containing proteins", Nat Cell Biol 2025) and is appropriate. PMID:37160800 is
"Peroxisomes: novel findings and future directions" (Histochem Cell Biol 2023) — an
18-author community overview with no abstract in PubMed, and typed by EuropePMC as meeting-report. It
carries ECO:0000269, "experimental evidence used in manual assertion". A meeting
report is not a source of experimental evidence, and this one predates the primary
paper by two years.

SULT1B1 — the negative control, annotated as a positive

GO:0006068 ethanol catabolic process, IDA, PMID:23207770, assigned by CAFA. The
paper is "Ethanol sulfation by the human cytosolic sulfotransferases: a systematic
analysis" (Biol Pharm Bull 2012) and its abstract says:

A systematic analysis revealed four ethanol-sulfating SULTs, SULT1A1, SULT1A2,
SULT1A3, and SULT1C4, among the eleven human SULT enzymes previously prepared and
purified.

SULT1B1 is one of the eleven. It is not one of the four. It was in the panel as a
negative, and the annotation asserts the opposite of the result. The cached record is
abstract-only, so the full-text panel table has not been checked — but the abstract's
enumeration is explicit, and the burden is on the annotation.

PNPLA3 — the paper reports the negative

This is the sharpest case, because it shows MISCITED and "verbatim quote" are
orthogonal. GOA gives PNPLA3 GO:0003841 (1-acylglycerol-3-phosphate
O-acyltransferase, i.e. LPAAT) three times over — once IEA, once IDA from
PMID:22560221, and once EXP from PMID:21878620, with no NOT qualifier.
PMID:21878620 (Huang, Cohen & Hobbs, JBC 2011) concludes:

Neither the wild-type nor mutant enzyme catalyzed transfer of oleic acid from
oleoyl-CoA to glycerophosphate, lysophosphatidic acid, or diacylglycerol,
suggesting that the enzyme does not promote de novo TAG synthesis.

"Lysophosphatidic acid" is the acceptor in GO:0003841. The paper assayed the exact
reaction, with CGI-58 as a positive control, and detected none. Every sentence of that
quote is verbatim; the verbatim check passes; the annotation still asserts the
reverse of what was measured. Had the row carried NOT, the citation would be
impeccable.

(Separately, the underlying biology is genuinely contested: PMID:22560221 reports the
LPAAT activity that PMID:21878620 could not find. The PNPLA3 review grades that pair
DISPUTED. The EXP/21878620 row is a miscitation regardless of how the dispute
resolves.)

Current state of the evidence

From the register, regenerated from the YAML
(see Reproducibility):

• 4513 reviewed gene files scanned; 2074 (46%) carry at least one
  reference_review block
• 14559 references have been manually adjudicated
• 559 (3.8%) are flagged as a citation or soundness problem
• 586 (4.0%) are explicitly UNVERIFIED — adjudication begun, this reference not
  yet checked

These figures are not a survey. The denominator is "references a reviewer chose to
adjudicate", and reviewers adjudicate a reference when they are already looking at it —
frequently because something about it looked wrong. The 3.8% is a flag rate on an
enriched, opportunistic sample, not a background error rate for GO citations, and it
should not be quoted as one. Nor is the register a complete list: 338 genes carry a
flagged reference out of 2074 adjudicated out of 4513 reviewed, and the great majority
of references in the repo have never been looked at this way at all.

The distribution across organisms reflects where curation effort has gone, not where
errors are: human accounts for 470 of the 559 flags, PSEPK for 32, DICDI 11, yeast 9,
DROME 7, SCHPO 7, and a long tail of single flags.

Patterns visible so far

A wrong identifier is rarely wrong once. Of the 26 WRONG_IDENTIFIER rows, 11 come
from just five PMIDs, each mis-attached to two or three related genes:

Two shapes are mixed in there. ELOVL1/2/3 and NAA10/NAA40 are paralog spread: one
family member's paper attached to its relatives, the classic failure of family-level
propagation. NPLOC4/UFD1, SERP1/SRPRB and UPF1/UPF2 are complex-partner spread:
one bad identifier scattered across the members of a complex or a functional pair,
which looks much more like a single import defect replicated than like independent
curation errors. Either way, finding one instance is worth checking its neighbours.

Gene-symbol collision is its own category. ADPRH is cited for a paper about "ARH1"
meaning autosomal-recessive hypercholesterolaemia; BRIP1 (sometimes "BACH1") is cited
for a paper about the bZIP transcription factor BACH1. These resolve, have matching
titles, and are about a different molecule that happens to share a string.

Not everything flagged is a PMID. 72 flagged references are file: citations into
the repo's own bioinformatics results, 28 are GO_REF:, and 27 are Reactome: — all
prefixes the reference validator skips wholesale. One Reactome: case is instructive:
P2RX7 takes a TAS localisation from Reactome:R-HSA-139855, an event titled
"P2X1-mediated entry of Ca++ from plasma" — a P2X1 event supporting a P2X7
annotation.

Method

1. Aggregate first. Run the extractor over every review; the register is the
   worklist. It costs nothing and it is already populated by work that has been done.
2. Verify before recording. A reference_review is a claim about the world, so it
   is checked against the world: resolve the PMID at NCBI, resolve accessions at
   UniProt REST, re-query GOA live at QuickGO rather than trusting the local -goa.tsv
   snapshot, check publication type at EuropePMC. Do not grade a reference VERIFIED
   on the strength of a deep-research summary asserting it is fine — that is the
   failure mode this project exists to catch, one level up.
3. Grade the reference, note the attachment. If the defect is in the reference,
   set correctness. If the defect is in how a source database attached the
   reference, the reference is VERIFIED and the defect goes in review_notes and in
   the annotation's review.reason — for now (see Status).
4. Check the neighbours. On a WRONG_IDENTIFIER, grep the flagged PMID across
   genes/ and check the gene's paralogs and complex partners.
5. Say where the error lives. Source-database defects are reported as such and are
   candidates for upstream reporting; review defects are fixed in place.

Reproducibility

Supporting material under MISCITATIONS/:

• MISCITATIONS/aggregate_miscitations.py —
  the extractor.
• MISCITATIONS/miscitation-register.md —
  the generated register (counts by correctness, by relevance, by organism; full
  table of every flagged reference).
• reports/miscitations.tsv — one row per adjudicated reference, untruncated notes.

uv run python projects/MISCITATIONS/aggregate_miscitations.py

STATUS

• [x] Aggregator over every references[].reference_review in genes/*/*/*-ai-review.yaml
• [x] Generated register + TSV (14559 adjudicated, 559 flagged)
• [x] Six seed cases independently re-verified (QuickGO, NCBI E-utilities, EuropePMC,
  UniProt REST) and written up
• [x] Two-kind taxonomy established: reference-level vs. evidence-attachment miscitation
• [ ] Open question: where does an evidence-attachment defect go? Four of six seed
  cases are not representable in reference_review, because the reference is
  genuinely fine. Options: a per-annotation flag analogous to correctness; reuse
  of FindingReviewStatusEnum; or an explicit evidence_review on
  ExistingAnnotation. Not yet decided — do not add a field before the shape of
  the problem is clearer than six cases.
• [ ] Triage the 274 MISCITED and 26 WRONG_IDENTIFIER rows: confirm, classify by
  pattern, and split source-database defects from review defects
• [ ] Neighbour sweep on the five recurring wrong PMIDs — check remaining ELOVL/NAA
  family members and other complex partners
• [x] MYH9 PMID:2732579 → PMID:27325790 (dropped digit) recorded; live GOA shows
  the same defect on RAB3A, SYTL4 and RAB10 (11 rows total)
• [ ] Record the RAB3A / SYTL4 / RAB10 rows when those genes are reviewed
• [x] ARATH/WIP1 PMID:20579133 (dental paper) on the WIP1-RANGAP1 IntAct pair
  recorded; the intended paper is probably PMID:17600715, but that is
  unconfirmed (IMEx IM-19345)
• [ ] Decide what, if anything, to report upstream to GOA / UniProt / IntAct, and in
  what form
• [ ] Consider whether reference_review should be required (currently a reviewer
  may simply omit it, which is indistinguishable from "no problem found")

Slides

• Slides (Marp source: MISCITATIONS-slides.md) — AI generated

Last updated: 2026-09-27

NOTES

2026-09-27

WIP1 IntAct wrong identifier. Found while reviewing Arabidopsis WIP1 for the
LINC complex module. IntAct's WIP1-RANGAP1 IPI pair cites PMID:20579133, a
2010 root canal sealer study. Checked at NCBI E-utilities and in live QuickGO
(2 rows). The IMEx record credits Xu et al. 2007 (PMID:17600715), but no
replacement is asserted because that paper has a separate IMEx record. This
kind of error should be reported to IntAct rather than GOA.

MYH9 dropped digit, spread across a complex. Found while reviewing MYH9 for the
nucleokinesis module. PMID:2732579 (a 1989 hypoparathyroidism case report) stands in
for PMID:27325790 (Rab3a/Slp4-a/NMHC-IIA lysosome positioning, 2016). Both titles
were checked at NCBI E-utilities, and the rows were checked in live QuickGO, not only
the local -goa.tsv. This is the second dropped-digit case after NLRP3. It is also the
first complex-partner spread of the dropped-digit kind: 11 rows over MYH9, RAB3A, SYTL4
and RAB10, all from one UniProt curation date. It is worth reporting to UniProt as one
group.

2026-09-17

Project creation. Built the aggregator and the register from all 4513 reviewed gene
files: 2074 carry reference_review, 14559 references adjudicated, 559 flagged (3.8%
of adjudicated — an enriched sample, not a survey).

Re-verified the six seed cases from the contested-functions review. All six stand.
Notes from doing so:

• NLRP3 needed the live query. The local -goa.tsv could have been stale; QuickGO
  returns the same two IDA rows against PMID:1189953 today, so the error is current
  GOA, not a snapshot artefact.
• The enum does not fit four of the six. ZBP1, GRID1, PEX39 and PNPLA3 all have a
  perfectly good reference; the defect is in the WITH/FROM column, the evidence code,
  or the absent NOT. All four are already graded VERIFIED in their reviews, with
  the real problem written into review_notes — correctly, but invisibly. This is the
  finding that most shapes what the project should do next, and it argues the register
  systematically under-counts attachment defects relative to reference defects.
  PNPLA3's own note states it exactly: "The citation itself is correct; what is wrong
  is GOA's use of it as EXP evidence for GO:0003841, a reaction this paper could not
  detect."
• Recurrence was the surprise. 11 of 26 WRONG_IDENTIFIER rows trace to five PMIDs.
  The complex-partner cases (UFD1/NPLOC4, SERP1/SRPRB, UPF1/UPF2) look like one import
  defect replicated across a complex rather than independent errors, which would make
  them cheap to fix upstream and worth reporting as a group.
• The validator's skip_prefixes are a real blind spot. 127 of the 559 flags are
  on prefixes (file:, GO_REF:, Reactome:) that are never snippet-checked. The
  file: ones matter for this repo specifically — SPKW's 2026-05-30 audit found ~48%
  of file: quotes non-verbatim across its plant reviews, all of which had passed
  just validate.