Wrong-Gene Leads

Bottom line: UniProt lists the papers it used for each entry. When a GOA row with
experimental or author-statement evidence cites a paper that the annotated protein's
own entry does not cite, but another reviewed entry in the same species does, the
paper may have studied that other protein. A sweep of every gene folder in this
repository finds 9,963 such rows in 2,640 genes. After scoring, 473 rows in 233 genes
rank in the top tier. These are leads for review, not verdicts: a quick look at the
top tier finds a mix of plausible wrong-gene annotations and papers that legitimately
studied several related genes.

This is the reference-level counterpart of the
function requirement rules: those ask whether a
protein can do what it is annotated with, this asks whether the cited paper was
about that protein. It follows
issue #5727.

Motivating cases

• human/CSGALNACT1. Ten rows cite PMID:11514575, which cloned the 802-residue
  chromosome-15 chondroitin synthase CHSY1 (Q86X52). CSGALNACT1 is 532 residues and
  was cloned a year later. Only CHSY1's UniProt entry cites the paper. The review
  records these rows as WRONG_GENE_PRODUCT source defects.
• human/ARHGDIB. A TAS cytoplasmic-vesicle row cites PMID:9113980, the
  RhoGDIγ (ARHGDIG, Q99819) paper; only ARHGDIG's entry cites it.
• ECOLI/traB (a known miss). The ATP hydrolysis IDA cites PMID:20172994, on
  plasmid pKM101's TraB (Q46698). That unreviewed entry does not cite the paper, so
  this signal cannot see the case.

Method

just wrong-gene-leads (src/ai_gene_review/tools/wrong_gene_leads.py):

1. For each gene folder, take GOA rows with evidence EXP, IDA, IPI, IMP, IGI, IEP,
   TAS or NAS that cite a PMID the gene's own *-uniprot.txt does not cite. NOT rows
   and bare protein binding (GO:0005515) are skipped, as are high-throughput codes.
2. Ask UniProt which reviewed entries cite each such PMID, in batches, cached under
   .cache/wrong-gene-leads/. UniProt's lit_pubmed: search matches only an
   entry's own references (its RX PubMed= lines), not the computationally
   mapped bibliography (GeneRIF, PRO, Reactome, GAD), so both sides of the
   comparison are the same kind of list. Checked 2026-11-06: for CSGALNACT1,
   ARHGDIB and SOCS3, lit_pubmed_id equals the flat file's RX set, and none of
   CSGALNACT1's eight mapped papers is matched by a lit_pubmed: search.
3. Drop citing entries that are the same gene in the same species (symbols are
   compared ignoring case, since UniProt writes rodent symbols in title case). Papers
   still cited by more than 20 other reviewed entries are treated as large-scale and
   skipped. An unreviewed or
   secondary entry annotated while the reviewed entry for the same gene cites the
   paper is not a wrong gene; before this filter, such cases dominated the top
   scores (for example Drosophila loqs, X2J5X6 against Q9VJY9).
4. Score each remaining row:

The name check matters because papers often study several members of a family; a
paper whose abstract names the annotated gene probably did study it. ARHGDIB's
paper mentions "GDI/D4", one of its synonyms, so that lead is kept but ranks low.
Gene names come from the whole GN block, and curated protein names (DE
Full=/Short=) are matched as phrases. Protein names whose only evidence is an
ARBA rule or submitter text are not used, and neither is a name shared by a citing
entry: a family name such as "Tubulin alpha chain" does not say which paralog a
paper studied. The check reads the cached title and abstract only, up to the next
section heading.

The report is written to reports/wrong-gene-leads/leads.tsv. It is generated and
git-ignored, never committed, because it aggregates every review. It joins each lead
to the existing review action and to any WRONG_GENE_PRODUCT source defect.

Results (2026-11-01)

9,963 leads in 2,640 genes:

Agreement with past reviews is low at every tier. That is expected rather than
reassuring: reviewers almost never checked whether a cited paper studied the
annotated protein, so most rows were accepted on the term alone.

The top tier (score 6 or more) is 473 rows in 233 genes, mostly human (319) and
mostly IDA, IMP, TAS and NAS rows; 57 are IPI. A random sample of 25 non-IPI rows
from that tier, judged from paper titles only (not yet checked against the papers),
splits roughly in two:

• Look like real candidates: FLNC "cytoplasm" (IDA) from a paper on filamin A and
  filamin B; PAPSS1 "skeletal system development" (TAS) from a PAPSS2 dysplasia
  paper; AURKC "chromosome passenger complex" (TAS) from an Aurora A review; NLRP3
  "inflammatory response" (IMP) from an NLRP4 paper; SOCS3 "protein kinase inhibitor
  activity" (IDA) from a SOCS2 paper.
• Probably legitimate multi-gene papers: CALM2 and CALM3 from a calmodulin-mutation
  paper about CALM1 (the three genes encode the same protein), TRIM13 from a TRIM
  screen, fibrinogen gamma from a fibrinogen Aα paper, and calcineurin isozyme papers.

The score-5 tier is large (3,864 rows) and dominated by "another gene in the same
species, few citing entries, gene not named in the abstract". Many of these are
complex or pathway papers where UniProt cited the paper only for the protein it
foregrounded.

First lead review (2026-11-01)

The six top-tier leads that looked most plausible from paper titles were checked
against the cached abstracts. Full text could not be retrieved for any of them
(Europe PMC has none, and the JBC pages returned 403), so CLAUDE.md's rule applies:
an experimental row is not called wrong-gene unless the abstract itself says so.

So one of six leads was confirmed, three need the full text, and two were detector
false positives. The two false-positive patterns, wrapped synonyms
and protein names written in words, are both name-matching gaps; the first is fixed.

Protein-name matching and second lead review (2026-11-01)

The name check now also matches UniProt protein names (DE Full=/Short=) as
phrases in the cached title and abstract, including family enumerations such as
"aurora kinases A, B, and C" for "Aurora kinase C". This penalised papers that spell
out the protein, and the top tier fell from 466 to 390 rows.

The next 20 top-tier leads (one row per gene, IPI rows skipped, genes already checked
excluded) were then checked against the cached abstracts, and the cached full text
where available:

Across both batches (26 leads): 3 rows are confirmed or probable wrong-gene
annotations (PAPSS1, tagC, AHK2), 1 is a wrong evidence code (POFUT1), 2 are wrong
references for true claims (GLN2, SYP121), 6 are unresolved queries pending full
text, 7 are false positives and 7 need no change. Remaining false positives
come mostly from synonyms UniProt does not list (DSP21) and from papers that study a
gene as an assay readout or as one of a named family ("five receptor isoforms").

Review fixes and re-sweep (2026-11-06)

Three changes from PR review, re-run over the current corpus (10,169 leads in 2,684
genes; the corpus grew since the first sweep, so the earlier tables are not directly
comparable):

• Gene symbols are compared ignoring case. UniProt writes mouse and rat symbols
  in title case (Socs3), so before this fix the same-gene-other-species signal
  could not fire for human against rodent entries. It now fires on 845 rows instead
  of 128, and the top tier grows from 406 to 448 rows.
• Protein names shared with a citing entry for another gene are not used. A
  family name such as "Tubulin alpha chain" does not say which paralog a paper
  studied. The guard changes the score of 13 rows and brings one back into the top
  tier: PAPSS1's skeletal-development row, the confirmed wrong-gene case above,
  which the protein-name penalty had buried. Names whose only evidence is an ARBA
  rule or submitter text are also dropped.
• The name check reads only the title and abstract. It stops at the next
  section heading and drops the PubMed citation line, author list (including
  bylines with "Jr" suffixes or group authors), affiliations and identifiers,
  which a symbol that is also a surname could otherwise match. 796
  abstract-only records keep their abstract under ## Content, and those are read
  too. Of the 299 full-text ## Content records, 242 open with the same PubMed
  block, and their abstract is read up to its DOI:/PMID: trailer. The other 57
  do not mark where the abstract ends, so only their title is used. The window
  change altered the name-check result for 19 rows, and the trailer rule for 5
  more.
• The review join keys on isoform and skips NOT rows, so isoform-specific rows
  no longer overwrite each other.

The top tier (score 6 or more) is now 444 rows in 225 genes; 52 of them carry the
same-gene-other-species signal. The 406, 448 and 444 here count score 6 or more on
the 2026-11-06 corpus. They are not comparable with the 466 (score 6 only) and 390
from the 2026-11-01 corpus above.

Third lead review (2026-11-10)

A fresh sweep after the byline fixes (bylines ending in NCBI's "New Collective
Author" placeholder or a lowercase group name, and "Collaborators:" paragraphs, are
now dropped) found 10,709 leads in 2,847 genes, with 460 rows in the top tier. The
next 20 top-tier genes were checked the same way as before: one gene per lead, IPI
rows skipped, genes already checked excluded, highest score first. All of a gene's
flagged rows were read, so some outcomes below come from a row other than the
highest-scoring one.

So of 20 leads: 4 are wrong-gene annotations (Rbpj, three Septins), 3 are wrong
evidence codes (PARD6B, DLG3, GNAQ), 2 are wrong references for true claims (JAK1,
PPP3CA, plus PARD6B's second row), 3 are queries, and 8 are false positives or need
no change. A new pattern this round: a direct-evidence code on a human protein from a
paper on the rodent ortholog. The same-gene-other-species signal, which the case fix
turned on, now surfaces this pattern routinely.

Across all three batches (46 leads): 7 wrong-gene annotations, 4 wrong evidence codes,
4 wrong references, 9 queries pending full text, and 22 false positives or no change.
No review action was changed on the strength of an abstract alone.

Fourth lead review (2026-11-12)

The sweep was re-run on current main: 10,811 leads. The next 20 genes were taken the same way
as before. These were the first score-6 genes not already checked (the score-7 and score-8
rows were all checked in earlier batches).

So of 20 leads: 1 is a probable wrong-gene annotation (ADAM17), 1 a wrong reference
(ALDOB), 6 are queries, and 12 are false positives. One of those false positives (spz)
turned up a confirmed wrong-term row. The false-positive rate is higher than in earlier
batches. Most of the misses come from how papers write gene names:

• spelling and diacritic variants (Spätzle, spaetzle);
• the yeast protein suffix (Cds1p) and allele marks (lsd1+/fas2+);
• a prefix on the symbol (c-Abl);
• an older name (ActRII for ACVR2A, Breathless for btl);
• a range (AKR1C1-AKR1C4);
• old genetic names that only PomBase lists (glu3 for idh1);
• a gene named only through a citation of its cloning paper (AGPAT2);
• a protein name worded differently from UniProt's (very-long-chain acyl-CoA
  dehydrogenase).

Complex-wide ComplexPortal rows and proteomics lists are a second, separate source of noise.

Across all four batches (66 leads): 8 wrong-gene annotations, 4 wrong evidence codes,
5 wrong references, 15 queries pending full text, and 34 false positives or no change.
No review action was changed on the strength of an abstract alone.

Limitations

• UniProt citation is incomplete. An entry not citing a paper does not mean the
  paper is irrelevant, and unreviewed entries (like pKM101 TraB) rarely cite papers at
  all, so cases like traB are invisible.
• The name check misses names UniProt does not list. An abstract that uses an
  unlisted synonym (DSP21 for rat Dusp21) is read as not naming the gene, which leaves
  some multi-gene papers ranked too high.
• IPI rows are noisy. An interaction paper is often cited by UniProt for only one
  partner.
• Uncached papers get no name-check adjustment.

Next steps

• Normalise names before the abstract check: fold diacritics, strip a trailing yeast p or +,
  accept c- prefixes, expand ranges such as AKR1C1-AKR1C4, and take SCHPO synonyms from
  PomBase's names file. The fourth batch suggests this would remove most of the remaining
  false positives.

• Down-weight rows whose source annotated the same term to every subunit of a complex or
  every protein in a proteomics list, which QuickGO shows by reference.

• Review the top tier, starting with the non-IPI human rows, and record confirmed
  cases as WRONG_GENE_PRODUCT source defects in the gene reviews.

• Consider scoring IPI rows separately, or requiring that the abstract names a citing
  entry's gene.
• Report the same-gene-other-species signal separately: a direct evidence code (IDA)
  on a human protein from a paper on the bovine or mouse ortholog is an evidence-code
  question (WRONG_EVIDENCE_CODE), not a wrong gene.