Gene model errors: fused UniProt entries

Bottom line: some UniProt entries are not one protein but two or more
neighbouring genes joined by a faulty gene prediction. Such an entry takes
its name, and often its GO terms, from one half, so the other gene's function
is mislabelled or lost. Working on Rhodotorula toruloides fitness data
(FUNGAL_PHENOTYPES/RHOTO), we tested 58
reference-proteome entries that are at least 1.3 times longer than the
matching protein of a second strain's assembly, and found 26 that are
mostly covered by two to four separate genes of that assembly. In all 26 the
genes are neighbours. In 15, transposon mutants show that the parts act
independently: mutating one part gives a strong growth defect in some
condition while mutating its neighbour has no effect. Examples include a
3-methylcrotonyl-CoA carboxylase named "Nuclear pore complex protein Nup85"
and a branched-chain keto acid dehydrogenase fused to an amino acid
transporter. The repository already had one fused gene model
(a Miscanthus "UMP-CMP kinase") and two truncated ones, all recorded only in
review prose. Gene reviews now have a structured gene_model_issues field
for such problems, and those three reviews have been back-filled.

Why it matters for annotation review

A gene review is keyed on a UniProt accession and assumes it names one gene
product. For a fused entry that assumption fails:

• the protein name, and GO terms transferred from it, may describe the wrong
  half (A0A0K3CAZ0 is annotated to the nuclear pore, from a Nup85-like
  segment, and to methylcrotonoyl-CoA carboxylase activity, from the other);
• IBA/IEA annotations propagated by family membership can come from either
  half, so the entry can carry an incoherent mixture;
• a phenotype or experiment on the real gene attaches to an entry most of
  whose sequence is a different protein.

Existing reviews with gene model problems

A text search of all *-ai-review.yaml files for gene-model language
("chimeric", "fused", "truncated gene model", "incomplete gene model" and
similar phrases) found these, each recorded only in the description or review
prose. All now carry gene_model_issues:

A fourth review, HORSE/CTDSP2, was first back-filled as a truncated model.
Its 2026-10 re-review shows the entry is the product of a real alternative
Ensembl transcript of a correctly modelled gene, with a premature stop and
full-length sibling entries (A0A9L0T5X1, A0A9L0R207). That is a probably
non-productive isoform, not a gene model error, so it carries no
gene_model_issues entry.

The Q2U1U6 analysis also found a fused model in another A. oryzae
assembly: A0ABQ6LHX9 (A. oryzae var. brunneus, 745 residues) carries the
Q2U1U6 sequence as residues 612–745 behind an unrelated beta-lactamase-like
domain.

For contrast, 9POAL/NCGR_LOCUS67308 is tagged chimeric-gene-model, but it
is a real biological fusion (the grass SDH2-RPS14 locus), not an annotation
error. That is the only gene-model tag in the corpus, and its meaning
differs from the cases here, which is part of the case for a controlled
field.

Scope. Entries that UniProt itself marks as fragments (DE Flags: Fragment, with NON_TER features; for example
9AVES/TAS1R1, 9AVES/TAS1R3 and MONBE/MONBRDRAFT_1233) were deliberately not
back-filled. UniProt already records that their sequence is incomplete, in a
machine-readable form; gene_model_issues is for problems the source record
does not flag, such as a mispredicted start, missing exons or fused
neighbours. A fragment entry should get a gene_model_issues entry only when
the review finds something beyond the flag.

Method

1. Find candidates. PROTEOME_REMOVAL/scripts/resolve_deleted_accessions.py
   maps each archived accession from the second assembly (here the IFO0880
   strain, UP000239560, deleted from UniProtKB) to its best match in the
   reference proteome (UP000199069). A current entry much longer than the
   sequence it replaced is the typical signature of a fusion.
2. Tile. scripts/tile_gene_model.py searches each candidate with phmmer
   against the second assembly's proteome, streamed from UniParc. It calls
   FUSION when at least two different proteins each match ≥ 100 residues at
   ≥ 90% identity, overlap by ≤ 30 residues, and together cover ≥ 70% of the
   entry. It also reports which InterPro domains fall in each segment.
3. Check the genetics. scripts/rhoto_fusion_genetics.py maps each part
   to its IFO0880 gene and RB-TDNAseq fitness profile, and records whether
   the parts are adjacent genes and whether they are genetically separated:
   in some condition one part's mutants have fitness ≤ −2.5 while another
   part's mutants have fitness > −1.

The tiling call is direction-neutral: it cannot by itself say whether the
reference entry is fused or the other assembly's genes are split. Two things
point to fusion. First, the parts often carry domains from unrelated families
(Nup85 and biotin carboxylase; E2 ubiquitin-conjugating enzyme and isocitrate
lyase; protein kinase and phosphoglucomutase). Second, genetic separation
shows that the parts act independently.

Results (R. toruloides)

All 26 fusions have adjacent component genes; 15 are also genetically
separated. These 15 are listed in this page's metadata.

The two isocitrate lyase-family entries are worth a closer look. The part of
A0A0K3CD57 ("methylisocitrate lyase") is needed on oleate and acetate, as
the glyoxylate-cycle isocitrate lyase would be; the part of A0A0K3CIB4
("Isocitrate lyase") is needed on valine, whose breakdown yields
propionyl-CoA and runs through the methylcitrate cycle. The phenotypes
therefore suggest the two names are swapped. This is a hypothesis from
growth data, not an established assignment.

Scope and caveats

• Only entries linked to a carbon-source fitness defect were screened (58 of
  them), so 26 fusions is not an estimate of the rate in the whole
  reference proteome; a proteome-wide screen is the obvious next step.
• INCONCLUSIVE (20) covers entries tiled by one protein over too little of
  their length, or by segments below the identity threshold. Some will be
  fusions with a component missing from the IFO0880 annotation.
• Genetic separation needs a condition where one part has a phenotype, so
  fusions of two unphenotyped genes cannot be confirmed this way.

Recording gene model problems: gene_model_issues

Gene reviews now have a gene_model_issues field (added 2026-10-09; schema
classes GeneModelIssue and GeneModelComponent). Each entry records:

• issue_type: FUSED_GENE_MODEL, SPLIT_GENE_MODEL, TRUNCATED_GENE_MODEL,
  WRONG_START_SITE or OTHER. A genuine biological fusion, such as the
  grass SDH2-RPS14 locus, is not a gene model issue.
• status: SUSPECTED (one line of evidence) or CONFIRMED (independent
  lines, such as tiling plus genetic separation).
• description, plus optional components (accession, segment and
  description of each underlying gene) and reviewed_segment (the residues
  the review's functional statements refer to).
• evidence_types: TILING, ADJACENT_GENES, GENETIC_SEPARATION,
  DOMAIN_ARCHITECTURE, ORTHOLOG_COMPARISON, TRANSCRIPT_EVIDENCE,
  GENOMIC_SEQUENCE or OTHER.
• supported_by: exact quotes, checked like any other supporting_text.

The rendered review shows the issues in a "Gene Model Issues" section under
the description. The three existing reviews above have been back-filled. For a fused R. toruloides entry, a review would look like:

gene_model_issues:
- issue_type: FUSED_GENE_MODEL
  status: CONFIRMED
  description: >-
    Joins the Nup85-like nucleoporin gene to the neighbouring
    3-methylcrotonyl-CoA carboxylase biotin-carboxylase gene; named after
    the Nup85 part. Only the carboxylase part is reviewed here.
  components:
  - accession: A0A2T0ACV7
    segment: 1-650
    description: Nup85-like nucleoporin (IFO0880 RTO4_12868)
  - accession: A0A2T0ACW0
    segment: 645-1419
    description: MCC biotin-carboxylase subunit (IFO0880 RTO4_12867)
  reviewed_segment: 645-1419
  evidence_types: [TILING, ADJACENT_GENES, GENETIC_SEPARATION]
  supported_by:
  - reference_id: file:projects/GENE_MODEL_ERRORS/data/rhoto_fusion_genetics.tsv
    supporting_text: "645-1419=RTO4_12867"

data/rhoto_fusion_genetics.tsv remains the pipeline record for the 26
R. toruloides fusions until they have gene reviews; this page's genes
metadata lists the three existing reviews and the 15 genetically separated
entries.

Next steps

1. Use gene_model_issues in the first R. toruloides gene reviews, since
   several candidates (BCKDH E1α, MCC α, the aromatic dioxygenase) sit in
   fused entries.
2. Run the tiling screen across the whole UP000199069 reference proteome
   against IFO0880, not just phenotype-linked entries.
3. Report the 15 genetically separated fusions to UniProt.
4. Look for fusions in reviewed species by the same route: any organism with
   two independent assemblies in UniParc can be screened.

Files

• scripts/tile_gene_model.py (needs uv run --with pyhmmer) →
  data/rhoto_fusion_screen.tsv
• scripts/rhoto_fusion_genetics.py → data/rhoto_fusion_genetics.tsv

Reproduce from the repository root after the FUNGAL_PHENOTYPES/RHOTO pipeline
(which caches the supplements and writes the resolution table):

uv run --with pyhmmer python projects/GENE_MODEL_ERRORS/scripts/tile_gene_model.py \
    --from-resolution projects/FUNGAL_PHENOTYPES/data/rhoto_accession_resolution.tsv \
    --comparison-proteome UP000239560 \
    --out projects/GENE_MODEL_ERRORS/data/rhoto_fusion_screen.tsv
uv run python projects/GENE_MODEL_ERRORS/scripts/rhoto_fusion_genetics.py