UniProt Proteome Removal Impact Assessment

Bottom line: UniProt planned to drop about 58M unreviewed (TrEMBL)
entries from redundant and non-reference proteomes by release 2026_02,
archiving them in UniParc, which would leave any gene review keyed on one of
those accessions pointing at an entry that no longer exists. We downloaded
UniProt's explicit removal list and on 2026-02-03 checked all 896 UniProt
IDs then in the repo against it:
15 reviews, all non-model microbes, phage, influenza and mosquito proteins,
were on the list. None of the 15 has been fixed or archived yet. A lookup
against the UniProt REST API on 2026-09-26 shows five of them are now
deleted from UniProtKB (PSEAI merB, ACEPA xdhB, RUMJO fae1A, ECO57 stx2A,
PSEEN Q1IFG0) and the other ten are still live. The repo has since grown to
about 4,990 gene folders, so the February scan no longer covers it and
should be re-run; the just targets listed under Tooling are not in the
current justfiles and would need to be restored first.

A replacement finder (PROTEOME_REMOVAL/scripts/resolve_deleted_accessions.py,
added 2026-10-09) now finds a current entry for three of the five deleted
reviews: stx2A and Q1IFG0 have identical-sequence active entries, and xdhB
has a near-identical reference-proteome match. merB and fae1A have none.

We did this because a review's primary key is its UniProt accession: when
the entry disappears, just fetch-gene and the validators can no longer
refresh the review, and the curated judgement is stranded.

Overview

UniProt is removing TrEMBL entries from non-Reference Proteomes by release 2026_02.

Sources:
- Protein removal list: https://ftp.ebi.ac.uk/pub/contrib/UniProt/proteomes/proteins_to_remove_from_UniProtKB.txt
- Proteins retained: https://ftp.ebi.ac.uk/pub/contrib/UniProt/proteomes/proteins_retained_in_UniProtKB.txt
- Help doc: https://www.uniprot.org/help/proteome_redundancy

What happens:
- TrEMBL entries from "Redundant Proteomes" and "Other Proteomes" will be removed from UniProtKB
- They will be archived in UniParc (still accessible via UniParc, but not in UniProtKB)
- SwissProt entries are NOT affected
- ~253M entries dropping to ~141M entries (~58M being removed)

Impact Summary

VERIFIED against proteins_to_remove_from_UniProtKB.txt (58M entries):

We have 15 entries that WILL BE REMOVED from UniProtKB.

Entries Being Removed

Action Required

• [ ] Find alternative entries or archive affected reviews (replacements
  found for 3 of the 5 deleted entries; see "Finding replacements" below)
• [ ] Consider if UniParc references are acceptable for these use cases

Tooling

Download the removal list and check entries:

just fetch-uniprot-removal-list       # Download ~578MB file to cache/uniprot/
just check-uniprot-removal ID1 ID2    # Check specific IDs
just check-all-uniprot-removal        # Check all gene reviews

Finding replacements

scripts/resolve_deleted_accessions.py takes any list of accessions, or a
TSV column of them, and for each reports, in order of preference:

1. active: still in UniProtKB.
2. identical_active: UniParc lists an active UniProtKB entry with the
   identical sequence. UniParc marks archived members with a version suffix
   (A0A2S9ZZK0.1) and active ones without (P09385); reviewed entries are
   preferred, then the same taxon.
3. same_gene (≥ 90% identity over ≥ 80% of the query),
   same_gene_partial_model (≥ 95% identity over a shorter span of at least
   100 residues, or half the query if that is shorter, as when the gene models
   differ in length) or weak: best phmmer hit of the archived
   sequence against the reference proteome(s) of its species, found
   automatically (taxonomy_id:<species> AND reference:true) or given with
   --proteome.
4. no_hit, no_reference_proteome, no_sequence or not_found.

uv run --with pyhmmer python projects/PROTEOME_REMOVAL/scripts/resolve_deleted_accessions.py \
    A0A1V0M5B3 Q1IFG0 --out resolved.tsv

The 15 affected reviews (2026-10-09)

Output: data/removed_review_replacements.tsv. Ten accessions are still
active. For the five deleted ones:

Re-keying stx2A, Q1IFG0 and xdhB onto these accessions is a curation
decision (a new id, just fetch-gene, re-checking the existing
annotations) and has not been done here. merB and fae1A can only stay keyed
on UniParc or be archived.

Other users of the script

• projects/FUNGAL_PHENOTYPES/RHOTO.md uses it to map 370
  Rhodotorula toruloides genes whose published accessions (IFO0880,
  UP000239560) were removed in the cleanup; 346 resolve.
• projects/GENE_MODEL_ERRORS.md uses its output to find fused gene models:
  a replacement much longer than the archived sequence is the typical
  signature.

STATUS

• [x] Download protein removal list (58M entries, 578MB)
• [x] Set up cache directory and just targets
• [x] Comprehensive check of all 896 gene review UniProt IDs
• [x] Identify 15 entries being removed
• [x] Script to find replacement entries (scripts/resolve_deleted_accessions.py)
• [ ] Find alternative entries or archive affected reviews (3 of 5 deleted have replacements)

NOTES

2026-02-03

Methodology

1. Downloaded proteins_to_remove_from_UniProtKB.txt (578MB, 58M entries)
2. Created cache/uniprot/ directory (gitignored)
3. Added just targets for automated checking
4. Extracted all 896 UniProt IDs from gene reviews
5. Used LC_ALL=C comm -12 for fast intersection (locale matters for sort order!)
6. Found 15 entries in removal list

Lessons Learned

• Proteome-based inference is unreliable - use the explicit protein removal list
• Sort locale matters: must use LC_ALL=C for consistent results with comm
• The removal list is already sorted (alphanumeric), so comm is O(n) - much faster than grep

Full Verification Output

$ just check-all-uniprot-removal
Found 896 unique UniProt IDs in our reviews
Checking against removal list (58M entries)...

Entries being removed from UniProtKB:
  A0A1S7IWC7: 9INFA/M2
  A0A1V0M5B3: PSEAI/merB
  A0A1Y0Y121: ACEPA/xdhB
  A0A2Z5TSL2: RUMJO/fae1A
  A0A9Q6Z964: ECO57/stx2A
  D2STP8: ANOGA/PGRPS3
  D2SU82: ANOGA/PGRPS2
  D9J262: NITRP/amoA
  O30741: CERSP/dorR
  Q1IFG0: PSEEN/Q1IFG0
  Q6DTY2: CLOCL/Q6DTY2
  Q9RGE6: CLOCL/Q9RGE6
  Q9RGE7: CLOCL/Q9RGE7
  Q9RGE8: CLOCL/Q9RGE8
  Q9XJG2: 9CAUD/darB

Summary: 15 will be removed, 881 safe