A mass-spectrometry protein group's leading identifier is not the interactor, and curating it puts dubious isoforms and unreviewed accessions into the interaction databases
also called MaxQuant protein group leading ID · razor protein trap · protein-group leading identifier
A mass spectrometer usually cannot tell apart the several versions of a protein that share the peptides it measured, so software reports a whole list and puts one of them first. Interaction databases store that first one as if it were the answer, and it is often a splice variant or an unreviewed entry rather than the real protein.
The same thing, for a scientist
A search engine such as MaxQuant reports a **protein group**: the set of database sequences that the observed peptides cannot distinguish. The group's leading identifier is chosen by peptide count and ordering heuristics, **not by evidence**. IMEx curators, correctly and by convention, record the leading identifier as the interactor. The consequence for a consumer is that an interaction database's accession is frequently not the canonical reviewed protein. **Three of the six DOT1L partners from PMID 26496610 demonstrate three different flavours of it in one pulldown** [E1]: NIN is recorded as `Q8N4C6-10`, which IntAct annotates on the same row as *'Dubious isoform produced through aberrant splice sites'*, from a thirteen-member group that contains the canonical Q8N4C6; BRINP1 is recorded as `O60477-3`, isoform 3, from a two-member group containing canonical O60477; IBTK is recorded as **`E7EPI0`, an UNREVIEWED TrEMBL entry of 1,338 aa**, from a four-member group containing the reviewed Q9P2D0 of 1,353 aa [E1][E2]. **This project already carried E7EPI0 as IBTK's accession** in `analyses/directed-interactome-coverage-recheck`, one automated write away from a gene entry keyed on an unreviewed identifier [E3].
Why the existing defences do not catch it
This graph has a long series of identifier rules and none of them fires on this class:
verify-accessions.py(accession pluslength_aa) passes. The trap does not produce a mismatched pair - E7EPI0 really is 1,338 aa. The accession and the length agree with each other and both are wrong for the protein the entry means. This is the same weakness brief rule 69 part 2 describes, arriving from a new direction.- Rule 62 (several reviewed records, take the primary) does not apply. There is only one reviewed record; the problem is that the database is not pointing at it.
- Rule 65 (read the protein name back) passes cleanly. E7EPI0's protein name is 'Inhibitor of Bruton tyrosine kinase' and Q8N4C6-10's parent is ninein. The readback is correct and the identifier is still not the one you want.
- Rule 66's length discriminator is unavailable. Q9P2D0 and E7EPI0 differ by 15 residues, well inside the range rule 69 measured as undiscriminating.
The only check that catches it is reading the accession's reviewed status and the full protein group, which means looking at MITAB columns 1-4 together rather than taking the accession alone.
What to do
- Never key an entry on an accession taken from an interaction database without checking its reviewed status. One UniProt query with
reviewedinfieldsanswers it. - Record the group, not just the representative. Store the source database's accession in a field such as
data.source_database_accessionwith astatusand awhy, and put the canonical reviewed accession indata.uniprot. Both entries written in this beat that hit the trap do exactly that. - Treat an isoform suffix in an interaction record as a flag, not as information.
-3,-10and so on in a MITAB interactor column usually mean 'this was the group leader', not 'this isoform was identified'. - When the group spans isoforms with different compartments, say the compartment is unresolved. IBTK's group contains Q9P2D0-2, which UniProt annotates as nuclear while the canonical protein is cytoplasmic - so any compartment argument about that row is undecidable from the data.
- The general form, which is the part worth keeping: an identifier chosen by software ordering carries no evidence, and a downstream system that treats it as a choice will make confident statements about objects nobody measured. Same family as brief rule 69b's fallback that constructs the value it is looking for.
| the check that does fire | request `reviewed` in the UniProt fields and read the whole protein group from MITAB columns 1-4 or the source table's prey.IDs, not the representative accession alone |
|---|---|
| rate measured | 3 of 9 spokes in one pulldown; corpus-wide rate NOT measured |
instances found — the numbers behind the chart
[
{
"gene": "NIN",
"source_accession": "Q8N4C6-10",
"flavour": "UniProt isoform annotated by IntAct as a dubious isoform produced through aberrant splice sites",
"group_size": 13,
"canonical": "Q8N4C6",
"canonical_length_aa": 2090
},
{
"gene": "BRINP1",
"source_accession": "O60477-3",
"flavour": "UniProt isoform",
"group_size": 2,
"canonical": "O60477",
"canonical_length_aa": 761
},
{
"gene": "IBTK",
"source_accession": "E7EPI0",
"flavour": "unreviewed TrEMBL entry",
"group_size": 4,
"canonical": "Q9P2D0",
"canonical_length_aa": 1353,
"trembl_length_aa": 1338
}
]checks that do not fire
[ "verify-accessions.py - accession and length agree with each other, so the pair is internally consistent", "brief rule 62 - only one reviewed record exists; the problem is that the database does not point at it", "brief rule 65 - reading the protein name back returns the correct protein name", "brief rule 66 length discriminator - Q9P2D0 vs E7EPI0 differ by 15 aa" ]
- Stated too strongly this reads as a criticism of IntAct curation, which would be wrong and would discourage use of the best interaction resource available. The failure is in naive consumption.
- A rule of 'never use an unreviewed accession' would be worse than the problem, because for many proteins TrEMBL is the only record and refusing it silently drops them - the same shape as brief rule 78's wholesale refusal of multi-mapping symbols.
- How many of this graph's partner entries carry an accession inherited from an interaction database rather than resolved from UniProt? A sweep of data.uniprot against reviewed status would answer it in one pass and has never been run.
- Does IntAct expose the full protein group anywhere, or only the representative? If only the representative, then the group can be recovered solely from the source paper's supplementary table, which makes the check dependent on the paper being readable.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
In IntAct's MITAB export for DOT1L, three of the six partners from PMID 26496610 are recorded under non-canonical accessions that are the leading identifiers of MaxQuant protein groups: NIN as Q8N4C6-10 carrying the row annotation 'Dubious isoform produced through aberrant splice sites', BRINP1 as O60477-3, and IBTK as E7EPI0; the corresponding source protein groups are Q8N4C6-10;...;Q8N4C6;... (13 members), O60477-3;O60477 (2 members) and E7EPI0;Q9P2D0;Q9P2D0-2;E9PDR5 (4 members), each containing the canonical accession.
IntAct PSICQUIC MITAB 2.7 export for id:Q8TEK3 (DOT1L), 49 rows, fetched 2026-09-21, read alongside Hein et al. 2015 supplementary Table S2 prey.IDs column · EMBL-EBI IntAct 2026What this evidence cannot show
Three instances in one pulldown is a demonstration, not a rate. Nobody has measured how often this happens across this graph's partner entries, and the honest bound is that it happened in 3 of 9 spokes of one experiment. The MaxQuant group ordering heuristic is described here from general knowledge of the tool, not from the paper's methods, which are unreadable.
Evidence that argues against this (1)
- IMEx curation of the leading group identifier is deliberate and defensible - recording four accessions as four interactors would inflate every AP-MS record several-fold. The databases are behaving correctly and the failure is entirely in naive downstream consumption, so a rule phrased as a criticism of IntAct would be wrong. EMBL-EBI IntAct 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
UniProt returns E7EPI0 as an UNREVIEWED (TrEMBL) human IBTK record of 1,338 aa whose only subcellular-location evidence is automatic ARBA annotation, against the reviewed Q9P2D0 (IBTK_HUMAN) of 1,353 aa with curated location comments including a separate nuclear call for isoform 2 - a 15-residue difference that no length-based check can discriminate.
UniProtKB Q9P2D0 (reviewed, 1353 aa) and E7EPI0 (unreviewed, 1338 aa), both human IBTK · UniProt Consortium 2026What this evidence cannot show
Both records fetched 2026-09-21 in a single query requesting accession, reviewed status, protein name, gene_primary, length and subcellular location, and read back. TrEMBL records are automatically annotated and are not wrong - they are simply uncurated, and for many proteins they are the only record available, so 'unreviewed' is a caution rather than a disqualification.
Evidence that argues against this (1)
- For proteins with no Swiss-Prot record at all, a TrEMBL accession is the correct and only identifier, so a blanket rule against unreviewed accessions would be wrong. The rule has to be 'check, and say which you chose and why', not 'never use TrEMBL'. UniProt Consortium 2026
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
This project's own directed-interactome coverage recheck already carries E7EPI0 as IBTK's accession and O60477 as BRINP1's in its hand-written RESOLVED dictionary, so the unreviewed identifier was one automated entry-writing pass away from becoming a node's data.uniprot.
Is '100% of the directed-evidence interactome is written' still true? No. It is 68.5%. · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
The recheck's dictionary is hand-verified and explicitly flags its non-standard rows as 'the point, not an oversight', so this is not a defect in that analysis - it is an observation that the identifier it correctly inherited from IntAct is not the one an entry should be keyed on. No entry was in fact written from it; the hazard is prospective.
Evidence that argues against this (1)
- The recheck resolved E7EPI0 to the symbol IBTK and joined on symbol as well as accession, so its coverage figure is unaffected by this trap - the hazard is confined to entry-writing, and the analysis should not be read as having made an error. Oncofusion knowledge graph (internal analysis) 2026
INHERITS THE TRAP. Preys are joined to copy number on the first accession in the protein group that appears in the proteome, which is the MaxQuant leading-identifier convention this concept entry names. Where the leading id is not the interactor, this analysis joins the wrong protein's abundance - and that is a reason its per-row numbers are softer than its aggregate conclusion.
NIN is recorded in IntAct as Q8N4C6-10, a dubious splice isoform, because it led the protein group
IBTK is recorded in IntAct and in this project's coverage recheck as E7EPI0, an unreviewed TrEMBL accession, because it led the protein group
The identifier half of the same audit: three of nine spokes in this pulldown carry a non-canonical accession