Oncofusion

start / graph / ana-the-numbering-check-that-passed-was-the-wrong-check

Analysisunrated confidence 0 of 5: unrated

The residue-type numbering check passed a site that is in the wrong place

also called patient phosphosite numbering re-check · isoform offset audit

We checked whether a list of measured protein sites used the same numbering we do, by asking whether each position held the right kind of amino acid. It turned out that test is too easy to pass by luck.

The same thing, for a scientist

Re-checks all 46 phosphosites in `ana-phosphosites-observed-in-kmt2a-r-aml-patients` against canonical UniProt. 44 pass a residue-type test, but only 25 are annotated phosphosites at the exact position claimed, and MLLT10 S705 passes the type test while being wrong by an isoform offset of -16. MLLT10 and MEN1 are both numbered on non-canonical isoforms and are resolved by peptide-window matching, not by offset search, which is non-identifying for both (7 and 19 candidate offsets). MEN1's canonical displayed sequence is O00255-2, not O00255-1.

Go deeper
What the earlier check did, and why it was not enough

ana-phosphosites-observed-in-kmt2a-r-aml-patients checked 20 of its sites by asking one question: is the canonical residue at that position a serine or a threonine? Two failed - MLLT10 S674 is a leucine, MEN1 S538 is a proline - and eighteen passed, which was recorded as a 10% error rate.

The test is far weaker than that framing implies. Serine and threonine together are about 11% of human protein sequence, so a wholly wrong position passes roughly one time in nine. Eighteen passes are not eighteen confirmations.

The Q25 beat proved it on a specific site. MLLT10 S705 passed the check and is still wrong. The source table numbers on a non-canonical isoform with a constant offset of −16; the real site is canonical S689, which is independently an annotated phosphosite. Canonical 705 is a serine in an unrelated sequence context - QIRYDQPGNSSLENLPPVAASI against the true SSPRGSLSPRSPVSSLQIRYD. The site that failed the check revealed the offset; the site that passed it is at the wrong position.

What the stronger check finds across all 46 sites

The re-run covers every site in the entry, not the 20 first checked, and asks two questions instead of one: is it the right residue type, and is that exact position annotated as a phosphosite by UniProt itself?

proteinsitespass typeannotated in UniProtoutright failure
PSIP112129-
KMT2A11116-
AFF4773-
MLLT1442-
AFF1332-
MLLT3332-
ELL220-
MLLT10210S674
MEN1100S538
DOT1L111-
total4644252

Twenty-five of forty-six sites are corroborated by an independent database at the exact position claimed. That is the real support level, against the 44 the residue-type test reports. The eight proteins with no outright failure all number consistently with canonical, which is the reassuring half of the result: the offset problem is confined to MLLT10 and MEN1, not spread through the table.

ELL is the one quiet flag. Both its sites pass on type and neither is annotated anywhere in UniProt, and ELL has no outright failure to expose an offset if one exists. It is the protein in this table whose numbering is least attested.

An offset search cannot resolve a single-site protein; only the peptide window can

The natural repair for an isoform mismatch is to look for one constant offset that makes every site in that protein land on a serine or threonine. It works when there are several sites and fails silently when there is one.

  • MLLT10, two sites: 7 candidate offsets fit, including the correct −16.
  • MEN1, one site: 19 candidate offsets fit - −51, −32, −26, −17, −13, −5, −4, +3, +5, +14, +17, +30, +34, +35, +42, +45, +55, +56, +58.

Neither search identifies anything. My first version of this script picked the offset with the most UniProt-annotated hits and returned −51 for MEN1 - wrong, and wrong confidently, because canonical 487 happens to be a real phosphosite. The script now reports all candidates and marks whether the search is identifying at all.

Both offsets were settled by matching the source's peptide window against the sequence, which is unambiguous. For MEN1 the window AGTARGPEGGSTAQVPAPTASP occurs exactly once, at canonical 523-544, putting the site at S533 - offset −5.

And the isoform naming is a trap in its own right. The source numbers on O00255-1, which is not UniProt's canonical sequence: the displayed canonical for MEN1 is the 610-residue short isoform O00255-2, and O00255-1 is the 615-residue long form. The -1 suffix does not mean canonical. Anyone who assumes it does gets a five-residue shift and no warning.

Still open
  1. Resolve the remaining 21 residue-type-only sites by peptide window from MOESM3 rather than by residue type. The windows are in the supplementary file already downloaded.
  2. Are ELL S308 and S437 real? Both pass on type, neither is annotated in UniProt, and ELL has no failing sibling site to expose an offset if one exists.
  3. Does the source's search database use non-canonical isoforms systematically, or only for MLLT10 and MEN1? Two hits out of ten proteins suggests a mixed build rather than a policy.
Every source behind this entry (2)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    Re-checking all 46 patient phosphosites against canonical UniProt sequences finds 44 pass a residue-type test but only 25 are annotated as phosphosites by UniProt at the exact position claimed. Two fail outright - MLLT10 S674 and MEN1 S538 - and both belong to a constant isoform offset rather than to a fabricated site.

    What this evidence cannot show

    Run 2026-09-19 against UniProt REST, with each accession's primary gene name asserted against the expected symbol before use. **UniProt annotation is an imperfect gold standard**: it is curated from published low-throughput work and lags large proteomic surveys, so an unannotated position is not thereby wrong - ELL's two sites are unannotated and may simply be under-curated. The converse also holds: an annotated position can coincide with a wrong one by chance, as MEN1 487 does at offset −51.

    Evidence that argues against this (1)
    • **This analysis does not rescue the table's numbering; it bounds how much of it is checked.** Twenty-one of forty-six sites rest on the residue-type test alone, which passes a wrong position about one time in nine. The only sound way to resolve any of them is the peptide window from the source file, and that has been done for exactly two. Oncofusion knowledge graph (internal analysis) 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    UniProt's canonical displayed sequence for MEN1 is the 610-residue isoform O00255-2, while O00255-1 is the 615-residue long isoform. The isoform numbered -1 is not the canonical one.

    What this evidence cannot show

    Read from the entry's ALTERNATIVE PRODUCTS record 2026-09-19: isoform named '1' (synonym Short) carries isoformId O00255-2 with sequence status 'Displayed'; isoform named '2' (synonym Long) carries O00255-1, status 'Described'. So even the isoform *names* and the isoform *ids* disagree with each other here. This is a property of this entry at this date; UniProt can reassign a displayed isoform between releases, which would silently change every position in this table.

    Evidence that argues against this (1)
    • For most human proteins the -1 suffix *does* denote the canonical sequence, so the general habit is not irrational - it is a habit with a failure rate. The defence is not to memorise exceptions but to read back the length of whatever sequence was actually fetched, which would have caught this in one line. UniProt Consortium 2026
Connected to