The six unwritten DOT1L partners are not six findings - they are nine spokes of ONE GFP-DOT1L pulldown, and the source paper already graded them

Title, twelve-author list, journal, year and DOI copied character-by-character out of a live NCBI efetch XML record fetched 2026-09-21, reading only PubmedData/ArticleIdList/ArticleId (brief rule 81). Abstract read in full; the main text is NOT readable - no PMCID exists, Europe PMC reports isOpenAccess N, and the DOI resolves to a 13-character Elsevier redirect stub (brief rule 60, the loud Elsevier failure). Everything quantitative in this entry therefore comes from the supplementary workbooks, which ARE served, and not from the narrative.

• A proteome-scale atlas is not a weak paper - it is a strong paper being asked a question it was not designed to answer. Hein et al. set out to describe global network topology, for which a false-positive rate that is fatal to a single-partner claim is immaterial. Reading its lowest-confidence rows as six partner assignments is an error of use here, not a defect in the source. Hein MY et al. 2015

One database on one day. Spoke expansion is IntAct's honest representation of an n-ary AP-MS experiment, not a curation error - but it means the binary pair 'DOT1L binds X' was INFERRED by the curator from co-purification and never observed as a binary interaction. MI:0914 'association' is the weakest interaction-type term in the vocabulary; it is not MI:0915 'physical association' and not MI:0407 'direct interaction'. The other six publications in the DOT1L record were not audited to this depth.

• Appearing once in IntAct bounds the CURATED record, not the literature. A partner reported only in a supplementary table of an uncurated study is invisible to this count, the same blind spot brief rule 80 documents for PubMed abstracts. EMBL-EBI IntAct 2026

Downloaded 2026-09-21 from the Elsevier static CDN with the magic bytes checked (PK, a valid zip) after the article HTML itself returned 403. The table numbering is taken from IntAct's own annotation of this experiment ('Suppl. table S2. For bait details see supp. table S1'), not from the paper's narrative, which is unreadable. The classes A+/A/A-/B+/B/B-/C are the authors' composite of enrichment, p value and prey-bait correlation; the exact thresholds are defined in the main text, which I could not read, so the class is used here as an ORDINAL grade published by the authors and not as a quantity I can re-derive.

• A confidence class is the authors' judgement, not a measurement, and it was calibrated for a network-wide analysis in which a low-class edge is a tolerable cost. It is being used here in a way its designers did not intend - to argue about individual partners - and a B- edge that is real is graded exactly like a B- edge that is not. Hein MY et al. 2015

A reciprocal AP-MS failure is weaker than it looks. Tag position, bait expression level (IBTK was expressed roughly ten-fold higher than DOT1L in this dataset) and the stoichiometry of the pair all make asymmetric recovery common, and a 1:476 prey would sit at the detection floor in either direction. This bounds the claim to 'not recovered in the reciprocal arm of this study', not to 'does not interact'.

• The reciprocal arm is also the only place the DOT1L-IBTK pair could have been CONFIRMED, and the same design that makes a negative weak would have made a positive strong. Treating a failed reciprocal as uninformative while treating the forward spoke as evidence is not a defensible asymmetry. Cell 163(3), supplementary material 2015

Bounded by titles and abstracts only (brief rule 80): a co-IP shown in a supporting figure of a paper about something else is unreachable from any of these queries, and that is precisely where a confirmation would be likely to sit. querytranslation was read back on every query; the only surprise was that 'nucleolar protein p40' expanded loosely into a MeSH nuclear-proteins set, which is why the single hit was title-checked and rejected.

• Zero literature is the expected result for ANY prey of a 2015 proteome-scale atlas, including the true ones, because nobody writes a title about a single row of somebody else's screen. The absence discriminates weakly and would look identical if all six were real but unfollowed. NCBI PubMed (queries run 2026-09-21) 2026

A whole-proteome workbook of 8,805 groups is not exhaustive - absence from it is absence from a three-replicate HeLa measurement, not from the cell. Copy numbers are proteomic-ruler estimates and the workbook carries no error bars for them. The near-endogenous bait level is a genuine strength of this dataset and is the reason its stoichiometries mean anything at all; it does not transfer to AP-MS studies using CMV-driven overexpression.

• Copy number cuts both ways. An abundant nucleolar protein is the classic AP-MS background, but it is also genuinely more likely to encounter any given nuclear protein, so a high-abundance prey is not automatically a false positive - the step from 459-fold excess to 'contaminant' is a prior, not a measurement. Cell 163(3), supplementary material 2015

Not my analysis - it was run by the coordinator during this beat and is quoted here, not re-derived. Its own first pass misclassified five focused complex-purification papers as screens on a partner-count heuristic (MED26 21729782, AF4/ENL/P-TEFb 20153263, DotCom 20203130, PAF/MLL 20541477, Yokoyama 15199122); the quoted figure is from the hand-verified replacement. I independently verified only the part that bears on this entry - that DotCom (20203130) appears in IntAct as a separate publication and shares none of my six.

• 73.2% describes the partners that HAVE entries as much as the ones that do not, and a filter can be 73% sound while still being the wrong instrument for the 27% where all the unwritten work sits. The figure should not be read as licence to keep using bait status as a quality grade. Oncofusion knowledge graph (internal analysis) 2026

Resolved 2026-09-21 by accession query with the primary gene name read back as BRINP1 and the length checked (brief rules 62 and 65). The DBC1 synonym is a live collision with CCAR2, whose common name is also DBC1 (Deleted in Breast Cancer 1); the Hein table uses DBC1 as the gene name for this row, so a reader resolving that string to CCAR2 gets a completely different, nuclear, SIRT1-regulating protein.

• A cytoplasmic UniProt call is exactly the kind of annotation this graph has found to be miscalled - see localization-partner-baseline-compartment-is-often-miscalled. 'Cytoplasmic, therefore not a DOT1L partner' would be the same error in the opposite direction, which is why the refusal here rests on the prey's absence from the assayed proteome and not on its compartment. UniProt Consortium 2026