Counting laboratories instead of papers demoted nothing, because BioPlex 2.0 and 3.0 have different last authors - and the 3' partners' own neighbourhoods are the emptiest part of the graph

Read down that column and the 3' partners look better covered than KMT2A itself. They are not. A partner shared with the core scaffold was already written for the core, and it counts again for every bait it touches.

Partition by which baits a partner appears against and the illusion goes:

• partners appearing only against core-scaffold baits: 150, 10.0% covered
• partners appearing only against 3' partner baits: 100, 6.0% covered

So the private neighbourhood of MLLT1, MLLT3, MLLT10, AFF3, ELL and ELL2 - a hundred proteins - is the emptiest region measured so far. That is also where a genuine partner-specific mechanism would live, which is the entire premise of this graph's comparative beats: KMT2A::MLLT3 differs from KMT2A::AFF1 because MLLT3 is not AFF1.

MEN1 remains last of all twelve at 16.7%.

The first run ranked by folded last author and printed 0 of the paper-count queue are DEMOTED by the laboratory rule. That is not a null result about the biology; it is the instrument failing.

BioPlex 2.0 (PMID 28514442) ends Huttlin ... Harper. BioPlex 3.0 (PMID 33961781) ends Huttlin ... Gygi. Same first author, same platform, same collaboration, different last author - so the proxy scored two laboratories and every BioPlex-only partner kept its promotion. The queue even displayed the failure in plain sight: fifteen rows carried the justification "gygi sp, harper jw".

ana-source-independence-by-lab had already met this and solved it - it merges two papers whose author sets overlap above a threshold, which is what a rotating senior-author slot looks like from outside. Reusing its constants unchanged (0.34 Jaccard, 3 shared names) rather than tuning new ones, so this is the same instrument and not a second one: 50 last authors collapse to 47 groups, and PIP4K2A, CAMKV, DAXX, DENND2D, ARRB2 and the rest of the BioPlex block drop to labs=1 while still listing both last authors.

Eight rows demoted: EXOSC3, PBX2, RSBN1, RYBP, SNRPA1, SNX24, TCP10L, YAF2.

The transferable lesson is not about BioPlex. A proxy fails hardest on the case that motivated it, because that case is the one selected for being unusual - and a correction that reports no change is the shape of a correction that did not run, not of a correction that was unnecessary.

The queue after the merge - no entry of its own, and either 2+ distinct groups or 2+ baits - is short. CSNK2A2 leads it at 3 groups across 2 baits (Beisel, Gstaiger, Gygi), which is the strongest database evidence for any unwritten partner here, and gene-csnk2a1 exists while the paralogue does not. Behind it: CCNT2, EPB41L3, ALKBH4, FAM9A, HCFC2, HMOX2, LRRK2, APPBP2, MIIP, MYC, NDEL1, YWHAG at 2 groups each.

H3C1 sits at rank 2 and should be ignored, for the reason already given in the previous analysis: this graph files histones as marks and interfaces, which is the right unit for a chromatin reader, and a gene-h3c1 entry would be worse than its absence. It is worth naming again because the instrument will keep promoting it every time this runs, and an instrument that repeats a known false positive without a note is how a false positive eventually gets written.

PAX5 is the name worth chasing. It appears against four baits with no entry of its own - 17 entries mention it and ana-dot1l-in-the-pax5-proximity-interactome exists - and it is the master B-lineage transcription factor and one of the commonest lesions in B-ALL. Its rows resolve to a single group, so the database evidence is thin; the reason to write it is the disease, not the database.