Oncofusion

start / graph / ana-q118-the-cpg-island-falsifier-reproduces-in-human-at-fold-1-34-against-q117s-1-35

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q118-menin-at-the-human-index-fusion

Q117's CpG-island falsifier reproduces in the human index fusion almost to the decimal - gained bins 5.66% on an island against a class-matched null of 4.23%, fold 1.34, Fisher p = 0.49, where Q117 got fold 1.35 and p = 0.53 - so the CXXC-redistribution model fails in human RS4;11 as it did in mouse

also called CXXC falsifier, human arm · criterion C · the 1.34 against 1.35

One theory said the fusion protein, once knocked off its usual spots by the drug, would resettle on the unmethylated CpG islands its DNA-binding domain likes. Tested in human cells this comes out the same as it did in mouse last cycle: the new spots carry none of that signature, while the old spots are overwhelmingly islands.

The same thing, for a scientist

**The competing model to eviction-without-redistribution is CXXC-driven relocation onto free unmethylated CpG islands.** Q117 tested it in murine MLL-AF9 with a FLAG-tagged fusion and it failed: 1 of 52 gained bins on a UCSC mm10 island (1.92%) against a class-matched null of 1.42%, **fold 1.35, Fisher two-sided p = 0.53**, while 65.6% of the fusion-held bins were islands against a 1.62% background. **Run here in human RS4;11 (KMT2A::AFF1) on the MLL1-N arm, hg19, 2 kb bins, universe 1,389,514, class-matched null matched on MLL1n-DMSO decile AND input decile** [E1]: | | Q117, murine FLAG-MLLAF9 | Q118, human RS4;11 MLL1-N | |---|---|---| | gained set on a CpG island | 1/52 = **1.92%** | 3/53 = **5.66%** | | class-matched null rate | 1.42% | 4.23% | | **fold over matched null** | **1.35** | **1.34** | | **Fisher two-sided p** | **0.53** | **0.49** | | held set on a CpG island | 806/1228 = 65.6% | **1045/1390 = 75.18%** | | universe background | 1.62% | **2.52%** | | held-set enrichment over background | 40x | **30x** | **Two datasets, two species, two antibodies, two genome builds, two CpG-island annotations - and a fold of 1.35 against 1.34 with p of 0.53 against 0.49.** The absolute rates differ (5.66% against 1.92%) because hg19 has a denser island annotation and a higher background, which is exactly why the comparison has to be against a **matched null** and not against a genome-wide rate. **The falsifier does not fire. The CXXC-redistribution model is refuted in the human index fusion on its own predicted substrate.** **And the same test on the MENIN arm gives the opposite answer** - see `ana-q118-menins-own-gained-set-is-cpg-island-enriched-and-the-mll1-n-one-is-not`, where 21 of 209 menin-gained bins are islands against a matched null of 2.71%, fold 3.71, **p = 5.2e-07**. So the instrument can return a positive; this negative is not an instrument failure.

Go deeper
Why the positive control matters more than the negative here

A CpG-island enrichment test that returns p = 0.49 is worth nothing unless the same test, in the same universe, with the same matcher, can return a small p when the effect is real. It can, and it did, 10 lines of code away: the MENIN gained set in the same cells returns fold 3.71 at p = 5.2e-07.

So the negative for MLL1-N is a measurement and not the shape of a broken query. This is rule 67 applied to a genomic overlap test rather than to a literature search - never trust a null from an instrument you have not shown can return a positive.

What the test does NOT bound

The same limit Q117 stated, and it still applies. This bounds a LARGE redistribution onto islands. It does not bound a small one. Several hundred CpG islands each gaining a little signal would not enter a 53-bin top-0.1% set at all, and would be invisible here.

The cheap test that would close it is a CpG-island-unit analysis rather than a 2 kb-bin one: take all 27,537 hg19 islands on the analysed chromosomes, compute the MLL1-N and MENIN share ratio in each, and ask whether the distribution over islands that were UNOCCUPIED at baseline shifts up under drug. Q117 listed this as not done and it is still not done.

Facts and figures
scriptanalyses/q118-menin-at-the-human-index-fusion/binsB.py
statusrun
result summaryGained set 3/53 = 5.66% CpG island vs class-matched null 4.23%, fold 1.34, Fisher p = 0.492. Held set 1045/1390 = 75.18% vs 2.52% background. Q117's murine figures: 1.92% vs 1.42%, fold 1.35, p = 0.53; held 65.6% vs 1.62%.
genome buildhg19
n islands analysed chroms27537
bin bp2000
epigenetic readouthg19 CpG-island overlap as a proxy for the CXXC/unmethylated-CpG substrate; held set 75.18%, gained set 5.66%, universe 2.52%
shown for fusionKMT2A::AFF1 in human RS4;11 via an MLL1 N-terminus antibody that also sees wild-type MLL1; the murine comparator is FLAG-tagged KMT2A::MLLT3
inputs
[
  "GSM3635695/96 RS4;11 MLL1n",
  "GSM3635699 RS4;11 input",
  "UCSC hg19 cpgIslandExt",
  "ENCODE hg19 blacklist v2"
]
How this could bite us back
  • A refuted redistribution model is a closed question, and this project's own record says a negative that closes a question is the most dangerous kind, because no future beat is dispatched to find the counter-example. The island-unit test above is written down precisely so that acceptance does not become permanent.
  • If the fusion does not redistribute onto free islands, a CXXC-domain-directed agent would have nothing new to catch after menin inhibition, which removes one rationale for combining the two.
Still open
  1. Does a CpG-island-unit analysis (all 27,537 hg19 islands as the unit rather than 2 kb bins) show a small upward shift for baseline-unoccupied islands? That is the redistribution this test cannot see, and it is one afternoon's work on data already binned.
  2. The absolute island rates differ 3-fold between the mouse and human runs (1.92% vs 5.66%) while the matched-null folds agree to 0.01. Is that the annotation density difference it looks like, or does hg19 cpgIslandExt over-call relative to mm10 cpgIslandExt? Nobody compared the two tracks' island-length distributions here.
  3. KDM2B is the CxxC-ZF protein that Q117 found sitting at the ceiling at the fusion's sites before either drug. GSE127507 has no KDM2B sample, so the human version of Q117's KDM2B question cannot be asked in this series at all.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    In human RS4;11, 3 of the 53 MLL1-N gained 2 kb bins (5.66%) overlap an hg19 CpG island against a class-matched null rate of 4.23% - fold 1.34, Fisher two-sided p = 0.49 - while 1,045 of the 1,390 MLL1-N held bins (75.18%) overlap one against a universe background of 2.52%.

    What this evidence cannot show

    NEITHER ANTIBODY IN THIS SERIES IS FUSION-SPECIFIC AND THE ANTIBODY NAME SAYS SO. MLL1n is anti-MLL1 N-terminus (Bethyl A300-086A, lots 5 or 6); the N-terminus is retained in KMT2A::AFF1, so that track is the sum of the fusion and wild-type MLL1 in unknown proportion. MENIN is Bethyl A300-105A Lot 8; menin binds the fusion, wild-type MLL1 and non-KMT2A partners. Q117 could attribute its numbers because its fusion was FLAG-tagged; nothing here can. Every occupancy statement in this beat is about MLL1-N-reactive material or menin-occupied sites, never about the fusion. No scale factor exists for any sample. The depositors state that a Drosophila spike-in was run on one batch, that fly/human ratios were 'found to be small', and that NO spike-in normalisation was applied; no number for that measurement is reported and none is deposited. Every cross-condition figure here is therefore a WITHIN-SAMPLE share or a within-sample rank, which cancel an unknown per-sample scale factor exactly, and never a fold change. n = 1 per antibody per condition per line - there is no replicate anywhere in GSE127507 - so no number here carries a spread and none is precise. A 2 kb bin is wider than most CpG islands so every rate is inflated equally and only the contrast is meaningful; cpgIslandExt is a sequence annotation and not a methylation measurement in RS4;11. The test bounds a LARGE redistribution onto islands and not a small one. The gained-set size of 53 sits inside this beat's pre-registered indifference band of 40-65 for criterion G1, so the denominator of this test is itself at a threshold.

    Evidence that argues against this (2)
    • The held set's island enrichment partly reflects that ChIP peaks and CpG islands both concentrate at promoters, so the 30-fold contrast overstates how specifically CXXC-driven it is. The surviving claim is narrower: the gained set does not resemble the held set on this property. Iqbal Lab oncofusion knowledge graph 2026
    • The MLL1-N gained set is real by an independent test - it reproduces in MOLM13 at 1.77-fold against a matched null of 1.01 with a flat input, p = 5.0e-4. So 'not CXXC-driven' is not the same as 'not there', and an entry that read this falsifier alone would conclude the second from the first. Iqbal Lab oncofusion knowledge graph 2026
Connected to