Oncofusion

start / graph / ana-q118-fifty-three-non-island-bins-gain-mll1-n-signal-in-two-lines-with-a-flat-input

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q118-menin-at-the-human-index-fusion

Fifty-three 2 kb bins gain MLL1-N-reactive signal under VTP-50469 in RS4;11, and the same 53 coordinates gain 1.77-fold in MOLM13 against a class-matched null of 1.01 with the MOLM13 input flat at 1.01 - so there IS a small, reproducible, input-controlled redistribution, and it carries no CpG-island signature

also called the 53 gained bins · a real but small redistribution · G1 at the threshold

Fifty-three small stretches of DNA pick up more of the MLL1 signal when the drug is added, and the same stretches pick it up in a second, unrelated leukaemia line while the background control does not move at all. So something really does move - but it is about thirty times smaller than what stays put, and it is not where the CpG-island theory predicted.

The same thing, for a scientist

**Pre-registered criterion G1 fires at 53 bins (threshold 50, declared indifference band 40-65), so it is reported as AT THE THRESHOLD and the magnitudes are what is quoted.** Q117 watched the identical criterion move 42 to 52 across a threshold of 50 on an unrelated deviation, which is why the band was declared in advance. What makes these 53 bins different from Q117's 52 is that **they survive every control this series can apply** [E1][E2]: | test | result | reading | |---|---|---| | amplitude in the VTP track | median **15.28x** the genome median, vs 12.18x for the 1,390 held bins and 1.97x for themselves in DMSO | **not low-amplitude** - Q117's 52 bins topped out at 11x against real peaks at 12-215x | | G2 vs class-matched null | **+27.44** percentile points (threshold 10) | fires | | G3, Jaccard with the MENIN gained set | **0.036** (threshold < 0.25) | fires - not a shared library property | | CpG island | 5.66% vs matched null 4.23%, fold 1.34, p = 0.49 | **no CXXC signature** | | within 10 kb of an ENCODE blacklist interval | **0.0%** (held 0.7%, universe 0.6%, Fisher p = 1) | not a known artefact region | | clustering | **53 of 53 are isolated single bins** more than 20 kb apart | no broad domain | | input percentile | median **56.2** (held set 54.3) | not a mappability or copy-number artefact | | at a RefGene TSS | **17.0%** (held 76.1%, universe 6.4%) | **mostly intergenic/intronic** | | **the same 53 coordinates in MOLM13 MLL1-N** | share ratio **1.7716** | **reproduces in a second line** | | **class-matched null for that** (2,000 draws, matched on RS4;11 MLL1n-DMSO decile AND input decile) | median **1.0134**, 95% interval 0.893-1.147, one-sided **p = 5.0e-4** | 1.75-fold over null | | **the MOLM13 INPUT at the same 53 coordinates** | **1.0115** | **the background does not move** | | MOLM13 input-corrected | **1.7516** | the gain is not a library effect | **The amplitude framing, which is the honest size of the thing** [E1]. The 53 bins hold **0.0503%** of the RS4;11 MLL1-N VTP track; the 1,390 held bins hold **1.4070%** - a **1 : 28** ratio. In share terms the held set loses 0.1444% of the track and the 53 bins gain 0.0445%, so **the gain accounts for about 31% of the (small) MLL1-N loss and about 5% of the (large) MENIN loss at the same sites**. **So Q117's headline generalises in one half and not the other.** *Eviction* replicates exactly (menin arm, 0.1999 against 0.179). ***Without redistribution* does not hold as a blanket statement in the human index fusion**: there is a reproducible, input-controlled, two-line gain. It is simply not the gain the CXXC model predicted, and it is 28-fold smaller than what remains.

Go deeper
What the 53 bins are, and what nobody has established about them

Top bins by VTP signal (hg19): chr19:3,762,000 | chr5:158,788,000 | chr20:23,636,000 | chr5:169,260,000 | chr17:76,588,000 | chr1:121,258,000 | chr4:3,878,000 | chr2:101,944,000 | chr2:239,684,000 | chrX:2,546,000 | chr6:3,024,000 | chr5:142,382,000 | chr1:33,458,000 | chr10:4,374,000 | chr8:112,474,000. The full list of 53 is in results-instrumentB-addendum.json.

Nine of the 53 fall inside one of the series' own 21,461 RefGene TSS windows, and those nine touch ADGB, APBA3, EMBP1, HTATSF1P2, LOC100507632, MAGEF1, NLRP12, RNF130 and TYROBP 2. That is not a leukaemia programme - no HOX, no MEIS/PBX, no myeloid or B-lineage transcription factor - which is the same negative Q117 reported for its 52.

What was deliberately not done: no external gene-coordinate resolver was used anywhere in this beat, so the other 44 bins carry no gene name rather than a possibly-wrong one. Q117 lost time to the UCSC search API placing Hoxa9 on chr1, and the correction is recorded in this graph as concept-q117-the-ucsc-search-api-puts-hoxa9-on-chr1-if-you-take-the-longest-match-and-six1-on-chr17-if-you-trust-the-symbol. Annotating those 44 bins properly - against a downloaded hg19 RefSeq or GENCODE table, with repeat and segmental-duplication tracks alongside - is the obvious next step and it is not done here.

One locus deserves naming as a hazard: chr1:121,258,000 is pericentromeric chr1, 226 kb from chr1:121,484,000, the single worst blacklisted bin in all nine files. It is outside the 10 kb proximity test that returned 0.0% and the ENCODE v2 list is not exhaustive for hg19 satellite sequence. One bin of 53 does not change the verdict, and it is the first thing a sceptical reader should check.

Why the cross-line test is the strongest control available, and why it is post-hoc

The 53 bins were defined in RS4;11 by selecting the top of the RS4;11 VTP distribution, so the RS4;11 aggregate share ratio at those bins (8.67) is large by construction and is not evidence of anything. The evidence is the 1.77 in MOLM13, a line in which the set was not selected, against a class-matched null of 1.01 built from RS4;11's own occupancy and input deciles.

And the MOLM13 input at the same coordinates is 1.0115 - so whatever rises there is not the background rising. RS4;11 has no VTP input, so this control exists only because a second line in the same series has both. That is the one place in this beat where the input asymmetry works in the reader's favour rather than against.

The baseline that makes 1.77 readable is MOLM13's own held set, where the same controls give no movement at all: MOLM13 MLL1-N at its own top 0.1% has a share ratio of 0.9438, its input at the same bins 0.9542, and the input-corrected value 0.9891. So in the line used to score the gain, the equivalent held-set measurement is flat to within 1%, and the 1.77 at the 53 imported coordinates stands against that.

This test was not pre-registered - section 4a has no criterion for cross-line reproduction at the same coordinates - and it is recorded as DEVIATION 7. What is not post-hoc is the direction of the design: define in one line, score in another.

Facts and figures
scriptanalyses/q118-menin-at-the-human-index-fusion/binsC.py and binsD.py
statusrun
result summaryG1 = 53 bins (threshold 50, band 40-65: AT THE THRESHOLD). G2 +27.44 points, G3 Jaccard 0.036, criterion C fold 1.34 p = 0.49. Cross-line: MOLM13 1.7716 vs matched null 1.0134, p = 5.0e-4, MOLM13 input 1.0115. The 53 bins hold 0.0503% of the VTP track against the held set's 1.4070% - a 1:28 ratio.
n gained53
bin bp2000
genome buildhg19
pre registered threshold50
indifference band40-65
cell lineRS4;11 (defined) and MOLM13 (scored)
shown for fusionKMT2A::AFF1 in RS4;11 and KMT2A::MLLT3 in MOLM13, via an MLL1 N-terminus antibody in both - the gain is in MLL1-N-reactive material and is NOT attributed to the fusion
inputs
[
  "GSM3635695/96 RS4;11 MLL1n",
  "GSM3635699 RS4;11 input",
  "GSM3635685/86 MOLM13 MLL1n",
  "GSM3635689/90 MOLM13 input",
  "UCSC hg19 cpgIslandExt",
  "ENCODE hg19 blacklist v2"
]
How this could bite us back
  • If a menin inhibitor lays down 53 new MLL1-N-occupied sites and they reproduce across two MLL-r lines, then the drug is not purely subtractive, and a resistance mechanism could be seeded at loci nobody is watching because they are neither promoters nor CpG islands.
  • The gain accounts for about 31% of the MLL1-N share lost from the held set. If a fraction of that is redistributed fusion, then the fraction of the fusion that leaves its sites but stays on chromatin is measurable and non-trivial, and an occupancy-based response assay would score it as unchanged total signal.
  • Fifty-three bins from an n=1 comparison at a pre-registered threshold of 50, inside a declared band of 40-65, is exactly the kind of count that this project has watched move on an unrelated deviation. Anybody quoting '53' should quote the amplitude and the cross-line 1.77 instead.
Still open
  1. What are the other 44 bins? They need annotating against a downloaded hg19 RefSeq or GENCODE table with repeat and segmental-duplication tracks alongside, which is one download and was deliberately not done here to avoid the coordinate-resolver trap Q117 hit.
  2. Does an IgG or a no-antibody control at these 53 coordinates also rise under drug? That is the one artefact class the input cannot exclude, and GSE127507 deposits no IgG sample in any line.
  3. Is the gain transcription-coupled? 83% of the bins are not at a RefGene TSS, so an intronic or enhancer location is likely, and the H3K79me2 samples in this same series (GSM3635697/98 for RS4;11) would say whether those regions are transcribed. They were not downloaded.
Every source behind this entry (2)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    The 53 RS4;11 MLL1-N gained 2 kb bins have a median VTP amplitude of 15.28-fold the genome median (held set 12.18-fold, themselves 1.97-fold in DMSO), 0.0% within 10 kb of a blacklist interval, input percentile median 56.2, and 53 of 53 are isolated single bins more than 20 kb apart; the same 53 coordinates give a MOLM13 MLL1-N share ratio of 1.7716 against a class-matched null median of 1.0134 (95% interval 0.893-1.147, 2,000 draws, one-sided p = 5.0e-4) while the MOLM13 input share ratio there is 1.0115.

    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. The cross-line test was NOT pre-registered (DEVIATION 7) and its threshold is therefore post-hoc; what is not post-hoc is that the set was defined in RS4;11 and scored in MOLM13. The empirical p of 5.0e-4 is the FLOOR for 2,000 draws and is not a smaller number than the draw count can express. Fifty-three bins holding 0.0503% of a track is a small quantity and the 1.75-fold effect in the second line is modest. RS4;11 has no VTP input so the input control exists only in MOLM13, at the same coordinates but in different cells. One of the 53 bins (chr1:121,258,000) is pericentromeric and 226 kb from the worst blacklisted bin in the series, outside the 10 kb proximity test.

    Evidence that argues against this (2)
    • A 1.75-fold rise at 53 isolated bins in a second line could be a property of the VTP ChIP protocol shared between the two experiments rather than of occupancy - both are anti-MLL1n ChIPs on drug-treated cells, and an antibody-background artefact would not show in the input. The controls here exclude a library-depth, mappability and copy-number explanation; they do not exclude a shared immunoprecipitation background. Iqbal Lab oncofusion knowledge graph 2026
    • Q117's equivalent 52 bins in mouse were LOW-amplitude (median 6x the genome median against real peaks at 12-215x) and it used that to override its own pre-registered pass. These 53 are median 15.28x, HIGHER than the held set, so the amplitude argument that carried Q117's negative is not available here and the two beats reach different conclusions on the same criterion. Iqbal Lab oncofusion knowledge graph 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    9 of the 53 gained bins (17.0%) fall inside one of the deposited series' own 21,461 RefGene TSS windows, against 76.1% of the held set and 6.4% of the universe (Fisher gained-vs-held p = 1.2e-18), and the nine touch ADGB, APBA3, EMBP1, HTATSF1P2, LOC100507632, MAGEF1, NLRP12, RNF130 and TYROBP.

    What this evidence cannot show

    Annotation is by overlap with the series' own 6 kb ROSE RefGene TSS windows and nothing else - no external gene-coordinate resolver was used anywhere in this beat, so 44 of the 53 bins carry no gene name rather than a possibly-wrong one. The universe figure is from a 1-in-37 systematic sample of the 1,389,514 universe bins, not a full pass. RefGene is not a complete transcript set, so 'not at a TSS' means 'not at a RefGene TSS in this table'.

    Evidence that argues against this (1)
    • This is also a measurement of instrument A's blindness and not only of the bins: 83% of the gained set is invisible to the deposited TSS table, which found only 3 gained windows in the same comparison. A beat that had used the convenient processed table alone would have reported a clean null. Iqbal Lab oncofusion knowledge graph 2026
Connected to