Oncofusion

start / graph / ana-q117-the-pre-registered-gain-criteria-fire-on-52-low-amplitude-bins-with-no-coherent-identity

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q117-gained-fusion-occupancy

Asked for the first time whether a KMT2A fusion GAINS chromatin sites under a menin inhibitor: all three pre-registered criteria fire, and the set they fire on is 52 low-amplitude bins reaching 6x the genome median, 16 of them with no gene within 20 kb and 1 of 52 on a CpG island where 66% of the fusion's real sites are

also called Q117 gained-occupancy call · eviction without redistribution to a recognisable program

We looked for places on the DNA where the leukaemia fusion protein shows up only after the drug is given, and we wrote down the rules for what would count before looking. A small scattered set passed the rules, but one of its fifty-two sites is a CpG island where two thirds of the fusion's real sites are, so the drug knocks the fusion off without moving it anywhere it would normally bind.

The same thing, for a scientist

Pre-registered in `analyses/q117-gained-fusion-occupancy/PREREGISTRATION.md` on 2026-09-23 before any signal was read. Design: 2 kb bins, mm10, autosomes plus chrX, universe = input-covered and non-blacklisted (n = 1,227,708), held set = top 0.1% of FLAG-MLLAF9 DMSO (n = 1,228). Gain set G = bins in the top 0.1% of FLAG VTP50469 that sit **below the 90th percentile** in FLAG DMSO. Three criteria, all required, all fixed in advance [E1]: | criterion | threshold set in advance | observed | verdict | |---|---|---:|---| | **G1** size of G | >= 50 bins | **52** | **passes, by two bins** | | **G2** median VTP-DMSO percentile shift of G over a class-matched null | >= 10 points | **+29.9** (16.3 vs -13.6; Mann-Whitney p = 1.7e-27) | **passes** | | **G3** not reproduced in the SGC0946 arm | Jaccard < 0.25 | **0.0** (intersection 0 of union 54) | **passes** | **And the pre-registered prior, which was that the fusion gains nothing, is wrong in its stated form.** P1 predicted fewer than 5% of top-0.1%-VTP bins would fall below the 99th DMSO percentile; the observed figure is **557 of 1,228 = 45.4%** [E1]. P3 predicted the apparent gain would be reproduced in the SGC0946 arm; it is not - SGC yields **2** bins to VTP's 52, and the same asymmetry holds at every cutoff from the 80th percentile (18 vs 0) to the 99th (557 vs 108) [E1]. **The set that passes, read rather than counted, does not look like a gained program.** The 52 bins reach a median of 2,203 coverage-bases against a track genome median of 375 - about **6x**, with a maximum of 4,174 (11x) - while the fusion's genuine DMSO peaks run from 4,320 to 80,658 (12x to 215x) [E2]. **16 of the 52 have no refGene within 20 kb.** The genes near the remaining 36 are *Igsf21, Sorcs2, Lmx1b, Susd4, Trpv4, Cdh13, Mylk, Krt77, Rasgrf1, Gabrb1, Galnt7, Wwox, Zfp423, Wnt7b, Kcnip3, Slc24a4, Cyfip2, Rps6ka2* and similar [E2]. **No Hox gene, no Meis/Pbx, no myeloid transcription factor, no PRC1.1 target, and nothing from the fusion's own DMSO programme.** **And the decisive test is not a judgement at all.** The competing model - that the evicted fusion relocates onto free unmethylated CpG-island substrate through its retained CXXC domain - predicts a CpG-island-enriched gained set. Run against the UCSC mm10 `cpgIslandExt` track: **1 of the 52 gained bins overlaps an island (1.92%) against a class-matched null of 1.42%, fold 1.35, Fisher exact two-sided p = 0.53 - while 806 of the 1,228 fusion-held bins (65.6%) overlap one, against a universe background of 1.62%** [E4]. The fusion's real sites carry the CXXC signature 40-fold over background; the gained set does not carry it at all. So the answer to the beat's question is: **menin inhibition evicts the fusion and does not redistribute it to a recognisable new programme.** A formal gain call fires, on a weak, scattered, gene-poor set that is most consistent with the noise floor of a collapsed distribution rising into the top 0.1% - and the call is fragile enough that it flipped on an unrelated processing choice.

Go deeper
Why a formal pass is still reported as a negative

Five things point the same way and none of them was available to the pre-registered thresholds:

  1. Amplitude. The gained bins peak at 11x the genome median. The fusion's own sites run to 215x. A 6x bin is within the range a ChIP background produces, and the criterion that admitted them - top 0.1% of the treated track - is a rank criterion that must admit 1,228 bins no matter how far the real peaks have collapsed.

  2. Identity. 16 of 52 intergenic; the rest near genes with no shared biology and no connection to the MLLAF9 programme. A real redistribution to unoccupied CpG islands, which is the live alternative hypothesis, would look like a coherent set.

  3. Mechanism of the artefact is present and measurable. The VTP track's own 99.9th percentile is 2,025 coverage-bases against DMSO's 4,331 - the top of the distribution has halved, so the bar for entering the top 0.1% has halved with it.

  4. The call is threshold-fragile. Before the ENCODE blacklist was applied the same pipeline returned 42 bins and G1 failed. The blacklist was added for an unrelated reason and moved a pre-registered verdict from fail to pass.

  5. Substrate. 1 of 52 gained bins is a CpG island; 806 of 1,228 fusion-held bins are. A CXXC-domain protein binds unmethylated CpG islands, and the gained set does not look like anything a CXXC domain bound.

The SGC0946 control is the one thing that argues the other way and it argues strongly. SGC0946 collapses the fusion's held-set enrichment almost as far (share ratio 0.298 against VTP's 0.179) and produces 2 gain bins rather than 52. If the 52 were purely the noise floor rising, SGC should produce a comparable number. It does not, and that is a genuine, pre-registered, drug-specific asymmetry that this entry does not explain away.

The pre-registered prior, scored honestly
prediction, written before lookingoutcome
P1 fewer than 5% of top-0.1% VTP bins below the 99th DMSO percentileWRONG. 45.4% (557/1,228).
P2 median shift of G over its class-matched null under 10 percentile pointsWRONG. +29.9 points.
P3 the VTP gain set largely reproduced in the SGC arm, Jaccard > 0.25WRONG. Jaccard 0.0.

All three priors failed, and the conclusion the entry reaches - eviction without meaningful redistribution - is therefore reached despite the statistics rather than because of them. That is stated plainly because it is the part a later reader most needs: the numerical criteria said gain, and what overrides them is amplitude, identity and - decisively - the CpG-island test, which WAS pre-registered as a falsifier of the competing model and was run in the same cycle. The amplitude and identity arguments are judgements and should have been pre-registered as criteria; the substrate argument is not a judgement.

Facts and figures
scriptanalyses/q117-gained-fusion-occupancy/run117.py
statusrun
result summaryAll three pre-registered gain criteria fire (52 bins, +29.9 percentile points over a class-matched null, Jaccard 0.0 with the SGC0946 arm) on a set whose amplitude is ~6x the genome median against real peaks at 12-215x, 16/52 intergenic, no Hox and no myeloid TF. Reported as eviction without redistribution to a recognisable programme.
preregistrationanalyses/q117-gained-fusion-occupancy/PREREGISTRATION.md
absence classmeasured-not-analysed
absence class noteONE gap and the token describes it: the gained-occupancy analysis on deposited fusion ChIP-seq under a menin inhibitor. The data were deposited in 2026 (GSE260742) and the analysis had never been run on them. This entry runs it, so the class is recorded as the state BEFORE this beat. It does NOT describe the separate and still-open gaps named in open_questions - the human KMT2A::AFF1 equivalent, the replicate question, and the unopened supplementary files - which are different gaps of different classes.
absence class assigned2026-09-23 (Q117)
shown for fusionKMT2A::MLLT3 (MLLAF9) in a murine myeloid line with a FLAG-tagged fusion; nothing in human KMT2A::AFF1
inputs
[
  "GSM8123359",
  "GSM8123360",
  "GSM8123361",
  "GSM8549823",
  "ENCODE mm10 blacklist v2",
  "UCSC mm10 refGene"
]
criteria
{
  "G1_size_threshold": 50,
  "G1_observed": 52,
  "G1_observed_without_blacklist": 42,
  "G2_points_threshold": 10,
  "G2_observed": 29.896,
  "G2_mannwhitney_p": 1.71e-27,
  "G3_jaccard_threshold": 0.25,
  "G3_observed": 0,
  "G3_intersection": 0,
  "G3_union": 54,
  "G3_denominator_nonzero": true
}
priors scored
{
  "P1_predicted_under_5pct_below_99th": "WRONG - 45.4% (557/1228)",
  "P2_predicted_under_10_points_over_null": "WRONG - +29.9 points",
  "P3_predicted_reproduced_in_SGC_jaccard_over_0.25": "WRONG - Jaccard 0.0"
}
cutoff sensitivity VTP vs SGC
{
  "below_80th": [
    18,
    0
  ],
  "below_85th": [
    30,
    0
  ],
  "below_90th": [
    52,
    2
  ],
  "below_95th": [
    114,
    6
  ],
  "below_97th": [
    215,
    11
  ],
  "below_99th": [
    557,
    108
  ],
  "note": "post hoc; the pre-registered cutoff was the 90th percentile and was not changed"
}
gain set amplitude coverage bases
{
  "median": 2203,
  "min": 2025,
  "max": 4174,
  "track_genome_median": 375,
  "held_set_DMSO_range": [
    4320,
    80658
  ]
}
gain set identity
{
  "n_bins": 52,
  "n_with_no_refGene_within_20kb": 16,
  "n_with_input_percentile_over_95": 1,
  "no_hox": true,
  "no_meis_or_pbx": true,
  "no_myeloid_tf": true,
  "no_prc1_1_target": true,
  "n_on_cpg_island": 1,
  "fusion_held_set_cpg_island_rate": 0.6564,
  "cgi_fold_over_matched_null": 1.351,
  "cgi_fisher_p": 0.531
}
How this could bite us back
  • If menin inhibition really does evict without redistribution, that is the cleaner pharmacology and it removes one worry - but it also removes a candidate resistance mechanism that would have been monitorable. Resistance would then have to come from restored occupancy at the old sites, from MEN1 interface mutation, or from transcriptional rewiring that no occupancy assay would see.
  • A negative gain call is the kind of result that gets quoted without its amplitude caveat. 'Fifty-two gained sites under VTP50469' is a sentence that could be lifted out of this entry and used to argue the opposite of what it concludes, which is why the amplitude and the intergenic fraction are in the summary and not only in the caveats.
  • The criteria fired and the judgement overrode them. Anyone reusing this design should add an amplitude floor to the pre-registration - a gained bin must reach some multiple of the track's own genome median - because a rank criterion on a collapsed distribution will always return a full quota of bins.
Still open
  1. Do the 52 bins reproduce in a second FLAG ChIP? No replicate exists in GSE260742, and a single technical repeat would settle whether this is a stable drug-specific set or the top of a collapsed noise distribution.
  2. Why does the DOT1L inhibitor produce 2 gain bins when it collapses the fusion's held-set enrichment to 0.298 against the menin inhibitor's 0.179? A noise-floor explanation predicts a comparable number and does not get one, and nothing in this beat accounts for the asymmetry.
  3. The uCpG redistribution hypothesis was tested in cycle 68 and failed: 1 of 52 gained bins is on a CpG island. What remains is whether a SMALL redistribution - a few hundred islands each gaining a little - would have been detectable at all in a 52-bin top-0.1% set from an n=1 track. It would not, and nothing in this beat bounds it.
Every source behind this entry (4)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Under a pre-registered design, 52 of the 1,228 top-0.1% FLAG-VTP50469 bins lie below the 90th percentile of FLAG DMSO, their median VTP-minus-DMSO percentile shift exceeds a decile-matched null by 29.9 points, and the equivalent SGC0946 set has 2 bins with zero overlap.

    What this evidence cannot show

    n = 1 per condition; no replicate exists in this series for the FLAG ChIP, so the Mann-Whitney p-value is a statement about bins within one sample pair, not about biological reproducibility, and it should not be read as one. The class-matched null matches on deciles of DMSO-FLAG percentile crossed with deciles of input percentile, which is coarse at the top of the distribution. The input used for matching (GSM8549823) comes from the second processing batch while the FLAG samples come from the first; the FLAG arm's own four input samples carry no deposited processed file. The gain set is defined on a rank criterion that must return 1,228 bins whatever happens to the real peaks.

    Evidence that argues against this (1)
    • The identical pipeline without the ENCODE blacklist returns 42 bins and G1 FAILS. The verdict of the pre-registered conjunction therefore depends on a processing choice made for an unrelated reason, and a reader is entitled to treat the call as unresolved rather than as a pass. Both numbers are recorded in analyses/q117-gained-fusion-occupancy/results-v1-no-blacklist.json and results.json. Amemiya HM et al. 2019
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    The 52 gained bins reach a median of 2,203 and a maximum of 4,174 coverage-bases against a track genome median of 375, where the fusion's DMSO held set runs from 4,320 to 80,658; 16 of the 52 have no refGene within 20 kb and the genes near the rest include no Hox gene, no Meis or Pbx family member and no myeloid transcription factor.

    What this evidence cannot show

    Gene assignment is proximity within 20 kb to a refGene transcript, which is a weak notion of regulatory target and will miss distal enhancer targets entirely. 'No coherent identity' is a judgement about a list of 36 gene symbols and is not a statistical test; no GO or pathway enrichment was run, and an enrichment test on 52 bins would be underpowered in any case. The amplitude comparison uses coverage-bases from igvtools count, which is mean coverage per window times window length and is proportional to reads but is not a read count.

    Evidence that argues against this (1)
    • Absence of a recognisable programme is exactly what a genuinely NOVEL redistribution would look like on first inspection. The fusion's normal targets were discovered because somebody looked; a set of 36 genes nobody has connected to KMT2A biology is not evidence that the set is noise, only that it is unfamiliar. The amplitude argument is the stronger one and should carry the conclusion; the identity argument on its own would be a circular appeal to prior expectation. NCBI Gene Expression Omnibus 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The depositing paper's own analysis ranks genes by MLLAF9 occupancy in DMSO and asks what is lost, so a gained-site analysis has not been run on these data by anyone.

    DOT1L provides transcriptional memory through PRC1.1 antagonism. · Neville D, Ferguson DT, Heikamp EB, et al. · Nature Cell Biology 2026
    PMID 41634341DOI
    What this evidence cannot show

    Established from the retrieved full text (142,554 characters) on 2026-09-23: the figure legends rank genes 'from highest to lowest by H3K79me2 occupancy in DMSO' and the two occurrences of the word 'gained' both refer to H2AK119ub. Supplementary data files were NOT opened, so a gained-peak table in a supplement cannot be excluded - rule 80 applies, since a gain count is exactly the kind of quantity that appears only in a supporting panel.

    Evidence that argues against this (1)
    • The authors state their conclusion that PRC1.1 deposition is independent of MLL-FP displacement, which implies they examined fusion occupancy carefully. If their extended data contain a gained-peak count, this beat is a replication rather than a first analysis, and the absence it rests on is measured-not-reported rather than measured-not-analysed. Nature Cell Biology 2026
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    One of the 52 VTP50469-gained bins overlaps a UCSC mm10 CpG island (1.92%) against a class-matched null rate of 1.42% over 2,600 draws (fold 1.35, Fisher exact two-sided p = 0.53), while 806 of the 1,228 fusion-held bins do (65.6%) against a universe background of 1.62%.

    What this evidence cannot show

    A 2 kb bin is larger than most CpG islands, so all four rates are inflated equally and only the contrast carries information. cpgIslandExt is a computational sequence annotation, not a methylation measurement in these cells. n = 1 per condition. The test bounds a large redistribution and not a small one: a few hundred islands each gaining a little would not enter a 52-bin set. The matched-null denominator is printed and non-zero (1.42%, 2,600 draws).

    Evidence that argues against this (1)
    • Of the two SGC0946 gained bins, one is on a CpG island, which is a nominal 25-fold enrichment at Fisher p = 0.058. On n = 2 that is not a result and is recorded only so that it cannot be quoted as one - but if a replicate expanded that set and the fraction held, the conclusion would invert for the other drug. UCSC Genome Browser 2026
Connected to