Oncofusion

start / graph / ana-q117-the-cpg-island-test-settles-it-66-percent-of-the-fusions-real-sites-are-islands-and-2-percent-of-the-gained-set-is

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

The falsifier that decides the beat, run: 65.6% of the fusion's own top sites overlap a CpG island and 1 of the 52 sites it formally 'gains' under a menin inhibitor does - 1.35-fold over a class-matched null, Fisher p = 0.53 - so the gained set does not carry the signature of CXXC-driven binding and the redistribution model is refuted on its own predicted substrate

also called Q117 CpG-island falsifier · 1 of 52 gained bins is a CpG island

The fusion protein grips DNA at unmethylated CpG islands, and two thirds of its normal sites are islands. Of the fifty-two places it appears to gain under the drug, one is - which is what you would get by chance, so those places are not new fusion binding sites.

The same thing, for a scientist

The competing model to eviction-without-redistribution is that the evicted fusion relocates onto **free unmethylated CpG-island substrate** through its retained CXXC domain. That model makes a sharp, cheap prediction, and this beat pre-registered it as a scoreable falsifier: *the 52 gained bins overlap annotated CpG islands at more than 2-fold the class-matched rate, with at least 20 of 52 on an island.* **Run against the UCSC mm10 `cpgIslandExt` track, 15,997 islands on the analysed chromosomes, 19,870 of 1,227,708 universe bins (1.62%) overlapping one** [E1]: | set | n | on a CpG island | rate | |---|---:|---:|---:| | whole universe | 1,227,708 | 19,870 | **1.62%** | | class-matched null for the gained set | 2,600 draws | 37 | **1.42%** | | **the 52 VTP50469-gained bins** | 52 | **1** | **1.92%** | | **the fusion's own top-0.1% DMSO bins** | 1,228 | 806 | **65.6%** | **Fold enrichment of the gained set over its class-matched null: 1.35. Fisher exact two-sided p = 0.53. The falsifier does not fire, and it does not come close.** **The contrast is the result.** The fusion's genuine sites are CpG islands 65.6% of the time - a 40-fold enrichment over the genomic background and exactly the signature a CXXC-domain protein should leave. The formally gained set is **1.92%**, statistically indistinguishable from matched chromatin. **Whatever the 52 bins are, they do not carry the biochemical signature of fusion binding.** This is the cleanest evidence in the beat that menin inhibition **evicts the fusion without redistributing it**, and it is stronger than the amplitude and gene-identity arguments because it tests a property the alternative model specifically predicts rather than a property that merely looks unfamiliar.

Go deeper
Why this falsifier was worth writing as a falsifier

The pre-registered gain criteria all fired: 52 bins against a threshold of 50, +29.9 percentile points against a class-matched null, and zero overlap with the DOT1L-inhibitor arm. On the statistics alone, the beat's answer was the fusion gains sites.

Three independent arguments then said otherwise - amplitude (6x the genome median against real peaks at 12-215x), identity (16 of 52 intergenic, no Hox, no Meis, no myeloid transcription factor), and now substrate (1 of 52 on a CpG island against 806 of 1,228 for the real sites). The first two are judgements. The third is a test of a prediction the competing model makes, scored against a class-matched null, with the denominator printed and non-zero, and it is the one that settles it.

The transferable point: when a rank-based criterion fires on a collapsed distribution, do not argue with the number - find a property the real thing has and check whether the candidate set has it. A 40-fold difference in CpG-island occupancy between the fusion's real sites and its supposed new ones is not a close call.

The one bin that is on an island, and the SGC arm

One of the 52 VTP-gained bins overlaps a CpG island. One of the two SGC0946-gained bins also does, which gives a nominal 25-fold enrichment for that arm at Fisher p = 0.058 - on n = 2, which is not a result and is recorded here so that nobody quotes it as one. With two observations, one hit on a 1.4% background is simply what happens sometimes.

That asymmetry is worth one line of honesty: if a future replicate expanded the SGC0946 gained set and the CpG-island fraction held, the picture would be the opposite of the one drawn here, for the other drug. Nothing in this beat can distinguish that from a coin landing heads twice.

Facts and figures
scriptanalyses/q117-gained-fusion-occupancy/cgi117.py
statusrun
result summary1 of 52 VTP50469-gained bins overlaps a CpG island (1.92%) against a class-matched null of 1.42% (fold 1.35, Fisher p = 0.53). 806 of 1,228 fusion-held bins do (65.6%) against a background of 1.62%. The pre-registered redistribution falsifier does not fire.
falsifier scoredhyp-q117-menin-inhibition-evicts-the-fusion-without-redistributing-it-to-a-recognisable-programme, falsifier 1 (REDISTRIBUTION, CpG ARM)
verdictDOES NOT FIRE
n cgi intervals15997
universe bins1227708
universe bins on cgi19870
background cgi rate0.01618
epigenetic readoutCpG-island occupancy as the biochemical signature of CXXC-domain-driven binding
shown for fusionKMT2A::MLLT3 (MLLAF9) in a murine myeloid line, FLAG-tagged
fusion held set
rate0.656rate: 0.6564fold_over_background40.6fold_over_background: 40.6n_on_cgi806.0n_on_cgi: 806n1,228n: 1228
inputs
[
  "GSM8123359",
  "GSM8123360",
  "GSM8123361",
  "GSM8549823",
  "UCSC mm10 cpgIslandExt",
  "ENCODE mm10 blacklist v2"
]
gain set VTP
{
  "n": 52,
  "n_on_cgi": 1,
  "rate": 0.0192,
  "matched_null_draws": 2600,
  "matched_null_rate": 0.0142,
  "matched_null_denominator_nonzero": true,
  "fold": 1.351,
  "fisher_p_two_sided": 0.531
}
gain set SGC
{
  "n": 2,
  "n_on_cgi": 1,
  "fold": 25,
  "fisher_p_two_sided": 0.058,
  "note": "n = 2. NOT a result. Recorded so that it cannot be quoted as one."
}
fusion held set — the numbers behind the chart
{
  "n": 1228,
  "n_on_cgi": 806,
  "rate": 0.6564,
  "fold_over_background": 40.6
}
How this could bite us back
  • A falsifier that fails to fire is weaker evidence than one that fires, and this one closes a question. If the fusion does redistribute at a scale below this beat's detection floor - 52 bins is not a large sample and a real redistribution of a few hundred low-occupancy islands would be invisible here - then the closure is premature and no future beat will reopen it, because the question will be recorded as answered.
Still open
  1. Is the murine MLLAF9 line's methylome such that the CpG islands the fusion does not occupy are actually unmethylated and accessible? cpgIslandExt is a sequence annotation, not a methylation measurement, and no WGBS from this line was joined.
  2. Does the same contrast hold in human KMT2A::AFF1? If the fusion's held set in RS4;11 is similarly CpG-island-dominated and a gained set there is not, the conclusion generalises to the index fusion.
  3. What ARE the 52 bins, if they are not fusion sites and not CpG islands? Their input percentiles are unremarkable (median 45.9, only one above the 95th), so they are not obvious artefacts either.
Every source behind this entry (2)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    Of the 52 bins that formally gain FLAG-MLLAF9 signal under VTP50469, one overlaps a UCSC mm10 CpG island (1.92%) against a class-matched null rate of 1.42% (fold 1.35, Fisher exact two-sided p = 0.53), while 806 of the 1,228 bins the fusion occupies most strongly in DMSO overlap an island (65.6%) against a genome-wide background of 1.62%.

    What this evidence cannot show

    A 2 kb bin is larger than most CpG islands, so 'overlaps an island' is a generous criterion that inflates all four rates equally; the comparison between them is the meaningful quantity and the absolute rates are not. The class-matched null matches on decile of DMSO FLAG percentile crossed with decile of input percentile and draws up to 50 bins per gain bin, excluding the gain set and the treated top set; 2,600 draws were obtained and the null's CpG rate is 1.42%, printed and non-zero. n = 1 per condition in the underlying ChIP. cpgIslandExt is a computational annotation (Gardiner-Garden criteria) and is not a measurement of methylation in these cells - no murine WGBS or MeDIP map from this line was joined, so 'unmethylated' is an assumption carried by the annotation's name.

    Evidence that argues against this (1)
    • The fusion's held set is 65.6% CpG island partly BY CONSTRUCTION of how ChIP peaks and CpG islands both concentrate at promoters, so the 40-fold contrast overstates how specifically CXXC-driven it is. The claim that survives is narrower and is the one used: the gained set does not resemble the held set on the property that distinguishes fusion binding, whatever fraction of that property is CXXC and what fraction is promoter architecture. UCSC Genome Browser 2026
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    The redistribution model this test refutes is a live hypothesis in this graph with a stated substrate: the fusion occupies only about 20% of available unmethylated CpG sites at baseline against 89% for KDM2B in the same cells, so free substrate exists.

    What this evidence cannot show

    The 20% and 89% figures are taken from an existing graph entry, are measured in human SEM cells, and were not re-fetched in this beat. They establish that the competing model has a substrate; they are not used quantitatively here.

    Evidence that argues against this (1)
    • If the free uCpG substrate in the murine MLLAF9 line is in closed chromatin, the absence of redistribution onto it is a statement about accessibility rather than about the drug. The test performed here cannot distinguish 'the fusion did not move' from 'the fusion could not move there', and only the first of those generalises to another system. Krivtsov AV et al. 2019
Connected to