Oncofusion

start / graph / ana-q116-the-methyl-poor-kmt2a-domain-against-385-length-and-expression-matched-gene-bodies

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q116-h3k9-axis-at-the-locus

THE DEFLATING NULL, RUN: matching on gene length AND MCF-7 expression roughly HALVES the H3K9me3 depletion over the KMT2A gene body - from the 7.6th-15.0th percentile of class-matched CTCF anchors to the 18.2nd-23.4th of 385 matched transcribed gene bodies - so the methyl-poor breakpoint domain is partly a fact about transcription and does not dissolve into one

also called the transcription null · is the methyl-poor domain just a gene body? · N4

This project has said four times that the region where KMT2A breaks is unusually free of a silencing tag. Compared against other genes of the same size that are switched on just as much, about half of that unusualness disappears - but not all of it.

The same thing, for a scientist

**The claim at risk, and why this null was owed.** This graph records in four places that the *KMT2A* breakpoint region is a methyl-poor domain. Q114 measured the MCF-7 version and then wrote down the objection itself: the depletion covers **the whole 90 kb gene body more deeply than the BCR core**, *KMT2A* is a large broadly transcribed gene, and transcribed gene bodies are H3K9me3-poor as a class. **If H3K9me3 is depleted over large transcribed genes generally, the finding is about transcription and not about the breakpoint.** Q114 named this null and did not run it. This is it [E1]. **The null (N4), declared before it was computed.** 90,377 bp windows centred on the gene-body midpoint of every GENCODE v29 protein-coding gene whose span is within **±25% of 90,377 bp** *and* whose **MCF-7 TPM is within a factor of 2 of *KMT2A*'s 32.94** (ENCFF251DYL, ENCSR000CWQ, polyA+ RNA-seq in the same cell line). **Both matches held: n = 385**, median span 86,492 bp against *KMT2A*'s 90,369, median TPM 29.60 against 32.94, spread over all 23 chromosomes. The brief's fallback - match length only and say so - was not needed. Contrast set **N4b**: the same windows on **661 length-matched silent genes (TPM < 1)** [E2]. **The answer, in one line: the null takes about half of the effect and leaves the rest.** | track | vs class-matched CTCF anchors | vs 385 matched **transcribed** gene bodies | vs 661 matched **silent** gene bodies | *KMT2A* / median matched transcribed | |---|---:|---:|---:|---:| | GSM4585206 (-E2) | **15.00th** | **23.38th** | 15.28th | 0.938x | | GSM4585207 (+E2) | **9.22nd** | **21.04th** | 6.96th | 0.843x | | GSM4133820 (shCTL) | **7.60th** | **18.18th** | 7.72nd | 0.798x | | GSM4133821 (shCTL) | **7.70th** | **20.26th** | 6.51st | 0.766x | Matching on length and expression **roughly doubles** *KMT2A*'s percentile and **roughly halves** the fold deficit (0.62-0.82x of random 90 kb tiles becomes 0.77-0.94x of matched transcribed gene bodies). **In all four tracks *KMT2A* remains below the matched median and inside the bottom quartile of matched transcribed gene bodies** [E1]. **THE PRE-REGISTERED RULE RETURNS `N4 UNINFORMATIVE`, AND I AM REPORTING THAT RATHER THAN THE ANSWER I PREFER.** §6 declared the null uninformative unless matched **silent** gene bodies carry at least **1.2x** the H3K9me3 of matched **transcribed** ones. Measured: **1.053, 1.122, 1.110, 1.096** - the direction is right in 4 of 4 tracks and the magnitude is below my own threshold in 4 of 4. **The threshold was set too high for 90 kb windows and I did not know that when I set it.** With that clause set aside, the primary criterion - *KMT2A* below the 25th percentile of N4 in at least 3 of 4 tracks - returns **SURVIVES, 4 of 4**. **My pre-registered prior was DEFLATES at 0.65. It was wrong in that direction and roughly half right in magnitude**, and the honest label is **PARTLY DEFLATES**: the phenomenon is real against the right null and it is about half as deep as the wrong null made it look.

Go deeper
What the four existing methyl-poor entries should now say

Three corpus entries and one mechanism carry the methyl-poor-domain framing. None is refuted. All four are overstated by roughly a factor of two in effect size, in MCF-7, and the reason is the denominator rather than the measurement.

  • mechanism-h3k9me1-2-ctcf-balance-gates-kmt2a-rearrangement - NOT TESTED HERE, and I nearly wrote that it was. That mechanism is about H3K9me1 and H3K9me2, whose baseline claim is that the BCR is depleted relative to flanking sequence in RPE cells. This null measures H3K9me3 in MCF-7. Different mark, different cell, different comparator. What transfers is only the method: a length- and expression-matched gene-body null has never been applied to its me1/me2 numbers either, and until it is, 'methyl-poor' there carries the same unexamined denominator this analysis half-removed for me3.
  • ana-q114-in-mcf7-h3k9me3-is-depleted-across-the-kmt2a-locus-while-the-writer-looks-elevated - the 15.77th/10.30th percentiles are reproduced here (15.00/9.22, C6) and are correct against the null they name. They should be read alongside 23.38/21.04 against the transcription-matched null.
  • ana-q114-the-methyl-poor-kmt2a-breakpoint-domain-shows-up-again-in-a-third-lineage - not tested here. Its numbers are in SU-DIPG13/36 CUT&RUN and the transcription-matched null has never been run in those cells, where no matched MCF-7-style expression file exists.

And the sharper point for the breakpoint framing specifically. The depletion this null half-removes is over the 90 kb gene body. The 7,465 bp BCR core sits at the 63.61st, 61.69th, 44.21st and 22.67th percentile of width-matched CTCF anchors in the same four tracks - higher than the gene body it sits inside, and not depleted at all in the shCTL pair. The methyl-poor thing in MCF-7 is the gene, not the breakpoint, and that was already visible in Q114's numbers.

Why the CTCF-anchor null was the wrong denominator for a 90 kb question, and this is a general lesson

The N1 null is 90,376 bp windows centred on CTCF summits. A CTCF anchor is not a random piece of genome: it is nucleosome-depleted, accessible, and enriched at active regulatory regions and TAD boundaries. Centring a 90 kb window on one does not make the window CTCF-like, but it does bias the window set toward gene-dense, active, H3K9me3-poor neighbourhoods... and KMT2A still reads at the 7.6th-15.0th percentile of it.

What N4 shows is that the residual is smaller than it looked. And the reason is visible in the medians: the median matched transcribed 90 kb window carries 0.698-0.758 units of H3K9me3 in the -E2 tracks while the median CTCF-anchor 90 kb window carries less - so KMT2A scores low against anchors partly because anchors are already H3K9me3-poor and KMT2A is poorer still, and partly because 90 kb of transcribed gene body is simply not where this mark lives.

The transferable rule. A class-matched null must match on the class that generates the signal. For a 147 bp element the class is 'CTCF anchor' and N1 is right. For a 90 kb gene body the class is 'gene body of this length and this expression', and a CTCF-anchor null answers a different question. Q112, Q114 and this beat all scored a 90 kb interval against a CTCF-anchor null, and every one of those numbers should be read as 'against CTCF anchors' rather than as 'depleted'.

The clause that fired, and why it was the wrong clause

§6's N4 UNINFORMATIVE clause exists to stop a null being used when it has no discriminating power - if transcribed gene bodies are not H3K9me3-poorer than silent ones, the null cannot adjudicate. I set the bar at 1.2x and measured 1.05-1.12x.

The clause fires, and I am not re-tuning it. But two things say the null is not actually powerless. First, the direction is right in 4 of 4 tracks and in all six H3K9me3 tracks read this beat. Second, KMT2A's percentile among silent matched genes (6.51-15.28) is lower than among transcribed matched genes (18.18-23.38) in 4 of 4 - the exact ordering the transcription model predicts, on the same windows.

So the correct statement is: the transcribed-versus-silent contrast over 90 kb gene bodies in these MCF-7 tracks is real, consistent and small (5-12%), and I pre-registered a threshold that a 5-12% effect could not clear. A future run should set this clause on a distribution rather than a round number - for example, requiring that the transcribed and silent medians differ by more than the between-replicate difference in the same medians.

Facts and figures
scriptanalyses/q116-h3k9-axis-at-the-locus/prep_n4.py and run116.py and score.py
prereganalyses/q116-h3k9-axis-at-the-locus/PREREGISTRATION.md, section 6
statusrun
run date2026-09-23
run byQ116-h3k9-axis-at-the-locus
complete runtrue
prior recorded before runningDEFLATES, prior_probability 0.65
verdict by the preregistered ruleN4 UNINFORMATIVE - the silent-over-transcribed median ratio is 1.053, 1.122, 1.110 and 1.096, below the declared 1.2x in 4 of 4 tracks
verdict setting that clause asideSURVIVES - KMT2A is below the 25th percentile of the 385 matched transcribed gene bodies in 4 of 4 tracks
honest labelPARTLY DEFLATES. Matching on length and expression roughly doubles the percentile (7.6-15.0 -> 18.2-23.4) and roughly halves the fold deficit (0.62-0.82x of random 90 kb tiles -> 0.77-0.94x of matched transcribed gene bodies). The depletion is real against the right null and about half as deep as the wrong null made it look.
bcr noteThe 7,465 bp BCR core sits HIGHER than the 90 kb gene body it is inside, in all four tracks, and is not depleted at all in the shCTL pair (63.61st and 61.69th). In MCF-7 the methyl-poor thing is the gene, not the breakpoint.
c6 reproduction of q114PASSES. Interval means identical to four decimal places; percentiles within 1.1 points, attributable to deviation D1 (147 bp versus 146 bp null geometry).
what this does NOT testThe DIPG, RPE and K562 arms of the methyl-poor claim. N4 is built from MCF-7's own expression and cannot be transported. Nothing here touches a haematopoietic cell, a KMT2A-rearranged cell, or the damage state.
localizationchromatin at chr11q23.3 (the 90,377 bp KMT2A gene-body window and the 7,465 bp BCR core) against 385 matched transcribed and 661 matched silent 90,377 bp gene-body windows across all 23 chromosomes
subnuclear localizationnot measured
cell typeMCF-7 (ER+ breast adenocarcinoma) for the chromatin; MCF-7 (ENCSR000CWQ) for the expression match - a different culture in a different laboratory
cell statehormone-deprived culture, 40 min oestradiol, and shCTL-transduced hormone-deprived culture
epigenetic readoutH3K9me3 over the KMT2A gene body against a transcription-matched gene-body null
shown for fusionNO - MCF-7 carries no KMT2A rearrangement, and the transcription-matched null has never been run in any KMT2A-rearranged or haematopoietic cell
absence classmeasured-not-analysed
absence class noteThe data and the annotation were public; the null was named by Q114 and not computed. This entry computes it.
self correction 2026 09 23An earlier draft of this entry and of its edge file recorded mechanism-h3k9me1-2-ctcf-balance-gates-kmt2a-rearrangement as REFUTED by this null. That was wrong and was caught before hand-back by re-reading the mechanism entry: it is an H3K9me1/me2 mechanism and this null measures H3K9me3. The edge is now related_to with confidence weak, and the section says NOT TESTED.
per track
GSM45852070.462GSM4585207: 0.462GSM41338210.573GSM4133821: 0.5725GSM41338200.605GSM4133820: 0.6048GSM45852060.655GSM4585206: 0.6546
results
[
  "analyses/q116-h3k9-axis-at-the-locus/n4_windows.json",
  "analyses/q116-h3k9-axis-at-the-locus/results.json",
  "analyses/q116-h3k9-axis-at-the-locus/results_scored.json"
]
null N4
{
  "n": 385,
  "definition": "90,377 bp midpoint-centred windows on GENCODE v29 protein-coding genes with span within +/-25% of 90,377 bp AND MCF-7 TPM within a factor of 2 of KMT2A's 32.94",
  "median_span_bp": 86492,
  "median_tpm": 29.6,
  "span_range_bp": [
    67842,
    112966
  ],
  "tpm_range": [
    16.51,
    64.98
  ],
  "chromosomes_represented": 23,
  "expression_match_dropped": false,
  "identities_persisted": "analyses/q116-h3k9-axis-at-the-locus/n4_windows.json"
}
null N4b silent contrast
{
  "n": 661,
  "definition": "the same windows on length-matched genes with MCF-7 TPM < 1",
  "median_span_bp": 86502,
  "median_tpm": 0.03
}
kmt2a
{
  "ensg": "ENSG00000118058",
  "gencode_v29_span_bp": 90369,
  "mcf7_tpm": 32.94,
  "interval_scored": "chr11:118,436,456-118,526,832 (90,377 bp, the stored kmt2a_gene interval)"
}
per track — the numbers behind the chart
{
  "GSM4585206": {
    "arm": "baseline -E2",
    "mean": 0.6546,
    "pct_vs_ctcf_anchors": 15,
    "pct_vs_N4_transcribed": 23.38,
    "pct_vs_N4b_silent": 15.28,
    "over_N4_median": 0.938,
    "over_random_90kb_tile_mean": 0.817,
    "silent_over_transcribed": 1.053
  },
  "GSM4585207": {
    "arm": "baseline +E2",
    "mean": 0.462,
    "pct_vs_ctcf_anchors": 9.22,
    "pct_vs_N4_transcribed": 21.04,
    "pct_vs_N4b_silent": 6.96,
    "over_N4_median": 0.843,
    "over_random_90kb_tile_mean": 0.645,
    "silent_over_transcribed": 1.122
  },
  "GSM4133820": {
    "arm": "shCTL -E2",
    "mean": 0.6048,
    "pct_vs_ctcf_anchors": 7.6,
    "pct_vs_N4_transcribed": 18.18,
    "pct_vs_N4b_silent": 7.72,
    "over_N4_median": 0.798,
    "over_random_90kb_tile_mean": 0.636,
    "silent_over_transcribed": 1.11
  },
  "GSM4133821": {
    "arm": "shCTL -E2",
    "mean": 0.5725,
    "pct_vs_ctcf_anchors": 7.7,
    "pct_vs_N4_transcribed": 20.26,
    "pct_vs_N4b_silent": 6.51,
    "over_N4_median": 0.766,
    "over_random_90kb_tile_mean": 0.618,
    "silent_over_transcribed": 1.096
  }
}
bcr core percentiles same four tracks
[
  44.21,
  22.67,
  63.61,
  61.69
]
How this could bite us back
  • If the methyl-poor domain is half a transcription artefact, then an intervention designed to restore H3K9 methylation at the breakpoint would be attempting to methylate an actively transcribed gene body - a state the cell maintains for a reason - rather than repairing a lesion specific to this locus. The off-target cost of that intervention is every other transcribed gene of similar size, of which this null names 385 in one cell type.
  • This result does NOT license relaxing concern about the H3K9 axis at this locus. It narrows the effect size and says nothing about the 147 bp element, where H3K9me3 reads an explicit 0.0000 in five of six MCF-7 tracks - a number no gene-body null touches.
Still open
  1. Does the same null deflate the DIPG, RPE and K562 arms of the methyl-poor claim? Each needs its own matched expression file, all three exist in public archives, and none has been run. Until they are, 'four lineages' is one lineage tested against the right null and three tested against the wrong one.
  2. Would matching on Pol II occupancy or H3K36me3 rather than on steady-state TPM change the answer? This same deposit contains Pol II and H3K36me3 ChIP-seq in the same MCF-7 cells, which match on elongation rather than on mRNA abundance, and that is the better match. This beat did not use them.
  3. Is the residual deficit - 0.77x to 0.94x of matched transcribed gene bodies - large enough to be a domain at all, or is it what any gene with this particular chromatin neighbourhood would show? A flanking-domain statistic (the window versus its own 1 Mb surround) would answer it and was not computed here, and it is also hyp-q114's unrun fourth prediction.
  4. Why is the BCR core LESS depleted than the gene body that contains it, in every MCF-7 track measured? If the breakpoint region is the least methyl-poor part of a methyl-poor gene, the mechanism's framing has the scale backwards.
Every source behind this entry (3)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    In four MCF-7 H3K9me3 ChIP-seq tracks, the 90,377 bp KMT2A gene-body window reads the 15.00th, 9.22nd, 7.60th and 7.70th percentile of width-matched same-cell-type CTCF summit windows, and the 23.38th, 21.04th, 18.18th and 20.26th percentile of 385 length- and expression-matched transcribed 90,377 bp gene-body windows - 0.938x, 0.843x, 0.798x and 0.766x the median of that matched set. Against 661 length-matched SILENT gene bodies it reads the 15.28th, 6.96th, 7.72nd and 6.51st.

    GSE139196: H3K9me3 ChIP-seq in MCF-7, GSM4585206, GSM4585207, GSM4133820, GSM4133821, hg38 bedGraph · Yang F, Tanasa B, Micheletti R, Ohgi KA, Aggarwal AK, Rosenfeld MG · NCBI GEO 2019
    What this evidence cannot show

    ONE CELL TYPE. MCF-7 is an ER+ breast adenocarcinoma, is not KMT2A-rearranged, and is not haematopoietic; nothing here transfers to the disease. The matched-gene null is built from MCF-7's OWN expression, so it cannot be applied to the DIPG, RPE or K562 arms of the same claim without their own expression data, and it has not been. All four tracks pass the C1 dynamic-range gate (10.0x-13.4x over a printed non-zero denominator) and the C1b heterochromatin gate (3.07x-4.65x over the chr19q13 KRAB-ZNF cluster, 3.81x-5.42x over the ZNF274 3' window), so the instrument is working for this mark. HOMER depth normalisation means only the within-track percentile is comparable between samples; the fold ratios are within-track and are safe.

    Evidence that argues against this (2)
    • THE PRE-REGISTERED RULE RETURNS 'N4 UNINFORMATIVE', NOT 'SURVIVES'. Section 6 required matched silent gene bodies to carry at least 1.2x the H3K9me3 of matched transcribed ones before the null could adjudicate; measured, the ratio is 1.053, 1.122, 1.110 and 1.096 - below the threshold in 4 of 4 tracks. The SURVIVES reading requires setting aside a clause I wrote before seeing the data, and that is stated in the title of this entry's own section rather than buried. source
    • The chr2 gene desert, declared as a NEGATIVE control, reads 1.59x-2.42x the random-tile mean in these H3K9me3 tracks - it is not negative for this mark, because gene deserts are heterochromatic. A reader could argue that the right comparator for a 90 kb window is neither a CTCF anchor nor a matched gene but the local chromatin domain, and this run does not compute a flanking-domain statistic. hyp-q114's fourth prediction asks for exactly that and it remains unrun. source
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    KMT2A (ENSG00000118058) has a GENCODE v29 gene span of 90,369 bp and an MCF-7 TPM of 32.94. Of 19,864 protein-coding genes on chr1-22 and chrX with a TPM, 2,013 are within +/-25% of 90,377 bp in span after excluding KMT2A and everything within 1 Mb of the locus, and 385 of those are also within a factor of 2 of 32.94 TPM (median span 86,492 bp, median TPM 29.60, spread over all 23 chromosomes). 661 length-matched genes have TPM < 1.

    What this evidence cannot show

    Both files fetched 2026-09-23. The annotation is GENCODE v29 because that is the exact annotation of the quantification file, so the join is by construction exact and no cross-release gene-id mapping is involved. The MCF-7 RNA-seq is ENCODE's, not the ChIP deposit's, so the expression and the chromatin come from different MCF-7 cultures in different laboratories - a real mismatch, and the reason the expression match is a factor of 2 rather than anything tighter. The 385 identities ARE PERSISTED in analyses/q116-h3k9-axis-at-the-locus/n4_windows.json (rule 90) and can be re-derived. Windows are midpoint-centred and of fixed 90,377 bp width, so for genes shorter than that the window includes flanking sequence and for longer ones it truncates the gene - a deliberate choice so that every window has the identical width to the KMT2A interval.

    Evidence that argues against this (1)
    • Matching on gene-level TPM does not match on transcription-unit activity across the body. A 90 kb gene can carry its Pol II at the promoter and almost nowhere else, and KMT2A is a documented pause-release-regulated gene. This deposit contains Pol II and H3K36me3 ChIP-seq in the same cells (GSM4585208/09, GSM4133802-05) which would match on elongation rather than on steady-state mRNA, and this run did not use them - the budget went to the H3K9 panel. That is the better match and it exists. source
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    This run reproduces Q114's MCF-7 H3K9me3 numbers at the locus: GSM4585206 gives a KMT2A gene-body mean of 0.6546 (Q114: 0.6546) at the 15.00th percentile (Q114: 15.77) and exon-11 at 0.0000 (Q114: 0.0000); GSM4585207 gives 0.4620 (Q114: 0.4620) at the 9.22nd (Q114: 10.30) and exon-11 at 0.3900 (Q114: 0.3900) at the 50.03rd (Q114: 50.03).

    What this evidence cannot show

    The interval MEANS are identical to four decimal places, so the two runs read the same numbers from the same files. The PERCENTILES differ by 0.77 and 1.08 points because Q114's summit windows were 147 bp wide while the exon-11 interval it scored is 146 bp half-open, and this run derives the null width from the interval (deviation D1). This is the pre-registered C6 control and it PASSES at the declared +/-3 percentile-point tolerance.

    Evidence that argues against this (1)
    • A reproduction by the same author of the same code path against the same files is the weakest kind of reproduction. It establishes that this beat's instrument matches Q114's; it establishes nothing about whether either is right. Both share the same null, the same deposit, the same normalisation and the same unverified antibody. source
Connected to