Oncofusion

start / graph / risk-a-kdm3b-or-g9a-intervention-moves-ctcf-at-most-anchors-in-the-genome

Riskmoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q111-ctcf-anchor-ranking

The off-target surface, measured rather than feared: KDM3B depletion moves CTCF occupancy at 84.6% of 12,443 RPE anchors and halves it at 1,436 - so an H3K9-axis drug given to prevent one rearrangement is remodelling loop architecture genome-wide, and preferentially at the strongest anchors

also called 84.6% of anchors move · the borrowed-specificity problem, quantified · why prophylaxis is the wrong indication for this axis

The idea behind this drug is to keep one gene's loop anchor in place so the gene does not break. But when the enzyme is removed, the anchor protein shifts at most loop anchors in the whole genome - about eight in ten - and drops by half at more than a thousand of them. Giving a healthy infant a drug that rearranges the genome's loop structure everywhere, to protect one gene, is a very different proposition from the published picture.

The same thing, for a scientist

PMID 37788669 ends with *"epigenetic regulators that are tractable drug targets, which has clinical implications"*, and reports that G9a inhibition suppresses *MLL/KMT2A* events accompanying doxorubicin in human cells and in mice [E2]. This graph has carried that as a candidate **prophylactic** strategy for therapy-related *KMT2A* rearrangement. **The off-target surface of that strategy is now measured in the paper's own data** [E1]. Across 12,443 CTCF anchors called from Gray's own RPE control tracks, KDM3B depletion gives: | | anchors | share | |---|---:|---:| | ratio <= 0.9x | **10,521** | **84.6%** | | ratio <= 0.7x | 6,325 | 50.8% | | ratio <= 0.5x | **1,436** | 11.5% | | ratio <= 0.5x with both replicates consistent | 1,401 | 11.3% | | ratio <= 0.4x | 280 | 2.3% | Population median **0.696**. **And the effect is largest at the strongest anchors** - decile medians fall monotonically from 0.779 to 0.615 - so the anchors most likely to be load-bearing TAD and enhancer-domain boundaries are the ones that move most [E1]. **Three specific hazards follow, none of which is in the published framing.** 1. **Loop-boundary loss is a known oncogenic mechanism in its own right.** A CTCF anchor that weakens can allow an enhancer to reach a proto-oncogene it was insulated from; this graph already records CTCF/cohesin boundary biology at the *HOXA* TAD. Doing that at a thousand anchors to protect one locus is trading a specific risk for a diffuse one. 2. **The specificity is borrowed, not intrinsic.** The *KMT2A* anchor is rank 58 of 1,625 strength-matched anchors (3.6th percentile) and 31.9th percentile under the pre-declared bias control, so the drug is not acting selectively at *KMT2A* - its apparent selectivity comes from whatever downstream fragility makes 11q23.3 break when its anchor weakens [E1]. 3. **The indication may be inverted.** In CD34+ progenitors, granulocytes, K562 and 6 of 10 primary AML samples the *KMT2A* locus carries **no H3K9me2 domain at all**, while **4 of 10 AML samples do** - and the paper's own model makes the AML-accumulated H3K9me2 blocks the G9a-dependent ones [E3]. If the mark this drug removes is present in established leukaemia and absent in the healthy progenitor, the pharmacology points at treatment and not at prevention.

Go deeper
Why this is a risk entry and not a refutation

The mechanism may be entirely correct and the drug may still be wrong for prophylaxis. Nothing here shows that G9a inhibition fails to suppress KMT2A events - the paper shows it does, in cells and in mice, and that result stands. What is measured here is the denominator: how much else the axis moves. A therapeutic-ratio argument needs both, and the published framing supplies only the numerator.

The population matters more than usual. The proposed use is in a child receiving anthracycline chemotherapy, or - in the broader version of the argument - in an infant at risk of a de novo event. That is a developing organism in which TAD boundaries and loop anchors are actively partitioning lineage-specifying enhancer domains, and in which a diffuse, structured perturbation of loop architecture has decades of latency in which to matter. Against a therapy-related leukaemia risk on the order of a few percent, an intervention that shifts occupancy at ten thousand anchors needs a much better selectivity argument than 'the mark rises most at KMT2A'.

And the measurement is in the wrong cell, which cuts both ways. All of this is hTERT-RPE - retinal pigment epithelium, epithelial, non-haematopoietic, unrearranged. The anchor population of a CD34+ progenitor is not this anchor population, and the fraction that moves there is unmeasured. That is a reason to withhold a quantitative safety claim, not a reason to discount the hazard: an 84.6% figure in any human cell type is a statement about how the axis works, and there is no CTCF measurement in any haematopoietic cell to compare it with (absence-no-human-ctcf-or-cohesin-occupancy-measurement-in-a-blood-cell-at-kmt2a).

Mitigation, and the two tests that would change the assessment

Mitigations that are real:

  • Restrict the window. The therapy-related argument is time-limited - the exposure that creates the risk is a defined number of anthracycline doses - so a short co-administration has far less integrated off-target exposure than a chronic prophylactic. This is the strongest available mitigation and it is compatible with the published mouse result.
  • Prefer the arm with the smaller surface. The axis has two sides. KDM3B depletion is what produces the 84.6% figure; G9a inhibition is the intervention actually proposed, and its CTCF consequences have never been measured, because there is no siG9A CTCF arm and no double-knockdown CTCF arm anywhere in GSE210480. It is possible - untested - that the opposing perturbation has a much narrower anchor footprint.
  • Screen for the hazard directly. CTCF ChIP or CUT&RUN under the candidate compound, in the target cell type, with the anchor ranking from this beat as the pre-specified readout. The 12,443-anchor table is persisted, so the analysis is pre-registrable.

Two tests that would change the assessment:

  1. The overlay (exp-overlay-the-ctcf-loss-ranking-on-grays-own-amplified-regions). If the anchors that lose CTCF hardest do not become unstable, then CTCF loss at an anchor is mostly tolerated and the diffuse hazard shrinks - a 0.5x occupancy change with no consequence is not a toxicity. If they do, the hazard is much larger than stated here.
  2. A G9a-inhibitor CTCF profile. One CTCF CUT&RUN in drug-treated cells would say whether the intervention's footprint is the mirror of the depletion's or something much smaller. It is the single most decision-relevant missing measurement in this whole territory and it does not exist in any archive this beat queried.
Facts and figures
kindtoxicity
severityunquantified but structurally large: 84.6% of 12,443 CTCF anchors move and 11.5% halve on genetic depletion of one arm of the axis, in an epithelial cell line; the pharmacological arm's own footprint has never been measured
mitigationRestrict exposure to the anthracycline window rather than chronic prophylaxis - the strongest available mitigation and compatible with the published mouse result. Prefer the arm with the smaller measured footprint, which currently means measuring the G9a-inhibitor arm at all, since no siG9A or drug CTCF profile exists in any archive queried. And screen the hazard directly with CTCF CUT&RUN under the candidate compound in the target cell type, pre-registering the persisted 12,443-anchor ranking as the readout.
deliverynot applicable to a small-molecule methyltransferase inhibitor, but relevant to the population: the proposed use is in a child during or after anthracycline therapy, or in an infant at risk of a de novo event, which is a developing organism in which loop boundaries are actively partitioning lineage-specifying enhancer domains
localizationchromatin, the CTCF anchor population genome-wide; no cellular localisation is asserted
subnuclear localizationnot measured and not claimed - CTCF anchors imply TAD boundaries but no imaging or Hi-C supports any statement here
cell typehTERT-immortalised human retinal pigment epithelial (RPE) cells for the CTCF surface; normal human CD34+ bone-marrow progenitors, granulocytes, K562 and primary AML myeloblasts for the H3K9me2 substrate argument. There is no CTCF measurement in any haematopoietic cell, which is the gap that prevents a quantitative safety claim.
cell stateasynchronous proliferating culture 72 h after siRNA transfection for the CTCF data; resting primary cells for the H3K9me2 data
epigenetic readoutCTCF ChIP-over-control occupancy at 12,443 anchors as the off-target surface; H3K9me2 ENRICHED-domain presence at the KMT2A locus across normal and leukaemic haematopoietic cells as the substrate question
shown for fusionNot shown for any fusion. Every measurement is in cells with no KMT2A fusion, and the events the intervention targets - focal amplification and break-apart of the intact locus - are pre-fusion. The eight AML samples' fusion status is unrecorded in GEO.
affects
[
  "gene-ehmt2",
  "gene-kdm3b",
  "gene-ctcf"
]
the measured surface
{
  "n_anchors": 12443,
  "n_le_0.9": 10521,
  "n_le_0.7": 6325,
  "n_le_0.5": 1436,
  "n_le_0.5_replicate_consistent": 1401,
  "n_le_0.4": 280,
  "population_median_ratio": 0.696,
  "strength_decile_medians": [
    0.779,
    0.751,
    0.731,
    0.731,
    0.714,
    0.694,
    0.672,
    0.653,
    0.64,
    0.615
  ],
  "perturbation_measured": "siKDM3B, a GENETIC depletion of the demethylase - not the proposed G9a inhibition",
  "cell": "hTERT-RPE"
}
three hazards
[
  "loop-boundary loss is itself oncogenic - a weakened CTCF anchor can expose a proto-oncogene to an enhancer it was insulated from, and this is being done at over a thousand anchors to protect one locus",
  "the specificity is BORROWED: the KMT2A anchor is the 3.6th to 31.9th percentile of strength-matched anchors, so the drug is not selective at KMT2A and its apparent selectivity comes from downstream fragility at 11q23.3",
  "the indication may be inverted - the KMT2A locus carries no H3K9me2 domain in normal CD34+ progenitors, granulocytes or K562 but does in 4 of 10 primary AML samples, so the substrate the drug removes is present in disease and absent in the healthy target cell"
]
what would change the assessment
[
  "the overlay in exp-overlay-the-ctcf-loss-ranking-on-grays-own-amplified-regions: if hard-losing anchors do NOT become unstable, occupancy loss is largely tolerated and the diffuse hazard shrinks",
  "a CTCF profile under a G9a inhibitor: the intervention's own footprint, which may be far narrower than the depletion's and has never been measured",
  "an absolute-occupancy readout: a ratio-track change of 0.5x is not half the CTCF molecules, and no spike-in or input exists in this deposit"
]
How this could bite us back
  • This entry could be used to kill a line of work on a number measured in the wrong cell, for the wrong arm of the axis, with n = 2 and no spike-in - and a global ChIP-efficiency shift between arms would produce exactly this broad pattern and cannot be excluded from a deposit whose Input samples have no file. The 84.6% figure is a prompt to measure the intervention's own footprint, not a safety verdict.
  • The inverted-indication argument is the weakest part and it is the part most likely to be quoted, because it is neat. It rests on two separate observations in two different datasets - domain-scale H3K9me2 presence in 4 of 10 unkaryotyped AML samples, and a G9a-dependence conclusion drawn from a different experiment - and this graph's own balance model makes rising H3K9me1/2 the DESTABILISING arm, which points the other way. Treat it as a question, not a recommendation.
  • And the deepest problem is that this graph's three H3K9 measurements at this locus predict different signs for the same manipulation. A safety statement that picks one of them is not a safety statement. The defensible position is: the direction of the H3K9-axis effect on CTCF at this element in a haematopoietic cell is unmeasured, its magnitude in RPE spans 0.20x to 0.50x across replicate pairings, and the off-target surface is large in the one cell type where it has been counted.
Still open
  1. What does a G9a inhibitor do to CTCF occupancy genome-wide? The proposed intervention's own off-target footprint has never been measured in any cell type - GSE210480 has no siG9A and no double-knockdown CTCF arm, and no other archive queried this cycle holds a CTCF profile under a G9a inhibitor. One CUT&RUN would produce the single most decision-relevant number in this territory.
  2. Does a 0.7x change in a fold-over-control CTCF ratio alter insulation? If not, the 84.6% figure is a statement about a measurement and the real surface is the 1,436 anchors at or below 0.5x. Public CTCF-degron Hi-C data relate occupancy to insulation quantitatively and could calibrate this without any new experiment.
  3. What is the absolute magnitude of the therapy-related KMT2A risk this intervention would be weighed against? A prophylactic argument needs a denominator, and this graph has clinical entries on therapy-related leukaemia but has never placed the incidence beside the off-target surface in one place.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    Across 12,443 CTCF anchors called from the two siCT control CTCF tracks of GSE210480 in hTERT-RPE cells, KDM3B depletion leaves 10,521 anchors (84.6%) at or below 0.9x of control occupancy, 6,325 (50.8%) at or below 0.7x, 1,436 (11.5%) at or below 0.5x - 1,401 of those with both siKDM3B replicates below both controls - and 280 (2.3%) at or below 0.4x, with a population median of 0.696; the effect scales with baseline anchor strength, decile medians falling monotonically from 0.779 in the weakest to 0.615 in the strongest; and the KMT2A exon-11 anchor is rank 58 of 1,625 strength-matched anchors (3.57th percentile) in the primary analysis and the 31.9th percentile under a pre-declared split-replicate bias control, so it is not a separated outlier.

    GEO GSE210480 - four unique RPE CTCF ChIP bigWigs read genome-wide at 200 bp with pybigtools 0.3.0, anchors called from the control tracks only, pre-registered design, 2026-09-22; analyses/q111-ctcf-anchor-matched-ranking/design.py, run.py, results.json, anchors.tsv · Gray ZH, Chakraborty D, Duttweiler RR, Alekbaeva GD, Murphy SE, Chetal K, Ji F, Ferman BI, Honer MA, Wang Z, Myers C, Sun R, Kaniskan HU, Toma MM, Bondarenko EA, Santoro JN, Miranda C, Dillingham ME, Tang R, Gozani O, Jin J, Skorski T, Duy C, Lee H, Sadreyev RI, Whetstine JR · NCBI GEO (data accompanying Cell 186:4528-4545.e18) 2023
    PMID 37788669DOI
    What this evidence cannot show

    This measures a GENETIC depletion of KDM3B, not the proposed pharmacological inhibition of G9a - the opposite arm of the axis - and the series contains no siG9A CTCF samples, so the intervention's own footprint is unmeasured. hTERT-RPE cells, not haematopoietic, not rearranged; the anchor population of a progenitor is different and unmeasured. n = 2 versus n = 2 from one experiment, no spike-in, no input file, no FDR on the anchor calls, and 'anchor' means a local maximum of a deposited ratio track above a fixed fold cut. A ratio-track occupancy change is not an absolute occupancy change: 0.5x of a fold-over-control value is not 'half the CTCF molecules', and without an input or a spike-in the absolute scale is unavailable. Finally, a global reduction in ChIP efficiency between arms would appear as a broad shift of exactly this kind and is absorbed into the per-track mean normalisation, which is the single most plausible benign explanation for the 84.6% figure and cannot be excluded from this deposit.

    Evidence that argues against this (1)
    • The strongest benign reading: a modest occupancy change at an anchor may be biologically inert. CTCF binds and unbinds on a timescale of seconds and loop extrusion is dynamic, so a 0.7x or 0.9x change in a population-averaged ChIP ratio need not alter insulation at all - insulation is buffered, and cohesin-CTCF systems tolerate substantial occupancy variation. On that reading only the tail matters, the 1,436 anchors at or below 0.5x are the real off-target surface, and 11.5% is a much less alarming figure than 84.6%. The honest statement is that this entry's headline number counts anchors that MOVED, and no evidence here shows that moving matters. Iqbal Lab oncofusion knowledge graph 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    PMID 37788669 reports that G9a inhibition in human cells or in mice suppresses MLL/KMT2A amplifications and rearrangements accompanying doxorubicin treatment, that doxorubicin suppresses KDM3B and CTCF protein levels, that KDM3B and CTCF overexpression rescues the doxorubicin-induced alterations, and concludes that these are tractable drug targets with clinical implications.

    Epigenetic balance ensures mechanistic control of MLL amplification and rearrangement. · Gray ZH, Chakraborty D, Duttweiler RR, Alekbaeva GD, Murphy SE, Chetal K, Ji F, Ferman BI, Honer MA, Wang Z, Myers C, Sun R, Kaniskan HU, Toma MM, Bondarenko EA, Santoro JN, Miranda C, Dillingham ME, Tang R, Gozani O, Jin J, Skorski T, Duy C, Lee H, Sadreyev RI, Whetstine JR · Cell 2023
    PMID 37788669DOI
    What this evidence cannot show

    Taken from the abstract. The compound, dose, schedule, mouse strain, exposure duration and any toxicity readout in the mouse experiment were not re-read for this entry, and they are exactly what a therapeutic-ratio argument would need. The paper does not claim a prophylactic indication in a healthy child; the prophylactic framing is this graph's extrapolation, and it is the framing this risk entry is aimed at.

    Evidence that argues against this (1)
    • A mouse experiment showing suppression of the events with no reported toxicity is genuine evidence that the intervention is tolerable at an effective dose, and it is stronger evidence about safety than a CTCF occupancy distribution is. An 84.6% anchor-movement figure in an immortalised epithelial line under genetic depletion of the OPPOSITE enzyme is several inferential steps away from a safety claim about a G9a inhibitor in a child, and this entry should not be read as outweighing the in vivo result. Gray ZH et al. 2023
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    In GSE71809's deposited H3K9me2 domain calls the KMT2A locus carries no called H3K9me2 domain in 3 of 3 normal human CD34+ bone-marrow progenitor ChIPs, 4 of 4 granulocyte ChIPs, 2 of 2 K562 ChIPs and 2 of 2 UNC0638-treated K562 ChIPs, but 4 of 10 primary AML myeloblast H3K9me2 ChIPs DO place a domain over it - and the paper reporting that series concludes that the AML-enriched H3K9me2 blocks, in contrast to the heterochromatin-associated ones, are the blocks reduced by pharmacological G9a/GLP inhibition.

    PMID 28301528DOI
    What this evidence cannot show

    Domain calls, median width 12,750-17,000 bp, so this is a domain-scale statement and says nothing about the 1,048 bp element. The 6:4 AML split cannot be read against fusion status because GEO records no karyotype for any of the eight AML samples, and none is known to be KMT2A-rearranged. The G9a-dependence of the AML blocks is the paper's conclusion about its own K562 UNC0638 experiment and the AML blocks collectively, NOT a measurement at the KMT2A locus - at that locus the UNC0638 arm is uninformative because the locus was already domain-free. So the inverted-indication argument is an inference from two separate observations and is graded moderate for that reason.

    Evidence that argues against this (1)
    • The indication argument could be exactly backwards. If the four AML samples that carry H3K9me2 over KMT2A are the ones whose KMT2A locus is already stabilised by the mark, then removing the mark in established disease would destabilise it further - which is a hazard of treating, not an argument for it. Nothing in these data distinguishes 'the mark is there because the locus is being silenced' from 'the mark is there because the locus is protected', and the published balance model makes H3K9me1/2 the DESTABILISING side, which points the opposite way from this entry's reading. Iqbal Lab oncofusion knowledge graph 2026
Connected to