Both arrows on KDM3B depletion are broad: H3K9me1 rises genome-wide AND CTCF falls at 84.6% of 12,443 anchors - the KMT2A anchor's 0.40x is the 3.6th percentile among strength-matched anchors, not a unique event
also called the specificity is in the CTCF step · CTCF 0.30x at exon 11, 1.00x everywhere else · where the KMT2A brake gets its address
Knocking down one enzyme raises a silencing tag all over the genome, and it also loosens the anchor protein all over the genome - at most anchors measured, not just at KMT2A. The KMT2A anchor is among the harder-hit ones, but it is not picked out.
The same thing, for a scientist
Reading Gray 2023's own deposited tracks over eight loci in the same cells separates two steps of the published chain that the paper reports together [E1]. **The mark step is not locus-selective.** On KDM3B depletion H3K9me1 rises 2.67x over the *KMT2A* BCR - but also 1.85x on the BCR's own 20 kb telomeric flank, 1.76x at the *GAPDH* promoter, 1.64x at the *ACTB* promoter, 1.46x over *TCF3* and 1.38x over *MLLT1*, a locus the paper scores as cytogenetically **non-responsive** [E1]. **The CTCF step is sharply locus-selective.** In the same experiment CTCF occupancy falls to **0.30x** of control over canonical exon 11 (6.96x to 2.10x the track's chr11 mean), 0.32x at the summit and 0.40x over the 1,048 bp element, with both replicates of siKDM3B below both controls - while CTCF is **unchanged** at both 20 kb flanks (1.04x, 0.99x), over *MLLT1* (0.93x), *MLLT3* (1.01x), *TCF3* (0.98x), *AFF3* (1.03x), at the *GAPDH* promoter (0.89x) and at the DCAL functional cut point 3.8 kb away (1.02x) [E1]. **CORRECTED IN CYCLE 60 - THE CTCF STEP IS NOT LOCUS-SPECIFIC EITHER, AND THIS ENTRY'S OWN REFUTATION TEST IS WHAT SHOWED IT.** `data.the_check_that_would_refute_it` specified an anchor-matched genome-wide comparison in Gray's own tracks, predicted that hundreds of comparable anchors would also fall, and recorded NOT RUN. It was run in cycle 60 (`analyses/q111-ctcf-anchor-matched-ranking/`), and the prediction was correct. Anchors were called from Gray's own two siCT control tracks - never an external peak list - giving 12,443 anchors, of which **84.6% fall to <=0.9x, the genome-wide median is 0.696x, 1,436 fall to <=0.5x, 280 to <=0.4x and 20 to <=0.3x**. Against 1,625 anchors matched within +/-10% of the KMT2A anchor's baseline strength, the KMT2A anchor ranks **58th, the 3.57th percentile**, and the **matched median is 0.633x - not 1.0x**, with **308 of the 1,625 (19.0%) at least halving**. CTCF loss also **scales with baseline anchor strength** (decile medians 0.779 to 0.615), which is the confound the eleven hand-picked intervals below could not address: they compared a strong anchor with gene bodies and housekeeping promoters, not with anchors. So the eleven per-interval ratios in `data.ctcf_sikdm3b_over_control` remain correct as measurements and are retained, but they do not support locus specificity. What survives is that the KMT2A anchor is **unusually sensitive, in the top few percent of a continuous distribution with no separated outlier** - which is the prediction of `hyp-the-kmt2a-anchor-is-hypersensitive-not-uniquely-targeted`, now confirmed on its own stated falsifier. RPE cells only.
The eight loci, one experiment, two different kinds of answer
All figures are siKDM3B / control ratios of fold-over-chr11-mean, two replicates per arm, from the four-arm main series of GSE210480 1.
| locus | H3K9me1 | H3K9me2 | H3K9me3 | CTCF | paper's cytogenetic call |
|---|---|---|---|---|---|
| KMT2A canonical exon 11 | 3.85 | 0.44 | 1.04 | 0.30 | responsive |
| KMT2A CTCF element (1,048 bp) | 3.74 | 0.44 | 1.01 | 0.40 | responsive |
| KMT2A BCR core (7,466 bp) | 2.67 | 0.69 | 1.34 | 0.62 | responsive |
| KMT2A DCAL cut point | 2.37 | 0.62 | 1.55 | 1.02 | - |
| KMT2A 20 kb centromeric flank | 1.01 | 1.26 | 1.01 | 1.04 | - |
| KMT2A 20 kb telomeric flank | 1.85 | 1.06 | 1.23 | 0.99 | - |
| TCF3 | 1.46 | 1.23 | 1.03 | 0.98 | responsive |
| AFF3 | 0.86 | 1.06 | 1.10 | 1.03 | responsive |
| MLLT1 | 1.38 | 1.43 | 1.25 | 0.93 | non-responsive |
| MLLT3 | 0.90 | 0.97 | 1.09 | 1.01 | non-responsive |
| GAPDH promoter | 1.76 | 0.52 | 1.11 | 0.89 | - |
| ACTB promoter | 1.64 | 1.01 | 1.91 | 1.21 | - |
Read the CTCF column first. Twelve intervals, one of them moves. The three KMT2A rows that contain the anchor fall by 60-70%; every other interval, including the KMT2A cut point 3.8 kb away and both KMT2A flanks 20 kb away, sits between 0.89x and 1.21x. That is a clean, internally controlled, locus-specific effect and it is the strongest single number this beat produced.
Then read the H3K9me1 column. KMT2A is the largest value, and its BCR ranks 72 of 18,082 chr11 bins for the same ratio - so the mark change is genuinely strongest here. But it is not confined here: two housekeeping promoters and a locus the paper calls non-responsive all rise 1.4-1.8x. The paper reports 6,386 KDM3B peaks overlapping CTCF peaks genome-wide and 17,077 CTCF peaks losing intensity, so a broad mark response is entirely consistent with what the authors describe genome-wide; what is new here is the direct side-by-side with the CTCF column in the same intervals.
And two rows break the mark-to-cytogenetics mapping. MLLT1 is scored non-responsive by FISH yet its marks move (me1 1.38x, me2 1.43x). AFF3 is scored responsive yet its me1 does not move (0.86x). Of the four loci the paper FISH-scored, the mark response matches the cytogenetic call in two and contradicts it in two. The mark is neither necessary nor sufficient at those four loci - measured, in the paper's own data. See ana-q108-mark-response-versus-cytogenetic-response-at-four-loci.
What this does to the chain, and what would break it
The published chain is: KDM3B loss -> H3K9me1/2 up at KMT2A -> CTCF off at KMT2A -> copy gain and break-apart at KMT2A. Each arrow carries the locus label.
What the data support instead: KDM3B loss -> H3K9me1 up at many loci including KMT2A most strongly -> CTCF off at KMT2A specifically -> the cytogenetic event. The specificity enters at the second arrow. That is a materially different mechanism, because it means the KMT2A anchor is unusually sensitive to a general mark change rather than being the only place the mark changes. A sensitive anchor is a property of the element - its motif strength, its cohesin load, its flanking sequence, its replication timing - not of the enzyme's targeting.
Three things would break this reading.
- A CTCF ChIP with more anchors. This beat measured CTCF at twelve intervals, only one of which is a strong CTCF anchor: the KMT2A element itself. Gene bodies of MLLT1, MLLT3, TCF3 and AFF3 average over many non-anchor bases and would dilute any anchor-level change inside them. If a genome-wide analysis of Gray's own CTCF tracks shows hundreds of anchors losing CTCF by 0.3-0.4x, then the CTCF step is not locus-specific either and this entry is wrong. The paper's own figure of 17,077 CTCF peaks losing intensity makes that the likely outcome, and it is the single most important check this entry does not contain.
- Anchor-matched controls. The right comparison is KMT2A exon 11 against other CTCF anchors of similar strength, not against gene bodies. That analysis is one afternoon with the same four files and is not done here.
- A cell type. RPE. Not blood, not rearranged.
| mechanism class | step-resolution correction to a published chain: which link carries the locus address |
|---|---|
| stage of disease | GENESIS - before any fusion exists |
| the correction | the published chain labels every arrow with the locus; the measurement shows the mark arrow is broad (H3K9me1 rises 1.4-1.9x at housekeeping promoters and at a non-responsive locus) while the CTCF arrow is narrow at the intervals tested (0.30x at KMT2A exon 11, 0.89-1.21x at eleven others) |
| the check that would refute it | RUN IN CYCLE 60, AND IT REFUTED THIS ENTRY'S HEADLINE. The check specified here was an anchor-matched genome-wide comparison in Gray's own CTCF tracks, with the stated refutation condition 'if hundreds of CTCF anchors of comparable strength also fall to 0.3-0.4x, the CTCF step is not locus-specific either', and the stated expectation that the paper's own 17,077 losing peaks made this the likely outcome. Result: 308 of 1,625 strength-matched anchors fall to <=0.5x and 64 to <=0.4x; 1,436 of 12,443 anchors genome-wide fall to <=0.5x and 280 to <=0.4x. The refutation condition is met. See analyses/q111-ctcf-anchor-matched-ranking/. |
| localization | chromatin at eight genomic loci on chr2, chr7, chr9, chr11, chr12 and chr19; this is an occupancy comparison and asserts nothing about protein localisation |
| subnuclear localization | not measured. The KMT2A element is a CTCF/cohesin anchor, which implies a loop-boundary position, but no imaging or Hi-C was read here |
| cell type | hTERT-immortalised human retinal pigment epithelial (RPE) cells - epithelial, non-haematopoietic, no KMT2A rearrangement |
| cell state | asynchronous proliferating culture 72 h after siRNA transfection; no drug, damage or differentiation arm |
| epigenetic readout | H3K9me1/2/3 and CTCF ChIP-over-input coverage compared interval-by-interval across eight loci in the same cells; histone tracks at 1 kb, CTCF at 200 bp |
| shown for fusion | Not shown for any fusion. All measurements are in unrearranged RPE; the loci compared include two KMT2A fusion-partner genes (MLLT1, MLLT3) and two lymphoid rearrangement genes (TCF3, AFF3), but no fusion-bearing cell was assayed and no fusion protein is involved in the mechanism described. |
| absence class | measured-not-analysed |
| the ID is now misleading | The node id still contains 'only' and asserts a claim this entry no longer makes. It is retained deliberately because edges across the graph reference it and renaming would break them; the title, summary and this block carry the correction. |
| what survives | The eleven per-interval CTCF ratios and the H3K9me1 ratios are measurements and stand. What does not stand is the inference from them that the CTCF step is locus-specific: the intervals compared a strong anchor against gene bodies and promoters rather than against other anchors, and CTCF loss scales with baseline anchor strength. |
ctcf sikdm3b over control — the numbers behind the chart
{
"kmt2a_canonical_exon_11": 0.3,
"kmt2a_ctcf_summit_pm250": 0.32,
"kmt2a_ctcf_element_1048bp": 0.4,
"kmt2a_bcr_core_7466bp": 0.62,
"kmt2a_dcal_cut_point": 1.02,
"kmt2a_flank_centromeric_20kb": 1.04,
"kmt2a_flank_telomeric_20kb": 0.99,
"MLLT1": 0.93,
"MLLT3": 1.01,
"TCF3": 0.98,
"AFF3": 1.03,
"GAPDH_promoter": 0.89,
"ACTB_promoter": 1.21
}h3k9me1 sikdm3b over control — the numbers behind the chart
{
"kmt2a_canonical_exon_11": 3.85,
"kmt2a_bcr_core_7466bp": 2.67,
"kmt2a_flank_telomeric_20kb": 1.85,
"kmt2a_flank_centromeric_20kb": 1.01,
"GAPDH_promoter": 1.76,
"ACTB_promoter": 1.64,
"TCF3": 1.46,
"MLLT1": 1.38,
"MLLT3": 0.9,
"AFF3": 0.86
}the anchor matched result
{
"source": "analyses/q111-ctcf-anchor-matched-ranking/ (design.py pre-registered before any signal value was read)",
"anchors_called_from": "Gray's own two siCT control tracks, chr1-22+X, 200 bp grid, 1,000 bp windows, both control reps >=2x",
"n_anchors": 12443,
"kmt2a_anchor": {
"summit_grch38": 118488685,
"baseline_strength": 4.9162,
"strength_percentile": 0.9175,
"ratio": 0.3964
},
"matched_within_10pct": {
"n": 1625,
"rank": 58,
"percentile_pct": 3.57,
"median_ratio": 0.6327,
"n_le_0.5x": 308,
"n_le_0.4x": 64
},
"other_matched_sets": {
"nearest_500": {
"rank": 22,
"percentile_pct": 4.4
},
"strength_decile": {
"rank": 61,
"n": 1245,
"percentile_pct": 4.9
},
"all_anchors": {
"rank": 258,
"n": 12443,
"percentile_pct": 2.07
}
},
"genome_wide": {
"fraction_le_0.9x_pct": 84.6,
"median_ratio": 0.6957,
"n_le_0.5x": 1436,
"n_le_0.4x": 280,
"n_le_0.3x": 20
},
"ctcf_loss_scales_with_baseline_strength": "decile medians 0.779 -> 0.615",
"king_verification": "every figure in this block was recomputed by the king from the beat's persisted anchors_S.npy and anchors_R.npy and matched exactly: n=1625, rank 58, 3.57%, matched median 0.6327, 308 at <=0.5x, all-anchor rank 258 of 12,443, genome-wide median 0.6957, 1,436/280/20, 84.6% at <=0.9x",
"THE_BRACKET_NOT_THE_POINT_ESTIMATE": "3.57% is the favourable end of a range, and the unfavourable end must be quoted with it. The beat's own PRE-DECLARED bias control - call anchors and take baseline strength from control replicate 1, then take the denominator from replicate 2 alone - puts the KMT2A anchor at the 31.9th PERCENTILE. The reason is that the two control replicates disagree roughly two-fold at this element (exon 11: 9.243x vs 4.688x), so the four per-replicate ratios span 0.200-0.499x. With n=2 versus n=2 the bracket 3.6th-31.9th percentile cannot be narrowed, and the honest statement of degree is that bracket."
}- If the specificity lives at the CTCF step, then a G9a inhibitor given prophylactically is acting on the broad, non-selective arm of the chain - raising or lowering H3K9 methylation at housekeeping promoters and thousands of other loci to protect one. That is a much worse therapeutic ratio than the published chain implies, and it is an argument against the intervention that the paper's own framing hides.
- This entry corrects a published chain on the strength of eleven control intervals, only one of which is a CTCF anchor. The correction could itself be the artefact. It is written as moderate confidence with the refuting experiment named because the alternative - saying nothing until the anchor-matched analysis is done - would leave the graph asserting the locus label on an arrow the data do not support.
- ANSWERED IN CYCLE 60, AGAINST THIS ENTRY. Yes - 1,436 of 12,443 anchors fall to <=0.5x, 280 to <=0.4x and 20 to <=0.3x in Gray's own tracks, and 308 of the 1,625 anchors matched to the KMT2A anchor's strength at least halve. The question correctly identified itself as the one check that decides whether this entry is a real mechanistic correction or an artefact of comparing an anchor with gene bodies, and the answer is the latter. The remaining open part is narrower: the degree is a bracket from the 3.6th to the 31.9th percentile because the two control replicates disagree two-fold at this element, and separating those requires more than n=2 versus n=2.
- Is the KMT2A exon-11 anchor unusually SENSITIVE rather than unusually targeted? If the mark rises broadly and CTCF falls narrowly, the element's own properties - motif score, cohesin load, replication timing, flanking CpG state - are what select it. None of those is recorded for this element in this graph, and the motif has never even been scanned.
- Why does H3K9me1 rise 1.85x on the BCR's telomeric 20 kb flank and 1.01x on its centromeric flank? The paper records de novo breakpoints skewed centromeric and therapy-related breakpoints skewed telomeric within the BCR. A mark change confined to the telomeric side is the shape that mechanism would predict, and no entry in this graph records it.
- Does the H3K9me1 rise also scale with baseline level, the way the CTCF loss scales with baseline anchor strength? If it does, the rank-72-of-18,082 figure in ana-q108-chr11-rank-of-the-kmt2a-bcr-h3k9-response needs the same matching correction this entry just received, and the tracks have already been read.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Read from Gray 2023's deposited GSE210480 bigWigs in RPE cells: on KDM3B depletion CTCF occupancy falls to 0.30x of control over Ensembl canonical KMT2A exon 11, 0.32x at the CTCF summit and 0.40x over the 1,048 bp CTCF/RAD21/SMC3 element, with both siKDM3B replicates below both controls, while CTCF is unchanged (0.89x-1.21x) at the KMT2A 20 kb flanks, the DCAL cut point, MLLT1, MLLT3, TCF3, AFF3 and the GAPDH and ACTB promoters; in the same samples H3K9me1 rises 2.67x over the KMT2A BCR but also 1.85x on its telomeric flank, 1.76x at the GAPDH promoter, 1.64x at the ACTB promoter, 1.46x over TCF3 and 1.38x over MLLT1; and MLLT1, scored cytogenetically non-responsive by the paper, shows a 1.38x H3K9me1 and 1.43x H3K9me2 rise while AFF3, scored responsive, shows none (0.86x).
GEO GSE210480 - 4 unique RPE CTCF ChIP-seq bigWigs (GSM6431019/21/23/25) and 24 H3K9me1/2/3 bigWigs, read at interval resolution over HTTP range requests with pybigtools 0.3.0, 2026-09-22; scripts and full output at analyses/q108-gse210480-h3k9me3-kmt2a-bcr/ · 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) 2023What this evidence cannot show
The control loci are GENE BODIES (MLLT1, MLLT3, TCF3, AFF3) and 1 kb promoter windows, not CTCF anchors. A gene body averages over tens of kilobases of mostly non-anchor sequence, so an anchor-level CTCF change inside one would be heavily diluted and could be missed entirely. This is the weakest part of the design and it means the entry's claim is properly stated as 'CTCF fell at the KMT2A anchor and not at these particular intervals', NOT 'CTCF fell only at KMT2A genome-wide'. n = 2 per arm, no spike-in, histone tracks 1 kb binned and CTCF tracks 200 bp binned. Values are fold over each track's own chr11 mean; the chr11 mean itself shifts modestly between arms. hTERT-RPE cells - epithelial, non-haematopoietic, no KMT2A rearrangement. The paper's cytogenetic calls for MLLT1, MLLT3, TCF3 and AFF3 are taken from this graph's mechanism entry, which records them from the paper's FISH panels.
Evidence that argues against this (1)
- The paper itself reports that 17,077 CTCF peaks LOSE intensity on KDM3B depletion genome-wide, of which 1,005 are KDM3B co-occupied. If that figure is right, then CTCF loss is a genome-wide phenomenon affecting thousands of anchors and the KMT2A anchor is not special in kind - only, perhaps, in degree. This entry's locus-specificity claim is bounded by eleven non-anchor intervals and would be refuted by an anchor-matched genome-wide comparison, which has not been run. The paper's own number is the strongest argument against this entry's framing. Gray ZH et al. 2023
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The published chain places the locus specificity at the mark step: KDM3B binds across KMT2A with a strong peak in the breakpoint cluster region where H3K9me1 is baseline-depleted, KDM3B loss raises H3K9me1 across that region, CTCF occupancy at the exon-11 site then falls without any change in total CTCF protein, and the locus undergoes copy gain and break-apart; TCF3 and AFF3 respond the same way while MLLT1 and MLLT3 do not.
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 2023What this evidence cannot show
The chain as summarised here is taken from the paper's abstract and from this graph's mechanism entry rather than from a fresh reading of every figure panel; the KDM3B occupancy link comes from reused public data in colorectal cancer stem cells (GSE71885), not from RPE. The FISH calls for the four comparison loci are cytogenetic (probe separation and copy number), not sequenced junctions - the paper reports no sequenced junction anywhere.
Evidence that argues against this (1)
- The paper does not claim the mark change is KMT2A-exclusive; it reports genome-wide H3K9me1 and H3K9me2 skews on megabase scales (81% of 109 Mb reset for me1, 75% of 292 Mb for me2), which is a broad effect by construction. So the 'locus label on the first arrow' is partly this graph's reading of the paper rather than the paper's own claim, and that reading is what this entry corrects. Iqbal Lab oncofusion knowledge graph 2026
the cycle-60 refutation that prompted this sweep - its falsifier field said NOT RUN and predicted its own outcome
agrees from the other direction: the resting MARK depletion is broad-domain over the 7.5 kb BCR while the element is unresolved, so element-level specificity is not the mark's - but the two datasets measure different quantities and that caveat is in the entry
the locus label belongs on the CTCF arrow, not the mark arrow: me1 rises 1.4-1.8x at housekeeping promoters and at a non-responsive locus
CTCF occupancy falls to 0.30x at KMT2A canonical exon 11 on KDM3B depletion and is unchanged at eleven other intervals
KDM3B depletion is the perturbation measured
the anchor whose occupancy falls is the element measured there
same files, same beat; the numbers come from that analysis
the mark rises broadly, which is why it cannot carry the locus address alone
MLLT1's mark moves and its CTCF does not; consistent with the address living at the CTCF step
the CTCF result is n=2 versus n=2 from one experiment, and no siG9A CTCF arm exists to test the balance shape
the broad-mark / narrow-CTCF measurement is the pattern this hypothesis explains
the word 'only' fails: 308 of 1,625 strength-matched anchors also fall to <=0.5x and the matched median is 0.63x
the anchor-matched ranking that entry named as its own refuting experiment and predicted would fail
its eleven comparators are unmatched on the covariate that drives the effect, and the gradient is measured: decile medians 0.78 to 0.62
whether CTCF occupancy loss at an anchor is sufficient for a focal cytogenetic event