start / graph / ana-q114-the-four-mcf7-suv39h1-tracks-barely-agree-with-each-other-and-estradiol-changes-nothing

AnalysismoderateQ114-suv39h1-at-the-locus

Estradiol does not redistribute SUV39H1 - because the replicates do not agree well enough to see it: over 56,171 CTCF anchors the ACROSS-condition Spearman (0.358) is HIGHER than the within-condition replicate correlations (0.326 for -E2, 0.207 for +E2), and the same 147 bp anchor reads 51.8, 69.1, 90.8 and 96.3 in the four samples

also called the E2 question, answered · SUV39H1 replicate concordance · Q114 MCF-7 reproducibility

The one dataset that measured this enzyme before and after a hormone signal cannot show a change, because two repeats of the same condition agree with each other no better than a before agrees with an after. The repeats disagree so much that a single measurement from this kind of experiment cannot be trusted to tens of percent.

The same thing, for a scientist

**A corpus open question, answered, and the answer is about the instrument.** `exp-suv39h1-chip-seq-in-mcf7-with-and-without-estradiol` asks: *"Does SUV39H1 occupancy redistribute at 40 minutes of estradiol, and if so is the redistribution at promoters or in heterochromatin? The deposit contains both conditions and the comparison appears not to have been published as a SUV39H1 result in its own right."* **No detectable redistribution, because the replicate noise is larger than any condition effect** [E1][E2]. Spearman correlation of mean signal over 56,171 MCF-7 CTCF summit windows: | comparison | Spearman rho | |---|---:| | replicate vs replicate, **-E2** (GSM4133857 vs GSM4133858) | **0.326** | | replicate vs replicate, **+E2** (GSM4133859 vs GSM4133860) | **0.207** | | **-E2 mean vs +E2 mean** (across condition) | **0.358** | **The across-condition correlation is the highest of the three.** Two replicates of the same condition resemble each other *less* than the two conditions resemble each other. There is no room in that for a 40-minute redistribution to be measured, and the honest answer to the open question is that **this deposit cannot answer it**, not that the answer is no. **At the locus, the same thing in a single number.** The 147 bp *KMT2A* exon-11 anchor reads the **51.78th, 69.05th, 90.81st and 96.29th** percentile of the same 56,171 anchors in the four samples - a **44.5-percentile spread** - with the conditions interleaved (−E2: 90.81, 69.05; +E2: 96.29, 51.78). **Why this matters well beyond MCF-7.** This corpus reasons from **one track per protein per cell type** almost everywhere, because that is usually all the archives hold - Q35, Q37, Q110, Q112 and this beat's own DIPG arm all do it, and **not one entry states an uncertainty on such a percentile.** Here four samples of one antibody, one cell line, one lab, one submission give a 44.5-percentile spread at a 147 bp window and a between-replicate rho of 0.21-0.33 over 56,171 windows. **That is the only replicate estimate of this uncertainty the corpus has.** **It does not invalidate the earlier negatives.** A value at the bottom of such a spread is still a negative, and the DIPG arm's exon-11 zero is an explicit zero rather than a low number. What it does invalidate is any *positive* read near the 75th-90th percentile from a single track - and several of this graph's borderline occupancy claims sit exactly there.

Go deeper

What a low replicate correlation does and does not mean here

rho = 0.207-0.326 over 56,171 windows is not zero - it is a real, highly 'significant' positive correlation that says the two tracks are measuring something in common. It is simply far too low to support a per-window comparison.

Three candidate causes, none separable from the deposited files:

Genuine low signal-to-noise. SUV39H1's canonical address is pericentromeric satellite; a track whose real signal is mostly unmappable will look like noise over the mappable genome. Consistent with the same tracks showing no enrichment at the KRAB-ZNF cluster (0.93-1.15x) while the H3K9me3 tracks in the same deposit show 3.05-5.41x there.
Depth and duplicate handling. 44.3 to 49.8 million bedGraph intervals across the four, with '1 copy/read per genome position' - a hard de-duplication that interacts with depth.
The unexplained dsgf token. GSM4133857 and GSM4133859 carry it and are the two HIGH tracks (90.81, 96.29); GSM4133858 and GSM4133860 do not and are the two LOW ones (69.05, 51.78). The split is perfect across four samples. If dsgf denotes a crosslinker or protocol variant, then this is not replicate noise at all but a protocol effect - and a much more useful fact. The deposit does not say, and the corresponding author has not been asked.

Relationship to existing entries

exp-suv39h1-chip-seq-in-mcf7-with-and-without-estradiol posed the estradiol question; it is resolved in place this cycle with this outcome.
ana-q114-in-mcf7-the-whole-kmt2a-locus-is-elevated-and-the-anchor-is-not-picked-out carries the interval-by-interval numbers and the pre-registered verdict for the same tracks; this entry is the reproducibility measurement, which is a different thing and is the one that travels.
concept-one-deposited-file-under-two-accessions-doubles-a-replicate-count is the neighbouring hazard: that entry is about counting two copies of one file as two replicates. This one is about two genuinely distinct files that agree far less than a replicate pair is assumed to. Content-Length was compared for all eight files used in this beat and all eight differ.

Facts and figures

script	analyses/q114-suv39h1-at-the-kmt2a-locus/run_mcf7.py (question_C block)
status	run
run date	2026-09-22
run by	Q114-suv39h1-at-the-locus
result summary	Across-condition Spearman (0.358) exceeds both within-condition replicate correlations (0.326 for -E2, 0.207 for +E2) over 56,171 MCF-7 CTCF anchor windows, so no estradiol redistribution is measurable; the exon-11 147 bp window spans 51.78-96.29 percentile across the four samples.
spearman minusE2 replicates	0.32632806130641345
spearman plusE2 replicates	0.20683047346863936
spearman across condition	0.3576224460340083
n windows	56171
exon11 spread percentile points	44.51
dsgf token splits the four perfectly	GSM4133857 and GSM4133859 carry 'dsgf' and are the two high tracks; GSM4133858 and GSM4133860 do not and are the two low tracks. Meaning of the token not given in the deposit.
content length check	all eight deposited files used in this beat have distinct Content-Length, so no accession pair is one file twice
cell type	MCF-7, ER-positive luminal breast adenocarcinoma
cell state	hormone-deprived versus 100 nM 17-beta-estradiol, 40 minutes
lineage	breast epithelium - NOT haematopoietic, NOT KMT2A-rearranged
shown for fusion	NO - nobody has looked in any KMT2A fusion system
resolves	open question 0 of exp-suv39h1-chip-seq-in-mcf7-with-and-without-estradiol

exon11 percentile by sample

inputs

[
  "GSM4133857",
  "GSM4133858",
  "GSM4133859",
  "GSM4133860",
  "ENCFF942TCG"
]

exon11 percentile by sample — the numbers behind the chart

{
  "GSM4133857_minusE2_dsgf": 90.81,
  "GSM4133858_minusE2": 69.05,
  "GSM4133859_plusE2_dsgf": 96.29,
  "GSM4133860_plusE2": 51.78
}

znf control in the same deposit

{
  "SUV39H1_tracks_over_tile_mean": [
    0.93,
    1.15,
    1.04,
    1.11
  ],
  "H3K9me3_tracks_over_tile_mean_1Mb": [
    3.05,
    4.15
  ],
  "H3K9me3_tracks_over_tile_mean_znf274_3prime": [
    3.79,
    5.41
  ]
}

the spread explained cycle 74

{
  "finding": "the 44.5-percentile spread across the four MCF-7 SUV39H1 tracks is substantially a CROSSLINKER CONTRAST plus ordinary sub-kilobase noise, not four replicates disagreeing",
  "the_partition_I_verified_myself": "the four sample titles split 2/2 on a `dsgf` token, ACROSS the estradiol condition: GSM4133857 SUV39h1_minusE2_dsgf, GSM4133858 SUV39h1_minusE2, GSM4133859 SUV39h1_plusE2_dsgf, GSM4133860 SUV39h1_plusE2. A token that partitions a sample set along the axis being measured is an experimental variable, not a serial number.",
  "what_dsgf_expands_to_REPORTED_NOT_VERIFIED_BY_ME": "the cycle-68 beat reports `dsgf` = DSG + formaldehyde double crosslinking, read from the paper's own methods. I verified the PARTITION from the sample titles but did NOT read the methods section, so the expansion is agent-reported. Within-protocol spread is reported as 5.48 and 17.27 against 44.5 across protocols.",
  "and_the_residual_is_just_window_width": "measured genome-wide for the first time: the median within-condition absolute percentile difference at 146 bp is 23.48 points over 56,171 anchors and 12 pairs, with per-pair p90 of 56.9-82.7. The exon-11 element sits at the 35.8th percentile of that noise distribution - it is NOT unusually noisy; 146 bp windows are. At 90 kb the same comparison reproduces to 4.56 points.",
  "consequence_for_this_entry": "its headline - that the four tracks barely agree - is right about the number and wrong about the cause. Two of the four differ from the other two by crosslinking protocol, and what remains is the generic noise of a 146 bp window rather than a property of this locus or this antibody.",
  "resolved_by": "the cycle-68 beat; partition verified by the king 2026-09-23"
}

How this could bite us back

Any therapeutic argument in this graph that rests on a single occupancy track - which is most of them - rests on a number whose replicate uncertainty has never been stated, and the one estimate now available is tens of percentile points wide.
The tempting use of this entry is to dismiss the MCF-7 deposit. That would be wrong in a specific way: its H3K9me3 tracks pass the heterochromatin control at 3-5x while its SUV39H1 tracks do not, so the deposit is not uniformly poor - the SUV39H1 arm specifically is.

Still open

What is 'dsgf'? It splits the four samples perfectly into the two high and the two low tracks, which means the 44.5-percentile spread may be a protocol effect rather than replicate noise. One email to the corresponding author would settle it and would change how this deposit should be used.
Does the low replicate concordance recover at a coarser scale? Over 10 kb or 100 kb bins the same four tracks might agree well, in which case they are usable for domain-scale statements and unusable for element-scale ones. That recomputation is cheap and was not run.
How many occupancy percentiles in this corpus are single-track reads near the 75th-90th percentile? Those are the ones this spread puts at risk, and nobody has listed them.

Every source behind this entry (2)

E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The four GSE139196 SUV39H1 ChIP-seq samples in MCF-7 (two minus estradiol, two at 40 min estradiol) and the matched H3K9me3 pair (GSM4585206 -E2, GSM4585207 +E2) were read over 56,171 same-cell-type CTCF summit windows and 200,000 random tiles, all by one code path, build established from each file's own maximum coordinate per chromosome as GRCh38.
GSE139196: SUV39H1 (KMT1A) and H3K9me3 ChIP-seq in MCF-7 with and without estradiol - GSM4133857-60, GSM4585206-207, HOMER bedGraph on hg38 · Yang F, et al. · NCBI GEO 2019
What this evidence cannot show
No antibody clone stated for any target; no input or IgG track identified in the sub-series. The four SUV39H1 samples differ in depth (44.3-49.8 million bedGraph intervals) and two of them carry an unexplained 'dsgf' token in their titles. HOMER normalises to 10 million tags, so the minimum value in each file is one tag's worth - 0.39 to 0.55 depending on the file - which is why several intervals read exactly the file minimum.
Evidence that argues against this (1)
- Replicate concordance depends on depth and on the aligner's handling of duplicates ('1 copy/read per genome position' here), so a low correlation between replicates can be a processing artefact rather than a statement about the assay. Without the raw runs this cannot be separated, and they were not retrieved. source
E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
The comparison set is 56,171 primary-assembly 147 bp windows centred on MCF-7's own CTCF point-source summits, from ENCODE's optimal-IDR peak file for the same cell type.
ENCSR000AHD - CTCF ChIP-seq in MCF-7; optimal IDR thresholded peaks ENCFF942TCG (GRCh38), default_analysis preferred_default · Richard Myers, HudsonAlpha · ENCODE Portal 2026
What this evidence cannot show
Same cell line, different lab, different decade from the SUV39H1 samples; MCF-7 sublines drift and CTCF site usage could differ. 32 of 56,203 peaks on non-primary scaffolds were dropped and counted.
Evidence that argues against this (1)
- Using a cell type's own CTCF anchors as the null assumes CTCF anchors are the right comparison class for an H3K9 writer. They are the class this project's earlier positives died against, which is why they are used - but a heterochromatin enzyme's natural comparison class might be H3K9me3 domains instead, and that null was not built. source

Connected to

tests →GSE139196: SUV39H1 ChIP-Seq in MCF-7 with and without 40 minutes of estradiol - a ligand-inducible SUV39H1 occupancy experiment, filed under the retired symbol KMT1A in its own sample titles
answers its estradiol open question: no measurable redistribution, because replicate noise exceeds the condition effect
related to →SUV39H1 (O43463, KMT1A) - the H3K9me3 writer: a menin effector whose binding region OVERLAPS the drug pocket at its N-terminus, a tumour suppressor in MLL-AF9 AML, recruited to double-strand breaks by CTCF, and never localised at the KMT2A locus in any cell type
the only replicate-concordance estimate for any SUV39H1 occupancy data: rho 0.21-0.33
supports →The MCF-7 arm, and the clause that saved it: SUV39H1 at the exon-11 anchor reaches the 96.3rd percentile of 56,171 same-cell-type CTCF anchors in one replicate - but all six non-anchor KMT2A intervals are elevated too (medians 73-87), so the pre-registered FIRES clause does not fire, the verdict is NEITHER, and four replicates of one antibody in one cell line span 51.8 to 96.3 at the same 147 bp
the reproducibility measurement behind that arm's 44.5-percentile spread
related to →GSE210480 carries FOUR CTCF bigWigs under EIGHT sample accessions, byte-identical, under different filenames - so counting GEO sample records doubled the replicate count and would have halved every standard error
the neighbouring hazard: there two accessions were one file; here two distinct files agree at rho 0.21-0.33
related to →ESR1 / estrogen receptor alpha - menin's best-documented non-leukaemic partner, and the only one a menin inhibitor is being given to patients to disrupt on purpose
the estradiol contrast is an ER-signalling perturbation in an ER-positive line
← refutesANSWERED from the paper's own methods, not from the corresponding author: the `dsgf` token that splits GSE139196's four SUV39H1 tracks into the two high and the two low ones is DSG-plus-formaldehyde double crosslinking - so the 44.5-percentile spread Q114 reported is a protocol variable, not replicate noise
the 44.5-percentile spread that entry reports as replicate disagreement is substantially a crosslinker contrast: within protocol the spreads are 5.48 and 17.27 points
← refutesMeasured genome-wide, not just at the locus: the median WITHIN-CONDITION replicate disagreement between two ChIP-seq tracks at a 146 bp window is 23.48 percentile points across 56,171 CTCF anchors and 12 replicate pairs, rising to 57-83 points at the 90th percentile of anchors - and the KMT2A exon-11 element sits at the 35.8th percentile of that distribution, so it is NOT unusually noisy, 146 bp windows are
that 44.5-percentile spread is roughly what a 146 bp window does anywhere: measured genome-wide the median within-condition replicate spread at 146 bp is 23.48 points and the exon-11 element sits at the 35.8th percentile of that noise