start / graph / concept-q116-replicate-spread-in-a-chip-percentile-is-a-function-of-window-width-and-at-146-bp-it-reaches-93-points

ExplainermoderateQ116-h3k9-axis-at-the-locus

Measured 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

also called the width-spread curve · why a 96th-percentile single track means nothing at 147 bp · report the spread or report the wider window

Two repeats of the same experiment disagree by about a quarter of the whole scale about any 150-base-pair spot in the genome, and by only a few points about a 90,000-base-pair stretch. The spot this project cares about is no worse than average - small spots are simply not measurable this way from one experiment.

The same thing, for a scientist

**The question, and the reason it had to be asked genome-wide.** Q114 measured a 44.5-percentile spread across four SUV39H1 samples at the *KMT2A* exon-11 element and recorded it as this corpus's only replicate estimate of that uncertainty. **That number is uninterpretable on its own**, because nothing said whether a 146 bp window is noisy *at this element* or noisy *everywhere*. This measures it everywhere [E1]. **THE ANSWER: 146 bp windows are that noisy everywhere, and the exon-11 element is unremarkable among them.** For each of 12 within-condition replicate pairs (eight antigens, 28 tracks, one deposit), the absolute difference in percentile rank was computed at **every one of 56,171** same-cell-type CTCF summit windows: | | value | |---|---:| | median of the 12 pairs' own median \|spread\| | **23.48 percentile points** | | range of those per-pair medians | 0.00 - 39.53 | | per-pair **90th percentile** of \|spread\| | **56.9 - 82.7** | | **where the exon-11 spread sits in its own pair's distribution, median across pairs** | **the 35.8th percentile** | **The *KMT2A* exon-11 element is at the 35.8th percentile of the noise, not the 90th.** Two of twelve pairs put it above the 90th (H3K9me3 -E2 at 96.9, TRIM28 +E2 at 97.6) and four put it at exactly 0.0 because both tracks read 0.0000 there [E1]. **And the width curve, measured at the seven stored intervals in the same 12 pairs:** | window width | median \|within-condition difference\| | range | |---:|---:|---| | **146 bp** (exon11) | **14.98** | 0.00 - 93.36 | | 999 bp | 13.56 | 0.00 - 55.18 | | 1,000 bp | **27.71** | 2.58 - 64.85 | | 1,047 bp | **27.56** | 1.77 - 70.78 | | 2,000 bp | **5.73** | 0.00 - 77.17 | | 7,465 bp | 11.86 | 1.77 - 56.47 | | **90,376 bp** (gene body) | **4.56** | **0.08** - 21.14 | **A 90 kb window is reproducible to about 4.6 percentile points; a 1 kb window to about 28.** That is a factor of six, and it is the practical rule this entry exists for [E1]. **The single cleanest illustration, within one antibody's four files.** The 2 kb *KMT2A* promoter window reads **99.70, 99.89, 99.91, 99.68** in the four KAP1 tracks - a spread of **0.23** - while the 146 bp anchor window in **those same four files** reads **90.62, 0.00, 81.85, 89.45**. Same files, same antibody, same day, same null [E1]. **A CORRECTION TO MY OWN FIRST DRAFT OF THIS ENTRY.** It quoted 'per-antigen spread at the 146 bp anchor' as H3K9me3 93.4, KDM4B 92.3, TRIM28 90.6, KDM4C 61.5, EHMT2 60.8, CBX5 47.9. **Those are ACROSS-CONDITION ranges (max minus min over every track of that antigen, mixing -E2 and +E2), not within-condition replicate spreads.** The within-condition pair spreads are much smaller and much more variable: EHMT2 **19.18** (-E2) and **60.79** (+E2); TRIM28 **8.77** and **89.45**; KDM4B **7.48** and **0.00**; KDM4C **29.78** and **10.78**; H3K9me3 **93.36**, **50.03**, **0.00**, **0.00**. Both quantities are real and they answer different questions, and conflating them overstated the replicate noise for four antigens.

Go deeper

The rule

Never report a single-track percentile at a sub-kilobase window without a spread. The measured within-condition replicate uncertainty at 146 bp in this deposit has a median of 23.48 percentile points genome-wide. A lone 146 bp percentile is a draw from a distribution roughly a quarter of the scale wide.
Prefer the widest interval that answers the question. 90 kb reproduces to 4.56 points here; 1 kb to 27.7. If a claim can be made at gene-body scale, make it there.
Do not read the width curve as monotone, and know why. The 146 bp median (14.98) is below the 1 kb medians (27.7, 27.6) because of zero-collapse: eight of eight H3K9me3 tracks read exactly 0.0000 at the exon-11 element, so four pairs contribute a spread of exactly 0.00. Report the range and the zero fraction beside the median.
Distinguish a within-condition replicate spread from an across-condition range. They differ by a factor of two to twelve in this deposit and only the first is replicate noise. This entry got that wrong in its own first draft.
Before quoting a locus's replicate spread as remarkable, compute the same statistic genome-wide. Q114's 44.5 looked alarming and the exon-11 element turns out to sit at the 35.8th percentile of the noise. A spread is only informative against the distribution of spreads.

What this costs the corpus, and which entries to re-read

This project reasons about a 147 bp element constantly. Q35, Q37, Q110, Q112, Q114 and this beat all read it, and all but this beat and Q114's MCF-7 arm read it from one track per protein per cell type.

Every one of those single-track numbers carries a median replicate uncertainty of about 23 percentile points on this evidence, and the corpus states that nowhere. It does not make them wrong - a measurement with a wide error bar is still a measurement - but it makes several of the corpus's cross-cell-type comparisons at this element, which differ by 20-30 percentile points, indistinguishable from within-condition replicate noise.

Two places where this beat's own conclusions depend on it, and both survive:

TRIM28 at the anchor reads 90.62, 0.00, 81.85, 89.45. The NEITHER verdict rests on the six-non-anchor clause, not on the anchor numbers.
The C6 reproduction of Q114 succeeded because the interval means were identical to four decimal places. Means reproduce between runs of the same code on the same file; percentiles at 146 bp do not reproduce between replicate experiments. Those are different claims and keeping them apart is the point.

Facts and figures

prereg	analyses/q116-h3k9-axis-at-the-locus/PREREGISTRATION.md - replicate spread reporting is required by section 9
run date	2026-09-23
run by	Q116-h3k9-axis-at-the-locus
n replicate pairs	12
n antigens	8
n tracks	28
deposit	GSE139196, MCF-7, HOMER makeUCSCfile bedGraph
null for every percentile	56,171 windows of each interval's own width centred on ENCFF942TCG CTCF point-source summits, MCF-7
why the curve is not monotone	zero-collapse. Eight of eight H3K9me3 tracks read exactly 0.0000 at the 146 bp exon-11 element, so four pairs contribute a spread of exactly 0.00 and pull the 146 bp median below the 1 kb medians. The 146 bp figure averages perfect agreement on nothing with total disagreement on something; the RANGE is the honest summary.
absence class	measured-not-analysed
absence class note	Every track was public. What has never been measured is the genome-wide version of this curve - the same replicate pairs at the same widths at random locations - which would separate window width from signal level. That is compute-only and was not run here.

median abs within condition spread by width

within condition pair exon11 spreads

results

[
  "analyses/q116-h3k9-axis-at-the-locus/results_scored.json"
]

median abs within condition spread by width — the numbers behind the chart

{
  "146": 14.98,
  "999": 13.56,
  "1000": 27.71,
  "1047": 27.56,
  "2000": 5.73,
  "7465": 11.86,
  "90376": 4.56
}

range by width

{
  "146": [
    0,
    93.36
  ],
  "999": [
    0,
    55.18
  ],
  "1000": [
    2.58,
    64.85
  ],
  "1047": [
    1.77,
    70.78
  ],
  "2000": [
    0,
    77.17
  ],
  "7465": [
    1.77,
    56.47
  ],
  "90376": [
    0.08,
    21.14
  ]
}

per antigen exon11 spread

{
  "_WHAT_THIS_IS": "ACROSS-CONDITION range (max minus min over every track of that antigen, mixing -E2 and +E2). NOT a within-condition replicate spread - see within_condition_pair_exon11_spreads for those. An earlier draft of this entry labelled these as replicate spreads, which overstated the replicate noise for four antigens; corrected 2026-09-23.",
  "H3K9me3": 93.36,
  "KDM4B": 92.26,
  "TRIM28": 90.62,
  "KDM4C": 61.53,
  "EHMT2": 60.79,
  "CBX5": 47.89
}

the within file contrast

{
  "four KAP1 tracks, 2,000 bp promoter window": [
    99.7,
    99.89,
    99.91,
    99.68
  ],
  "spread": 0.23,
  "the same four files, 146 bp anchor window": [
    90.62,
    0,
    81.85,
    89.45
  ],
  "spread_": 90.62,
  "reading": "the instability is a property of the window, not of the experiment"
}

library depth range bytes

genome wide 146bp arm

{
  "n_pairs": 12,
  "n_anchors": 56171,
  "median_of_per_pair_median_abs_spread": 23.48,
  "per_pair_median_abs_spread_range": [
    0,
    39.53
  ],
  "per_pair_p90_abs_spread_range": [
    56.9,
    82.7
  ],
  "median_position_of_the_exon11_spread_within_its_own_pair_distribution_pct": 35.8,
  "pairs_where_exon11_is_above_the_90th_percentile_of_its_own_noise": [
    "H3K9me3 baseline -E2 (96.9)",
    "TRIM28 panel +E2 (97.6)"
  ],
  "pairs_where_exon11_spread_is_exactly_zero_because_both_tracks_read_zero": 4,
  "verdict": "the KMT2A exon-11 element is NOT unusually noisy; 146 bp windows in this deposit are that noisy everywhere",
  "results_file": "analyses/q116-h3k9-axis-at-the-locus/results_spread_vs_genomewide.json"
}

within condition pair exon11 spreads — the numbers behind the chart

{
  "EHMT2_minusE2": 19.18,
  "EHMT2_plusE2": 60.79,
  "H3K9me3_baseline_minusE2": 93.36,
  "H3K9me3_baseline_plusE2": 50.03,
  "H3K9me3_shCTL_minusE2": 0,
  "H3K9me3_shKDM4B_shKDM4C_minusE2": 0,
  "KDM4B_minusE2": 7.48,
  "KDM4B_plusE2": 0,
  "KDM4C_minusE2": 29.78,
  "KDM4C_plusE2": 10.78,
  "TRIM28_minusE2": 8.77,
  "TRIM28_plusE2": 89.45
}

Still open

PARTLY ANSWERED THE SAME DAY: the genome-wide arm at 146 bp shows the exon-11 spread sits at the 35.8th percentile of its own pair's noise, so the element is not unusually noisy. What remains open is the WIDTH curve genome-wide - the same 12 pairs at each of the seven widths at random locations - which needs the wider anchor vectors cached and was not run.
How much of the spread is library depth? The pairs span a 6-fold range in file size and the shallowest track in the deposit is the one that reads 0.00 at the anchor in three different antigens' tables. A spread-versus-depth regression over the 12 pairs is one line and would say whether the fix is deeper libraries rather than wider windows.
Should this corpus attach an uncertainty band to every stored exon-11 percentile? Six beats have read that element and all but two read it from one track per protein per cell type. If the answer is yes, the band is 40-90 percentile points on current evidence, which would make several cross-cell-type comparisons in the corpus unresolvable.

Every source behind this entry (2)

E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Over 12 within-condition replicate pairs of eight antigens in GSE139196 (MCF-7), the absolute within-condition difference in percentile rank across all 56,171 same-cell-type CTCF summit windows has a per-pair median of 0.00-39.53 percentile points (median of those medians 23.48) and a per-pair 90th percentile of 56.9-82.7. The KMT2A exon-11 spread sits at the 35.8th percentile of its own pair's distribution, median across pairs. At the seven stored intervals the median within-condition difference is 14.98 at 146 bp (range 0.00-93.36), 27.71 at 1,000 bp, 5.73 at 2,000 bp and 4.56 at 90,376 bp (range 0.08-21.14). In the four KAP1 tracks the 2 kb promoter window spans 0.23 points while the 146 bp anchor window in the same files spans 90.62.
analyses/q116-h3k9-axis-at-the-locus/results_scored.json, replicate_spread_by_window_width and per_antigen · Q116-h3k9-axis-at-the-locus · oncofusion-knowledge-graph / NCBI GEO GSE139196 2026
What this evidence cannot show
ONE DEPOSIT, one submitter, one HOMER pipeline, one cell line, and the pairs are of unequal library depth (61,995,812 to 374,334,050 bytes, a 6-fold range) - so part of the spread is depth and not window width. 'Replicate pair' means two tracks the deposit labels as the same antigen and the same condition; the deposit states no clone and no batch information, so a pair could be two antibodies or two batches. The GENOME-WIDE arm is measured only at 146 bp, because only the 146 bp anchor vectors were cached per track - so the width curve is still measured at the seven KMT2A intervals only, and width remains confounded with signal level there (the 2 kb window is an active promoter, the 146 bp window is an H3K9me3-empty CTCF anchor). The genome-wide arm removes the confound for the 146 bp number and not for the curve.
Evidence that argues against this (1)
- THE WIDTH CURVE IS STILL CONFOUNDED WITH SIGNAL LEVEL AND ONLY THE 146 bp NUMBER IS NOT. The genome-wide arm was computed only at 146 bp, because only the 146 bp anchor vectors were cached. So 'reproducibility improves with width' rests on seven intervals at one locus that differ in signal as well as in width - the most reproducible narrow window (2 kb, 5.73) is the one with the most signal, and the least reproducible (~1 kb, 27.7) surrounds an element that is empty for half the panel. A genome-wide version at each width would settle it, needs no new downloads if the wider vectors are cached on a future run, and was not run. source
E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Q114 measured a 44.5-percentile spread across four SUV39H1 samples of this same deposit at this same 146 bp window and recorded it as the only replicate estimate of that uncertainty in this corpus. Measured genome-wide here, the median within-condition replicate spread at 146 bp is 23.48 percentile points and the exon-11 element sits at the 35.8th percentile of that noise distribution - so Q114's number is roughly what a 146 bp window does anywhere, and is not a fact about the element.
ana-q114-the-four-mcf7-suv39h1-tracks-barely-agree-with-each-other-and-estradiol-changes-nothing · Q114-suv39h1-at-the-locus · oncofusion-knowledge-graph 2026
What this evidence cannot show
Q114's 44.5 is a RANGE across four samples spanning two conditions and, as concept-q116-the-dsgf-token... establishes, two crosslinking protocols - so it is not a within-condition replicate spread and is not directly comparable to the 12 within-condition pairs measured here. Within protocol Q114's spreads are 5.48 and 17.27. This entry's per-antigen figures are within-condition where the deposit labels them so.
Evidence that argues against this (1)
- Because Q114's 44.5 mixes conditions and crosslinking protocols, saying that six antigens here 'exceed it' compares a within-condition spread against a mixed one, which is not a like-for-like comparison and makes this beat's numbers look worse than the comparison warrants. The defensible statement is the absolute one: within-condition spreads of 47.9 to 93.4 percentile points at a 146 bp window, on their own terms. source

Connected to

refutes →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
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
related to →The CTCF site in KMT2A exon 11 - a conserved insulator element sitting exactly on the functional cut point of the breakpoint cluster region, whose occupancy decides whether the locus rearranges
a single-track percentile at this 146 bp element carries a median replicate uncertainty of about 23 percentile points, and the element is not unusually noisy - the window size is
related to →NEITHER, and it is the closest call this project has had: KAP1/TRIM28 reads the 90.62nd, 81.85th and 89.45th percentile of MCF-7 CTCF anchors at the exon-11 element in three of four tracks - satisfying the first clause of the pre-registered PRESENT rule - and the six-non-anchor-interval clause blocks it at 65.87
the four KAP1 files are the cleanest illustration: 0.23 percentile points of spread at the 2 kb promoter and 90.62 at the 146 bp anchor, same files same day
related to →A panel of many antigens in one cell type is the input track the deposit did not provide: eight H3K9-axis proteins in MCF-7 all read the whole KMT2A locus elevated (six-non-anchor medians 49.8-72.4) and all read the KMT2A promoter at the 87.6th-99.9th percentile - so the shared component has to be subtracted before any single antigen's number means anything
both are controls that a multi-track deposit makes free: one subtracts the shared component across antigens, the other estimates the noise across replicates
related to →The resolution floor: every deposited H3K9 track in GSE210480 is 1,000 bp binned, so the 147 bp exon-11 question cannot be answered from the processed files - and the planned analysis that asked for 500 bp windows was asking for something that is not there
same family of problem from the other direction: that entry is about a track whose bins are wider than the query, this one about a query narrower than the data can support reproducibly