Oncofusion

start / graph / ana-q153-two-window-deconvolution-bounds-what-is-attributable-at-a-shared-promoter

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q153-can-an-irx2-promoter-measurement-be-attributed

Two-window deconvolution: at a typical shared promoter only 1.8% of the signal can be proved to belong to the gene it is filed under, and at IRX2's one near-positive it is 19.7%

also called shared window deconvolution · attributable fraction promoter

When two overlapping windows both report a number, the difference between them can only come from the small stretches they do not share. That difference sets a floor on how much of a gene's signal is genuinely its own. Across this dataset the floor is usually under 2%, meaning almost all of a shared-window measurement is unattributable arithmetic.

The same thing, for a scientist

The two windows at this locus differ only in two 215 bp flanks, so with no modelling at all: `(mean_IRX2 − mean_IRX2DT) × 2000 = sum(flank_IRX2) − sum(flank_IRX2DT)`, and because the second term is non-negative, **the IRX2-only flank contributes at least `Δ/mean_IRX2` of IRX2's window signal**. This is a hard lower bound on the attributable fraction and it is the only attribution this data structure permits without base-resolution coverage [E1]. Applied to the one GSE252378 sample where the locus carries appreciable signal, **p454 KMT2A-N1: IRX2 3.8119, IRX2-DT 3.0629, Δ 0.7490, so ≥19.7% of IRX2's window signal comes from sequence IRX2-DT's window does not contain** [E1]. Applied as a general statistic over every ≥50%-sharing pair in the table, in every KMT2A-N library where the larger member reaches a share of 2.0, **the median attributable fraction is 0.018 and it is below 0.10 in 82.1% of 25,880 pair × library cases** [E1]. **So in four out of five occupied shared-promoter calls, over 90% of the signal is arithmetically unattributable**, and p454's IRX2 at 0.197 sits at the **91.6th percentile** — this locus is an unusually GOOD case and is still four-fifths unattributable [E1]. The twist worth reading twice: **the 215 bp that is IRX2's alone is `chr5:2,750,677-2,750,892`, which is 785–1,000 bp INTO IRX2's gene body on the minus strand — not its promoter.** The only part of the measurement that can be attributed to IRX2 is not promoter sequence at all [E2].

Go deeper
The arithmetic, in full, because it is the whole method
IRX2 window     chr5:2,750,677 - 2,752,677   (2,000 bp)
IRX2-DT window  chr5:2,750,892 - 2,752,892   (2,000 bp)
shared          chr5:2,750,892 - 2,752,677   (1,785 bp)
flank A, IRX2 only    chr5:2,750,677 - 2,750,892   (215 bp)
flank B, IRX2-DT only chr5:2,752,677 - 2,752,892   (215 bp)

mean_IRX2   * 2000 = sum(shared) + sum(A)
mean_IRX2DT * 2000 = sum(shared) + sum(B)
---------------------------------------------
(mean_IRX2 - mean_IRX2DT) * 2000 = sum(A) - sum(B)

Two equations, three unknowns. The shared term cannot be recovered, which is precisely the attribution failure. What can be recovered is the difference of the two flanks, and since sum(B) >= 0, a positive difference is a lower bound on sum(A).

The quantity reported is max(0, Δ) / mean_IRX2 — the fraction of the window's own mean that is provably not shared. It is conservative in two ways: it ignores any signal in flank B (which would raise sum(A)), and it says nothing about the shared 1,785 bp, which may or may not be IRX2's.

Every GSE252378 sample with any signal at this locus

KMT2A-N1 arm, within-sample shares:

samplefusionIRX2IRX2-DTΔflank-sum difference≥ this fraction is IRX2-only
p454AFF13.81193.0629+0.74901,4980.197
TB051939AFF10.61950.0168+0.60271,2050.973
148752ENL0.26630.2251+0.0413830.155
152985ENL0.98451.0403−0.0557−1110 (IRX2-DT side)
p186ENL0.23280.2433−0.0105−210 (IRX2-DT side)
R3595EPS150.02520.0358−0.0106−210 (IRX2-DT side)
A89691, A93949, R3652, p279—equalequal000

TB051939 is the striking row: 97.3% of a small IRX2 window signal is in the 215 bp flank, i.e. essentially all of it sits in IRX2's first intron and none of it near either TSS. p454, the near-positive that q151 reported honestly, is 19.7%. Neither reaches the occupancy gate, so this changes no call — it changes what a call would have meant.

Three samples put the excess on the IRX2-DT side. With n = 10 informative libraries and no direction pre-specified, the sign is not a finding.

The null, and why 0.018 is the number that travels

Over all 1,563 ≥50%-sharing pairs, in every KMT2A-N library where the larger member reaches a share of 2.0 (n = 25,880 pair × library):

statisticvalue
median attributable fraction0.018
interquartile range0.000 – 0.068
below 0.1082.1%
p454 IRX2's 0.19791.6th percentile

A median of 1.8% is the headline. For a typical occupied shared-promoter window in this deposit, 98% of the signal is in sequence two genes own jointly. That is not a threshold artefact and not a property of any particular locus — it is what a 2 kb mean does to a 200 bp offset.

Facts and figures
scriptanalyses/q153-can-an-irx2-promoter-measurement-be-attributed/q153_deconv.py
statusrun
result summaryExact two-window arithmetic gives a hard lower bound on the attributable fraction. Median 0.018 over 25,880 shared-pair cases; p454 IRX2 0.197 at the 91.6th percentile; the attributable 215 bp is gene body, not promoter.
identity(mean_a - mean_b) * 2000 = sum(flank_a) - sum(flank_b); sum(flank_b) >= 0 so the bound is one-sided and conservative
measurements
min_attributable_fraction0.018min_attributable_fraction: 0.018min_attributable_fraction0.197min_attributable_fraction: 0.197min_attributable_fraction0.973min_attributable_fraction: 0.973
measurements — the numbers behind the chart
[
  {
    "quantity": "min_attributable_fraction",
    "subject": "IRX2's 2 kb window signal in p454 KMT2A-N1, fraction provably from sequence IRX2-DT's window does not contain",
    "value": 0.197,
    "units": "fraction of the window mean",
    "n": 1,
    "universe": "p454 KMT2A-N1 of GSE252378; IRX2 share 3.8119, IRX2-DT share 3.0629; 215 bp unique flank each side of a 1,785 bp shared core",
    "null": 0.018,
    "note": "null is the median over 25,880 pair x library cases at >=50%-sharing pairs; 91.6th percentile, so this locus is an unusually GOOD case and is still four-fifths unattributable"
  },
  {
    "quantity": "min_attributable_fraction",
    "subject": "median across all >=50%-sharing promoter pairs in libraries where the larger window reaches a share of 2.0",
    "value": 0.018,
    "units": "fraction of the larger window's mean",
    "n": 25880,
    "universe": "1,563 pairs x up to 54 KMT2A-N libraries, restricted to cases where max(share) >= 2.0",
    "null": null,
    "note": "IQR 0.000-0.068; below 0.10 in 82.1% of cases. In four of five occupied shared-promoter calls, over 90% of the signal is arithmetically unattributable"
  },
  {
    "quantity": "min_attributable_fraction",
    "subject": "IRX2's window signal in TB051939 KMT2A-N1",
    "value": 0.973,
    "units": "fraction of the window mean",
    "n": 1,
    "universe": "TB051939 KMT2A-N1; IRX2 share 0.6195, IRX2-DT share 0.0168",
    "null": 0.018,
    "note": "almost all of a SMALL signal sits in the 215 bp IRX2-only flank, i.e. in IRX2's first intron, not at either TSS. Well below the 2.0 occupancy gate, so it changes no call"
  }
]
How this could bite us back
  • The bound is one-sided, so it can only ever say a signal IS partly attributable, never that it is not. An intervention programme cannot use it to clear a target - a 0.018 median means 'we could not prove attribution', which is exactly the state in which a wrong gene gets nominated and nothing in the analysis objects.
Still open
  1. Can the shared term be recovered with a third window? A window centred between the two TSSs, or a set of tiled 200 bp windows across chr5:2,750,677-2,752,892, would turn two equations into many and could localise the summit - and it needs the bigWig, which this beat was instructed not to download.
  2. Is the flank excess at TB051939 and p454 a real intronic feature or coverage noise at a window edge? A 215 bp mean from a single library has no error bar here, and none was computed.
  3. Does the median attributable fraction of 0.018 depend on the 2.0 share cutoff used to define 'occupied enough to attribute'? The statistic was computed at one cutoff and its sensitivity was not tested.
Every source behind this entry (2)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Because the IRX2 and IRX2-DT 2 kb windows differ only in two 215 bp flanks, the difference of their means times 2000 equals the difference of the two flank sums, giving a lower bound on the attributable fraction; that bound is 19.7% for IRX2 in p454 KMT2A-N1 and has a median of 0.018 with 82.1% below 0.10 across 25,880 pair-by-library cases over all 1,563 pairs sharing at least half their window.

    q153 two-window deconvolution at the IRX2 / IRX2-DT locus with a shared-pair null, GSE252378 · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026
    What this evidence cannot show

    The identity is exact arithmetic on window means and involves no model, but it rests on the windows being what the table says they are - asserted at load: all 19,404 windows are exactly TSS+/-1000 with end-start = 2000. The bound is one-sided and conservative (it assumes zero signal in the neighbour's unique flank), so the true attributable fraction is at least this large and possibly much larger; it is NOT an estimate of the attributable fraction, only a floor. The shared 1,785 bp is unrecoverable by construction, so no upper bound is available. The null is computed only where the larger window reaches a share of 2.0, so it describes occupied-or-nearly-occupied windows and not the whole matrix. Values are within-sample shares from a deposit with no spike-in. n = 10 informative libraries for the IRX2 locus itself, which is far too few to support any claim about the sign of the excess.

    Evidence that argues against this (1)
    • A low attributable fraction is partly a consequence of the window being wide relative to the offset, and is therefore partly a criticism of the 2 kb convention rather than of the data. The same bigWigs at base resolution would very likely resolve a summit to within 215 bp, because CUT&RUN is a high-resolution assay by design. The deconvolution measures the information a 2 kb mean destroyed, not information the experiment never had. Skene PJ et al. 2017
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    The 215 bp unique to IRX2's window is chr5:2,750,677-2,750,892, which on the minus strand lies 785 to 1,000 bp downstream of IRX2's TSS and therefore inside IRX2's gene body rather than in its promoter; the 215 bp unique to IRX2-DT's window, chr5:2,752,677-2,752,892, is likewise inside IRX2-DT's body.

    UCSC Genome Browser REST API, hg38 refGene track, chr5:2,735,000-2,765,000 · UCSC Genome Browser, api.genome.ucsc.edu · UCSC Genome Browser 2026
    What this evidence cannot show

    Arithmetic on fetched refGene coordinates: IRX2's body is chr5:2,745,844-2,751,677 on the minus strand, so its TSS is the higher coordinate and the window's LOW end runs into the body. The statement 'inside the gene body' is about the refGene transcript span, not about a functional element; the region could still contain a promoter-proximal or intronic regulatory element, and none was looked up.

    Evidence that argues against this (1)
    • That the only attributable sequence is gene-body rather than promoter sequence weakens the attribution in one direction and strengthens it in another: intragenic fusion signal at 800-1000 bp into the first intron is a real and commonly reported pattern for KMT2A fusions, so a reader could argue this flank signal IS the interesting part and the shared TSS region is the distraction. Nothing in this deposit's promoter-scale data can settle which. Iqbal Lab oncofusion knowledge graph et al. 2026
Connected to