Oncofusion

start / graph / ana-q150-the-quantity-two-cycles-called-library-coverage-is-enrichment-and-correlates-with-true-sequencing-depth-at-0-29

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q150-the-authors-intervals-and-the-depth-matched-discriminator

The 'raw library promoter mean' that two cycles treated as library coverage tracks the in-promoter fragment FRACTION at Spearman +0.993 and true sequencing depth at only +0.292, and the oncoprotein share correlates with true depth at -0.255

also called the +0.659 is a correlation with FRiP, not with depth

A recent conclusion said one patient sample looked low on fusion protein only because it had been sequenced shallowly. It was not sequenced shallowly - it sits in the middle of the pack for read depth. The number that was called 'coverage' is really a measure of how concentrated a sample's signal is at genes, which is a different thing.

The same thing, for a scientist

`ana-whether-a73051-is-globally-low-oncoprotein-cannot-be-answered-in-this-deposit` rests on a single identification: that the column mean of `C:\Temp\q147part\promoter_matrix.tsv` is **library coverage**, and that the `Oncoprotein_Target` share correlating with it at Spearman **+0.659** is therefore a depth confound. **I reproduce the +0.659 exactly and the identification is wrong.** GSE252378's deposited fragment BEDs give the true per-library depth directly (`#total_fragments`, counted in cycle 134 for all 133 libraries and on disk at `C:\Temp\q144mag\counts`). Against that: | pair, 133 libraries | Spearman | p | |---|---:|---:| | promoter-matrix column mean vs **in-promoter fragment FRACTION** | **+0.993** | 1.5e-123 | | promoter-matrix column mean vs **true total fragments** | **+0.292** | 6.5e-04 | | true total fragments vs in-promoter fraction | +0.252 | 3.4e-03 | The column mean **is** the fraction of enrichment, to three decimal places. It is not depth. And on the 54 KMT2A-N libraries the confound reverses sign once depth is measured properly: | | Spearman | p | |---|---:|---:| | `Oncoprotein_Target` share vs column mean | **+0.659** | 6.2e-08 | | `Oncoprotein_Target` share vs **true total fragments** | **-0.255** | 0.062 | **A73051's two KMT2A-N libraries hold 9,279,052 and 9,101,995 fragments - ranks 21 and 20 of 54, the 37th and 39th percentiles, against a cohort median of 10,300,517.** They are 7th and 10th lowest only on the enrichment axis. **q148's description of them as 'mid-range' was correct on depth**, and the cycle-135 correction that called them 'in the bottom fifth' is correct only about enrichment.

Go deeper
What survives of the unresolvability argument, because something does

The chain scarce protein -> fewer reads -> flattened share is refuted: sequencing depth is set at the machine, not by the protein, and the data say so.

The chain that is real is scarce protein -> lower signal-to-background -> share diluted toward 1.0. A within-sample share divides by the mean over all 19,404 promoters, and uniform background enters numerator and denominator alike, so as FRiP falls every share converges on 1.0 by construction. That is a technical mechanism and it is what the +0.659 shows.

But it is a different mechanism with a different remedy, and the difference decides what to do next. Depth is fixable by down-sampling and is not the problem. Background dilution is not fixable by down-sampling and is arithmetically correctable - see ana-q150-background-subtraction-makes-the-oncoprotein-share-coverage-invariant-in-forty-of-fifty-four-libraries-and-a73051-is-not-one-of-them. 23

Why this was easy to get wrong, and it is not a careless error

The promoter-matrix column mean is a perfectly reasonable thing to call 'coverage' - it is a mean coverage, in bigWig units, over promoters. The word is right and the referent is not. Nothing in the file says which of the two quantities it is, and the only way to tell them apart is to hold a true depth beside it, which required a different file set that the brief did not name.

The generalisable form: a quantity named after its units is not identified by its units. mean coverage over promoters is depth x enrichment x window, and which of the three factors drives its variance is an empirical question with a one-line answer. 2

The reagent check nobody had run

The bigWig promoter matrix and my own fragment counts are independent extractions of the same libraries. The Oncoprotein_Target share computed from each correlates at Spearman +0.994 (p = 3.4e-51, 54 libraries), and the fragment-count version reproduces the confound at +0.656 against the column mean and -0.262 against true depth. So none of this is a bigWig artefact and q148's reagent was sound. 2

Facts and figures
scriptanalyses/q150-the-authors-intervals-and-the-depth-matched-discriminator/m2a.py
statusrun
depositGSE252378
n libraries133
n kmt2a n libraries54
result summaryThe promoter-matrix column mean is the in-promoter fragment fraction (rho +0.993), not sequencing depth (rho +0.292). The Oncoprotein_Target share correlates with true depth at -0.255. A73051 is at the 37th-39th percentile of depth, not the 13th-19th.
correctsana-whether-a73051-is-globally-low-oncoprotein-cannot-be-answered-in-this-deposit
measurements
spearman_rho-0.255spearman_rho: -0.255spearman_rho0.292spearman_rho: 0.292spearman_rho0.659spearman_rho: 0.659spearman_rho0.993spearman_rho: 0.993spearman_rho0.994spearman_rho: 0.994
measurements — the numbers behind the chart
[
  {
    "quantity": "spearman_rho",
    "subject": "promoter-matrix column mean vs in-promoter fragment fraction",
    "value": 0.993,
    "units": "Spearman rho",
    "n": 133,
    "universe": "all 133 GSE252378 KMT2A-fusion-patient libraries",
    "null": 0,
    "note": "p = 1.47e-123; the column mean IS the enrichment fraction"
  },
  {
    "quantity": "spearman_rho",
    "subject": "promoter-matrix column mean vs TRUE total fragments",
    "value": 0.292,
    "units": "Spearman rho",
    "n": 133,
    "universe": "all 133 GSE252378 KMT2A-fusion-patient libraries",
    "null": 0,
    "note": "p = 6.5e-04; the pre-registered reading required |rho| < 0.30 and it is 0.292"
  },
  {
    "quantity": "spearman_rho",
    "subject": "Oncoprotein_Target within-sample share vs TRUE total fragments",
    "value": -0.255,
    "units": "Spearman rho",
    "n": 54,
    "universe": "the 54 KMT2A-N libraries, 827-promoter Oncoprotein_Target class",
    "null": 0,
    "note": "p = 0.062; the brief predicted a strong positive"
  },
  {
    "quantity": "spearman_rho",
    "subject": "Oncoprotein_Target share vs promoter-matrix column mean",
    "value": 0.659,
    "units": "Spearman rho",
    "n": 54,
    "universe": "the 54 KMT2A-N libraries",
    "null": 0,
    "note": "reproduces the cycle-135 figure to three decimals; the number was right"
  },
  {
    "quantity": "true_total_fragments",
    "subject": "A73051 KMT2A-N1 and KMT2A-N2",
    "value": [
      9279052,
      9101995
    ],
    "units": "fragments",
    "n": 2,
    "universe": "ranks 21 and 20 of 54 KMT2A-N libraries; cohort median 10,300,517",
    "null": null,
    "note": "the 37th and 39th percentiles, not the 13th and 19th"
  },
  {
    "quantity": "spearman_rho",
    "subject": "Oncoprotein_Target share from bigWig means vs from fragment counts",
    "value": 0.994,
    "units": "Spearman rho",
    "n": 54,
    "universe": "the 54 KMT2A-N libraries",
    "null": 0,
    "note": "the two independent reagents agree; this is not a bigWig artefact"
  }
]
How this could bite us back
  • If a patient were to be selected, or deselected, for a menin inhibitor on the grounds that their blast is 'low oncoprotein', the measurement behind that call is a within-sample share whose denominator includes the background. Two samples with identical fusion biology and different antibody performance will get different calls, and nothing in the assay flags it.
  • The remedy is not deeper sequencing. A clinic that responded to this finding by sequencing more deeply would spend money and change nothing, because depth is not what varies here.
Still open
  1. Does the same misidentification affect any other entry in this corpus that reasons from a 'coverage' or 'depth' figure derived from a promoter mean rather than from a fragment total? The promoter matrix is used by q145, q147 and q148.
  2. The in-promoter fraction spans 1.09% to 19.97% across 133 libraries, an 18-fold range. Is that range biological (target abundance) or technical (cell input, antibody lot, batch)? The deposit records cell numbers for some samples in the sample titles and not others.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    GSE252378 deposits AutoCUT&RUN and CUT&Tag for KMT2A-rearranged patient samples with no exogenous spike-in in any of its 171 GSM records.

    What this evidence cannot show

    Read from the SOFT records, which describe what was deposited, not what was done at the bench. A spike-in used and not deposited would be invisible here.

    Evidence that argues against this (1)
    • Spike-in normalisation is the established way to make a global-abundance claim in ChIP-seq and CUT&RUN, and its absence is exactly why a within-sample share has to carry the argument in this deposit. The same review warns that a single scalar normaliser is fragile, which cuts both ways: even a deposited spike-in would not have made this automatic. Patel LA et al. 2024
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    In GSE252378's 133 libraries the promoter-matrix column mean correlates with the in-promoter fragment fraction at Spearman +0.993 and with true total fragments at only +0.292, and over the 54 KMT2A-N libraries the Oncoprotein_Target share correlates with the column mean at +0.659 and with true depth at -0.255.

    What this evidence cannot show

    The fragment counts were produced in cycle 134 by a streaming midpoint counter over the deposited fragment BEDs and were not re-derived from raw reads here; I asserted that each file's per-promoter counts sum exactly to its own #in_promoter_fragments header, which they do for all 133. The two Spearmans against depth (-0.255, -0.262) have p = 0.062 and 0.056 and should be read as 'not positive', not as 'negative'.

    Evidence that argues against this (1)
    • A defender of the depth reading could argue that #total_fragments is itself downstream of the library's complexity, so a sample with little chromatin recovered yields both fewer fragments and lower FRiP. That is true in principle, and it is why I report rho(depth, FRiP) = +0.252 rather than suppressing it. But +0.252 is far too weak to make depth the driver of a +0.993 relationship, and the sign of share-vs-depth is negative. Q150 (Oncofusion Knowledge Graph) 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    A73051's two KMT2A-N libraries hold 9,279,052 and 9,101,995 fragments, ranks 21 and 20 of the 54 KMT2A-N libraries, against a median of 10,300,517 and a range of 413,733 to 17,835,153.

    What this evidence cannot show

    Rank of 54 libraries, not of 27 samples; A73051 contributes two of the 54. The percentile figures (37th, 39th) are rank/54 and are not interpolated.

    Evidence that argues against this (1)
    • A73051's in-promoter FRACTION is genuinely low - 2.56% and 3.04% against a cohort median of 6.29% - so 'A73051 is not an outlier' would be false. It is an outlier on enrichment and not on depth, and that is the whole distinction this entry is making. Q150 (Oncofusion Knowledge Graph) 2026
Connected to