The score's amplitude correlates with sequencing depth at Spearman -0.474 (p = 0.012) - a real and significant depth dependence that cleared the pre-registered voiding threshold of 0.6 only because the threshold was set loosely
Samples sequenced more deeply give a smaller spread on this measure, which means part of the spread is measurement noise rather than biology. The pre-registered check for this passed, but only because the bar was set low.
The same thing, for a scientist
**Pre-registered clause V2** voided the magnitude verdict if the amplitude statistic correlated with library depth at |Spearman| > 0.6. **Measured: -0.474, p = 0.012, across 27 samples** [E1]. **It passes, and it should not be reported as a clean pass.** A significant negative correlation of this size means a material part of the between-sample amplitude spread is a property of how deeply each library was sequenced. **The sign is the expected one and that is what makes it credible as noise.** A log ratio of two small counts has larger variance than a log ratio of two large ones, so shallower libraries inflate the SD of the score. Deeper sequencing shrinks it. That is measurement error behaving exactly as it must, and it is the same fact that makes the counting-noise floor non-trivial. **Why this strengthens rather than weakens the magnitude conclusion.** The conclusion is drawn from the **call count** and not from the amplitude, and the call count was explicitly depth-matched by seeded binomial thinning to a common fragment total per arm. Its correlation with depth is **-0.165 (p = 0.411)** before matching and the between-sample spread **widens** after matching [E2]. The depth-sensitive statistic is the one this beat declines to draw a verdict from, and the depth-insensitive one is the one it uses. **The generalisable half.** A pre-registered threshold that a result passes is not the same as a result that is clean, and the difference is visible only if the measured value is printed rather than the verdict. **0.474 with p = 0.012 against a bar of 0.6 is a pass whose number would have failed a stricter pre-registration**, and the honest report is the number, not the word PASS.
The two deepest libraries are the two lowest-amplitude samples
148752 and 152985 are the two most deeply sequenced samples in the cohort and carry the lowest and second-lowest amplitude. They are also the pre- and post-menin-inhibitor aliquots of the patient whose oncoprotein the paper reports as undetectable after treatment. Both explanations are present in the same two samples, so their amplitude rank should not be read as biology, while their call counts - one of which is exactly zero, and depth raises rather than lowers call power - should.
| script | analyses/q144-unbounded-magnitude/run.py |
|---|---|
| status | run |
| pre registered outcome | V2 PASS at |rho| <= 0.6, measured rho = -0.474, p = 0.012 |
| result summary | Amplitude is depth-dependent; the call count is not, and the magnitude verdict rests on the call count. |
- How much of the amplitude spread survives thinning every library to a common depth? The depth-matched amplitude is computed and is smaller, but it was not the pre-registered statistic and is not used for a verdict.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Across 27 samples in GSE252378, the SD of the per-promoter N-over-C score correlates with log10 library depth at Spearman -0.474 (p = 0.012).
GSE252378 - KMT2A oncoproteins induce epigenetic resistance to targeted therapies (171 GSMs, AutoCUT&RUN / AutoCUT&Tag, hg38 fragment BEDs and bigWigs) · Janssens DH, Henikoff S (Fred Hutchinson Cancer Center / Van Andel Institute) · NCBI Gene Expression Omnibus 2024What this evidence cannot show
n = 27 samples, so the correlation's confidence interval is wide; the point estimate is reported with its p-value rather than an interval. Depth is the summed fragment total over all four KMT2A libraries of a sample.
Evidence that argues against this (1)
- The correlation could also be biological: a sample whose oncoprotein occupies few sites may have been sequenced more deeply on purpose, or deeper libraries may come from higher-input samples with better chromatin. Neither is testable from the deposit, and the noise explanation is preferred here only because the sign matches what counting statistics require. Janssens DH et al. 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The call-count statistic used for this beat's magnitude conclusion correlates with depth at Spearman -0.165 (p = 0.411) and its between-sample spread widens rather than narrows after seeded binomial thinning to a common depth.
GSE252378 - KMT2A oncoproteins induce epigenetic resistance to targeted therapies (171 GSMs, AutoCUT&RUN / AutoCUT&Tag, hg38 fragment BEDs and bigWigs) · Janssens DH, Henikoff S (Fred Hutchinson Cancer Center / Van Andel Institute) · NCBI Gene Expression Omnibus 2024What this evidence cannot show
Thinning is a single seeded draw rather than an average over draws, so the depth-matched counts carry their own sampling noise.
Evidence that argues against this (1)
- Thinning equalises fragment number but not signal-to-background, which differs between samples for reasons of cell input and antibody batch, so a depth-matched comparison is not a fully controlled one. Janssens DH et al. 2024
the pre-registered clause this reports in full
shows why the verdict rests on the call count and not on the amplitude
library depths from the deposited fragment BEDs
the depth dependence is why the null must match on coverage and the calls must be depth-matched
an absolute rule's set size tracks depth, which is why depth matching is mandatory with it
depth dependence exists in the deposit but is negligible for this pair