start / graph / ana-q144-the-unbounded-statistic-works-and-the-magnitude-test-is-still-void-but-for-the-opposite-reason-to-q142

AnalysisstrongQ144-unbounded-magnitude

The unbounded oncoprotein score spans 1.73-12.32 log2 units within a sample - it is not the compressed statistic q142 was stuck with - and the pre-registered magnitude verdict is STILL void, because the score's amplitude is no larger than the disagreement between two antibodies against the same end of KMT2A

I rebuilt the measurement with a scale that can show big differences, and it does show them. But a control built from the same data shows the scale moves just as much when nothing biological changes, so the headline question still cannot be answered.

The same thing, for a scientist

**The statistic.** `S(p,s) = log2( ((nN+1)/TN) / ((nC+1)/TC) )` on TSS +/- 1 kb promoters, from raw fragment counts in the deposited BEDs. It is unbounded: no denominator contains its numerator [E1]. **U1 and U2, the pre-registered unboundedness checks, both PASS.** Within-sample P99-P1 spread of S runs **1.73 to 12.32 log2 units** (3.31x to 5107.45x) across the 27 samples, against a pre-registered floor of 2 log2 units, with **2 of 27 samples failing**. Between samples the amplitude A (the SD of S) runs **0.347 to 2.259 log2 units, a 6.51x spread** with no algebraic ceiling. **This is the thing q142 could not have: its top-decile share moved 1.11-1.36x and could not move further** [E1]. **V1, the pre-registered voiding clause, FIRES.** A must be at least 1.5x both same-protein floors in a majority of samples; it is so in **0 of 27**. Median A = **1.001**, median A(N1 vs N2) = **0.905**, median A(C1 vs C2) = **1.457**. **The oncoprotein score's dynamic range at promoter resolution is not bigger than the range you get by comparing two antibodies raised against the same end of the same protein in the same nuclei** [E1]. **V2, V3 and V4 all pass**, so the void is specifically about the reagent floor and not about depth, promoter count or IgG: A against log10 depth gives Spearman **-0.474** (p = 0.012); **4646** promoters survive the pre-registered n_N+n_C >= 20 filter in every sample; median Spearman of IgG promoter shares against KMT2A-N shares is **0.252**, above 0.6 in **1 of 25** samples [E1]. **What the numbers would have said if V1 had not fired, reported because suppressing them would be selective.** A(max)/A(min) = **6.51** between samples against a within-sample antibody-pair ratio of **1.32** (median), a ratio of **4.91**. That points at H-MAGNITUDE-DOES-NOT-TRANSFER. **It is not reported as the verdict**, because the pre-registration says a statistic that does not clear its own noise floor cannot deliver one, and this one does not. **The void is not the same void as q142's, and the difference is the finding.** q142's test failed because the statistic could not express a difference. This one expresses differences perfectly well - it is the *control* that expresses them too.

Go deeper

The counting-noise floor, which separates two different objections (POST HOC)

A seeded binomial null - resampling each promoter's N and C counts from its own total under no promoter-specific N/C difference - gives a median amplitude of 0.200 log2 units against the observed median 1.001, a ratio of 5.67 (minimum across samples 1.17). So S is not counting noise. It carries roughly four-fold more structure than Poisson sampling at these depths. The problem V1 identifies is therefore reagent-specific, not statistical: what the score cannot separate itself from is where a different antibody against the same protein binds, not random sampling.

The filter, and what it removed

The pre-registered common-promoter filter (n_N + n_C >= 20 in every sample) keeps 4646 of 19404 promoters and removes 14758 (76.1%). That is a large removal and it is not neutral: it removes the low-coverage promoters, which are exactly where a fusion that is absent would look absent. The sensitivity ladder is {'10': 6291, '20': 4646, '50': 1649} at thresholds 10 / 20 / 50. Requiring every one of 27 samples to clear the threshold is what makes the intersection this small; the filter is an intersection, so it tightens with each sample added.

Signal-to-floor improves with promoter coverage, and part of that rise is free (POST HOC)

Raising the coverage threshold raises the ratio of A to the same-protein floor: thr 10: 6291 promoters, A/A(N1vN2) = 0.90, A/counting = 5.44, between-sample Spearman = 0.197 ; thr 20: 4646 promoters, A/A(N1vN2) = 0.94, A/counting = 5.63, between-sample Spearman = 0.231 ; thr 50: 1649 promoters, A/A(N1vN2) = 0.99, A/counting = 6.05, between-sample Spearman = 0.341 ; thr 100: 261 promoters, A/A(N1vN2) = 0.96, A/counting = 6.83, between-sample Spearman = 0.407. Some of that rise is guaranteed rather than informative: both A and the floor contain counting noise, so removing counting noise from both raises their ratio whenever A has any non-counting component at all. The gradient therefore confirms that a biological component exists; it does not size it.

Facts and figures

script	analyses/q144-unbounded-magnitude/run.py
status	run
pre registered outcome	U1 PASS, U2 PASS, V2/V3/V4 PASS, V1 FIRES -> magnitude verdict on the amplitude statistic is VOID
result summary	Unbounded statistic achieved (1.73-12.32 log2 units within sample); V1 voids the amplitude verdict; directional numbers favour magnitude-does-not-transfer but are not reported as the verdict.
A max over min between samples	6.505054593450988
A ratio within sample between antibody pairs median	1.324736474737719
n samples	27

Still open

Would the amplitude statistic clear V1 on the authors' own called intervals rather than on promoters, and is the interval scale the whole of the difference?
Is A(N1 vs N2) inflated by the mouse/rabbit host difference? A same-host pair of N-terminal antibodies would answer it and none is deposited.

Every source behind this entry (2)

E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
In GSE252378, a log-ratio oncoprotein score computed from deposited fragment counts spans 1.73-12.32 log2 units within a sample and 0.347-2.259 log2 units of amplitude between 27 samples, but fails a pre-registered requirement to exceed 1.5x the amplitude of a same-protein two-antibody comparison in a majority of samples (0 of 27).
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 2024
What this evidence cannot show
TSS +/- 1 kb refGene promoters, not the paper's own peak intervals; the published criterion was never intended for this interval scale. Within-sample shares only - there is NO SPIKE-IN in this deposit, so nothing here is an occupancy-per-cell statement. The 1.5x factor in V1 was chosen before computing and is arbitrary in the sense that any threshold is; the underlying numbers (A vs A_NN vs A_CC) are reported so a reader can apply their own.
Evidence that argues against this (1)
- The same-protein floor may be too harsh a control. N1 is a mouse and N2 a rabbit antibody against the same region, and CUT&RUN efficiency differs between host species, so A(N1 vs N2) contains a species term that A(N vs C) - which pools both - partly averages out. On that reading V1 fires because the control is inflated, not because the signal is weak. The counting-noise floor, which has no reagent term at all, sits four-fold below the signal and supports that reading. Janssens DH et al. 2024
E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The published magnitude claim for this system is that the genomic enrichment of the oncoprotein is highly variable between samples and subject to dynamic regulation.
MLL oncoprotein levels influence leukemia lineage identities · Janssens DH, Duran M, Otto DJ, Wu W, Xu Y, Kirkey D, Mullighan CG, Yi JS, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Communications 15:9341 2024
PMID 39472576 DOI
What this evidence cannot show
Quoted from the abstract, whose measurement is the authors' oncoprotein score on their own called intervals with a significance term, not a promoter-level SD.
Evidence that argues against this (1)
- Because the published claim is made on selected, significance-filtered intervals and this beat's is made on all promoters, the two are not the same measurement and a disagreement between them is not evidence against the paper. Where they can be compared - the between-sample spread of the call count - they agree in direction and this beat's spread is the larger. Janssens DH et al. 2024

Connected to

depends on →GSE252378 fragment BEDs: a complete KMT2A N-terminal / C-terminal antibody set in 27 patient samples across six fusion partners, deposited and until now unopened by this corpus
the fragment BEDs this was computed from
supports →The pre-registered magnitude test failed because the statistic was bounded by construction - top-decile share spreads only 1.1-1.4x between samples and cannot express what the published oncoprotein score says varies
answers that entry's open question and reaches a void for the opposite reason
depends on →A statistic whose denominator contains its numerator cannot test a magnitude claim, and the tell is that its between-sample spread sits near 1 whatever the biology does
the diagnostic this statistic was built to satisfy
related to →KMT2A / MLL1 — the gene that breaks
N-terminal over C-terminal KMT2A signal
← related toScored in the published criterion's own currency, the number of oncoprotein-target promoters runs 0 to 2282 across 27 KMT2A-rearranged patient samples - 20.01x between the 10th and 90th percentiles - and the spread widens to 61.10x after depth matching
the companion amplitude statistic, which is voided
← supportsAt promoter resolution the published N/C oncoprotein criterion calls 1043 promoters in fusion-negative CD34+ HSPCs - above the median across 27 KMT2A-rearranged patients - while the same-reagent controls in that sample call 66 and 11
an independent control pointing the same way as the voiding clause
← supportsIn 18 of 27 patient samples, comparing two antibodies raised against the SAME end of KMT2A calls at least as many oncoprotein-target promoters as the real N-over-C comparison does, and swapping the antibody pair moves a sample's own call count by a median of 2.57x
the criterion-space version of the pre-registered voiding clause
← supportsThe fusion-specific score's per-promoter reliability inside one patient is r = 0.183, and the median correlation between two DIFFERENT patients is 0.196 - so the ordering question cannot be resolved at promoter resolution, and the corpus's 0.82-0.97 figures cannot be carried to this readout
the same reliability problem that fires the pre-registered voiding clause
← depends onThe 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
the pre-registered clause this reports in full