Oncofusion

start / graph / ana-q144-two-different-patients-agree-about-the-promoter-ordering-slightly-better-than-two-antibody-pairs-in-one-patient-do

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q144-unbounded-magnitude

The 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

Measured twice in the same patient's cells with two sets of antibodies, this readout barely agrees with itself. Two different patients agree about as well, which means the method cannot tell whether different patients really rank their genes the same way.

The same thing, for a scientist

**The corpus's standing claim** is that *per-gene ordering transfers and magnitude does not*, supported by pairwise Spearman figures in this same deposit of **AFF1 0.820-0.893, KMT2A-N 0.834-0.920, H3K4me3 0.932-0.970** against an IgG floor of 0.121-0.175 [E2]. **Every one of those is a single-channel measurement** - total AFF1, total KMT2A-N (fusion plus wild-type), or a histone mark. **None is fusion-specific.** The N-over-C score is, and it is a difference of two correlated channels, so the shared component cancels and what is left is the part the two channels do not share. **The ceiling, measured inside each patient.** `S_A` is the score from the N1/C1 antibody pair and `S_B` from N2/C2, in the **same nuclei, from the same chromatin**. Their correlation is the reliability of the score, and it is **median 0.183** (range -0.056 to 0.372 across 27 samples) [E1]. Under the classical attenuation relation, no between-sample correlation of this quantity can exceed roughly that value. **The between-patient correlation sits AT the ceiling, slightly above it.** Median Pearson between two different patients is **0.196** and median Spearman **0.231**, over **351** pairs; **199 of 351 pairs** exceed the within-patient median [E1]. **Two different children's leukaemias agree about the promoter ordering of the fusion-specific score as well as two antibody pairs in one child's nuclei do.** **So the pre-registered H-NEITHER branch fires for the ordering question.** The data cannot distinguish *ordering transfers* from *ordering does not transfer*, because the instrument's own repeatability is at the same level as the between-patient agreement. Reporting the 0.23 as evidence that ordering fails would be a claim about the instrument. **What this damages, said plainly.** The corpus's *ordering transfers* half is measured on channels that are not fusion-specific, and there is currently **no measurement of ordering transfer on a fusion-specific readout at promoter resolution that is above its own noise floor**. The figures 0.82-0.97 should not be quoted for the fusion's ordering. **What it supports:** ordering agreement does rise monotonically with promoter coverage - threshold 10 -> 6291 promoters, median rho 0.197; threshold 20 -> 4646 promoters, median rho 0.231; threshold 50 -> 1649 promoters, median rho 0.341; threshold 100 -> 261 promoters, median rho 0.407 - which is what a real correlation attenuated by noise does, and not what an absent one does.

Go deeper
Why the reliability is so low, and why that is not a flaw in this beat

At the median promoter in the common set the N and C counts are tens of fragments, so a log ratio of two such counts is a noisy quantity, and the two antibody pairs use disjoint halves of the data. The same arithmetic applies to any promoter-resolution N/C score computed from this deposit, including one computed by anyone else - which is the point. The published analysis avoids it by scoring selected, high-coverage intervals with a significance term, not by scoring all promoters.

Why the magnitude answer survives this and the ordering answer does not

Reliability bounds a correlation, not a count. The number of promoters clearing a fixed significance threshold is reproducible in rank between the two antibody pairs (Spearman 0.397, Pearson 0.679 on log(1+n)) and its between-patient spread is one to two orders of magnitude, far above the 2.57x the reagent swap produces. So the magnitude statement stands on a quantity whose noise floor it clears, and the ordering statement does not.

Facts and figures
scriptanalyses/q144-unbounded-magnitude/run.py
statusrun
pre registered outcomeH-NEITHER for the ordering question: the instrument cannot separate the hypotheses
result summaryWithin-patient reliability r = 0.183; between-patient Pearson 0.196, Spearman 0.231; between exceeds within in 199 of 351 pairs.
reliability ceiling
{
  "within_sample_r_SA_vs_SB_median": 0.18275879633835793,
  "within_sample_r_min": -0.05647157298845456,
  "within_sample_r_max": 0.3717407885038158,
  "between_sample_pearson_median": 0.19602816912587795,
  "between_sample_spearman_median": 0.23128981407491167,
  "between_exceeds_within": true,
  "n_pairs_between_above_within_median": 199,
  "n_pairs": 351,
  "note": "S_A = log2 ratio from the N1/C1 antibody pair, S_B from N2/C2, in the SAME nuclei.  Their correlation is the ceiling on any between-sample correlation of S under the classical attenuation formula.  If the between-sample correlation is at or above it, two patients agree about the promoter ordering as well as two antibody pairs in one patient do, and the instrument cannot resolve a between-patient difference."
}
pairwise
{
  "n_pairs": 351,
  "spearman": {
    "median": 0.23128981407491167,
    "min": -0.02437674012308966,
    "max": 0.551861331234824,
    "frac_gt_0p6": 0
  },
  "ols_slope": {
    "median": 0.2580360865342012,
    "p10": 0.04743384890090302,
    "p90": 0.5730255779450627,
    "frac_in_0p8_1p25": 0.05413105413105413
  },
  "ols_slope_disattenuated": {
    "median": null,
    "frac_in_0p8_1p25": 0.06837606837606838,
    "note": "median is null: the disattenuation divides by a per-sample reliability that is near zero or negative in several samples, so the corrected slope is undefined for those pairs. Reported as null rather than imputed."
  },
  "sma_slope": {
    "median": 1.0887281193024254,
    "median_abs_log2": 0.6236131169411205
  }
}
Still open
  1. What is the reliability and the between-sample correlation within the eight KMT2A::AFF1 samples alone? That is one filter away and is the fair comparison to the corpus's cell-line figure.
  2. Does the N channel alone reproduce the corpus's 0.834-0.920 on these fragment counts, which would confirm that the collapse is specific to the ratio and not to the fragment-BED pipeline?
Every source behind this entry (2)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    In GSE252378, the per-promoter N-over-C score computed from the N1/C1 antibody pair correlates with the same score from the N2/C2 pair in the same nuclei at a median r of 0.183 (range -0.056 to 0.372 over 27 samples), while the median correlation between two different patients is 0.196 (Pearson) and 0.231 (Spearman) over 351 pairs, with 199 pairs above the within-patient median.

    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

    The two antibody pairs are not parallel measures in the strict psychometric sense - they differ by reagent and for the N arm by host species - so their correlation UNDERSTATES reliability, and the true ceiling is somewhat higher than 0.183. That makes the conclusion conservative in one direction and not in the other: the between-patient correlation may sit below a true ceiling rather than at it. Promoters, within-sample shares, no spike-in, 4646 common promoters after removing 14758 of 19404.

    Evidence that argues against this (1)
    • The cohort spans six fusion partners and three lineages, so a low between-patient correlation is partly biological heterogeneity rather than instrument noise, and a within-partner subset (the eight KMT2A::AFF1 samples) would be the fair comparison to the corpus's cell-line result. That subset was not scored separately here, which is the largest single gap in this entry. Janssens DH et al. 2024
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    A prior beat in this corpus measured pairwise Spearman in this same deposit at 0.820-0.893 for AFF1, 0.834-0.920 for KMT2A-N and 0.932-0.970 for H3K4me3, against an IgG floor of 0.121-0.175.

    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

    Those are this corpus's own computations on the deposited bigWigs, re-read from the q142 entries rather than recomputed here, and they are single-channel promoter signals rather than ratios.

    Evidence that argues against this (1)
    • Because those are single-channel measurements and this is a ratio, the two are not in direct competition and neither refutes the other. What is challenged is not the measurement but the generalisation drawn from it - that the fusion's target ordering transfers between patients. Janssens DH et al. 2024
Connected to