Oncofusion

start / graph / ana-q130-no-matched-null-draw-of-twenty-reached-the-observed-span-and-the-residual-keeps-the-sign

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q130-does-the-eviction-gradient-exist-in-h3k79me2

The H3K79me2 eviction gradient against a DMSO-decile-matched null drawn from outside the fusion's sites at the same quartile sizes: the null is flat at 1.030-1.057, none of 20 draws reached the observed span of 0.6005, and the residual retains Spearman -1.000 - but the restricted-population trap is real and the baseline IS graded 2.2-fold

also called Q130's matched null · the regression-to-the-mean control · restricting a population widens its null

The four groups of DNA stretches being compared are not alike to begin with - the group that loses the most fusion also starts with the most of the tag. To check that the pattern is not just that difference reappearing, the same comparison was run on matched stretches elsewhere in the genome, twenty times. None of the twenty produced anything like the observed pattern.

The same thing, for a scientist

**The trap this control exists for.** The four eviction quartiles differ systematically in their own H3K79me2 baseline: the DMSO share runs **0.013029 / 0.008957 / 0.006825 / 0.005918** across Q1..Q4, a **2.20-fold decline** [E1]. Bins selected for being highest in one sample regress towards the mean in any second sample, drug or not - so a ratio that is lowest where the baseline is highest is exactly what regression produces for free. This is the same shape as the failure recorded last cycle, where restricting a correlation from 1,208 cell lines to 227 raised the observed mean 2.9-fold and the null rose in lockstep. **The null, and it is Q125's construction with Q125's seed so the two beats are comparable.** For each contrast, bins from OUTSIDE `S_FLAG` matched to each quartile on the decile of *that track's own DMSO share*, drawn at the **same quartile sizes**, scored with the **same statistic**, 20 draws, seed `20260923` [E2]. | | Q1 | Q2 | Q3 | Q4 | Spearman | span | |---|---:|---:|---:|---:|---:|---:| | observed, VTP, control genotype | 0.5860 | 0.8188 | 0.9280 | 1.1865 | -1.000 | -0.6005 | | matched-null median | 1.0301 | 1.0398 | 1.0566 | 1.0528 | +0.400 | -0.0200 | | **residual** | **-0.4441** | **-0.2210** | **-0.1286** | **+0.1337** | **-1.000** | -0.5778 | **0 of 20 null draws reached the observed span** (empirical two-sided p = 1/21 = **0.0476**, the smallest value 20 draws can produce), the largest null span in absolute value was 0.1620, and the residual keeps Spearman -1.000 [E2]. **One addition to Q125's construction, and it is the reason this entry exists separately.** Q125's deviation 3 reported the null's MEDIAN ratio per quartile and one Spearman of those medians. That is a null for the *inputs* of the statistic. Here each of the 20 draws is carried through as a complete four-quartile vector, giving **20 null Spearmans and 20 null spans** - a null OF THE STATISTIC. It matters: **3 of the 20 null draws produced |Spearman| = 1.000**, against a chance expectation of 1.67 in 20. A perfect monotone in a null draw is a common event, and that is why the SPAN and not the Spearman is the scored quantity [E2][E3].

Go deeper
What this null can and cannot rule out

Can rule out: that the gradient is regression to the mean from the quartiles' unequal baselines, because the matched bins have the same baseline structure by construction and return a flat profile. Can rule out: that it is a coverage, blacklist or depth property of the two libraries, because the null draws come from the same two libraries.

Cannot rule out: an artefact specific to the S_FLAG bins themselves that is absent from matched bins elsewhere - a mappability or copy-number feature of the fusion's strongest sites. Something of that family is demonstrably present: all three deposited inputs carry their own monotone gradient across the four quartiles at Spearman -1.000, 1.13 to 1.17-fold. That cancels in a within-genotype drug/vehicle ratio, and it is still a reason to distrust any single-track comparison across these quartiles.

Cannot rule out a class-level antibody effect. The panel entry handles that and does not close it either.

The p-value's floor, and why it is quoted as a floor

With 20 draws and the conventional (1+k)/(1+n) estimator, 0.0476 is the smallest p this design can return. So the number does not mean 'p = 0.048'; it means 'no draw came close, and 20 draws is as far as this can be pushed without re-running'. The informative statement is the ratio: the observed span is 3.71 times the largest of 20 null spans and 30 times the null median span. Both are in the results and the conservative one is the one the pre-registered criterion was scored on - see the deviation entry on that criterion, which did not say which summary it meant.

Facts and figures
scriptanalyses/q130-does-the-eviction-gradient-exist-in-h3k79me2/run.py
statusrun
result summaryMatched null flat at 1.0301-1.0566; 0 of 20 draws reached the observed span 0.6005; residual Spearman -1.000; 3 of 20 null draws gave |Spearman| = 1.
seed20260923
n draws20
n outside S FLAG1226480
absence classmeasured-not-analysed
absence class noteThe matched-null machinery was written by Q125 in cycle 88 and committed. Applying it to H3K79me2 required changing four track names.
absence class assigned2026-09-23 (Q130)
epigenetic readoutH3K79me2 eviction-depth gradient against a baseline-matched null
inputs
[
  "GSM8123438/440 H3K79me2 NT DMSO and VTP",
  "the 1,226,480 bins of U_m outside S_FLAG"
]
matched null median
[
  1.0301,
  1.0398,
  1.0566,
  1.0528
]
residual
[
  -0.4441,
  -0.221,
  -0.1286,
  0.1337
]
null spans all draws
[
  0.0626,
  -0.0849,
  0.0662,
  -0.0139,
  -0.0811,
  0.0256,
  -0.0274,
  -0.1145,
  0.0063,
  -0.0422,
  -0.0261,
  0.0508,
  0.0246,
  -0.1002,
  -0.0092,
  0.0793,
  0.0451,
  -0.162,
  -0.0604,
  -0.0674
]
null spearmans all draws
[
  0.4,
  -0.8,
  1,
  -0.6,
  -0.6,
  0.8,
  -0.4,
  -0.8,
  0.6,
  -0.6,
  -0.6,
  0.4,
  0,
  -1,
  0,
  0.8,
  0.4,
  -1,
  -0.8,
  -0.4
]
n4 exact null
{
  "n_orderings": 24,
  "attainable_rho": [
    -1,
    -0.8,
    -0.6,
    -0.4,
    -0.2,
    0,
    0.2,
    0.4,
    0.6,
    0.8,
    1
  ],
  "counts": {
    "-1.0": 1,
    "-0.8": 3,
    "-0.6": 1,
    "-0.4": 4,
    "-0.2": 2,
    "0.0": 2,
    "0.2": 2,
    "0.4": 4,
    "0.6": 1,
    "0.8": 3,
    "1.0": 1
  },
  "P_rho_eq_plus_1": 0.0417,
  "P_abs_rho_eq_1": 0.0833,
  "P_at_least_one_abs_1_in_4_contrasts": 0.2939,
  "expected_n_perfect_monotones_in_4_contrasts": 0.3333,
  "note": "stated in PREREGISTRATION section 4 BEFORE any H3K79me2 quartile number was computed. A +-20%% band around 0.80 contains exactly one attainable value, 0.80 itself."
}
dmso baseline by quartile
[
  0.013029,
  0.008957,
  0.006825,
  0.005918
]
How this could bite us back
  • If the null had to be built from the same libraries, then any systematic drug-independent difference between the two libraries - depth, duplication rate, fragment length - is present in the null and in the observation alike, and is therefore invisible rather than controlled. That is the honest limit of every within-series control in this beat.
Still open
  1. Would 2,000 draws change the reading? The empirical p is floored at 0.0476 by the draw count and nothing else; the compute is minutes and was not spent because the ratio to the largest null draw is the more informative number.
  2. Is there an S_FLAG-specific artefact the matched null structurally cannot see? The inputs' own 1.15-fold monotone gradient shows the bins differ; whether anything larger is hiding there would need a DMSO-versus-DMSO contrast at S_FLAG, which GSE260742 does not deposit for H3K79me2.
Every source behind this entry (3)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    The four eviction quartiles are not baseline-equal: H3K79me2's own DMSO within-sample share runs 0.013029 / 0.008957 / 0.006825 / 0.005918 across Q1..Q4, a 2.20-fold decline, so the most-evicted quartile is also the quartile with the highest baseline coverage of the mark being scored.

    What this evidence cannot show

    NO SPIKE-IN NORMALISATION IS POSSIBLE. Q117 verified zero hits for the entire spike-in vocabulary across all 419,824 characters of GSE260742's 158 sample records and no scale factor is deposited, so every cross-condition number here is a WITHIN-SAMPLE SHARE. It cancels the unknown per-sample scale factor exactly and IS BLIND TO A UNIFORM CHANGE BY CONSTRUCTION: nothing here measures H3K79me2's abundance, everything measures its distribution. EXPOSURE ASYMMETRY: DMSO 72 h and SGC0946 5 uM 72 h are matched, VTP-50469 500 nM is 48 h, so every VTP ratio is a LOWER BOUND on the 72-hour effect - identical in both genotypes, so the genotype DIFFERENCE is unaffected. REPLICATION, CORRECTED MID-BEAT: the NTvKO arm deposits ONE H3K79me2 ChIP per genotype per drug, so ITS spans carry no technical error bar and its bin-level bootstrap intervals measure sampling within one library, NOT library-to-library variation. A REPLICATE PAIR DOES EXIST for a different culture - GSM8123372/374 and GSM8123386/388, both titled MLLAF9 Control, the second carrying `rep2` - already binned by Q122 and scored in Q130 DEVIATION 4: the SIGN and the monotonicity reproduce (Spearman -1.000 in all four murine VTP contrasts in this deposit) while the SPAN reproduces only to within a factor of 1.99 (-0.2438 against -0.4847). EVERY SPAN IN THIS BEAT SHOULD BE READ WITH THAT 2-FOLD REPRODUCIBILITY. THE KNOCKOUT is one CRISPR guide, one clone C1, no second clone, no rescue, and NO deposited PCGF1 protein or transcript measurement. S_FLAG is defined by a DYKDDDDK-tag ChIP on a RETROVIRALLY OVER-EXPRESSED tagged MLLAF9 construct, so it is the over-expressed molecule's sites and not an endogenous fusion's. THE H3K79me2 ANTIBODY IS NOT RESOLVED - `antibod` occurs zero times in the 158 GEO records - and is recorded unresolved rather than guessed. SYSTEM: murine MLL-AF9 (KMT2A::MLLT3) myeloid blood cells, NOT the infant B-lineage KMT2A::AFF1 disease this graph is about.

    Evidence that argues against this (2)
    • A published methods study establishes that STANDARD ChIP-seq normalisation FAILS TO DETECT a genome-wide decrease in a histone mark under an inhibitor of its writer: standard methods see no H3K27me3 reduction under an EZH2 inhibitor, and a substantial reduction appears only with a Drosophila chromatin spike-in. Every number in this entry is exactly the kind of normalisation that paper shows to be blind. Egan B et al. 2016
    • The system is murine MLL-AF9 (KMT2A::MLLT3) myeloid cells carrying a retrovirally over-expressed FLAG-tagged construct. This graph exists for human infant B-lineage KMT2A::AFF1, and this corpus's commonest documented synthesis failure is generalising a murine MLL-AF9 myeloid result to that disease. Neville D et al. 2026
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Against a DMSO-decile-matched null drawn from outside S_FLAG at the same quartile sizes with the same statistic (20 draws, seed 20260923), the matched-null median profile is flat at 1.0301 / 1.0398 / 1.0566 / 1.0528, none of the 20 draws reached the observed span of 0.6005 (largest null span 0.1620, empirical two-sided p = 0.0476), and the residual observed-minus-null is -0.4441 / -0.2210 / -0.1286 / +0.1337 with Spearman -1.000.

    What this evidence cannot show

    20 draws floors the empirical p at 0.0476. The draws are not independent of one another (they are drawn from overlapping pools) nor of the observed data (same libraries), which is deliberate - the point is shared artefacts - but it means the p is a permutation-like diagnostic and not a frequentist test with a defensible type-I rate. 3 of 20 draws gave |Spearman| = 1.000, so the Spearman is not the scored statistic.

    Evidence that argues against this (1)
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    A Spearman over four ordered quartiles has 24 possible orderings and 11 attainable values; P(rho = +1.000) = 1/24 = 0.0417 and P(|rho| = 1.000) = 2/24 = 0.0833 under a uniform random ordering, so with 20 null draws the expected number of perfect monotones is 1.67 - and 3 were observed, which is why a perfect Spearman cannot carry the result on its own.

    What this evidence cannot show

    The exact enumeration is of the Spearman statistic under a uniform random ordering. The four quartile ratios are not exchangeable in reality - they share libraries, bins and a monotone baseline - so 1/24 is the null for 'a random ordering', not for 'these four numbers under no drug effect'. The matched null is the empirical answer to the latter and the enumeration is the arithmetic bound.

    Evidence that argues against this (1)
    • Quoting the permutation null makes the single Spearman look weak, which understates the evidence: six independent VTP contrasts in this beat (two genotypes plus four timecourse timepoints) all returned -1.000. Iqbal Lab oncofusion knowledge graph 2026
Connected to