start / graph / finding-q154-six-gene-bodies-are-aff1-exclusive-and-recurrent-against-a-permutation-null-of-0-86

AnalysisstrongQ154-gene-scale-occupancy-null

At gene-body scale the AFF1-exclusive recurrent class is 6 against an exact institution-stratified null of 0.86 - but the singleton class is 63 against 74, so recurrence carries the signal and exclusivity alone does not

also called q154 T2 · the gene-scale label permutation

Six genes are hit by the fusion protein in three or more of the eight AFF1 leukaemias and in none of the other eighteen patients. Shuffling which patients count as the AFF1 group gives about one such gene, so six is about seven times more than chance. But genes hit in only ONE AFF1 patient and nobody else number sixty-three, and shuffling gives seventy-four - so that weaker pattern is exactly what chance produces.

The same thing, for a scientist

The label permutation the brief asked for, run at gene-body scale on a 19,402 x 26 binary occupancy matrix (27 samples minus the drug-exposed 152985), holding the matrix fixed and re-drawing the 8-sample index arm. **Recurrent class (>=3 of 8 AFF1, 0 of 18 non-AFF1): observed 6.** Null mean **0.88** unstratified (p_upper 0.0026), **0.96** call-rate-matched (0.0036), **0.86** under an **exact** institution-stratified permutation over all 4,410 assignments (**0.0011**). Restricted to the 16 B-lineage samples at k>=2: **20 against 8.18 (p 0.0002)**. **Identical class membership under all three floor definitions** [E1]. **Singleton class (>=1 of 8 AFF1, 0 of 18): observed 63, null 73.9 (p_upper 0.612), call-rate-matched 76.9 (0.658), institution-stratified 49.4 (0.199).** **At or below chance** - the same direction q152 found at promoter scale (332 against 425) [E1][E2]. **Patient level, which is the conservative reading: 12 against 6.35, p_upper 0.116 - not significant.** Collapsing `p279`/`p318` to one patient costs the class half its members and all of its significance [E1]. **So the gene-scale AFF1-exclusive signal is real at sample level and does not survive patient collapse**, which is the pre-registered U5 escape and is why this beat's verdict is NEITHER rather than SURVIVES.

Go deeper

Every null, both tails, and the draw count

class	stratification	observed	null mean	SD	5th	95th	p_upper	p_lower	draws
k>=3 of 8, 0 of 18	S1 unstratified	6	0.88	1.03	0	3	0.0026	0.9996	10,000
k>=3	S2 pass-rate matched +-15%	6	0.96	1.07	0	3	0.0036	0.9980	7,963
k>=3	S3 institution, EXACT	6	0.86	0.93	0	3	0.0011	1.0000	4,410 (all)
k>=2 of 7, 0 of 9	S4 B-lineage only	20	8.18	2.75	4	13	0.0002	1.0000	10,000
k>=2 of 7 patients, 0 of 17	patient level	12	6.35	3.98	2	14	0.1164	0.9108	10,000
k>=1 of 8, 0 of 18	S1 unstratified	63	73.91	26.19	36	120	0.6120	0.4038	10,000
k>=1	S2 pass-rate matched	63	76.94	26.38	39	124	0.6584	0.3552	7,967
k>=1	S3 institution, EXACT	63	49.44	13.56	26	71	0.1986	0.8252	4,410
k>=1 of 7 patients, 0 of 17	patient level	63	70.03	25.08	34	114	0.5612	0.4536	10,000

Monte-Carlo standard errors of the null means are 0.010-0.30, so no verdict here is within 2 MC SE of its null and the pre-registered U6 precision escape does not fire. The AFF1 arm's mean pass rate is 0.642% against the non-AFF1 arm's 0.558% - a 1.15-fold difference, which is why call-rate matching barely moves the null, unlike in the authors' own call table where the same matching raised it five-fold 1.

Why the two classes point in opposite directions, and which one to believe

A singleton AFF1-only gene body needs one AFF1 sample to be above the gate and eighteen comparators below it. With a 0.608% pass rate that is easy, happens 63 times, and happens 74 times by chance - the label carries no information. A recurrent one needs three AFF1 samples to agree, which the 0.608% rate makes very unlikely for an arbitrary gene: the null is 0.88.

So the informative statistic at gene-body scale is recurrence, not exclusivity - the exact opposite of the reading that has been applied to the promoter-scale 332, where the singleton class was the headline. At promoter scale the recurrent class was 17 against 9.1 (p 0.112); at gene-body scale it is 6 against 0.86 (p 0.0011). The gene-body rule is the one that can see recurrence, because its pass rate is 17.7-fold lower.

And the patient collapse is what decides it. Twelve against 6.35 at patient level, p 0.116, means the class cannot be defended as a programme across independent patients. The brief asked whether the pattern is what chance gives; at sample level it is not, at patient level it is not distinguishable, and the difference is p279/p318 being one child.

Leave-one-out and the reference class

Dropping each AFF1 sample and recounting at the same >=3 threshold: TB051939 6, TB920488 5, TB986292 5, p179 5, p247 5, p318 5, p279 4, p454 4 1. No single sample carries the class - the worst case loses two of six. That is the pre-registered C5 control and it passes.

The reference class - gene bodies occupied in >=3 of the 8 AFF1 samples regardless of the non-AFF1 arm - is 148, identical under all three floors. So 6 of 148 recurrent-in-AFF1 gene bodies are also absent from all 18 comparators; under permutation that fraction is 0.86 of a similar-sized class. The pre-registered U2 escape (fewer than 30 in the reference class) does not fire 1.

Facts and figures

script	analyses/q154-gene-scale-occupancy-null/analyze.py
status	run
result summary	Recurrent AFF1-exclusive class 6 vs 0.86 exact institution-stratified null (p 0.0011); singleton class 63 vs 73.9 (p 0.612); patient level 12 vs 6.35 (p 0.116). Stable across three floors. U5 fires: sample and patient level disagree.
reference class size ge3 of 8 aff1	148

leave one out class size

measurements

inputs

[
  "Q154 gene-body occupancy matrix, 19,402 x 133, built from streamed GSE252378 fragment BEDs",
  "analyses/q154-gene-scale-occupancy-null/nulls.py (the four stratifications)"
]

class members

[
  "ELF1",
  "GNAQ",
  "H3C15",
  "IRX1",
  "PAN3",
  "SPAG7"
]

leave one out class size — the numbers behind the chart

{
  "TB051939": 6,
  "TB920488": 5,
  "TB986292": 5,
  "p179": 5,
  "p247": 5,
  "p318": 5,
  "p279": 4,
  "p454": 4
}

measurements — the numbers behind the chart

[
  {
    "quantity": "count_recurrent_partner_exclusive_gene_bodies",
    "subject": "gene bodies occupied in >=3 of the 8 KMT2A::AFF1 samples and 0 of the 18 non-AFF1 samples",
    "value": 6,
    "units": "gene bodies",
    "n": 19402,
    "universe": "19,402 hg38 refGene gene bodies x 26 GSE252378 samples (27 minus the drug-exposed 152985), two-antibody 2.0x-plus-floor rule against each library's own mean promoter density",
    "null": 0.86,
    "note": "null is the mean of the EXACT institution-stratified permutation over all 4,410 assignments; p_upper 0.0011. Unstratified null 0.88 (0.0026), call-rate-matched 0.96 (0.0036)."
  },
  {
    "quantity": "count_singleton_partner_exclusive_gene_bodies",
    "subject": "gene bodies occupied in >=1 of the 8 AFF1 samples and 0 of the 18 non-AFF1 samples",
    "value": 63,
    "units": "gene bodies",
    "n": 19402,
    "universe": "the same matrix",
    "null": 73.91,
    "note": "p_upper 0.612, p_lower 0.404 - at or below chance, reproducing the direction of q152's 332-against-425 at promoter scale"
  },
  {
    "quantity": "count_recurrent_partner_exclusive_gene_bodies_patient_level",
    "subject": "gene bodies occupied in >=2 of the 7 KMT2A::AFF1 patients and 0 of the 17 non-AFF1 patients",
    "value": 12,
    "units": "gene bodies",
    "n": 19402,
    "universe": "the same matrix collapsed to 24 independent patients (p279/p318 and R3595/R3652 are pairs; 152985 dropped so 148752 stands alone)",
    "null": 6.35,
    "note": "p_upper 0.1164 - NOT significant. This is the conservative reading and it is why the beat's verdict is NEITHER."
  },
  {
    "quantity": "count_recurrent_partner_exclusive_gene_bodies_b_lineage_only",
    "subject": "gene bodies occupied in >=2 of the 7 B-lineage AFF1 samples and 0 of the 9 B-lineage non-AFF1 samples",
    "value": 20,
    "units": "gene bodies",
    "n": 19402,
    "universe": "the 16 B-lineage samples of the 26; the 8 Fred Hutch paediatric AMLs, the one T-ALL and p318 are excluded",
    "null": 8.18,
    "note": "p_upper 0.0002 - the class is not a lineage artefact at sample level"
  }
]

How this could bite us back

A six-gene 'AFF1-specific programme' that loses its significance when two samples from one child are counted once is not a basis for a stratified trial. Any partner-specific therapeutic argument from this deposit has seven patients behind it, and two of them are the same child in two lineages.

Still open

The same permutation with a within-family index arm - the 4 KMT2A-AFDN samples, or the 6 KMT2A-ENL - would say whether a recurrent exclusive class of about six is a property of AFF1 or of any partner group of that size in this deposit. It has never been run at any scale and it is the cheapest remaining control.
Why is the recurrent class significant at gene-body scale and not at promoter scale, when the singleton class behaves the same way at both? The pass-rate difference is the obvious answer but it has not been demonstrated - re-thresholding the promoter rule to a 0.608% pass rate would test it directly.
Patient collapse takes the class from 6 (p 0.001) to 12 of a larger class at p 0.116. What would the patient-level p be with one more independent AFF1 patient, and is that the smallest experiment that settles the beat?

Every source behind this entry (2)

E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
On a 19,402 gene-body x 26 sample occupancy matrix, the class occupied in >=3 of 8 AFF1 and 0 of 18 non-AFF1 has 6 members against permutation nulls of 0.88 (unstratified, p_upper 0.0026), 0.96 (call-rate matched, 0.0036) and 0.86 (exact institution-stratified over all 4,410 assignments, 0.0011); the singleton class is 63 against 73.9 (p_upper 0.612); patient-level is 12 against 6.35 (p_upper 0.116); B-lineage-only at k>=2 is 20 against 8.18 (0.0002); leave-one-AFF1-out gives 4 to 6; the reference class of >=3-of-8-AFF1 regardless of comparators is 148.
Q154 genome-wide gene-body occupancy matrix and label permutations · Iqbal Lab oncofusion knowledge graph, beat Q154 · this project 2026
What this evidence cannot show
No analytic independence calculation is used anywhere: at promoter scale an independence bound gave 150.9 against a permutation mean of 9.1, a 16.6-fold over-estimate, because joint absence from 18 comparators is made harder by clustering, not easier. Every null here holds the occupancy matrix fixed and permutes only labels. The institution-stratified permutation is exhaustive over its 4,410 assignments and has no Monte-Carlo error; the others are 10,000 draws with MC SEs of 0.010-0.30.
Evidence that argues against this (1)
- A label permutation cannot separate a partner-specific ADDRESS from a partner-correlated GRADIENT in signal strength. Five of the six class members have a non-AFF1 sample within a factor of 2.2 of the gate and three of them are called in non-AFF1 samples by the deposit paper's own caller, so most of this class is plausibly a threshold effect that the permutation preserves rather than removes. Janssens DH et al. 2024
E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The promoter-scale analogue, computed by q152 on the same 27 samples, gives 332 AFF1-only promoters against a label-permutation null of 425.4 (5th-95th 288-580, p_lower 0.1465) and a recurrent core of 17 against 9.1 (p_upper 0.1119) - the singleton class below chance and the recurrent class not distinguishable from it.
Q152 results - the 332-promoter AFF1-only set, the 17-promoter core and their label-permutation nulls · Iqbal Lab oncofusion knowledge graph, beat Q152 · this project 2026
What this evidence cannot show
q152's comparator arm is all 19 non-AFF1 samples including the drug-exposed 152985; this beat's is 18. q152's own drop-both-treated variant gives 342 AFF1-only and does not change the direction.
Evidence that argues against this (1)
- The two scales use different intervals over the same libraries, so the comparison is not a replication and not a refutation. It is the same data cut two ways, and the honest summary is that the answer depends on the interval, which is itself the finding. Iqbal Lab oncofusion knowledge graph et al. 2026

Connected to

← related toThe two instruments' opposite verdicts decompose: the CALLER explains the 16-fold magnitude gap and the UNIVERSE explains the patient-level survival - and inside the shared universe the only gene both call AFF1-exclusive is IRX1
reproduces the class exactly as a positive control, then varies the universe
← supportsPartner-exclusivity does discriminate: of four comparator partner groups only AFDN fires, it is 4-of-4 AML, and lineage-matching collapses it to exactly chance while AFF1's class survives the same restriction
runs the pseudo-index control that finding asked for; the class discriminates
← refutesThe menin-dependence test I proposed in the cycle-146 log cannot fire by construction - and what the one pair CAN answer is that 85% of gene-body occupancy is menin-dependent while histone clusters preferentially survive
H3C15, a class member, is a histone-cluster gene and histone clusters behave like background
← refutesThe reciprocal-antibody ratio splits the six-gene AFF1-exclusive class in two: H3C15 and SPAG7 sit at or below background, the three most enriched are the three the authors' caller says are not exclusive, and IRX1 is the only member that is both
H3C15 at 1.04x and SPAG7 at 0.49x of background show no fusion-arm enrichment
← refutes42% of every gene-body occupancy call is at a locus where wild-type MLL1 is also bound - but the AFF1-exclusive class is not, HOXA9 and MEIS1 are not, and SPAG7 is
SPAG7 is C-occupied wherever N-occupied and should be removed from the class
depends on →The same occupancy rule passes 0.61% of gene bodies and 10.75% of promoters, so IRX1 at 3 of 27 is 18x its gene-scale expectation where at promoter scale it was below it
the pass rate is the null's denominator
supports →The 332-promoter AFF1-only set is BELOW the count a random eight-sample arm produces (null 425)
reproduces the direction at gene-body scale for the SINGLETON class (63 vs 73.9) while the RECURRENT class goes the other way (6 vs 0.86)
supports →A partner-exclusive promoter set needs a label-permutation null, not an independence one
four stratified label permutations, no analytic independence bound anywhere
related to →KMT2A::AFF1 (MLL-AF4), t(4;11)(q21;q23)
six gene bodies are occupied in >=3 of 8 KMT2A::AFF1 samples and none of 18 comparators, against a null of 0.86
← refutesSix gene bodies are AFF1-exclusive and recurrent at gene scale; the authors' own pipeline confirms one of them, refutes three and has never measured two - and IRX1 is the one it confirms
the class is significant as a count and mostly threshold-driven as a set: 5 of 6 members have a non-index sample within a factor of 2.2 of the gate and 3 are called in non-AFF1 samples by the deposit paper's own caller
← supportsRun on the reciprocal KMT2A-C antibody instead of the fusion-side KMT2A-N, the same rule gives zero AFF1-exclusive recurrent gene bodies against a null of 0.52 - the gene-scale class is fusion-arm specific
the same rule on the reciprocal KMT2A-C antibody gives 0 against a null of 0.52, so the class is fusion-arm specific
← refutesIgG passes this project's occupancy gate at TWICE the rate of the fusion antibody - because the gate is normalised to each library's own promoter mean and IgG has almost none
the pre-registered IgG FALLS condition fired against this class; it is overridden by an 8.45-fold promoter-density difference that makes the IgG unit incomparable
← depends onVerdict: q147's IRX1 headline survives as a measurement and not as an inference - the pre-registered branch is NEITHER, on U5 and on a FALLS condition that fired
the class null is the primary test the branch is scored on
← related toIn the deposit paper's own target-gene table, the AFF1-exclusive class is 64 genes against a permutation null of 2.3, and it survives call-rate matching, institution stratification, B-lineage restriction and patient collapse
the same test on the authors' own 1,091-gene call table gives 64 against 2.3, so the direction agrees across instruments while the magnitude does not