Most gene-body 'AFF1-exclusive' calls in GSE252378 are a fold-change threshold cutting an oncoprotein-level gradient - and IRX1 is the one member where the non-AFF1 samples are genuinely at background
also called the gradient-versus-address model · q154 threshold hypothesis
Because patients' leukaemia cells carry different amounts of the fusion protein, a test that asks whether a gene has twice the average signal will pass in the high patients and just fail in the low ones, which looks like the gene being specific to the high group. Five of the six genes that look AFF1-specific behave exactly like that, and IRX1 does not - in the other eighteen patients its signal is essentially zero rather than just under the bar.
The same thing, for a scientist
**Statement.** Partner-exclusive occupancy calls in GSE252378 arise predominantly because a fixed fold gate (2.0x an average promoter, both KMT2A-N antibodies) is applied to samples whose oncoprotein signal differs in magnitude. Where the non-index samples sit just below the gate, exclusivity is manufactured; where they sit at background, it is an address. **The deposit's own paper is titled *MLL oncoprotein levels influence leukemia lineage identities*, so the gradient is the authors' own headline and not an invented confound.** **Origin: data pattern.** Of the six gene bodies that are occupied in >=3 of 8 KMT2A::AFF1 samples and 0 of 18 non-AFF1, five have a non-AFF1 sample within a factor of 2.2 of the gate (H3C15 within 1.4%, SPAG7 9.7%, ELF1 20%, PAN3 33%, GNAQ 120%) and three of them - ELF1, GNAQ, PAN3 - are called in non-AFF1 samples by the authors' own more sensitive, FDR-based caller. **IRX1's nearest non-AFF1 sample is at 0.252x an average promoter, 7.9-fold below the gate, and the authors' caller agrees it is AFF1-only.** PAN3 is the worked counter-example: it already sits on this project's own list of canonical, partner-general KMT2A-fusion targets. **Prior probability 0.7** that the gradient explains the majority of the class; **0.25** that it also explains IRX1.
Predictions, with the ones already observed marked
- ALREADY OBSERVED. At the class members other than IRX1, the non-AFF1 distribution is a continuum immediately below the gate rather than a mass at zero. Measured: the highest non-AFF1 min-of-replicates shares are 1.973, 1.807, 1.605, 1.337 and 0.911 against a gate of 2.0.
- ALREADY OBSERVED. A more sensitive caller applied to the same libraries calls those genes in non-AFF1 samples. Janssens et al.'s pipeline calls ELF1, GNAQ and PAN3 in KMT2A-ENL and KMT2A-AF9 samples.
- ALREADY PARTLY OBSERVED. Moving the gate turns the class over. Class sizes at gates 1.25 / 1.5 / 1.75 / 2.0 / 2.5 / 3.0 are 8 / 6 / 9 / 6 / 6 / 3, with only GNAQ and PAN3 present throughout. IRX1's exclusivity is stable at every gate; only its recurrence tier moves, because p318's IRX1 is 2.167x and fails a 2.5 gate.
- NOT YET TESTED. If the gradient model is right, a per-sample quantile-normalised score - each gene's rank within its own library rather than its fold over an average promoter - should dissolve the class for the five and keep IRX1.
- NOT YET TESTED. The AFF1 samples should have systematically higher oncoprotein signal at shared targets too, not only at the exclusive ones. The authors' 'average oncoprotein score' per sample is the natural test and is published per sample.
What follows if it is right
Every partner-specificity claim this project has made from a fold gate - at promoter scale and at gene-body scale - is a claim about detection, not about addressing, unless the non-index samples are shown to be at background rather than under the bar. The cheap general fix is a margin column: report, beside every exclusivity call, the highest non-index value and its ratio to the gate. On the six-member class that single column separates IRX1 from the rest at a glance.
It also means the class-level permutation result - 6 observed against 0.86 under an exact institution-stratified null - is true but not about partner specificity: a label permutation cannot distinguish an address from a gradient, because the gradient is a real property of the real labels.
| statement | Gene-body partner-exclusive occupancy calls in GSE252378 are produced mainly by a fixed fold gate cutting a between-sample oncoprotein-level gradient; IRX1 is the one member of the six-gene class whose non-index samples are at background rather than under the bar. |
|---|---|
| origin | data-pattern |
| prior probability | 0.7 |
predictions
[ "ALREADY OBSERVED: the non-AFF1 values at five of six class members lie between 0.911x and 1.973x against a 2.0x gate.", "ALREADY OBSERVED: the deposit paper's own caller calls ELF1, GNAQ and PAN3 in non-AFF1 samples.", "ALREADY PARTLY OBSERVED: class membership turns over across gates 1.25-3.0 (sizes 8, 6, 9, 6, 6, 3); IRX1's exclusivity does not.", "NOT YET TESTED: a within-library rank score rather than a fold score should dissolve the class for the five and keep IRX1.", "NOT YET TESTED: the AFF1 arm should show higher signal at SHARED targets too, measurable against the authors' published per-sample average oncoprotein score." ]
experiments
[ "exp-q154-score-the-six-gene-bodies-in-a-second-institution-untreated-kmt2a-aff1-blast" ]
analyses
[ "finding-q154-irx1-is-the-only-one-of-the-six-aff1-exclusive-gene-bodies-with-a-clean-margin-and-an-independent-confirmation", "finding-q154-the-gene-body-occupancy-rule-passes-0-61-percent-of-gene-bodies-against-10-75-percent-of-promoters" ]
falsifiers
[ "If the non-AFF1 shares at the class members are BIMODAL - a mass near zero plus one or two samples just under the gate - rather than a continuum filling 0.9x to 2.0x, the gradient model is wrong and the calls are addresses with occasional leakage. Scoreable on the 18 non-AFF1 samples x 6 genes already in results_gate_sweep.json; the model predicts no member other than IRX1 and GNAQ has its non-AFF1 maximum below 1.2x.", "If a within-library rank-based score (each gene's percentile inside its own library's gene-body distribution, no fold gate) leaves at least four of the six exclusive at a matched call volume, the gradient model is wrong.", "If the authors' per-sample average oncoprotein score is NOT higher in the 8 AFF1 samples than in the 18 non-AFF1 samples, the gradient does not exist in the direction this hypothesis needs and the model fails at its first premise.", "If IRX1's non-AFF1 maximum rises above 1.0x an average promoter in any re-measurement - a different interval, a different build, a different caller - then IRX1 is not the exception and the hypothesis loses the only case it exempts." ]
- If partner-exclusivity is mostly a level effect, then a therapy targeted at an 'AFF1-specific' gene will engage the same gene in other KMT2A-rearranged leukaemias at lower occupancy - which is a widened indication if the drug works and a widened toxicity if it does not.
- A biomarker built on an exclusivity call with a 1.4% margin, as H3C15 has here, will flip between laboratories on sequencing depth alone.
- Is 'the AFF1 samples have more oncoprotein on chromatin' a technical statement about these libraries or a biological statement about the fusion? The deposit has no spike-in, so the two cannot be separated within it.
- If exclusivity is a threshold effect, is there any fold gate at which a partner-specific address can be demonstrated at all in a deposit with this much between-sample variance - or does the question require a spike-in-normalised experiment?
- The same argument applies to every promoter-scale exclusivity claim in q152 and q151. What is the margin distribution over the 332 AFF1-only promoters, and how many have a non-AFF1 sample above 1.8x?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Five of the six gene-body AFF1-exclusive recurrent genes have a non-AFF1 sample within a factor of 2.2 of the 2.0x gate, three of them are called in non-AFF1 samples by the deposit paper's own pipeline, and class membership turns over across gates of 1.25 to 3.0 - while IRX1's nearest non-AFF1 value is 0.252x and its exclusivity is stable at every gate.
Q154 gene-body class membership, margins and threshold sweep · Iqbal Lab oncofusion knowledge graph, beat Q154 · this project 2026What this evidence cannot show
A margin is not a p-value and a small margin does not prove a gene is not exclusive - it says the call is fragile, and five fragile calls out of six is a pattern rather than a proof. Six members is a small class on which to rest a general model.
Evidence that argues against this (1)
- The alternative reading is that the authors' caller is over-sensitive rather than that this project's gate is over-strict, in which case the three 'refutations' are false positives in their pipeline and the class is real. Nothing here adjudicates: both instruments are thresholds on the same libraries, and neither has an external truth set. Janssens DH et al. 2024
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The deposit's own paper establishes that MLL oncoprotein expression level varies between these patient samples and that genome-wide binding-site selection tracks that level, so a magnitude gradient across samples is the paper's own finding rather than an assumed confound.
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 2024What this evidence cannot show
The paper's gradient claim is about lineage identity and binding-site number; it does not itself say that partner-exclusive calls are artefacts, and this hypothesis extends it further than the authors do.
Evidence that argues against this (1)
- If oncoprotein level drives binding-site selection, then a partner group with higher levels genuinely occupies more sites, and 'exclusivity produced by level' is a biological fact about that partner rather than a measurement artefact. The distinction between 'the artefact explains it' and 'the level IS the partner effect' is not resolved by any measurement in this deposit. NCBI Gene Expression Omnibus 2024
the margin and cross-instrument audit that motivates the model
- related to →IRX1 — Iroquois homeobox 1, the gene that names the HOXA-negative half of infant t(4;11) ALL
IRX1 is the exempted case: its non-index samples are at background, not under the bar
one added AFF1 patient scores all six class members with a mandatory margin column