The published peak classes do transfer from 20 kb peaks to 2 kb promoters - the GMP-over-B-lineage ratio swings 8.56-fold across the pair that defined them - but only in the between-sample direction the classes were built in
Before using someone else's gene lists you have to show your own version of their measurement still sees what they saw. It does: in the one patient whose before-and-after samples produced the lists, the myeloid-to-lymphoid balance flips nearly ninefold.
The same thing, for a scientist
The classes were constructed as a **between-sample** contrast: intervals with z > +1.96 in a p279-minus-p318 comparison become `ALL_specific`, intervals with z < -1.96 become the AML-high pool, and that pool restricted to non-canonical targets becomes `AML_specific`. Re-measured at 2 kb promoters in the same pair: | class | p279 (infant B-ALL) | p318 (post-switch AML) | p279/p318 | |---|---:|---:|---:| | `ALL_specific` | **2.014** | **0.426** | **4.72** | | `AML_specific` | 2.064 | **3.739** | 0.55 | | `Oncoprotein_Target` | 2.614 | 2.378 | 1.10 | **AML/ALL goes 1.025 -> 8.770, an 8.56-fold swing**, while the canonical class barely moves (1.10). That is the reagent working exactly as advertised: the classes are differential between the two states and the shared class is not. **What does NOT hold is the within-sample ordering**, and my pre-registration wrongly required it: in p279, `ALL_specific` (2.014) is not above `AML_specific` (2.064) - they are within 2.5% of each other. The class construction never claimed otherwise. So the pre-registered gate G1 failed on a test the reagent was never built to pass, and the corrected between-sample gate passes. **Both are reported; the failing pre-registered gate is not deleted.**
What a promoter window loses, counted
From the authors' 2,685-row signal table:
- 1,692 unique intervals - 52
ALL_specific, 45AML_specific, 1,595Oncoprotein_Target. The 52 and 45 reproduce the q143 beat's peak-scale n exactly, which is the cross-check that the interval parse is right. - 284 intervals (16.8%) carry no annotated gene and are lost entirely to a promoter-window reagent.
- 29 genes are annotated to more than one class and are dropped from all three by the pre-registered exclusivity rule: ANTXR2, BANK1, C11orf74, CLEC2A, COL23A1, DSE, EBF1, FOXP1, GALNT2, HNRNPC, IGF2BP2, IGF2BP2-AS1, KCNK12, LOC101928269, MED13L, MSH6, MSI2, NFE2L2, NSD3, PIK3AP1 and nine more. EBF1, FOXP1, BANK1 and PIK3AP1 are B-lineage genes and MSI2 and NSD3 are canonical KMT2A-fusion targets, so the exclusivity rule is not free - it removes some of the most interpretable members.
- 300 genes fail to join the 19,404 promoter symbols: 5, 6 and 289 by class, overwhelmingly
LOC*,-AS1,LINC*andMIR*symbols that a RefSeq canonical-TSS set does not carry.
Final: 37, 40 and 827 promoters, from 66, 51 and 1,145 raw gene names.
The 40 GMP-like promoters, named so the next reader does not re-derive them
ADD1, AGPS, BEST1, BNIP2, BTG1, CDCA7L, CHSY1, CPNE1, CSGALNACT1, DDX3X, DIDO1, ELF2, FNDC3B, FRAT2, FTH1, HDAC4, HIPK1, HNRNPF, IRS2, KLF6, LETM2, LYST, MAPK1, NAA16, NEK9, NR5A2, PIGBOS1, PTPN22, RAB27A, RANBP9, RBM12, SAMHD1, SIMC1, SNRK, SPRY1, SRGN, TLE4, TMED10, TP53BP2, ZFAND5
KLF6 is one of the three loci the paper's own genome-browser panels name in the ENL-persistence figure (MBNL1, RUNX2, KLF6). Of those three, only KLF6 lands in AML_specific; MBNL1 and RUNX2 both land in Oncoprotein_Target - so the figure's named loci straddle two classes, and treating the figure and the GMP-like supplementary table as the same object treats three loci as one class when they are two. SRGN, SAMHD1, FTH1, HDAC4 and TLE4 are myeloid/monocytic in character; DDX3X, HNRNPF, RBM12 and MAPK1 are housekeeping-adjacent, which is part of why the class scores near a coverability-matched null.
| rows in author csv | 2685 |
|---|---|
| unique intervals | 1692 |
| intervals with no gene | 284 |
| ambiguous genes dropped | 29 |
| script | analyses/q148-what-the-hoxa-negative-blast-occupies/genesets148.py |
| status | run |
| result summary | the classes transfer between samples (8.56x swing) and not within one |
intervals per class — the numbers behind the chart
{
"ALL_specific": 52,
"AML_specific": 45,
"Oncoprotein_Target": 1595
}raw genes per class — the numbers behind the chart
{
"ALL_specific": 66,
"AML_specific": 51,
"Oncoprotein_Target": 1145
}genes failing promoter join — the numbers behind the chart
{
"ALL_specific": 5,
"AML_specific": 6,
"Oncoprotein_Target": 289
}final promoters — the numbers behind the chart
{
"ALL_specific": 37,
"AML_specific": 40,
"Oncoprotein_Target": 827
}aml specific members
[ "ADD1", "AGPS", "BEST1", "BNIP2", "BTG1", "CDCA7L", "CHSY1", "CPNE1", "CSGALNACT1", "DDX3X", "DIDO1", "ELF2", "FNDC3B", "FRAT2", "FTH1", "HDAC4", "HIPK1", "HNRNPF", "IRS2", "KLF6", "LETM2", "LYST", "MAPK1", "NAA16", "NEK9", "NR5A2", "PIGBOS1", "PTPN22", "RAB27A", "RANBP9", "RBM12", "SAMHD1", "SIMC1", "SNRK", "SPRY1", "SRGN", "TLE4", "TMED10", "TP53BP2", "ZFAND5" ]
measurements — the numbers behind the chart
[
{
"quantity": "fold_swing_in_AML_over_ALL",
"subject": "the p279 -> p318 pair that defined the classes",
"value": 8.56,
"units": "fold",
"n": 2,
"universe": "40 AML_specific over 37 ALL_specific promoters, KMT2A-N1 and N2 of each sample",
"null": 1,
"note": "1.025 in p279 to 8.770 in p318; the canonical class moves 1.10-fold over the same pair; CIRCULAR positive control"
},
{
"quantity": "intervals_lost_for_having_no_annotated_gene",
"subject": "the authors' 1,692 unique peak intervals",
"value": 284,
"units": "intervals",
"n": 1692,
"universe": "p148752_152985_Menin_Signal.csv unique (chrom,start,stop)",
"null": null,
"note": "16.8%; invisible to a promoter-window reagent"
},
{
"quantity": "genes_dropped_for_multi_class_membership",
"subject": "the three peak classes combined",
"value": 29,
"units": "genes",
"n": 1262,
"universe": "66 + 51 + 1145 raw gene names across the three classes",
"null": null,
"note": "includes EBF1, FOXP1, BANK1, PIK3AP1, MSI2, NSD3 - some of the most interpretable members"
}
]- Reusing a published peak class without reproducing its defining contrast is how a gene list becomes a reagent nobody has calibrated. Passing that test, as here, is necessary and very far from sufficient.
- Do the 284 gene-less intervals carry the GMP-like signal? They are 16.8% of the interval set and a promoter reagent cannot see them; scoring them directly from the bigWigs is the obvious next step and was not done here.
- MBNL1 and RUNX2 sit in Oncoprotein_Target while KLF6 sits in AML_specific. Does the programme the figures illustrate differ from the one the supplementary table defines, and if so which does the published claim rest on?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Re-measured at 2 kb promoters, the AML-over-B-lineage class ratio swings from 1.025 in p279 to 8.770 in p318, a 8.56-fold change, while the canonical class ratio moves only 1.10-fold.
analyses/q148-what-the-hoxa-negative-blast-occupies - results148_controls.json C1_G1_corrected · Q148 (Oncofusion Knowledge Graph) · this repository 2026What this evidence cannot show
p279 and p318 defined the classes, so this is a positive control and not a finding. It shows the promoter proxy can see the contrast; it does not show the classes mean anything in any other sample.
Evidence that argues against this (1)
- A between-sample contrast defined by a z-score threshold will always reproduce in the samples that defined it, at any window size that overlaps the intervals. The transfer test is therefore necessary and very weak, and passing it is not evidence that the classes generalise. Janssens DH 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The interval parse recovers exactly the 52 ALL_specific and 45 AML_specific unique intervals that an independent earlier beat scored at peak scale, from 2,685 CSV rows and 1,692 unique intervals.
analyses/q148-what-the-hoxa-negative-blast-occupies - genesets148.py output; C:/Temp/q148prog/q148_genesets.json · Q148 (Oncofusion Knowledge Graph) · this repository 2026What this evidence cannot show
Agreement on interval counts does not imply agreement on signal, and the two beats use different windows and different normalisations.
Evidence that argues against this (1)
- The earlier beat mapped peaks to 5 kb bins by bin-centre containment; this one maps them to annotated gene promoters. Two parses agreeing on n says the de-duplication is the same, not that the measurements are comparable. Q148 (Oncofusion Knowledge Graph) 2026
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
Of the three loci the deposit's paper names in its ENL-persistence genome-browser panels, only KLF6 falls in the GMP-like AML_specific class; MBNL1 and RUNX2 both fall in the shared Oncoprotein_Target class.
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 three loci are named in the paper for ENL persistence over oncoprotein-target genes, which the abstract says includes GMP-like-programme genes; the paper does NOT label all three as GMP-like and this entry does not claim it does. The class assignments are the authors' own Peak_type column, joined to promoters.
Evidence that argues against this (1)
- Two of the three named loci sit in the shared class rather than the GMP-like one, so the 40-promoter class is not a faithful stand-in for the figure panels the paper's ENL claim is anchored on, and a negative about the class is not a negative about those loci. Janssens DH et al. 2024
the reagent calibration that measurement rests on
- supports →p279 and p318: one infant, two diseases, and an oncoprotein target map that does not carry over
an 8.56x swing in programme balance across the same infant's two samples
the corrected between-sample gate and its numbers