A filter-free promoter ranking with no peak caller and no target list recovers 40 of the top 50 published oncoprotein-target genes in the drug-free infant sample, and 0 of 50 in the drug-exposed one
Before believing anything else, we checked that our way of measuring 'where the fusion protein sits' actually finds the genes the original authors said it sits on. In the sample taken before the drug it found 40 of its top 50 on their list. In the sample taken during the drug it found none - which is itself the result.
The same thing, for a scientist
The primary instrument here counts fragment midpoints from the deposited hg38 BEDs into a fixed 19,404-gene promoter set (RefSeq canonical TSS +/- 1 kb), with no peak caller, no author signal floor and no target list - so nothing about the denominator is chosen. Its positive control is the authors' independently derived 947-symbol oncoprotein-target gene set (`All_OPTG_Genes_set.txt`), of which 687 map into the promoter universe (3.5%). Ranking 148752 promoters on the mean of the two KMT2A N-terminal antibodies gives **40 of the top 50 on that list - 22.6-fold enrichment, hypergeometric p = 2.2e-49**; at k = 200, 83 hits, 11.7-fold; at k = 1000, 186 hits, 5.25-fold. The menin channel of the same sample gives **14 of 50, 7.9-fold, p = 1.2e-09**. The KMT2A **C**-terminal channel gives **1 of 50, 0.56-fold, p = 0.84 - not enriched at all.** In 152985 both the KMT2A-N and the menin channels give **0 of 50**, as does the out-of-sample IgG reference.
The top of the ranking, read by a human
Rule: print the top of any ranking with readable labels and read it. The top 50 KMT2A-N promoters in 148752, in order:
SUPT3H, ANXA2R, PAX5, MLLT10, APOLD1, SENP6, CDK6, GUCY1A1, SKIDA1, MEIS1, MBNL1, HOXA7, ARID2, ZEB2, HOXA9, JMJD1C, MYB, HOXA3, NRIP1, CDKN1B, REEP3, TAPT1, STT3B, CPEB2, G3BP1, LEF1, CHD1, CLEC2B, COMMD3-BMI1, COMMD3, LNPEP, ELF1, SATB1, PHIP, NFKB1, ZNF131, RNF220, SOCS2, CDK19, ZBTB20, HOXA10, ARID4B, PAN3, PPM1B, ZBED5, IKZF2, BANK1, GFI1, MLH1, ARID1B
That is the canonical KMT2A-fusion program - HOXA3/7/9/10, MEIS1, MYB, CDK6, JMJD1C, MBNL1, SKIDA1, ZEB2, GFI1, SATB1 - recovered from raw fragments with no prior. PAX5, LEF1, IKZF2 and BANK1 are the B-lineage half the paper attributes to high oncoprotein levels in B-ALL.
The same ranking in 152985: MTRNR2L8, MTRNR2L2, TSEN34, FARP2, TMEM87B, FAM86B2, MBOAT7, FBRS, G3BP1, ZBTB20, ... - nuclear-mitochondrial pseudogenes and housekeeping loci. It is a background list.
What the C-terminal result means and does not mean
The C-terminal channel not ranking the oncoprotein targets (1/50, p = 0.84) is exactly what the taspase-1 biology predicts and is the strongest single piece of internal validation in this pair - see ana-q143-the-c-terminal-antibody-does-not-rank-the-oncoprotein-targets-and-the-n-terminal-one-does. It does not mean the C-terminal libraries failed: at k = 1000 they are 2.32-fold enriched (p = 9e-13), and the two C reagents agree with each other at Spearman 0.935, the highest agreement of any pair in this deposit.
The honest reading of the 0 of 50 in the drug arm
Two explanations fit and this design cannot separate them. Either the specific signal is gone (the drug worked, or the relapse clone has little oncoprotein - the paper's own reading), or the 152985 libraries are poor. The KMT2A CUT&RUN libraries in 152985 carry 14-16 million placed fragments, comparable to 148752's 15-17.5 million, so it is not a depth failure. But ana-q143-the-152985-cut-and-tag-replicates-agree-at-0-11-to-0-25 shows the CUT&Tag channels of that sample have essentially no internal reproducibility, which a pure-background library would also produce. The 0 of 50 is reported as a joint statement about the sample and the libraries, not as a measurement of the drug.
| instrument | fragment midpoints into RefSeq canonical TSS +/- 1 kb, 19,404 genes, filter-free |
|---|---|
| status | run |
| script | analyses/menin-under-revumenib-in-an-infant-patient-pair/run3_positive_control.py |
| preregistration | analyses/menin-under-revumenib-in-an-infant-patient-pair/PREREGISTRATION.md |
positive control set
{
"file": "All_OPTG_Genes_set.txt",
"symbols": 947,
"mapped_to_universe": 687,
"universe": 19404,
"base_rate_pct": 3.5
}results
{
"148752_KMT2A_N": {
"k50": {
"hits": 40,
"fold": 22.6,
"p": 2.21e-49
},
"k200": {
"hits": 83,
"fold": 11.72,
"p": 3.23e-67
},
"k1000": {
"hits": 186,
"fold": 5.25,
"p": 1.39e-84
}
},
"148752_MENIN": {
"k50": {
"hits": 14,
"fold": 7.91,
"p": 1.23e-9
},
"k1000": {
"hits": 124,
"fold": 3.5,
"p": 1.11e-35
}
},
"148752_KMT2A_C": {
"k50": {
"hits": 1,
"fold": 0.56,
"p": 0.835
},
"k1000": {
"hits": 82,
"fold": 2.32,
"p": 8.98e-13
}
},
"152985_KMT2A_N": {
"k50": {
"hits": 0,
"fold": 0,
"p": 1
}
},
"152985_MENIN": {
"k50": {
"hits": 0,
"fold": 0,
"p": 1
}
},
"IgG_out_of_sample": {
"k50": {
"hits": 0,
"fold": 0,
"p": 1
}
}
}top50 148752 KMT2A N head
[ "SUPT3H", "ANXA2R", "PAX5", "MLLT10", "APOLD1", "SENP6", "CDK6", "GUCY1A1", "SKIDA1", "MEIS1", "MBNL1", "HOXA7", "ARID2", "ZEB2", "HOXA9", "JMJD1C", "MYB", "HOXA3" ]
top50 152985 KMT2A N head
[ "MTRNR2L8", "MTRNR2L2", "TSEN34", "FARP2", "TMEM87B", "FAM86B2", "MBOAT7", "FBRS", "G3BP1", "ZBTB20" ]
- This control licenses a set of occupancy claims in one infant patient. It does not license transferring the MAGNITUDES to KMT2A::AFF1, where this project has already measured that ordering transfers and magnitude does not.
- A ranking that recovers HOXA9/MEIS1 so cleanly will tempt a reader to treat the 152985 zero as a direct measurement of drug effect. It is not - the two samples are different clones.
- Does the same filter-free ranking recover the published target set in the untreated KMT2A::AFF1 samples of this deposit (p279, p454, SEM)? That is the index-fusion version of this control and it was not run here.
- Is the 0-of-50 in 152985 a property of the sample or of the libraries? Nothing in this deposit separates them.
- Why is SUPT3H, not a canonical fusion target, top of the 148752 ranking? It shares a promoter region with the RUNX2 locus in some annotations and may be a mapping artefact of the canonical-TSS choice.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Ranking 148752 promoters on KMT2A N-terminal signal alone, with no peak caller and no target list, puts 40 of the authors' 947 published oncoprotein-target genes in the top 50 (22.6-fold, p = 2.2e-49).
analyses/menin-under-revumenib-in-an-infant-patient-pair - pre-registered re-analysis of GSE252378 samples 148752 and 152985 · Q143-menin-under-drug-in-infants (Oncofusion Knowledge Graph) · this repository 2026What this evidence cannot show
One sample, one ranking. The hypergeometric p-value tests enrichment against a random draw from the promoter universe; it is not a replication statistic and there is no second patient. The authors' gene set was derived from their own data in overlapping samples, so this is an independent DEFINITION rather than an independent replication.
Evidence that argues against this (1)
- The authors' oncoprotein-target set was built across their whole sample panel including this patient, so recovering it from this patient's own data is partly circular. The non-circular part is that my ranking uses no peak call, no floor and no author file, and still lands on the same genes. Janssens DH 2024
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The authors' oncoprotein-target gene set contains 947 symbols, of which 687 map into the 19,404-gene RefSeq promoter universe used here.
MLL_oncoprotein_levels_NatComms - processed data tables and the figure-generation notebook for Janssens et al. 2024 · Janssens DH · GitHub / Zenodo (the paper's Code Availability statement) 2024What this evidence cannot show
260 symbols do not map, mostly non-coding and deprecated symbols; the drop is reported rather than silently absorbed, and the enrichment is computed against the 687 that map so the unmapped ones cannot inflate it.
Evidence that argues against this (1)
- Dropping 260 of 947 symbols is a 27% loss and this project has measured that a symbol-keyed join can drop exactly the genes a question is about; here the dropped symbols are mostly non-coding and deprecated, but that was not verified gene by gene, so the enrichment is computed on a remainder whose composition is only partly characterised. Janssens DH 2024
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The KMT2A C-terminal channel of the same sample shows no enrichment for those genes at k = 50 (1 hit, 0.56-fold, p = 0.84), while agreeing with its own second reagent at Spearman 0.935.
analyses/menin-under-revumenib-in-an-infant-patient-pair - pre-registered re-analysis of GSE252378 samples 148752 and 152985 · Q143-menin-under-drug-in-infants (Oncofusion Knowledge Graph) · this repository 2026What this evidence cannot show
A null result at k = 50 on a channel that IS enriched at k = 1000 (2.32-fold, p = 9e-13); the claim is about the top of the ranking, not about the channel being blank.
Evidence that argues against this (1)
- A C-terminal antibody can be blocked at a busy promoter for reasons other than the epitope being absent - crowding, fixation, or the C180 fragment being present but less accessible - so a low C signal at fusion targets is consistent with, but does not prove, the fusion interpretation. Olsen SN et al. 2022
the control is run on the pair
- related to →HOXA9 — required for survival of established MLL-rearranged leukemia, and dispensable for starting it
recovered in the top 15 of the filter-free fusion ranking
recovered in the top 10 of the filter-free fusion ranking
- related to →KMT2A::MLLT1 (MLL-ENL), t(11;19)(q23;p13.3)
the fusion whose targets are recovered
the two instrument checks that gate everything downstream
the positive control that licenses the ranking
the observation the hypothesis rests on