The obvious result was that VTP-50469 cuts H3K4me3 at HOXA9 three-fold in SEM. The pre-registered housekeeping control says GAPDH, RPL13A and TUBB fell just as far, so the obvious result was the normalisation
Our first look said the menin inhibitor slashed the chemical mark at the classic leukaemia genes by about three-fold - exactly the expected answer. But genes that have nothing to do with leukaemia fell by the same amount, which means the drop was an artefact of how the data are scaled. We had written down in advance that this exact pattern would mean 'not biology', so we threw the result away.
The same thing, for a scientist
Ratio of mean CPM at a 4 kb promoter window, VTP-50469 arm over DMSO arm, in the SEM H3K4me3 libraries of GSE159608: **HOXA9 0.368, MEIS1 0.342, CDK6 0.411, JMJD1C 0.298, FLT3 0.345, BCL2 0.373, PBX3 0.467**. Pre-specified housekeeping controls, same measure: **GAPDH 0.461, RPL13A 0.326, TUBB 0.335**. The controls are inside the range of the targets. `PREREGISTRATION.md` P3 states: *'Housekeeping promoters move less than 1.3-fold under either drug. If they move as much as the HOX loci, the effect is normalisation, not biology.'* It fired. The uniform depression is fully accounted for by the VTP libraries holding 2.5-fold less of their reads in H3K4me3 peaks combined with depth normalisation, and no target-specific claim can be made from this measure.
Why this entry exists at all
Because the discarded result is the one that would have been published. It is the expected sign, the expected magnitude, at the canonical loci, in the right cell line, with replicates agreeing - HOXA9 99.4 and 124.5 CPM under VTP against 291.3 and 317.2 under DMSO, both replicates, same direction. Nothing about it looks wrong. The only thing standing between it and the graph was a control written down before the data were touched, and the control is three lines long. 1
The two loci that moved the other way, and what they show
HOXA10 (1.736), ACTB (1.368) and PGK1 (1.668) went up under VTP on the same measure. At 5 kb bins these are short or neighbour-crowded genes whose promoter window overlaps something else, and their behaviour is a further sign that the measure is dominated by window composition rather than by the mark. A result whose sign flips with gene length is a result about the ruler.
What replaced it
The measure that survives is a ratio of two site classes read within one library, so the global scale cancels. It gives 0.766 rather than 0.35, with the housekeeping-driven part removed. See ana-q140-vtp-50469-depletes-h3k4me3-at-menin-bound-sites-by-23-percent-in-sem-and-epz-5676-does-not. The three-fold number is not a smaller version of the 23% number; it is a different quantity, and most of it was never biology.
| script | analyses/sem-arm-identity-from-signal/run.py |
|---|---|
| status | run |
| measure | mean CPM in TSS +/- 2 kb, VTP arm over DMSO arm |
| prereg clause | P3 |
| verdict | FIRED - result discarded as normalisation |
| coordinate source | MyGene.info genomic_pos_hg19, fetched 2026-09-23 |
targets — the numbers behind the chart
{
"HOXA9": 0.368,
"HOXA10": 1.736,
"MEIS1": 0.342,
"PBX3": 0.467,
"CDK6": 0.411,
"JMJD1C": 0.298,
"FLT3": 0.345,
"BCL2": 0.373
}housekeeping controls — the numbers behind the chart
{
"GAPDH": 0.461,
"ACTB": 1.368,
"RPL13A": 0.326,
"TUBB": 0.335,
"PGK1": 1.668
}- A three-fold H3K4me3 drop at HOXA9 is exactly the pharmacodynamic biomarker a trial would reach for. This entry is the reason not to: without a within-sample internal control the same number appears whenever a sample's signal-to-noise drops, which in a clinical specimen it routinely does.
- How many H3K4me3 drug-effect figures in the wider literature are this measure on compositionally normalised data with no housekeeping control shown?
- Would a 1 kb binning rescue the promoter measure, or does the peak-share confound survive any bin size? The confound is multiplicative and global, so it should survive - but that is an argument, not a measurement.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
VTP-50469/DMSO promoter CPM ratios in SEM H3K4me3 are 0.368 at HOXA9 and 0.342 at MEIS1, and 0.461, 0.326 and 0.335 at GAPDH, RPL13A and TUBB respectively.
sem-arm-identity-from-signal: run.py and results.json, locus_ratios (this repository) · Q140-sem-menin-arm · Oncofusion knowledge graph 2026What this evidence cannot show
hg19 promoter windows are TSS +/- 2 kb, with coordinates fetched live from MyGene.info rather than typed, and stored in results.json. 5 kb binning means a 4 kb window spans one or two bins, which is coarse; that coarseness is part of why this measure fails and is the reason the replacement measure uses site classes rather than single windows.
Evidence that argues against this (1)
- A housekeeping promoter is not a perfect null for a menin inhibitor. Menin has fusion-independent functions at many promoters, so a real drug effect at GAPDH is not impossible - but a real effect of the same size at GAPDH, RPL13A and TUBB as at HOXA9 and MEIS1 would mean the drug is not fusion-selective, which contradicts its established selectivity for MLL-rearranged cells. Krivtsov AV et al. 2019
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The depositors report from this same dataset that aberrant H3K4me3 enrichment in gene bodies is sensitive to menin inhibitors.
Automated CUT&Tag profiling of chromatin heterogeneity in mixed-lineage leukemia · Janssens DH, Meers MP, Wu SJ, Babaeva E, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Genetics 53:1586-1596 2021What this evidence cannot show
Read from the abstract. We were unable to test this specific claim: a body:promoter ratio at 5 kb resolution is meaningless for genes shorter than a bin (GAPDH returns exactly 1.000 for that reason), so our attempt is reported as uninformative, not as disagreement.
Evidence that argues against this (1)
- We could not test the depositors' gene-body claim at 5 kb resolution, so its agreement or disagreement with this entry is unknown. It is listed here so a later reader does not mistake our silence for a refutation. Janssens DH et al. 2021
The housekeeping control killed the raw locus measure and forced the within-library contrast A3 reports
- related to →HOXA9 — required for survival of established MLL-rearranged leukemia, and dispensable for starting it
HOXA9 promoter H3K4me3 fell 2.7-fold on the discarded measure and the housekeeping controls fell with it
MEIS1 promoter H3K4me3 fell 2.9-fold on the discarded measure
A uniform peak-share drop under depth normalisation is what produced the spurious three-fold locus effect