Cycle 104's open question - 'why do two marks break the floor at D10 and not at D1?' - has a false premise. Against the measured same-condition floor of 2.3% at D10, THREE marks break it there (H3K79me2 16.2%, H3K4me3 12.6%, H3K27me3 6.3%), and the same three break it at D1 too. There is no D1-versus-D10 contrast to explain; there is a continuous bulk gradient that a single borrowed 10% figure had chopped in two
also called the D10 question dissolves · three marks, not two · the bulk gradient is continuous
Last cycle noticed that two chromatin marks looked changed in the emptiest part of the genome but not in the busier part just below the peaks, and asked why. With a correctly measured yardstick the marks are changed in both places, and a third mark joins them - so the puzzle was created by the yardstick, not by the biology.
The same thing, for a scientist
**Cycle 104 closed with a numbered open question:** *"Why do two marks break the floor at D10 and not at D1? H3K4me3 0.8744 and H3K79me2 0.8383 in the sparsest decile. Either a real depletion in the bulk or the noisiest stratum behaving noisily - the input pairs there (1.0256, 0.9931) mildly favour the former."* **The premise is false.** The contrast between D10 and D1 exists only because a single 10% figure, measured at the top 0.1%, was applied at both. The measured same-condition floor is **2.42% at D1** and **2.28% at D10** - essentially the same - and the marks exceed it in both places. | mark, VTP/DMSO (NT) | D1 | vs 2.42% | D10 | vs 2.28% | |---|---:|---|---:|---| | H3K79me2 | 1.0777 (**7.8% up**) | **breaks** | 0.8383 (**16.2% down**) | **breaks** | | H3K4me3 | 1.0982 (**9.8% up**) | **breaks** | 0.8744 (**12.6% down**) | **breaks** | | H3K27me3 | 1.0478 (**4.8% up**) | **breaks** | 0.9375 (**6.3% down**) | **breaks** | | H2AK119ub `PAIRED` | 1.0524 | unscoreable | 0.9846 | unscoreable | | H2AK119ub `REP2` | 0.9784 | unscoreable | 1.0027 | unscoreable | **Cycle 104 counted two marks at D10 because it scored against 10%. Against the measured floor it is three**, and H3K27me3's 6.3% is comfortably outside 2.28%. **What replaces the question is a better one.** All three scoreable marks are UP at D1 and DOWN at D10, monotonically in between on cycle 104's own tables. That is a **continuous redistribution within the bulk, from the sparsest regions toward the regions just below the peaks**, and it is the same shape for three chemically unrelated marks. Shares are compositional so part of any such pattern is arithmetic, but a shared monotone direction across three independent antibodies is not forced by arithmetic. **And the selection artefact cannot explain it**, because cycle 104's strata came from the external parental FLAG universe, and a same-condition input pair on an external cut of that kind is flat to 0.73% with no monotone trend.
Why D10 is nevertheless the weakest stratum to argue from
D10 is the sparsest decile and its per-bin ratios are the widest in the genome: in this deposit the central 90% interval of a single D10 bin's vehicle-vs-vehicle ratio runs roughly 0.38 to 2.79, against 0.78 to 1.17 inside the top stratum. About 1% of D10 bins are zero in one arm and drop out entirely.
The aggregate is still well determined - it pools ~140,000 bins - which is why the stratum ratio's floor is 2.3% while a single bin's is a factor of seven. But any claim about which regions in D10 changed is unsupportable at 2 kb, and the honest statement is about the decile as a whole.
The one number that does not survive its own decile
Cycle 104 reported the input pairs at D10 as 1.0256 and 0.9931 and read them as mildly favouring a real depletion. Against the same-condition floor measured here, 1.0256 is itself outside 2.28% - marginally, by 0.28 points.
So GSE260742's own inputs are, at D10, at or just past the same-condition floor. That does not overturn the mark results, which are 3 to 7 times larger, but it means D10 is the one stratum where the borrowed input control is not comfortably clean, and any future D10 claim should say so.
| script | analyses/the-same-condition-noise-floor/run.py |
|---|---|
| status | run |
| result summary | Three marks, not two, break the same-condition floor at D10, and the same three break it at D1. The D1-versus-D10 contrast cycle 104 asked about is an artefact of applying one TOP-stratum figure at both. |
| d10 floor measured | 0.0228 |
| d1 floor measured | 0.0242 |
marks breaking at d10
[ "H3K27me3", "H3K4me3", "H3K79me2" ]
marks breaking at d1
[ "H3K27me3", "H3K4me3", "H3K79me2" ]
d10 per bin central90 MEN1
[ 0.3805, 2.7902 ]
top per bin central90 MEN1
[ 0.7796, 1.167 ]
- If the bulk mark changes are a compositional shadow of a total-abundance change rather than a redistribution, then the peak-stratum mark results this thread has built on are also partly compositional, and the drug's apparent gain of H3K4me3 at the peaks could be a loss everywhere else. That reading is not excluded by anything on disk and it inverts the mechanism.
- A mark that moves 12-16% in the sparsest decile of the genome under a menin inhibitor is a whole-genome chromatin perturbation, not a targeted one. In a normal haematopoietic cell with the same complexes that is a toxicity surface, and no normal-cell chromatin arm exists in either deposit.
- All three scoreable marks are up at D1 and down at D10 with a monotone run between. Is that one process or three? A mark-by-mark correlation of the per-bin changes across the bulk would separate a shared compositional shift from three independent ones, and every bin is already on disk.
- Within-sample share cannot see a uniform global change in mark abundance. If a menin inhibitor reduced total H3K4me3 everywhere, the share analysis would show exactly this shape - redistribution toward the peaks - with no real redistribution at all. A spike-in-normalised menin-inhibitor histone ChIP would settle it and none is known in this thread.
- Nobody has measured any mark in a deposit with a same-condition histone replicate. Until one exists, every mark floor in this project is imported from a factor ChIP.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The measured same-condition antibody-inclusive share-ratio floor in GSE299100 is 2.42% at D1 and 2.28% at D10, and cycle 104's control-genotype VTP/DMSO ratios at D10 are 0.9375 (H3K27me3), 0.8744 (H3K4me3) and 0.8383 (H3K79me2), all of which exceed it, as do their D1 values of 1.0478, 1.0982 and 1.0777.
What this evidence cannot show
The floor is from a human MV4;11 menin-complex factor ChIP pair; the marks are mouse mm10 histone ChIPs in an MLL-AF9 line. n = 1 per mark per treatment. MV4;11 is KMT2A::AFF1 but presents as acute MONOCYTIC leukaemia with FLT3-ITD, not the infant B-ALL this project studies.
Evidence that argues against this (1)
- Cycle 104's own D10 input controls read 1.0256 and 0.9931, and the first of those is itself marginally outside the 2.28% floor measured here. If the true D10 floor for a histone-mark library in that deposit is nearer 2.6% than 2.3%, H3K27me3's 6.3% still breaks it and nothing in this entry's verdict changes - but the claim that D10 is a clean stratum does. source 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
In GSE299100 the per-bin vehicle-versus-vehicle share ratio at D10 has a 5th-95th percentile interval of 0.3805-2.7902 (MEN1) and 0.3934-2.7692 (MLL1) with about 1% of bins zero in one arm, against 0.7796-1.1670 and 0.7534-1.1275 inside the top 0.1% stratum.
What this evidence cannot show
These are per-bin distributions at 2 kb in one deposit. They bound what a single bin can say; they do not bound the aggregate, which pools ~140,000 bins and is far better determined.
Evidence that argues against this (1)
- A wide per-bin distribution with a well-determined mean is exactly what Poisson counting at low depth produces, and would be narrower at a coarser bin size. If a future beat repeats this at 10 kb the D10 per-bin interval will shrink and the stratum ratio will barely move - so the per-bin figure should not be read as a limit on the biology, only on bin-level claims. source 2026