In human KMT2A::AFF1 MV4;11 (monocytic, not B-lineage), the vehicle-versus-vehicle noise floor is 9.4% at the top 0.1% of menin/MLL1 sites and 0.1-2.6% in every other stratum - so cycle 102's borrowed 4-10% band is about right AT THE PEAKS and four to five times too generous everywhere else, and it was applied everywhere
also called the same-condition floor · the floor is stratum-dependent · GSE299100 shLuc-1 versus shLuc-2 floor
To tell whether a drug really changed something in a cell, you need to know how much two identical, untreated samples differ just by chance - and for four recent results this project had to guess that number from the wrong kind of comparison. Two genuinely identical control samples have now been measured three times over, and the answer is that the guess was roughly right at the handful of strongest sites and about five times too forgiving everywhere else.
The same thing, for a scientist
**In human MV4;11 cells carrying KMT2A::AFF1 - the index fusion, but presenting as acute MONOCYTIC leukaemia with FLT3-ITD, not the infant B-ALL this project studies - GSE299100 supplies the vehicle-versus-vehicle pair three times over.** Two independent non-targeting hairpins (`shLuc #1`, `shLuc #2`), both DMSO 72 h, both sequenced 230927, uniformly processed to `_hg38_rmdup.tdf`, for MEN1, MLL1 and input. 1.87 GB downloaded under bin-and-delete; nothing left on disk but the binned numbers. **2 kb hg38 bins, ENCODE hg38 blacklist v2 removed (98,227 bins), within-sample share (no spike-in exists), strata cut on an EXTERNAL reference containing no track of the pair being scored** - byte-for-byte the aggregation and the design cycles 102/103/104 used. | stratum | input | MEN1 | MLL1 | **antibody-inclusive floor** | input-only | |---|---:|---:|---:|---:|---:| | **TOP 0.1%** (1,402 bins) | 1.0057 | **0.9063** | **0.9117** | **9.4%** | **0.6%** | | D1 (140,085) | 0.9933 | 0.9758 | 0.9918 | 2.4% | 0.7% | | D2 | 0.9947 | 1.0181 | 0.9877 | 1.8% | 0.5% | | D3 | 0.9985 | 1.0230 | 0.9992 | 2.3% | 0.2% | | D4 | 0.9978 | 1.0241 | 1.0077 | 2.4% | 0.2% | | D5 | 0.9999 | 1.0200 | 1.0150 | 2.0% | 0.01% | | D6 | 1.0009 | 1.0181 | 1.0175 | 1.8% | 0.1% | | D7 | 1.0032 | 1.0177 | 1.0214 | 2.1% | 0.3% | | D8 | 1.0048 | 1.0155 | 1.0233 | 2.3% | 0.5% | | D9 | 1.0038 | 1.0127 | 1.0255 | 2.6% | 0.4% | | D10 | 1.0073 | 1.0145 | 1.0228 | 2.3% | 0.7% | **The pre-registered decision rules, scored.** `F1` (the borrowed floor is too generous at the peaks) **does not fire** - 9.4% sits just inside 10%. `F2` (the borrowed floor survives) **fires**: no assay exceeds 10% at any stratum. `F3` (the floor does not transfer) **does not fire** - the input arm replicates. `P1`, my own prediction that the antibody-inclusive floor would be wider than the input-only floor, **is confirmed and by a wide margin at the peaks: 9.4% against 0.6%, sixteen-fold.** **So the headline is neither "wider" nor "the same". It is that the floor is STRONGLY STRATUM-DEPENDENT and cycle 102 measured it in the one stratum where the borrowed number happens to be right.** 4-10% is a good description of the top 0.1%. It is a four- to five-fold overstatement of every decile below, and cycles 103 and 104 applied it at D1 and D10. **The per-bin distribution says something separate and is the reason `F_STRAT` is the primary statistic.** Genome-wide at 2 kb, a single bin's vehicle-vs-vehicle ratio has a central 90% interval of roughly **0.45 to 2.40** in all three assays (log2 SD 0.74-0.77). **At 2 kb resolution one bin's ratio carries almost no information**; what is reproducible is the pooled share over tens of thousands of bins. Inside the top 0.1%, where the ChIP actually has counts, per-bin reproducibility is far better - MEN1 median 0.9264 with a 5th-95th of 0.7796-1.1670, MLL1 median 0.9148 with 0.7534-1.1275 - while the *input* over those same peak bins is the widest of the three (median 1.0250, 5th-95th 0.5411-2.1508) because it has little signal there.
Why the two arms are the right comparison, and what they are not
shLuc #1 and shLuc #2 are two different non-targeting shRNA constructs, independently transduced, selected and grown, then treated with DMSO for 72 h and processed on the same day through the same pipeline 1. That makes this a same-condition BIOLOGICAL plus TECHNICAL floor, not a technical floor, and the distinction was declared before the run. It is the right quantity: a vehicle-versus-vehicle contrast is exactly the null for a drug-versus-vehicle contrast, and it is if anything conservative, because two hairpins carry a little more variation than one sample split in two.
What it is not is a replicate of the same library. Nothing here bounds sequencing-run or batch variation beyond what one day's processing contains.
The stratification was changed after pre-registration, and the reason is itself a result
The pre-registration made a symmetric internal cut primary, reasoning that strata cut on the mean of the two arms "cannot bias the ratio in either direction". That reasoning is wrong when the two arms differ in depth, and the input pair differs 1.74x (9.52e9 against 5.47e9 coverage bases). Selection on the mean then preferentially picks bins where the noisier arm fluctuated up, pushing the ratio below 1 at the top and above 1 at the bottom.
It duly did: the self-stratified input pair produced a monotone decile gradient from 0.9040 to 1.2349 out of two biologically identical samples. That is recorded as DEVIATION 1 in results.json and has its own entry 2. The external cut is primary here, and it is also, independently, the design cycles 102/103/104 used, so it is the comparable one. Both cuts are in results.json in full.
What this does and does not license
Licensed: using ~9% as the same-condition floor for a share ratio taken over a top-0.1% peak stratum, and ~2-3% for a share ratio taken over a decile of the bulk, in a 2 kb within-sample-share design.
Not licensed: quoting a single global floor. Not licensed: using this floor for a per-bin quantity - at 2 kb one bin's ratio spans 0.45-2.40 at the 5th-95th. Not licensed without the sqrt(2) correction: applying it to a ratio-of-ratios, which combines four tracks rather than two 3.
And it bounds THIS deposit. GSE260742 is mouse mm10, MLL-AF9 and PCGF1-knockout lines, a different lab. That the input arms agree (F3 does not fire) is the evidence that the transfer is reasonable, not a proof of it.
The four floor-dependent verdicts, re-tested at the stratum each was taken at
| claim | cycle | stratum | effect | new floor | verdict |
|---|---|---|---|---|---|
| VTP ratio-of-ratios inside the floor (1.0134) | 102 | TOP | 1.34% | 13.5% | SURVIVES, and more comfortably |
| SGC ratio-of-ratios outside the floor (0.7724) | 102 | TOP | 22.76% | 13.5% | SURVIVES, margin 9.3 pts |
| PCGF1 loss lowers baseline fusion occupancy (0.8374) | 102 | TOP | 16.26% | 9.4% | SURVIVES, margin 6.9 pts |
M1 - no mark moves at D1 | 104 | D1 | up to 9.82% | 2.4% | DOES NOT SURVIVE 4 |
| two marks break the floor at D10 | 104 | D10 | 6.3-16.2% | 2.3% | SURVIVES, and a third mark joins them |
| the drug acts only at the peaks | 103 | TOP + D1 | 24.5% / 0.63% | 9.4% / 2.4% | both halves SURVIVE |
Three of cycle 102's verdicts stand and cycle 104's M1 does not, and the reason is the same fact in both directions: the floor is wide at the peaks, where cycle 102 worked, and narrow in the bulk, where cycle 104 applied the peak figure.
| script | analyses/the-same-condition-noise-floor/run.py |
|---|---|
| status | run |
| result summary | Antibody-inclusive same-condition floor 9.4% at the top 0.1% and 1.8-2.6% across deciles D1-D10; input-only floor 0.6-0.7% throughout. Borrowed 4-10% is right at the peaks and 4-5x too generous in the bulk. |
| bin size bp | 2000 |
| genome | hg38 |
| cell line | MV4;11 (CVCL_0064) |
| bytes downloaded | 1865230942 |
| peak disk bytes | 344523744 |
inputs
[ "GSE299100 GSM9032977 input shLuc-1", "GSE299100 GSM9032983 input shLuc-2", "GSE299100 GSM9032979 MEN1 shLuc-1", "GSE299100 GSM9032985 MEN1 shLuc-2", "GSE299100 GSM9032981 MLL1 shLuc-1", "GSE299100 GSM9032987 MLL1 shLuc-2", "ENCODE hg38 blacklist v2" ]
floor by stratum antibody inclusive — the numbers behind the chart
{
"TOP": 0.0937,
"D1": 0.0242,
"D2": 0.0181,
"D3": 0.023,
"D4": 0.0241,
"D5": 0.02,
"D6": 0.0181,
"D7": 0.0214,
"D8": 0.0233,
"D9": 0.0255,
"D10": 0.0228
}floor by stratum input only — the numbers behind the chart
{
"TOP": 0.0057,
"D1": 0.0067,
"D2": 0.0053,
"D3": 0.0015,
"D4": 0.0022,
"D5": 0.0001,
"D6": 0.0009,
"D7": 0.0032,
"D8": 0.0048,
"D9": 0.0038,
"D10": 0.0073
}per bin genome wide central90
{
"input": [
0.4742,
2.4387
],
"MEN1": [
0.4465,
2.3962
],
"MLL1": [
0.4436,
2.3973
]
}decision rules
{
"F1_borrowed_floor_too_generous_at_the_peaks": false,
"F2_borrowed_floor_survives_everywhere": true,
"F3_floor_does_not_transfer": false,
"P1_antibody_floor_wider_than_input_floor": true,
"P2_input_arm_replicates_cycle_102": true
}- A floor this stratum-dependent is a trap for pharmacodynamic readouts. A trial-grade occupancy assay that pools across the genome would be reading a stratum where the same-condition floor is 2%, and would call a 5% change significant; the same assay restricted to the top sites would need 10% to say anything. The stratum has to be declared with the threshold.
- The floor at the peaks is ~9% for a menin ChIP in the absence of any drug. A menin inhibitor whose true effect on occupancy is under 10% at the peaks is therefore unmeasurable by this design in a single pair, however real it is - so a negative occupancy result for a weaker menin inhibitor must not be read as no target engagement.
- Because the floor here is measured from two non-targeting hairpins, it includes whatever variation lentiviral transduction and selection add. A study without a hairpin arm has a narrower floor than this and should not import this number uncritically.
- Is the peak-stratum floor of ~9% an antibody property or a peak-sharpness property? MEN1 (9.4%) and MLL1 (8.8%) are two different antibodies against two different proteins and agree to within 0.6 points, which argues for the latter - a third, unrelated ChIP in the same deposit (UBR5, IKAROS, MEIS1, NRF1 or ZNF692) would settle it and is already in GSE299100.
- Does the floor change with bin size? 2 kb was chosen for comparability, not because it is optimal. At 200 bp the peak stratum would be sharper and the floor might be narrower or wider; nothing here says which.
- The raw reads are deposited under BioProject PRJNA1272575. Re-processing one pair from FASTQ would test whether any part of the 9.4% is submitter pipeline rather than biology plus IP, and is the only way to close counter-evidence on E1.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
GEO series GSE299100, 'UBR5 Mediates Menin Inhibitor Dependent Menin/KMT2A Complex Stability [ChIP-seq]', public 2025-08-01, contains GSM9032977/GSM9032983 (input), GSM9032979/GSM9032985 (Menin, Bethyl A300-105A) and GSM9032981/GSM9032987 (KMT2A/MLL1, Bethyl A300-086A), all MV4;11, all treatment DMSO, timepoint 72 hours, shRNA shLuc #1 or shLuc #2, all supplied as *_230927_*_hg38_rmdup.tdf.
GEO series GSE299100 and its six control samples, series record and all six GSM records fetched individually 2026-09-23 via acc.cgi targ=self form=text; treatment, timepoint, cell type, chip target, chip antibody, antibody vendor and catalogue number read off the live record · Cutler JA, Tsai J, Armstrong SA, Hatton C · NCBI Gene Expression Omnibus 2025What this evidence cannot show
The deposit has NO associated publication indexed as of 2026-09-23. Europe PMC free-text for "GSE299100" returns hitCount 0; PubMed `"UBR5"[Title/Abstract] AND "Menin"[Title/Abstract]` returns 0 with a non-empty querytranslation, and `"UBR5"[Title/Abstract] AND ("KMT2A"[Title/Abstract] OR "MLL"[Title/Abstract] OR "MEN1"[Title/Abstract])` returns exactly one record, PMID 38940486, a parathyroid whole-exome paper that is NOT this study. The GEO record carries only a BioProject relation (PRJNA1272575), no PubMed link. So this entry rests on a deposited dataset that has not been peer reviewed, and the submitters' processing (read filtering, duplicate removal, any normalisation inside igvtools) is undocumented.
Evidence that argues against this (1)
- The tracks are submitter-processed .tdf, not raw reads. If the two arms of a pair were normalised differently the ratio is contaminated and the floor is wrong in an unknown direction. Recorded against this: the two input arms differ 1.74x in total coverage bases and the two MEN1 arms only 1.08x, which is the pattern of differing library depth rather than differing normalisation - a normalised pair would agree in total. The raw reads under PRJNA1272575 were not re-processed and that check remains open. source
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Over 2 kb hg38 bins with the ENCODE hg38 blacklist v2 removed, scored on an external peak-set cut, the shLuc-1/shLuc-2 within-sample share ratio at the top 0.1% stratum is 1.0057 for input, 0.9063 for MEN1 and 0.9117 for MLL1; at deciles D1-D10 the input ratio stays within 0.9933-1.0073 and the two antibody ratios within 0.9758-1.0255.
What this evidence cannot show
n = 1 pair per assay, so the floor itself has no confidence interval; three assays are three independent draws and none is averaged with another. The strata are cut on a menin/MLL1 peak reference, so 'TOP' means the strongest menin-complex sites, not the strongest sites of whatever is being scored. 98,227 of ~1.55M bins were dropped by the blacklist and are not represented anywhere in these numbers.
Evidence that argues against this (1)
- The ENCODE blacklist removes exactly the regions where a coverage artefact would be largest, so excluding it makes the floor look TIGHTER than an unfiltered analysis would. Cycles 102-104 used the mm10 blacklist in the same way, so the two are comparable, but an unfiltered floor has not been computed here and would be wider. Amemiya HM et al. 2019
- E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
MV4-11 (Cellosaurus CVCL_0064) is annotated as childhood acute monocytic leukaemia / acute monoblastic-monocytic leukaemia from a 10-year-old male, carrying the gene fusion KMT2A-AFF1 from t(4;11)(q21;q23) and additionally an FLT3 internal tandem duplication.
Cellosaurus entry MV4-11 (CVCL_0064), fetched 2026-09-23 · Cellosaurus, SIB Swiss Institute of BioinformaticsWhat this evidence cannot show
Right fusion, wrong lineage presentation. Every claim drawn from this deposit must say so in its first clause (rule 73): this is a human KMT2A::AFF1 MONOCYTIC line with FLT3-ITD, not infant B-lymphoblastic leukaemia. Whether a noise floor is lineage-dependent is not obviously yes or no, but the cell line also carries a second driver lesion the index disease usually does not.
Evidence that argues against this (1)
- A noise floor is a property of the assay and the library, so a lineage mismatch is a weaker objection here than it would be for a biological claim. Against that: peak sharpness differs between cell types, and the floor measured here is strongly concentrated in the peak stratum, so a line with broader or weaker menin peaks would give a different TOP-stratum figure. source
- E4established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
Cycle 102 computed its 4-10% floor from two input pairs over the reconstructed S_FLAG top-0.1% stratum (q125 C1/NT = 1.0967, q124 C1/NT = 1.0360) and declared, before running, that 'an input pair carries no antibody, so it bounds library and coverage variation but NOT ChIP efficiency; the floor is a LOWER BOUND on a ChIP ratio's variability'.
What this evidence cannot show
This is a prior entry in this repository, not an external source. It is cited because the present measurement is the test of a caveat that entry wrote about itself.
Evidence that argues against this (1)
- Cycle 102's two 'input pairs' are NT versus PCGF1-knockout, i.e. a GENOTYPE contrast, not a same-condition pair. The same-condition input floor measured here is 0.6% at the peak stratum against their 3.6-9.7%, which is consistent with their figure containing real genotype biology rather than pure noise. That reading is developed in its own entry and would, if correct, mean cycle 102's floor was calibrated partly on the effect it was meant to bound. source 2026
its 2.42% is a real measurement in MV4-11 with MEN1/MLL1 antibodies, imported into a murine MLL-AF9 H2AK119ub comparison; the same ratios give 0 or 4 breaking marks depending only on which floor is imported
- related to →KMT2A / MLL1 — the gene that breaks
- related to →KMT2A::AFF1 (MLL-AF4), t(4;11)(q21;q23)