Peak-call retention tracks peak strength, not fusion occupancy, and that inverts the answer
Counting whether a 'peak' is still detected after treatment mostly measures how strong the peak was to begin with. Because the fusion sits on the strongest sites, that counting method says menin is MORE stable where the fusion was - the exact opposite of what the amount of menin actually does there.
The same thing, for a scientist
The peak-based instrument and the signal-based instrument **disagree in sign** on the same question, and the pre-registration named that a 'neither' trigger. Diagnosis: summit-containment retention is a monotone function of control-peak strength. For MEN1 the retention-by-q-value-decile curve runs **7, 10, 10, 17, 22, 26, 29, 40, 53, 83%** (weakest to strongest decile). Because HA-occupied sites are the strongest sites in every channel, the raw occupancy contrast is positive for **every** channel measured, including two the authors report as unchanged at 3 h: MEN1 +35.0 points, RNA Pol II +42.4, H3K79me2 +27.3, DOT1L +18.8. Matching within strength decile collapses most of these to noise (Pol II +4.2, H3K79me2 +4.1, DOT1L **-8.2**) while MEN1 stays at **+23.1**. A contrast that survives for menin but reverses for DOT1L, on an instrument whose output is dominated by peak strength, cannot arbitrate the biological question - and that is the finding.
The numbers, with the width ratio beside each as required
Summit containment, 3 h, control-arm peaks that contain at least one treated-arm summit. widthR is median treated peak width / median control peak width; the pre-registered re-score trigger is 1.3.
| channel | nCtrl peaks | widthR | all | HA-occupied | non-occupied | strength-matched gap |
|---|---|---|---|---|---|---|
| MEN1 | 4,641 | 0.93 | 29.7% | 50.8% | 15.8% | +23.1 pts |
| DOT1L | 3,167 | 1.13 | 29.0% | 37.4% | 18.6% | -8.2 pts |
| RNA Pol II | 14,035 | 1.01 | 66.1% | 88.8% | 46.4% | +4.2 pts |
| RNA Pol II pS2 | 5,429 | 1.04 | 48.2% | 70.8% | 30.7% | not computed |
| RNA Pol II pS5 | 10,776 | 0.93 | 30.8% | 48.1% | 15.2% | not computed |
| H3K79me2 | 15,424 | 0.96 | 46.7% | 69.5% | 42.2% | +4.1 pts |
| CDK9 | 1,220 | 2.46 FLAGGED | 20.6% | 34.8% | 17.4% | not computed |
| AFF4 | 1,784 | 1.05 | 36.5% | EXCLUDED | EXCLUDED | no batch-matched HA library |
Why the sign flips and what it costs
MACS returns a call when enrichment clears a significance threshold. A strong site that loses half its signal clears the threshold anyway; a marginal site that loses nothing may fail to clear it on a shallower treated library. So the binary measure is insensitive in the regime where the biology is (2-fold, at strong sites) and hypersensitive in the regime where it is not (marginal sites, sampling noise). The consequence for this corpus is specific: the genome-wide MENIN peak-retention figure is not evidence that menin stays, because it could not have detected the loss that is there.
Honouring the pre-registration
Pre-registered outcome A required 'the same contrast in the peak-based measure has the same sign'. In its binary form that criterion FAILS. In its quantitative form - the fold-enrichment of retained peaks - it passes (see the companion entry). Under the strictest reading of the pre-registered conjunction, the verdict is C (neither); the diagnosis above is why the primary instrument is nevertheless the one to believe, and it is recorded here rather than argued away.
Scale and limits
No spike-in exists anywhere in GSE173599, so every number is a within-sample share, and the treated arms are scaled by a 21-gene housekeeping panel whose own raw 3 h H3K27ac ratio averages 0.809 and is forced to 1. A housekeeping normalisation DEFINES the global change to be zero, so only the DIFFERENCE between strata is measurable; a uniform genome-wide loss of any of these proteins would be invisible here. n = 1 per arm for every channel except HA, which has two independent 3 h batches. The cells are HCB1, transformed by KMT2A::MLLT3 (MLL::AF9) - NOT the index fusion of this project, KMT2A::AFF1 - and a prior beat in this corpus found that per-gene ordering transfers between MLLT3 and AFF1 lines while magnitude does not.
| script | analyses/menin-retention-by-ha-occupancy/analyze_peaks2.py |
|---|---|
| status | run |
| result summary | Summit retention is a monotone function of control-peak strength (MEN1 deciles 7->83%). Raw HA-occupancy contrast positive for all channels; strength-matched it collapses (Pol II +4.2, H3K79me2 +4.1) or reverses (DOT1L -8.2) while MEN1 stays +23.1. The binary peak instrument cannot arbitrate this question. |
| peak width ratio note | MEN1 0.93, DOT1L 1.13, Pol II 1.01, pS2 1.04, pS5 0.93, H3K79me2 0.96, AFF4 1.05, CDK9 2.46 (FLAGGED), HA self-controls 2.79 / 1.93 / 2.18 (all FLAGGED). |
inputs
[ "GSE173599 MACS peak tables, 62 files" ]
- A trial or preclinical study using peak-call retention as a pharmacodynamic endpoint would report a working menin inhibitor as having no chromatin effect, because the assay is blind in exactly the effect-size range the drug produces.
- What is the minimum fold-change in occupancy that a MACS peak-call census at this depth can detect at a strong site? Without that number, every 'peak retention' figure in this corpus is a bound of unknown tightness.
- Would a quantitative differential-binding method (DiffBind/csaw on the raw reads) agree with the signal table? The BAMs are 23 GB and were not downloaded.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Summit-containment retention of MEN1 control peaks rises monotonically with control-peak strength from 7% in the weakest decile to 83% in the strongest, and every channel shows a positive raw HA-occupancy contrast including channels the authors report as unchanged.
GSE173599 - Chromatin state changes induced by targeted degradation of MLL-AF9 are phenocopied by combined DOT1L and MENIN inhibition [ChIP-Seq] · Olsen SN, Hatton C (Armstrong lab, Dana-Farber) · NCBI GEO (SubSeries of GSE173600) 2022What this evidence cannot show
Computed from the deposited MACS peak tables (62 files). n=1 per arm; peak calling is sensitive to library depth, which differs between arms and is not spike-in controlled. Occupancy strata for MEN1 are defined by the HA DMSO 4 d peak file from the same sequencing batch (1912x2), because that is the only batch-matched HA library.
Evidence that argues against this (1)
- Retention differences across strength deciles could in principle be biological rather than technical if weak sites are genuinely more labile; the DOT1L sign reversal under matching argues against a single biological reading. Olsen SN et al. 2022
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The authors' own quantitative panel, on the same cells and timepoint, reports MENIN and DOT1L loss at target genes and no change at non-target genes - the opposite stratum ordering to the binary peak measure.
MLL::AF9 degradation induces rapid changes in transcriptional elongation and subsequent loss of an active chromatin landscape · Olsen SN, Godfrey L, Healy JP, Choi YA, Kai Y, Hatton C, Perner F, Haarer EL, Nabet B, Yuan GC, Armstrong SA · Molecular Cell 82(6):1140-1155.e11 2022What this evidence cannot show
Figure 4B is a boxplot of fold change of total reads in a TSS window, a different instrument from a peak-call census; the disagreement is between instruments, not between datasets.
Evidence that argues against this (1)
- The authors' panel is restricted to genes passing their signal floor, so it cannot see the weak sites that dominate the peak census; the two instruments are not sampling the same sites, which is an alternative to the saturation explanation. Olsen SN et al. 2022
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
CDK9's treated/control median peak width ratio is 2.46, above the pre-registered 1.3 re-score trigger, so its peak-based figure is reported flagged.
GSE173599 - Chromatin state changes induced by targeted degradation of MLL-AF9 are phenocopied by combined DOT1L and MENIN inhibition [ChIP-Seq] · Olsen SN, Hatton C (Armstrong lab, Dana-Farber) · NCBI GEO (SubSeries of GSE173600) 2022What this evidence cannot show
Only 1,220 CDK9 control peaks and 646 treated peaks; the CDK9 libraries are the shallowest in the deposit and were sequenced in their own batch (200211).
Evidence that argues against this (1)
- Width inflation is the expected MACS behaviour for a genuinely collapsing channel, so flagging CDK9 on width may be discarding a true positive rather than an artefact. Olsen SN et al. 2022
measured the saturation directly: MEN1 summit retention 7% to 83% across control-peak strength deciles
the 80% figure reproduces (83.4% here) and is a saturated measure, so it does not bound the menin loss
the genome-wide peak instrument behind that entry is dominated by peak strength; every channel shows the same raw stratum contrast, including ones the authors report as unchanged
- ← related toSummit containment passes the positive control that the bin-overlap measure failed, in all three HA batches
summit containment fixes the width trap but not the dynamic-range trap
names the one channel that fires the pre-registered width re-score trigger
the build guard came back clean, so the peak intersections in this beat are not cross-build artefacts
- ← related toNo HA library exists in the AFF4 sequencing batch, so AFF4 is excluded from the peak partition by name
the one channel excluded by name from the peak partition, and why