A menin peak that keeps its call has still lost a sixth of its enrichment - and only where the fusion was
Looking only at the menin sites that were still detected after the fusion was destroyed, the amount of menin at them had fallen by about a sixth if the fusion had been there, and not at all if it had not. So even the sites that 'kept their peak' lost menin.
The same thing, for a scientist
The reconciliation test, run entirely inside the deposited MACS tables so that it does not depend on the authors' processed table. Among MEN1 control peaks whose call is retained (a treated summit falls inside the control interval), the median treated/control **fold_enrichment** ratio is **0.832 at HA-occupied sites (n=934)** and **0.995 at non-occupied sites (n=444)**. DOT1L gives 0.720 (n=657) vs 0.798 (n=262); H3K79me2 0.782 (n=1,753) vs 0.831 (n=5,449); the HA positive control gives 1.051 at the 1.9% of its own sites that survive, which is the survivorship signature. **This measure has the same sign as the authors' signal panel and the opposite sign to the binary peak census, from the same files.** It is heavily survivorship-biased - conditioning on the peak still being called selects the sites that lost least - so 0.832 is an upper bound on retention, i.e. a lower bound on loss.
Why this settles the instrument disagreement
The binary census and the signal panel disagree in sign. This measure is quantitative like the signal panel but is computed from the peak files like the census, so it isolates which property of the census produces the inversion. The answer is the binarisation, not the file, not the normalisation and not the gene annotation.
The bias runs against the finding, which is the useful direction
Conditioning on a retained call throws away every site where menin fell far enough to lose its peak. Those are by definition the sites that lost most. So the true loss at HA-occupied menin sites is larger than the 17% measured here, and the gap against the 0.5% at non-occupied sites is larger than 16 points. This is consistent with the 49.7% loss the signal panel reports over the whole target-gene set.
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. MACS fold_enrichment is computed against that library's own input and lambda model, so it is a within-sample share like everything else here; it is not comparable in level across channels or batches, only within a channel between strata.
| script | analyses/menin-retention-by-ha-occupancy/analyze_peaks3.py |
|---|---|
| status | run |
| result summary | At RETAINED MEN1 peaks, treated/control fold_enrichment 0.832 (HA-occupied, n=934) vs 0.995 (non-occupied, n=444). Same sign as the authors' signal panel; opposite sign to the binary peak census. Survivorship bias runs against the finding. |
| peak width ratio note | MEN1 treated/control median peak width ratio 0.93 - below the 1.3 pre-registered trigger, so the MEN1 peak figures are not flagged. DOT1L 1.13, H3K79me2 0.96; HA positive controls 2.79 and 1.93, both flagged. |
inputs
[ "GSE173599 MACS peak tables" ]
- If the sites that keep a menin peak have lost only a sixth of their menin while the sites that lose the call have lost much more, then the menin residue after a fusion-directed intervention is concentrated at the sites that matter most - the strongest ones - which is the worst distribution for a therapeutic aiming to clear the complex.
- What does a proper differential-binding analysis on the raw reads give at these sites? The BAMs exist in the deposit (23 GB) and were not downloaded.
- How much of the 17% is lambda-model drift rather than occupancy? A spike-in arm would answer this and none exists.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Among retained MEN1 peaks, the median treated/control MACS fold_enrichment ratio is 0.832 at HA-occupied sites and 0.995 at non-occupied sites.
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
Survivorship-biased by construction: only peaks that kept their call contribute, so the measured loss is a lower bound. n=1 per arm. HA occupancy is defined by the batch-matched HA DMSO 4 d peak file (1912x2); a 4 d DMSO arm is used because no 3 h HA library exists in the MEN1 batch.
Evidence that argues against this (1)
- MACS fold_enrichment depends on the local lambda estimate, which itself shifts when global signal changes, so a fold_enrichment ratio is not a clean occupancy ratio. Olsen SN et al. 2022
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The authors' independent quantitative panel gives a MENIN loss of the same sign and larger magnitude (0.503 at target genes) on the same cells and timepoint.
GSE173599_TSS_gene_body_signal_Figure3.xlsx - per-gene housekeeping-normalised TSS and gene-body ChIP signal, DMSO vs dTAG-VHL 3 h and 24 h, 16 channels · Olsen SN et al. · NCBI GEO supplementary file 2022What this evidence cannot show
Different window (TSS -1 kb/+3 kb vs MACS peak interval), different normalisation (housekeeping panel vs MACS lambda), same underlying libraries.
Evidence that argues against this (1)
- The two measures use different windows (a MACS peak interval versus TSS -1 kb/+3 kb), so agreement in sign is weaker evidence than it looks: a peak-centred and a TSS-centred window can move together for reasons of promoter architecture rather than of occupancy. Olsen SN et al. 2022
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
The HA positive control gives a fold_enrichment ratio of 1.051 at the 1.9% of HA sites that retain a call, which is the expected survivorship signature when a protein is essentially gone.
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
n=229 retained peaks of 12,017; a ratio computed on 1.9% of sites says nothing about the other 98.1%, which is the point.
Evidence that argues against this (1)
- A fold_enrichment ratio above 1 at surviving sites is also what would be seen if those sites were never fusion-bound and the HA signal there were background, which would make the survivorship reading unnecessary. Olsen SN et al. 2022
independent quantitative readout from the peak files alone, same sign: 0.832 at HA-occupied retained menin peaks vs 0.995 elsewhere
shows the inversion is caused by binarisation, not by the file, the normalisation or the annotation
menin sites that keep a peak call have still lost a sixth of their enrichment, but only where the fusion was