Oncofusion

start / graph / ana-q120-the-two-component-masking-model-has-no-solution-at-the-tagged-fusions-own-sites

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q120-h3k79me2-second-line-and-the-mll1-epitope

The shared-epitope masking calibration, run against a FLAG tag and then refuted by its own control: the MLL1 antibody reports 2.23-fold more retained signal than the tag at the fusion's own sites, but MLL1-dominated non-fusion sites fall HARDER than the fusion does (0.1069 against 0.1793), so no mixture of fusion and wild-type MLL1 can produce the observed 0.3995

also called the masking model fails · M4 has no solution · 2.23-fold masking

The obvious way to work out how much of an antibody's signal comes from the cancer protein is to compare it with a clean tag on that protein in the same cells. Doing it gives a tidy-looking answer of about three quarters - and the control built in to check the assumption shows the arithmetic cannot be right.

The same thing, for a scientist

Murine FLAG-MLLAF9, GSE260742, 2 kb bins, universe 1,227,708 (input > 0, ENCODE mm10 blacklist v2 removed), top 0.1% = 1,228 bins [E1]: | set | n | FLAG (the fusion) | MLL1 (A300-086A) | MLL1 / FLAG | |---|---:|---:|---:|---:| | **FLAG's own top 0.1%** | 1,228 | **0.1793** | **0.3995** | **2.23x** | | MLL1's own top 0.1% | 1,228 | 0.1651 | 0.3303 | 2.00x | | **MLL1-positive, FLAG-negative** (MLL1 top 0.1% minus FLAG top 1%) | 120 | 0.4584 | **0.1069** | 0.23x | | genome-wide mid-occupancy floor | ~500,000 | 0.9295 | 0.9165 | - | **M3 FIRES.** The pre-registered masking criterion was `R_MLL1 / R_FLAG >= 2.0` with band 1.8-2.2; observed **2.2281**, i.e. above the band by 0.028 and read as *at the band's upper edge* rather than as a comfortable pass. **This is the first measurement of shared-epitope masking against a tag anywhere in this project.** Q118's human estimate of 3.54-4.00x used menin as the fusion proxy; with a real tag the factor is 2.23x in mouse. **M4, the naive calibration, gives a clean-looking number.** Under *MLL1 = fusion + wild-type, and the wild-type component does not move*, the fusion's share of the MLL1 signal is `f = (1 - R_MLL1)/(1 - R_FLAG)` = **0.7317**, bootstrap 95% CI over 2,000 bin-level resamples **[0.7025, 0.7586]**. So 73% fusion, 27% masking, with a tight interval. **M6, the pre-registered control for that assumption, destroys it.** At the 120 bins where MLL1 is in its own top 0.1% but the fusion is not even in its top 1% - wild-type-dominated MLL1 sites by construction - the MLL1 share ratio is **0.1069**. Wild-type-dominated MLL1 sites are evicted *harder* than the tagged fusion's own sites. M6 demanded `>= 0.85` and **DOES NOT FIRE**; falsifier **F7 FIRES**. **And then the model has no solution at all.** A two-component mixture's ratio must lie between its components' ratios. The observed 0.3995 lies **above both** candidates - fusion 0.1793 and wild-type-dominated 0.1069. **No weighting of those two components produces the observed value**, so the model is refuted at these bins rather than merely uncertain, and the tidy 0.7317 with its tight confidence interval is an artefact of an assumption that its own control rejects.

Go deeper
Three readings of the impossible number, none excluded

1. Wild-type MLL1 at the fusion's sites is not the same population as wild-type MLL1 elsewhere. The 120-bin control set is defined by the absence of the fusion, so it selects a chromatin class - and if wild-type MLL1 at fusion-co-occupied promoters is menin-independent while wild-type MLL1 elsewhere is menin-dependent, the mixture is fine and the control is measuring the wrong wild-type.

2. A third species. A300-086A is a polyclonal anti-MLL1-N reagent; KMT2B shares extensive homology with KMT2A and this project already records that MLL2/MLL4 naming collides across two genes. If the track contains a menin-independent third component the mixture opens up.

3. Tag loss. NEB 14793S DYKDDDDK (D6W5B) reads the FLAG-tagged MLLAF9 construct only, so it is ground truth for the fusion and blind to wild-type MLL1. Neither it nor the MLL1 antibody can see a fusion molecule whose FLAG tag has been proteolysed, which would inflate the MLL1/FLAG ratio and mimic masking. A fusion population whose FLAG epitope is proteolysed or occluded is invisible to FLAG and visible to MLL1-N, and it would raise R_MLL1 above R_FLAG exactly as observed. Nothing in this design excludes it and no anti-FLAG western of the treated chromatin is deposited.

What all three share is that the masking factor is not identifiable from these two tracks, which is the honest conclusion and the opposite of what a beat setting out to calibrate masking would like to report.

The number that does survive, and it is a measurement rather than a model

2.23-fold. At the tagged fusion's own top 0.1% of bins, an MLL1 N-terminal ChIP reports a drug effect 2.23 times smaller than an anti-FLAG ChIP of the fusion in the same nuclei. That statement needs no mixture model - it is two ratios measured on the same bins - and it is the first time in this project that a shared-epitope antibody has been scored against ground truth.

For comparison, Q118's human estimate using menin as the fusion proxy is 3.54-4.00x in RS4;11, and the same proxy calculation on this beat's MOLM13 data gives a fusion fraction of 0.2373 against the mouse's 0.7317. The human MLL1-N channel behaves as though it is about one fifth fusion and the mouse channel as though it is about three quarters - which is what a retroviral over-expression model should do, and is the reason the mouse number must not be carried across.

Facts and figures
scriptanalyses/q120-h3k79me2-second-line-and-the-mll1-epitope/arm2-mouse-epitope.py
statusrun
result summaryM3 fires at 2.2281 (band 1.8-2.2, so at the upper edge). M4's naive fusion fraction is 0.7317 [0.7025, 0.7586]. M5 does not fire. M6 does not fire and F7 fires: wild-type-dominated MLL1 sites fall to 0.1069, harder than the tagged fusion's 0.1793, and the two-component model then has no solution.
cell linemurine FLAG-MLLAF9
fusionMLL-AF9, FLAG-tagged, retrovirally expressed
drugVTP50469 500 nM, 48 h against a 72 h DMSO control
genome buildmm10, verified from each file's own chromosome extents
bin bp2000
n bins universe1227708
n held1228
masking fold measured against a tag2.2281
human proxy fusion fraction rs4110.212
human proxy fusion fraction molm130.2373
antibodyMLL1 Bethyl A300-086A; FLAG NEB 14793S DYKDDDDK D6W5B
shown for fusionMLL-AF9 in a murine FLAG-tagged over-expression model; the corresponding human measurement does not exist because no human KMT2A-rearranged line with a tagged fusion has been ChIPped and deposited
absence classnever-measured
absence class noteThe quantity this entry tried and failed to calibrate - the fusion's share of an MLL1-N ChIP signal in a HUMAN KMT2A-rearranged cell - has never been measured by anyone, and no archive holds it: it needs a tagged allele in a human line, i.e. wet work. The murine substitute exists, was analysed here, and its own control refutes the model needed to transfer it.
absence class assigned2026-09-23 (Q120)
How this could bite us back
  • If wild-type MLL1 occupancy collapses to 0.11 under a menin inhibitor while the fusion's sites hold at 0.18, then the drug's on-target effect on normal haematopoietic COMPASS function is larger than its effect on the fusion at the fusion's own sites - which is a mechanism for differentiation-syndrome-like toxicity and for the loss of normal progenitor function.
  • A masking factor that cannot be identified is worse for the field than one that is large, because every published MLL1-N occupancy statement then carries an unbounded correction rather than a known one.
Still open
  1. Is the impossible mixture explained by tag loss? An anti-FLAG western of chromatin from treated and untreated cells, or a second tag at the other terminus of the same construct, would settle it, and neither exists in this deposition.
  2. Does A300-086A cross-react with KMT2B? A polyclonal raised against an MLL1 N-terminal region in a family with extensive homology is the obvious third component, and a KMT2B-knockout control ChIP would bound it. This project already records two separate naming collisions in this gene family.
  3. Why do wild-type-dominated MLL1 sites lose share harder than the fusion's own sites under a menin inhibitor? If that is real it says menin inhibition is at least as disruptive to wild-type KMT2A complexes as to the fusion, which is a toxicity argument nobody in this project has made from occupancy data.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    In murine FLAG-MLLAF9 (GSE260742), at the FLAG-DMSO top 0.1% of 1,227,708 mm10 2 kb bins (n = 1,228) the within-sample share ratios VTP50469/DMSO are FLAG 0.1793 and MLL1 (Bethyl A300-086A) 0.3995, a ratio of 2.2281; at the MLL1-DMSO top 0.1% they are FLAG 0.1651 and MLL1 0.3303; at the 120 bins in the MLL1 top 0.1% but not the FLAG top 1% the MLL1 ratio is 0.1069 and the FLAG ratio 0.4584. The naive fusion fraction (1-R_MLL1)/(1-R_FLAG) at the FLAG held set is 0.7317, bootstrap 95% CI [0.7025, 0.7586] over 2,000 bin-level resamples. Genome-wide mid-occupancy floors are FLAG 0.9295 and MLL1 0.9165. The largest single bin carries 0.026% of either MLL1 track and 1.74-1.89% of MLL1 coverage is in blacklist intervals, so the pre-registered instrument check M0 passes.

    What this evidence cannot show

    VTP50469 here is 500 nM against 330 nM in GSE127507 - NOT matched doses. Within GSE260742 both arms are 48 h VTP against a 72 h DMSO control, so the FLAG-versus-MLL1 comparison is exposure-matched to itself. The murine fusion is a RETROVIRALLY OVER-EXPRESSED FLAG-MLLAF9 construct and its fusion-to-wild-type ratio has no reason to match an endogenous human fusion - which is the single largest obstacle to transferring any number here to human data. NEB 14793S DYKDDDDK (D6W5B) reads the FLAG-tagged MLLAF9 construct only, so it is ground truth for the fusion and blind to wild-type MLL1. Neither it nor the MLL1 antibody can see a fusion molecule whose FLAG tag has been proteolysed, which would inflate the MLL1/FLAG ratio and mimic masking. Bethyl A300-086A, deposited as 'anti-MLL1n', an MLL1 N-TERMINUS epitope. The N-terminus is retained in every KMT2A fusion, so this track is the sum of the fusion and wild-type MLL1 in unknown proportion and CANNOT distinguish them. The immunogen residue range was not verified: three vendor routes were tried this session (fortislife.com two URL forms, bethyl.com/product/A300-086A) and all returned HTTP 404, so 'N-terminus' rests on the depositors' own target label and on the standard fact that MLL-FPs lack the C-terminal SET domain. n = 1 per antibody per condition per line in both series. No number in this beat carries a spread and none is precise. The 120-bin control set is small (9.77% of the MLL1 held set) and is defined by the absence of FLAG signal, which is a selection on the very quantity the model is about; a set defined that way can differ from 'wild-type MLL1 at fusion sites' in ways the measurement cannot see.

    Evidence that argues against this (1)
Connected to