Oncofusion

start / graph / ana-q140-a-menin-high-bin-class-is-an-accessibility-class-until-you-divide-by-igg

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q140-sem-menin-arm

Ranking bins by raw menin signal builds a class that the IgG control piles into eight to thirty-six fold - so the first version of our menin-bound contrast was measuring accessibility, and its own negative control said so before we believed it

To ask whether a drug strips a chemical mark from the places a protein sits, you first have to define those places, and we defined them by where the protein's signal was strongest - which turned out to mean 'where every experiment piles up', because the no-antibody control piled up there hardest of all. Dividing by that control before defining the places fixed it, and the answer changed from 'no effect' to a real one.

The same thing, for a scientist

First attempt: within the top-0.1% DMSO H3K4me3 bin set, split into menin-high and menin-low tertiles by **raw menin CPM**, and compare each arm's bound/free read-share ratio. Result: DMSO 1.265, EPZ 1.216, **VTP 1.227** - VTP/DMSO = 0.970, a clean null. The control that killed it: the **IgG libraries show a bound/free ratio of 8.02, 8.58, 7.75 and 35.80** against H3K4me3's 1.22-1.27. A negative control cannot be eight to thirty-six fold enriched in a class that means 'menin is here'; the class means 'reads accumulate here'. Redefining bound on **(menin+1)/(IgG+1)** and matching the classes decile-by-decile on DMSO H3K4me3 gives an IgG bound/control ratio of **0.076 and 0.078** - the control is now depleted, as it must be - and the answer changes from 0.970 to **0.766**.

Go deeper
The null was the comfortable answer, and it was wrong

The contaminated contrast returned 'the menin inhibitor does not preferentially deplete H3K4me3 at menin-bound sites' - a publishable negative, consistent with the housekeeping control that had just fired, and consistent with the story we had already half-written. It was the answer we were on our way to. What stopped it was a number printed beside it that had no business being large. Printing the negative control's value in the same line as the result is the whole mechanism here; had it been logged elsewhere, or not computed, the null would have gone into the graph. 1

Why the contamination is structural, not accidental

CUT&Tag pA-Tn5 tagmentation is biased toward accessible chromatin regardless of the antibody, so every antibody's raw signal contains an accessibility component, and the strongest raw signal is disproportionately at the most accessible sites. Ranking by raw signal therefore ranks partly by accessibility for any target. The correction is not specific to menin: any class defined by raw CUT&Tag occupancy needs an IgG denominator, and the IgG library's own behaviour in that class is the test of whether it worked. 12

The second fix, which mattered independently

Menin-bound sites carry more H3K4me3 than menin-free sites (263 vs 208 CPM in the contaminated version). A class with more signal to lose loses more under any global depression, so regression to the mean alone could manufacture the predicted result. Matching the classes within deciles of DMSO H3K4me3 removes it: realised means 235.7 vs 229.0 CPM, while menin enrichment separates 44.6-fold vs 5.2-fold. The classes differ in menin and not in H3K4me3, by construction.

Facts and figures
scriptanalyses/sem-arm-identity-from-signal/run3.py
statusrun
contaminated definition
{
  "basis": "raw menin CPM tertiles within the H3K4me3 peak set",
  "arm_ratios": {
    "untreated": 1.4815,
    "DMSO": 1.2647,
    "EPZ5676": 1.2163,
    "VTP50469": 1.2265
  },
  "VTP_over_DMSO": 0.9698,
  "IgG_bound_over_free": [
    8.0235,
    8.5831,
    35.8014,
    7.7534
  ],
  "verdict": "REJECTED - negative control enriched 8-36x"
}
corrected definition
{
  "basis": "(menin+1)/(IgG+1), classes matched within DMSO H3K4me3 deciles",
  "VTP_over_DMSO": 0.7657,
  "IgG_bound_over_control": [
    0.0763,
    0.0781
  ],
  "verdict": "ACCEPTED - negative control depleted"
}
How this could bite us back
  • The correction moved the answer toward the hypothesis this project favours. That is the direction in which a methodological fix should be trusted least, and it is why the corrected contrast carries an explicit negative-control value in its own data payload rather than a claim that it is clean.
Still open
  1. How many occupancy-based site classifications in this graph were built on raw signal rather than on enrichment over a matched control? The IgG-in-the-class test is cheap and would bound it.
  2. Is the residual 0.076 IgG ratio in the corrected class low enough, or does it still leave a few percent of accessibility contamination that scales with the peak-share confound?
Every source behind this entry (2)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    IgG libraries show a menin-high/menin-low read-share ratio of 8.02-35.80 when the classes are defined on raw menin signal, and 0.076-0.078 when defined on menin enrichment over IgG with H3K4me3-matched controls; the VTP/DMSO answer moves from 0.970 to 0.766 between the two definitions.

    What this evidence cannot show

    The menin libraries' own 9.3 ratio in the contaminated version is circular - they defined the split - and is not evidence either way; only the IgG value is diagnostic. Four IgG libraries across two protocols were used and all four agree in direction.

    Evidence that argues against this (1)
    • Dividing by IgG has its own hazard: where IgG counts are near zero the ratio is unstable and can select bins on control noise rather than on signal. The pseudocount of 1 CPM damps this, but a bin set defined on an enrichment ratio is not automatically cleaner than one defined on raw signal - the evidence that this one is cleaner is the IgG value inside it (0.076), not the fact that a ratio was used. Janssens DH et al. 2021
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    CUT&Tag uses antibody-tethered Tn5 tagmentation, which integrates adapters at the targeted site; the method's controls are IgG libraries run through the identical protocol, which is why an IgG track is the appropriate denominator for occupancy enrichment.

    Automated CUT&Tag profiling of chromatin heterogeneity in mixed-lineage leukemia · Janssens DH, Meers MP, Wu SJ, Babaeva E, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Genetics 53:1586-1596 2021
    PMID 34663924DOI
    What this evidence cannot show

    The deposit includes IgG libraries for exactly this purpose and the paper uses them; our contribution is only the observation that a class defined on raw signal fails the IgG test while one defined on the ratio passes it.

    Evidence that argues against this (2)
    • That IgG is the protocol's own control does not make it the right DENOMINATOR. An IgG library reports where tethered Tn5 can reach, which is accessibility plus sequence bias; dividing by it removes accessibility but also removes any genuine signal that happens to sit in open chromatin - and fusion target promoters are exactly that. The ratio can therefore understate a real effect, and this entry's own IgG class spanning 8.0-35.8x shows how large the correction is.
    • The alternative reading of the same data is that the raw menin class was not an artefact at all: menin genuinely occupies accessible promoters, so an 'accessibility class' and a 'menin class' are not separable in this system by any normalisation, and the choice of denominator decides the answer rather than revealing it.
Connected to