Oncofusion

start / graph / hyp-q125-the-menin-inhibitor-pharmacodynamic-signal-worth-measuring-is-the-eviction-gradient-not-any-single-site-average

Hypothesisspeculative confidence 1 of 5: Our own inference or analogy. Not shown to be true.Q125-the-h2ak119ub-arm-of-the-pcgf1-knockout

Every chromatin readout of a menin inhibitor measured in this corpus is graded by how much fusion the drug removed from a site, so a within-sample average over the fusion's occupancy set is the wrong statistic and will report near-null in a responding cell - the signal is the SLOPE against eviction depth, not the level

also called measure the slope, not the level · the eviction-gradient hypothesis · why five committed aggregate ratios were near 1.0

The drug's effect on the chromatin is strong where the drug works hardest and absent where it barely works, so averaging over all the sites cancels it out. The proposal is that the useful measurement is how steeply the effect changes across sites, not its average.

The same thing, for a scientist

**Observation.** Every mark this beat measured is monotone in fusion-eviction depth at Spearman +/-1.000 and near-null on average: H2AK119ub aggregates to 1.1694 and 0.9071 while running 1.4450 and 1.0151 at the most-evicted quartile; H3K27me3 aggregates to 1.3612 while reaching 2.0007; H3K4me3 aggregates to 1.1276 with a gradient of the opposite sign that the average cannot express at all [E1]. **Claim.** The quantity that carries the drug's effect in a within-sample-normalised chromatin readout is the **regression of the mark's share ratio on eviction depth**, and the aggregate share ratio over an unstratified occupancy set is a statistic whose expectation is near 1.0 whatever the drug does, because the set spans the whole range of eviction and the numerator and denominator are both normalised within sample [E1] [E2]. **Why it matters beyond this beat.** This corpus has committed at least five aggregate share ratios of exactly this shape - Q122's `R_K119ub(VTP)` = 0.8473 with criterion D1 *DOES NOT FIRE*, Q124's ARM-2 H3K79me2 nulls, and this beat's own S1, S2 and V1 - and in every case the conclusion drawn was 'the drug does little here'. **If this hypothesis is right, those conclusions are statements about the statistic and the H3K79me2 arm should be re-run stratified before `hyp-q124` is believed** [E2].

Go deeper
What makes this testable rather than a reinterpretation

The stratifying variable - FLAG VTP/DMSO share ratio per bin - is measured in a different channel from every readout it stratifies, comes from a different beat's committed file, and is available for every arm of this series. So the test is mechanical: re-score every committed aggregate share ratio in this corpus against eviction depth and see whether a gradient appears where the average did not.

The most valuable single case is hyp-q124's. It concluded that H3K79me2's drug response is unchanged in a PCGF1 knockout, from aggregate ratios at S_FLAG. The bins are already on disk in C:/Temp/q124 and the quartiles are already defined. It is a script, not a download.

Facts and figures
statementIn within-sample-normalised chromatin readouts of menin inhibition, the drug's effect is monotone in the depth of fusion eviction at each site, so the aggregate share ratio over an unstratified occupancy set has an expectation near 1.0 regardless of the drug's effect, and the informative statistic is the gradient against eviction depth.
origindata-pattern
prior probability0.6
absence classmeasured-not-analysed
absence class noteEvery input this hypothesis needs is binned and on disk in C:/Temp/q117, q120, q122, q124 and q125, and every committed aggregate it questions is in a results.json in this repository. Nothing needs downloading and nothing needs wet work; all three falsifiers are scripts.
absence class assigned2026-09-23 (Q125)
predictions
[
  "Re-scoring Q124's H3K79me2 ARM-2 bins (already on disk at C:/Temp/q124) against Q117's eviction quartiles reveals a gradient in the control genotype that its aggregate did not show",
  "Q122's F3 and F3_NT H2AK119ub arms, re-stratified, show the same +1.000 gradient as the NT and C1 cultures measured here, and its 0.8473 becomes a Q1 value above 1.0",
  "In any arm of this series, the Spearman of a mark's share ratio against eviction depth has a larger absolute value than |aggregate - 1| has, scaled by the same null",
  "A site set pre-selected for deep eviction gives a larger and more reproducible drug effect than the full top-0.1% set, in every channel and both drugs"
]
experiments
[
  "exp-q125-restratify-every-committed-share-ratio-in-this-corpus-against-eviction-depth"
]
analyses
[
  "ana-q125-the-menin-inhibitor-does-silence-the-sites-it-evicts-the-fusion-from-and-the-effect-is-graded-by-how-much-fusion-left",
  "ana-q125-the-two-second-priority-marks-carry-the-largest-effect-and-the-control-that-rules-out-an-artefact"
]
falsifiers
[
  "F1 - THE GRADIENT IS AN ARTEFACT OF THE STRATIFIER, NOT OF THE DRUG. Population: murine MLL-AF9 (GSE260742, cell line MLLAF9, taxid 10090), the 1,228-bin FLAG-defined set, Q117's four eviction quartiles. Fires if the same quartile split built from the FLAG SGC0946/DMSO ratio instead of the VTP/DMSO ratio - a stratifier from a DIFFERENT drug in the same channel - gives the same Spearman +1.000 for the VTP contrast of every mark. That would show the gradient tracks a property of the bins (their FLAG dynamic range) rather than the specific drug's action. Consequence: the hypothesis becomes 'the readout is graded by a bin property' and its clinical corollary dies. Instrument: the FLAG SGC track GSM8123360 is already binned in C:/Temp/q117; no download. NOT YET RUN AND IT IS THE FIRST THING THAT SHOULD BE.",
  "F2 - RESTRATIFICATION DOES NOT RESCUE THE H3K79me2 NULL. Population: murine MLL-AF9 NT and C1 cultures, H3K79me2 (GSM8123438/440 and GSM8123435/437), Q117's four eviction quartiles, 500 nM VTP-50469 48 h against DMSO 72 h. Fires if the control genotype's H3K79me2 Spearman against eviction depth is below 0.60 in absolute value, i.e. the mark is genuinely flat across eviction depth where H2AK119ub, H3K27me3 and H3K4me3 are not. Consequence: hyp-q124's asymmetry is real at the gradient level as well as the aggregate level, and this hypothesis's scope narrows to the polycomb channels. Instrument: ABCAM ab3594 polyclonal, which cannot separate me2 from me3 and reads a mark, not an enzyme. Cost: one script over bins already on disk. NOT YET RUN.",
  "F3 - THE AGGREGATE IS NOT NEAR 1.0 BY CONSTRUCTION. Population: all committed within-sample share ratios at S_FLAG across Q117, Q118, Q120, Q122, Q124 and Q125 - 20 or more values. Fires if their distribution is NOT centred near 1.0, i.e. if aggregates routinely take extreme values in this corpus. Consequence: the 'expectation near 1.0 whatever the drug does' premise is wrong and the hypothesis is arguing from a selected sample of near-null results. Instrument: the committed results.json files; no new computation. NOT YET RUN, and it is the cheapest of the three."
]
How this could bite us back
  • If the gradient is the signal, then a pharmacodynamic assay in a patient sample - where no FLAG track exists and eviction depth per site is unmeasurable - cannot compute it, and the practical readout has to be a fixed set of sites pre-selected for deep eviction in a model system. That set would be model-derived and its transfer to a human tumour is exactly the partner-import problem this corpus keeps finding.
  • The hypothesis also cuts against this beat's own strongest result: if the aggregate is uninformative by construction, then so is the aggregate H2AK119ub gain, and the quartile analysis carries all the weight at n = 4 strata with Spearman able to take six values.
Still open
  1. Does the gradient exist in a human line? Q120's MOLM13 and Q118's RS4;11 arms have menin or MLL1-N tracks that could define an eviction stratifier in the same way, and neither has been stratified.
  2. Is eviction depth the right stratifier, or is it a proxy for baseline fusion occupancy? The two are correlated and this beat did not separate them.
Every source behind this entry (2)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    In GSE260742's NTvKO control genotype, three histone marks are monotone in fusion-eviction depth at Spearman +1.000 (H2AK119ub, both conventions), +1.000 (H3K27me3) and -1.000 (H3K4me3) while their aggregate share ratios over the unstratified 1,228-bin set are 1.1694, 0.9071, 1.3612 and 1.1276, and the corresponding matched nulls are flat to within 0.007.

    What this evidence cannot show

    NO SPIKE-IN NORMALISATION IS POSSIBLE. Q117 verified zero hits for the entire spike-in vocabulary across all 419,824 characters of GSE260742's 158 sample records and no scale factor is deposited, so every cross-condition number here is a WITHIN-SAMPLE SHARE, which cancels the unknown per-sample scale factor exactly and IS BLIND TO A UNIFORM CHANGE BY CONSTRUCTION. Nothing here measures any mark's abundance; everything measures its distribution. EXPOSURE ASYMMETRY: DMSO 72 h and SGC0946 5 uM 72 h are matched, VTP-50469 500 nM is 48 h, so every VTP ratio is a LOWER BOUND on the 72-hour effect - identical in both genotypes, so the genotype DIFFERENCE is unaffected. REAGENTS: H2AK119ub is Ubiquityl-Histone H2A (Lys119) (D27C4) XP Rabbit mAb, NEB 8240S, and it CANNOT distinguish PRC1.1 from canonical PRC1 - RING1A/B is the ligase in both and PCGF1 is one of six PCGF paralogs - nor deposition from loss of a deubiquitinase. The H3K27me3 and H3K4me3 reagents are NOT RESOLVED and are recorded unresolved rather than guessed; `antibod` occurs zero times in the 158 GEO records. S_FLAG is defined by DYKDDDDK Tag (D6W5B) Rabbit mAb, NEB 14793S, on a RETROVIRALLY OVER-EXPRESSED tagged MLLAF9 construct, so it is the over-expressed molecule's sites and not an endogenous fusion's. THE KNOCKOUT is one CRISPR guide, one clone C1, no second clone, no rescue and NO deposited PCGF1 protein or transcript measurement. SYSTEM: murine MLL-AF9 (KMT2A::MLLT3) myeloid blood cells, NOT the infant B-lineage KMT2A::AFF1 disease this graph is about. The gradient is measured across FOUR strata of ONE set in ONE experiment and one cell system; Spearman at n = 4 is coarse and can only take six magnitudes. The stratification is Q117's, recorded there as a post-hoc deviation.

    Evidence that argues against this (1)
  2. E2weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.dataset

    Q122 measured R_H2AK119ub(VTP/DMSO) = 0.8473 at this exact set in a different murine culture and scored its criterion D1 'DOES NOT FIRE', concluding no drug engagement in that channel; the same aggregate statistic in two further cultures here is 1.1694 and 0.9071 and hides a +1.000 gradient, so the earlier conclusion is a property of the statistic rather than of the channel.

    What this evidence cannot show

    This is an inference about an earlier beat's conclusion and NOT a re-measurement of its data: Q122's F3 culture files were not re-binned here and its aggregate was not re-stratified. The claim that its conclusion would change under stratification is UNTESTED and is the first experiment this hypothesis proposes. Confidence is deliberately weak.

    Evidence that argues against this (1)
Connected to