Oncofusion

start / graph / ana-q124-the-deposited-vtp-timecourse-cannot-settle-whether-f-is-a-lower-bound-and-five-statistics-say-so

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q124-the-timecourse-and-the-pcgf1-knockout

The only deposited H3K79me2 kinetic series for a menin inhibitor is non-monotone - a minimum at 16 h and a recovery at 24 h that survives common support and two rank statistics - and its 24 h point shows MORE redistribution than the 48 h main-arm sample, so the 48 h/72 h asymmetry in cycle 82's F stands and F keeps the words lower bound

also called the timecourse arm's negative · K0 failed and RK2 fired · F keeps the words lower bound

A public dataset follows the H3K79 tag for 4, 8, 16 and 24 hours after the menin drug goes on, which could have told us whether a number this project committed yesterday is near-final or only a floor. It does not: the tag is lowest at 16 hours and back up at 24, five ways of measuring it disagree about which way that last step goes, and the number keeps its floor.

The same thing, for a scientist

**Cycle 82's headline** - with direct DOT1L inhibition bounding the achievable effect, a menin inhibitor buys `F = (1 - V)/(1 - G) = 0.0947 to 0.2500` of it - **carries one stated weakness: in the murine arm the numerator is measured at 48 h and the denominator at 72 h, so `F` is a lower bound.** GSE260742 deposits a murine VTP-50469 H3K79me2 kinetic series that could bound how much that matters. **It does not, and the failure is informative** [E1]. | within-sample share ratio against DMSO 24 h | 4 h | 8 h | 16 h | 24 h | |---|---:|---:|---:|---:| | at `S_FLAG`, the tagged fusion's own top 0.1% | 0.8996 | 0.8835 | **0.6181** | **0.6790** | | at `S_K79tc`, the mark's own top 0.1% | 0.8511 | 0.8250 | **0.6087** | **0.7269** | **Both series RISE between 16 h and 24 h, so pre-registered guard K0 fails on both sets, refutation RK2 fires, and the declared consequence is that criteria K1, K2 and K3 are SUPPRESSED rather than reported.** **The guard earned its place immediately, because K1's arithmetic fired for the wrong reason.** K1's statistic is `L = [A(24) - A(16)]/A(24)` where `A(t) = 1 - V(t)`, and K1 fires if `L <= 0.10`. Observed `L` is **-0.1897** at `S_FLAG` and **-0.4328** at `S_K79tc`. A negative `L` satisfies `L <= 0.10`, so the scoring code printed *"K1 FIRES - PLATEAUED BY 24 h"*. **It is not a plateau; a threshold written for a one-sided quantity was applied to a two-sided one.** Both strings are left verbatim in `results.json` next to a `SUPPRESSED_BY_RK2` field, because the size of a criterion's error is a result. **Five statistics were then computed post hoc, and they do not license a plateau either** [E2]: | statistic | 4 / 8 / 16 / 24 h | 16-to-24 step over 8-to-16 step | direction of the last step | |---|---|---:|---| | aggregate share over `U_m` at `S_FLAG` | 0.8996 0.8835 0.6181 0.6790 | 0.2295 | less effect | | aggregate share over **common support** | 0.9441 0.9211 0.6530 0.6656 | 0.0470 | less effect | | median **per-bin** share ratio at `S_FLAG` | 0.9176 0.9168 0.6684 0.7265 | 0.2339 | less effect | | fraction of `S_FLAG` bins **losing rank** | 0.4642 0.4609 0.7647 0.7345 | 0.0994 | less effect | | **enrichment residual** (DEVIATION 1) | +0.0265 +0.0213 -0.0543 **-0.1837** | **1.7116** | **more effect** | **Four say the last interval is 5-23% of the previous one and points to less effect; the fifth says it is 171% and points to more. The sign is not consistent, so no signed increment can be projected to 48 h.** **And a second, independent reason the arm cannot be carried to `F` at all: the two experiment batches are not on one scale.** Pre-registered criterion **K4** asks whether the 24 h timecourse point has moved the mark no further than Q122's 48 h F3 point. **It DOES NOT FIRE on either set**: `V_tc(24 h) = 0.6790` against `V_F3(48 h) = 0.7734` at `S_FLAG` (threshold 0.7234), and 0.7269 against 0.9117 at `S_K79` (threshold 0.8617). **The 24 h sample from the timecourse experiment shows MORE redistribution than the 48 h sample from the main experiment** - impossible if the two were one time series, and straightforward if they are two batches with their own vehicles, inputs and libraries. Prior PR2 was the one this beat said it was least sure of, and it is the one that failed. **So cycle 82's `F = 0.09-0.25` keeps the words "lower bound", and cycle 82's next-step item 4 - *"does GSE260742 hold a 72 h VTP H3K79me2 sample? If so, one download closes arm 1's 48/72 h asymmetry"* - is answered: the series holds something better in principle and it does not close the asymmetry.**

Go deeper
The one positive that survives, reported at weak

The enrichment residual - V at S_FLAG minus the track's own genome-wide enrichment change - is the only one of the five statistics that is monotone: +0.0265, +0.0213, -0.0543, -0.1837. And its 24 h value lands inside the range the same quantity takes in four independent 48 h VTP cultures: -0.1584 (NT), -0.1972 (PCGF1-KO), -0.2015 (F3), -0.2040 (F3_NT).

Read generously that says the mark's relocation is essentially complete by 24 h and the 48 h numerator is near-final. It is reported at weak and no numeric correction to F is proposed from it, for three reasons. It is post hoc. It rests on the one statistic of five that disagrees with the other four about the direction of the last step. And the index it subtracts, C = the share of a track's own coverage in its own top 0.1%, is not a purely technical quantity - a genuine redistribution lowers it too, which the SGC0946 contrasts prove by giving residuals of only -0.004 to -0.025 for a drug whose effect is enormous. A statistic that cannot separate technical from biological in the control it was validated against cannot be load-bearing in the arm it was built for.

What the 16 h minimum is, and what it is not

DEVIATION 1 proposed a library explanation and DEVIATION 2 partly refuted it, which is why neither may be quoted alone.

The five timecourse tracks span 2.208-fold in non-zero bin count within the universe - DMSO 501,520, 4 h 789,737, 8 h 973,715, 16 h 1,107,059, 24 h 986,555 - and Spearman(V, non-zero count) over the four timepoints is exactly -1.000. The 16 h track is the one poorly-enriched library: its within-sample enrichment index is 0.0653 against 0.0849, 0.0838 and 0.0838. A null built from that index alone spans 0.6724-0.8731, against V's observed 0.6181-0.8996 - the observed range is not wider than the null's.

But the minimum survives the controls. Over the common-support universe - bins non-zero in all five tracks, 319,091 bins, 25.99% of U_m, where a spread in non-zero count cannot act through the denominator - V is 0.9441, 0.9211, 0.6530, 0.6656: the 16 h minimum and the 24 h recovery are both still there, with the upward step shrunk from +0.0609 to +0.0126. And both rank statistics, which are largely invariant to a monotone rescaling of a track, are non-monotone in the same direction.

So depth explains most of the recovery's SIZE and none of its EXISTENCE. The honest description is a steep fall to 16 h and a small, sign-ambiguous movement afterwards, in a series of four unreplicated timepoints whose libraries differ more than a kinetic experiment's should.

What would have made this arm work

Three things, none of which is in the deposit.

  1. A spike-in. Q117 verified zero spike-in vocabulary across all 419,824 characters of the 158 sample records. Without one, a plateau in a within-sample share cannot be distinguished from continued uniform loss, and the published literature has demonstrated exactly that failure for H3K27me3 under an EZH2 inhibitor - see the counter-evidence on E1.
  2. A replicate at any one timepoint. With n = 1 per timepoint, the 16-to-24 h step has no noise estimate other than the statistic-to-statistic spread, which is what this entry had to use instead.
  3. An SGC0946 arm inside the timecourse. The timecourse deposits none, so G cannot be measured in this batch and F cannot be recomputed within it at any timepoint - which is why K4's batch-scale test was necessary and why its failure is terminal for the arm's usefulness to F.
Facts and figures
scriptanalyses/q124-the-timecourse-and-the-pcgf1-knockout/run.py, deviation1_enrichment_null.py, deviation2_common_support.py, deviation3_four_statistic_steps.py
statusrun
result summaryV at S_FLAG is 0.8996, 0.8835, 0.6181, 0.6790 at 4/8/16/24 h - non-monotone, so K0 fails on both sets, RK2 fires and K1/K2/K3 are suppressed. Five post-hoc statistics disagree about the sign of the last step, so no plateau is supported. K4 does not fire on either set: the 24 h timecourse point (0.6790) is below Q122's 48 h F3 point (0.7734), so the two batches are not on one scale. Cycle 82's F = 0.09-0.25 keeps the words lower bound.
cell linemurine MLL-AF9 F3, timecourse batch (filename family ChIP_K79_F3_TIMECOURSE), GEO cell line 'MLLAF9', cell type 'Myeloid', tissue 'Blood', taxid 10090
fusionKMT2A::MLLT3 (MLL-AF9), FLAG-tagged, retrovirally over-expressed
lineagemyeloid (AML model); NOT infant B-lineage KMT2A::AFF1
drugVTP-50469 500 nM at 4, 8, 16 and 24 h against DMSO 24 h; the 24 h point is exposure-matched and is the only exposure-matched murine VTP/DMSO H3K79me2 contrast in this series
genome buildmm10, verified from each of the six files' own per-chromosome declared extents against mm10, mm39, hg38 and hg19
bin bp2000
n bins universe1227708
n held1228
antibodyH3K79me2 ABCAM ab3594-100UG rabbit polyclonal; FLAG NEB 14793S clone D6W5B
epigenetic readoutH3K79me2's within-sample share ratio at the tagged fusion's own top-0.1% sites under 500 nM VTP-50469 is 0.8996 at 4 h, 0.8835 at 8 h, 0.6181 at 16 h and 0.6790 at 24 h against a 24 h vehicle - non-monotone, and its 24 h value is below the 0.7734 the same quantity gives at 48 h in a different batch of the same line
mark writtenH3K79me2, by DOT1L
writer of that markgene-dot1l
shown for fusionKMT2A::MLLT3 in a murine FLAG-tagged over-expression model; nothing human and nothing KMT2A::AFF1. The mark cannot be attributed to the fusion by this antibody, and the fusion enters only through the FLAG-defined set
absence classmeasured-not-analysed
absence class noteThe kinetic series has been public since 2025-11-23 and Q122's section 7 called it 'the single most valuable thing in the series that this beat does not touch', naming four of its five accessions. It is now analysed and the answer is that it does not settle the question it was expected to settle. THE ABSENCE THAT REMAINS IS DIFFERENT AND NARROWER: a duration-matched VTP-50469 H3K79me2 sample in the SAME batch as the 72 h SGC0946 arm does not exist anywhere in GSE260742, and that is never-measured rather than measured-not-analysed. Cost to close: a wet experiment, not a download.
absence class assigned2026-09-23 (Q124)
inputs
[
  "GSM8123452 Timecourse Control - DMSO H3K79me2, 24hrs",
  "GSM8123449 / 8123450 / 8123447 / 8123448 Timecourse Control - VTP50469 500nM at 4 / 8 / 16 / 24 h H3K79me2",
  "GSM8123451 Timecourse Control - DMSO input, 24hrs (robustness universe, 1,194,614 bins)",
  "GSM8549823 input batch 2 defining U_m; GSM8123359 FLAG DMSO defining S_FLAG",
  "Q122 results.json for V_F3(48 h) = 0.7734 and V_F3(S_K79) = 0.9117, read in code",
  "mm10-blacklist.v2.bed, 3,435 intervals; mm10 chromosome sizes re-fetched live"
]
How this could bite us back
  • This arm was dispatched to remove a caveat from a committed number and it strengthened the caveat instead. Anyone quoting F = 0.09-0.25 as a point estimate after this cycle is using a figure whose only available kinetic check failed its own monotonicity guard.
  • The batch-scale failure is the more general hazard. Two experiments in one GEO series, one cell line, one lab, one antibody, one instrument, and a 24 h drug sample from one batch shows more effect than a 48 h sample from the other. Every cross-batch comparison this project makes inside GSE260742 - and there are several - inherits that.
  • If the vehicle track's library complexity is half that of the drug tracks it normalises, then every V in this arm is computed against a denominator measured differently from its numerators. That is not fixable by a within-sample share, and the series has no spike-in.
Still open
  1. Is the 16 h minimum real? It survives common support and two rank statistics, so it is not purely a library artefact, but with n = 1 per timepoint and a 2.2-fold spread in library complexity nothing in this arm can establish it. A single replicate at 16 h and 24 h would settle it and none exists in GSE260742.
  2. If the minimum is real, what recovers between 16 and 24 h? A partial restoration of H3K79me2 at the fusion's own sites 24 h into continuous drug is an adaptation, and the depositing paper's own thesis is transcriptional memory. This beat named it and did not pursue it because the statistics disagree about whether it happened.
  3. The H2AK119ub and H3K27me3 timecourses exist but only at 8 h and 24 h (GSM8123441/443/442 and GSM8123444/446/445) - no 4 h and no 16 h point, so they cannot resolve the 16 h minimum and can only bracket it. Would the polycomb channel show a minimum at the same place?
  4. Why does the 24 h vehicle have 501,520 non-zero bins when every drug track has 790,000 to 1,107,000? A vehicle with half the library complexity of every sample it normalises is a hazard for every ratio in this arm, and this beat measured it without explaining it.
Every source behind this entry (2)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    In murine MLL-AF9 F3 cells (GEO cell line MLLAF9, myeloid, blood, Mus musculus taxid 10090; filename family ChIP_K79_F3_TIMECOURSE), 2 kb bins, universe 1,227,708 (GSM8549823 input > 0, mm10 blacklist v2 removed, mm10 verified from each file's own extents), top 0.1% = 1,228 bins: the within-sample H3K79me2 share ratio under 500 nM VTP-50469 against DMSO 24 h is 0.8996 at 4 h, 0.8835 at 8 h, 0.6181 at 16 h and 0.6790 at 24 h at the FLAG-DMSO top-0.1% set, and 0.8511, 0.8250, 0.6087 and 0.7269 at the mark's own top-0.1% set. Both series rise between 16 h and 24 h. Over the timecourse's own input universe (GSM8123451 > 0, 1,194,614 bins) the values are 0.8988, 0.8817, 0.6165, 0.6757 and 0.8507, 0.8248, 0.6084, 0.7262. The 24 h value at the FLAG set, 0.6790, is below Q122's 48 h F3 value of 0.7734 at the same set.

    What this evidence cannot show

    THE PRE-REGISTERED KINETIC CRITERIA ARE SUPPRESSED, NOT REPORTED: K0 failed on both sets, RK2 fired, and K1, K2 and K3 do not count. K1's arithmetic fired on a NEGATIVE increment, which is a defect in the criterion and is recorded as one. THE FOUR VTP TIMEPOINTS SHARE ONE 24 h VEHICLE, so the 4, 8 and 16 h contrasts are not exposure-matched to their control and a vehicle-duration effect on the mark would have to be assumed absent; sensitivity criterion K5 reports V(4 h) = 0.8996 and 0.8511, i.e. 31% and 55% of the 24 h effect is already present at 4 h, which is fast for a chromatin mark and is compatible with part of V being a batch offset. THE FIVE TRACKS SPAN 2.208-FOLD IN NON-ZERO BIN COUNT and Spearman(V, non-zero count) over the four timepoints is exactly -1.000, so the series is confounded with library enrichment; the minimum survives a common-support restriction and two rank statistics, so the confound explains the recovery's size and not its existence. THE TIMECOURSE DEPOSITS NO SGC0946 SAMPLE, so G cannot be measured in this batch and no F is computed here. n = 1 per timepoint, no replicate anywhere in the arm. 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 the mark's abundance; everything measures its distribution. n = 1 per antibody per condition per culture; no number carries a spread and none is precise. REAGENTS AND WHAT EACH CANNOT DISTINGUISH: H3K79me2 is ABCAM ab3594-100UG, 'Rabbit polyclonal to Histone H3 (di methyl K79) - ChIP Grade', read verbatim from the paper's key-resources sheet 41556_2025_1859_MOESM6_ESM.xlsx this session - a POLYCLONAL with no specificity document in the sheet, so it cannot distinguish me2 from me3, and it reads a MARK, so it cannot distinguish halted deposition from active demethylation from replicative dilution and cannot attribute any of them to the fusion. `antibod` occurs ZERO times in the 158 GEO records, so the spreadsheet is the only route to any reagent in this series. 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, and it cannot see a fusion whose tag has been proteolysed. VTP50469 disrupts an INTERFACE, so its H3K79me2 effect is at least four steps removed; SGC0946 inhibits DOT1L's CATALYSIS, so its effect says nothing about DOT1L occupancy. SYSTEM: murine MLL-AF9 (KMT2A::MLLT3) myeloid blood cells, NOT the infant B-lineage KMT2A::AFF1 disease this graph is about.

    Evidence that argues against this (2)
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Post hoc, five statistics computed on the same four timepoints give the 16-to-24 h step as a fraction of the 8-to-16 h step of 0.2295 (aggregate share over U_m at S_FLAG), 0.0470 (aggregate share over the common-support universe of 319,091 bins), 0.2339 (median per-bin share ratio at S_FLAG), 0.0994 (fraction of S_FLAG bins losing within-track rank) and 1.7116 (the enrichment residual), and the direction of that step is 'less drug effect' for the first four and 'more drug effect' for the fifth; the enrichment residual is +0.0265, +0.0213, -0.0543, -0.1837 across the four timepoints and its 24 h value lies inside the range -0.1584 to -0.2040 that the same quantity takes in four independent 48 h VTP-50469 cultures.

    What this evidence cannot show

    EVERY STATISTIC IN THIS ITEM IS POST HOC and none was pre-registered; the verdict sentence was written by the scoring script from the numbers rather than chosen afterwards, and the script's own conclusion is 'No plateau statement is supported'. The enrichment index C that the residual subtracts is NOT purely technical - a genuine redistribution lowers it too, demonstrated by the SGC0946 contrasts whose residuals are only -0.0043 to -0.0248 for a drug whose share-ratio effect is 0.09-0.15 - so the residual is a difference of two correlated quantities and not a decomposition. With n = 1 per timepoint the statistic-to-statistic spread is the only noise estimate available, and it is not a substitute for a replicate. The cross-culture concordance at 24 h compares one timepoint of one batch against four cultures of two other batches.

    Evidence that argues against this (1)
Connected to