Oncofusion

start / graph / ana-q130-the-timecourse-gradient-span-is-monotone-in-exposure-time-where-the-aggregate-that-q124-suppressed-is-not

Analysisweak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.Q130-does-the-eviction-gradient-exist-in-h3k79me2

In the VTP-50469 timecourse Q124 declared uninterpretable, the eviction-gradient SPAN grows monotonically 0.104 / 0.198 / 0.214 / 0.335 at 4/8/16/24 h with Spearman -1.000 at every timepoint - while the aggregate it was suppressed for is non-monotone and perfectly anti-correlated with library complexity. The span is better behaved, and it is NOT clean.

also called G6 · the timecourse gradient · the statistic that survives a suppressed arm

In a separate experiment following the drug over a day, the pattern deepens steadily the longer the drug is on. The overall average in the same experiment jumps around and was written off as unreliable. The pattern is steadier than the average, but it is not fully free of the same technical problem, and this entry says so.

The same thing, for a scientist

**Pre-declared as SUPPORTING ONLY, before it was computed.** Q124's DEVIATION 1 measured a **2.208-fold spread in non-zero bin count** across these libraries, its own criterion K0 **FAILED** (the aggregate V series is non-monotone with a +0.0609 upward step between 16 h and 24 h), and it **suppressed K1, K2 and K3** in consequence. The pre-registration therefore stated that no headline in this beat would rest on this arm in either direction [E1]. **What the arm says when stratified** - murine MLL-AF9 F3_TIMECOURSE, 500 nM VTP-50469 at 4/8/16/24 h against DMSO 24 h, scored at the same `S_FLAG` quartiles [E2]: | | Q1 | Q2 | Q3 | Q4 | Spearman | span | aggregate | |---|---:|---:|---:|---:|---:|---:|---:| | 4 h | 0.8697 | 0.8860 | 0.9102 | 0.9738 | -1.000 | -0.1041 | 0.8996 | | 8 h | 0.8191 | 0.8680 | 0.9104 | 1.0167 | -1.000 | -0.1976 | 0.8835 | | 16 h | 0.5438 | 0.6070 | 0.6521 | 0.7577 | -1.000 | -0.2139 | 0.6181 | | 24 h | 0.5626 | 0.6662 | 0.7269 | 0.8973 | -1.000 | **-0.3347** | 0.6790 | **Spearman -1.000 at all four timepoints, and the span grows monotonically with exposure.** The aggregate does not: it falls, falls, then RISES at 24 h, which is exactly Q124's K0 failure [E2]. **And here is the part this entry exists to state rather than hide.** The obvious next question is whether the span tracks the confound too, and it partly does [E3]: | | Spearman vs exposure hours | Spearman vs non-zero bin count | |---|---:|---:| | aggregate V | -0.800, **non-monotone** | **-1.000** | | gradient span | **-1.000, monotone** | -0.800 | Library complexity is itself **non-monotone** in time (789,820 / 973,841 / 1,107,216 / 986,679 non-zero bins at 4/8/16/24 h - it turns at 24 h). **The aggregate turns exactly where the complexity turns. The span does not.** But the span is still anti-correlated with complexity at -0.800, and at n = 4 the difference between -1.000 and -0.800 is one adjacent transposition. **G6 corroborates and does not establish, exactly as pre-declared.**

Go deeper
Why this is worth an entry even though it settles nothing

Because the statistic behaves differently from the aggregate under a known confound, and that is a fact about the statistic rather than about this arm. Q124 suppressed three criteria because the aggregate was non-monotone. The same libraries, the same bins, the same normalisation, stratified, give a monotone series. If the eviction-depth span is systematically less sensitive to library complexity than a share ratio over an unstratified set, that is worth knowing before the next beat suppresses an arm.

And it is worth knowing the other way round too. If a beat had run only the stratified analysis on this arm and reported a clean monotone time-response, it would have reported a result from an arm its own project had already declared batch-affected. That is why the suppression was written into the pre-registration before the numbers existed.

The 4 h point, and a caveat Q124 already raised

Q124's K5 noted that V(4h) = 0.8996 means a substantial share of the 24 h effect is already present at 4 hours, which is fast for a chromatin-mark redistribution and raises the possibility that part of V is a batch or vehicle-duration offset rather than drug time. The stratified view makes that reading less likely rather than more: at 4 h the gradient span is already -0.1041 with Spearman -1.000, so whatever is present at 4 h is already eviction-shaped, and a vehicle-duration offset has no reason to be. That is an argument, not a measurement, and it is offered as one.

Facts and figures
scriptanalyses/q130-does-the-eviction-gradient-exist-in-h3k79me2/run.py
statusrun
result summarySpearman -1.000 at all four timepoints; span -0.1041 / -0.1976 / -0.2139 / -0.3347 monotone in exposure; aggregate non-monotone; span vs non-zero bin count Spearman -0.800 against the aggregate's -1.000.
weightSUPPORTING ONLY, pre-declared
absence classmeasured-not-analysed
absence class noteQ124 binned all five timecourse tracks in cycle 83 and scored them as aggregates only, then suppressed three criteria when the aggregate series proved non-monotone. The stratified view of the same files was never computed.
absence class assigned2026-09-23 (Q130)
epigenetic readoutH3K79me2 eviction-depth gradient as a function of menin-inhibitor exposure time
per timepoint
4h-1.0004h: -18h-1.0008h: -116h-1.00016h: -124h-1.00024h: -1
inputs
[
  "GSM8123452 TC DMSO 24 h",
  "GSM8123449/450/447/448 TC VTP 4/8/16/24 h",
  "GSM8123451 TC input",
  "GSM8123359/GSM8123361 FLAG (define the quartiles)"
]
per timepoint — the numbers behind the chart
{
  "4h": {
    "R_by_quartile_Q1_to_Q4": [
      0.8697,
      0.886,
      0.9102,
      0.9738
    ],
    "spearman_vs_eviction_depth": -1,
    "span_Q1_minus_Q4": -0.1041,
    "R_aggregate_S_FLAG": 0.8996
  },
  "8h": {
    "R_by_quartile_Q1_to_Q4": [
      0.8191,
      0.868,
      0.9104,
      1.0167
    ],
    "spearman_vs_eviction_depth": -1,
    "span_Q1_minus_Q4": -0.1976,
    "R_aggregate_S_FLAG": 0.8835
  },
  "16h": {
    "R_by_quartile_Q1_to_Q4": [
      0.5438,
      0.607,
      0.6521,
      0.7577
    ],
    "spearman_vs_eviction_depth": -1,
    "span_Q1_minus_Q4": -0.2139,
    "R_aggregate_S_FLAG": 0.6181
  },
  "24h": {
    "R_by_quartile_Q1_to_Q4": [
      0.5626,
      0.6662,
      0.7269,
      0.8973
    ],
    "spearman_vs_eviction_depth": -1,
    "span_Q1_minus_Q4": -0.3347,
    "R_aggregate_S_FLAG": 0.679
  }
}
span series
[
  -0.1041,
  -0.1976,
  -0.2139,
  -0.3347
]
confound check
{
  "deviation": "DEVIATION 3 - is the timecourse span confounded by library complexity? POST HOC.",
  "hours": [
    4,
    8,
    16,
    24
  ],
  "nonzero_bin_counts_from_q124": [
    789820,
    973841,
    1107216,
    986679
  ],
  "nonzero_count_max_over_min_including_dmso": 2.2077,
  "nonzero_count_monotone_in_time": false,
  "span_series": [
    -0.1041,
    -0.1976,
    -0.2139,
    -0.3347
  ],
  "aggregate_series": [
    0.8996,
    0.8835,
    0.6181,
    0.679
  ],
  "spearman_span_vs_hours": -1,
  "spearman_aggregate_vs_hours": -0.8,
  "spearman_span_vs_nonzero_count": -0.8,
  "spearman_aggregate_vs_nonzero_count": -1,
  "span_monotone_in_time": true,
  "aggregate_monotone_in_time": false,
  "verdict": "THE SPAN IS NOT CLEAN OF THE CONFOUND AND THIS ENTRY SAYS SO. Library complexity is NON-monotone in time ([789820, 973841, 1107216, 986679], rising to 16 h and falling at 24 h). The AGGREGATE is non-monotone in time ([0.8996, 0.8835, 0.6181, 0.679]) and perfectly anti-correlated with library complexity at Spearman -1.0 - it turns exactly where the complexity turns, which is Q124's K0 failure and the reason it suppressed K1-K3. The SPAN is monotone in time at Spearman -1.0 and anti-correlated with complexity at -0.8. So the span is BETTER behaved than the aggregate - it does not turn where the complexity turns - but it is not independent of it, and with four points and no replicate the difference between -1.000 and -0.800 is one adjacent transposition. G6 REMAINS SUPPORTING ONLY, exactly as pre-declared, and this deviation exists so that the monotonicity is not quoted as though the confound had been excluded."
}
How this could bite us back
  • If the gradient deepens through 24 h and the 48 h NTvKO arm sits deeper still (span 0.6005), the effect has not plateaued by any timepoint measured, which means every VTP number in this beat and in Q124 is a lower bound and the true steady-state gradient is unknown. A pharmacodynamic sampling time chosen from this data would be chosen from a curve that has not levelled off.
Still open
  1. Is the eviction-depth span systematically less confound-sensitive than an unstratified share ratio? Two arms is not a demonstration; the claim is testable across every arm in this deposit and would change how future beats decide whether to suppress one.
  2. Why does the 24 h point's aggregate rise while its span deepens? If both came from the same libraries the two statistics must be reading different features of them, and which feature is not identified here.
  3. Does the timecourse culture even carry the same S_FLAG sites? S_FLAG is defined in the MLLAF9 Control FLAG culture and the timecourse is a third culture again, so the quartiles are imported into it rather than measured in it.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    Q124 measured a 2.208-fold spread in non-zero bin count across the GSE260742 timecourse H3K79me2 libraries, its criterion K0 failed on both scored sets with a +0.0609 upward step between 16 h and 24 h, and it suppressed K1, K2 and K3 - so this arm was declared batch-affected by the project before this beat scored it.

    What this evidence cannot show

    Read from the committed JSON. Q124's suppression was a pre-registered consequence of its own guard criterion RK2 firing, which is the correct behaviour and is the reason this entry is scoped as supporting only.

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

    In the GSE260742 murine MLL-AF9 timecourse, the H3K79me2 VTP/DMSO share ratio at the four eviction quartiles gives Spearman -1.000 at every one of 4, 8, 16 and 24 h, with spans growing monotonically at -0.1041 / -0.1976 / -0.2139 / -0.3347, while the aggregate over the same bins is non-monotone at 0.8996 / 0.8835 / 0.6181 / 0.6790.

    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. It cancels the unknown per-sample scale factor exactly and IS BLIND TO A UNIFORM CHANGE BY CONSTRUCTION: nothing here measures H3K79me2'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. REPLICATION, CORRECTED MID-BEAT: the NTvKO arm deposits ONE H3K79me2 ChIP per genotype per drug, so ITS spans carry no technical error bar and its bin-level bootstrap intervals measure sampling within one library, NOT library-to-library variation. A REPLICATE PAIR DOES EXIST for a different culture - GSM8123372/374 and GSM8123386/388, both titled MLLAF9 Control, the second carrying `rep2` - already binned by Q122 and scored in Q130 DEVIATION 4: the SIGN and the monotonicity reproduce (Spearman -1.000 in all four murine VTP contrasts in this deposit) while the SPAN reproduces only to within a factor of 1.99 (-0.2438 against -0.4847). EVERY SPAN IN THIS BEAT SHOULD BE READ WITH THAT 2-FOLD REPRODUCIBILITY. THE KNOCKOUT is one CRISPR guide, one clone C1, no second clone, no rescue, and NO deposited PCGF1 protein or transcript measurement. S_FLAG is defined by a DYKDDDDK-tag ChIP on a RETROVIRALLY OVER-EXPRESSED tagged MLLAF9 construct, so it is the over-expressed molecule's sites and not an endogenous fusion's. THE H3K79me2 ANTIBODY IS NOT RESOLVED - `antibod` occurs zero times in the 158 GEO records - and is recorded unresolved rather than guessed. SYSTEM: murine MLL-AF9 (KMT2A::MLLT3) myeloid blood cells, NOT the infant B-lineage KMT2A::AFF1 disease this graph is about. ARM-SPECIFIC AND DECISIVE: this arm was declared batch-affected by Q124 before this beat ran, its libraries differ 2.208-fold in non-zero bin count, the 4/8/16 h points share a single 24 h vehicle by design, and confidence is set to WEAK for that reason and not because the numbers are noisy.

    Evidence that argues against this (2)
    • A published methods study establishes that STANDARD ChIP-seq normalisation FAILS TO DETECT a genome-wide decrease in a histone mark under an inhibitor of its writer: standard methods see no H3K27me3 reduction under an EZH2 inhibitor, and a substantial reduction appears only with a Drosophila chromatin spike-in. Every number in this entry is exactly the kind of normalisation that paper shows to be blind. Egan B et al. 2016
    • The system is murine MLL-AF9 (KMT2A::MLLT3) myeloid cells carrying a retrovirally over-expressed FLAG-tagged construct. This graph exists for human infant B-lineage KMT2A::AFF1, and this corpus's commonest documented synthesis failure is generalising a murine MLL-AF9 myeloid result to that disease. Neville D et al. 2026
  3. E3weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.dataset

    The gradient span is better behaved than the aggregate under the known confound but is not free of it: non-zero bin count is non-monotone in time, the aggregate is non-monotone and anti-correlated with it at Spearman -1.000, and the span is monotone in time and anti-correlated with it at -0.800.

    What this evidence cannot show

    POST HOC and n = 4: at four points the difference between Spearman -1.000 and -0.800 is a single adjacent transposition, so this comparison is suggestive at best. Non-zero bin count is one proxy for library complexity and not the only one; sequencing depth and duplication rate are not recoverable from the deposited TDF files.

    Evidence that argues against this (1)
    • Choosing non-zero bin count as the confound proxy after seeing the result is itself a researcher-degree-of-freedom, and a different proxy could give a different comparison. Iqbal Lab oncofusion knowledge graph 2026
Connected to