Oncofusion

start / graph / ana-d1-is-not-resolvable-from-gse260742-and-three-published-verdicts-were-measuring-the-ruler

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.

Scoring cycle 104's mark panel against a floor measured in its OWN arm: 9 of 10 mark-by-arm cells at D1 land inside the floor's own spread, so D1 is not resolvable from this deposit at all - and the same mark ratios give 0, 4 or 4 breaking marks depending only on which floor was imported, so cycle 104's pass, cycle 107's F1 and the GSE299100 beat's failure were all measuring the ruler rather than the marks

also called D1 is not resolvable · three verdicts, one ruler · the floor spread swamps the effect · stop scoring D1 from GSE260742

To say a chemical mark changed, you need to know how much two repeats of the same experiment differ - the noise floor. Three previous rounds of this project each picked a different floor and reached a different conclusion about the same measurements. Measuring the floor separately for each arm of the experiment shows why: within a single arm the floor itself varies as much as the effects being measured. The honest answer is that this dataset cannot settle the question, and the disagreement was never about biology.

The same thing, for a scientist

**Statement.** Cycle 104 asked whether any chromatin mark moves outside the fusion's peaks under a menin inhibitor and answered **no** (`M1`), scoring every mark against a borrowed global **10%**. Two later beats attacked that number from opposite directions — GSE299100 measured **2.42%** and concluded `M1` fails; cycle 107 measured **3.03%** and concluded the borrowed floor was 3.3× too generous. Cycle 110 showed cycle 107's number was a median over two genotypes whose floors differ 5.5×, and Q135 showed the floor spans **~20× across arms of one deposit**. **So each arm was scored against its own floor. Zero downloads; nothing recomputed from bins.** **Arm floors, from same-condition H2AK119ub replicate pairs:** | arm | D1 median | D1 spread | D10 median | D10 spread | |---|---:|---|---:|---| | NT, non-targeting control | 6.76% | **3.03–10.49** | 11.09% | 9.25–12.92 | | C1, PCGF1-KO clone | 1.89% | **0.40–9.06** | 1.73% | 1.18–13.21 | **The pre-registered `NEITHER` branch fired. `A1`, `A2` and `A3` are all wrong.** At D1, **9 of 10 mark × arm cells land INSIDE the floor's own spread** — neither above it nor below it. **The spread within a single arm is as large as the effects being scored.** **And the three published verdicts differ only in which point estimate they imported.** The same mark ratios, in the NT arm alone: | floor used | source | marks breaking | |---|---|---:| | 10.00% | cycle 104, borrowed | **0** | | 3.03% | cycle 107, median over both genotypes | **4** | | 2.42% | GSE299100, different deposit and cell line | **4** | **None of them was measuring the marks. All three were measuring the ruler.** **D10 is a different story and is resolvable.** There, two marks exceed even the widest condition floor in their own arm — **H3K79me2 at 16.17% against a 12.92% maximum in NT**, and **H3K4me3 at 14.47% against 13.21% in C1**. Different marks, different arms. `M1` fails at D10 in both arms, which no previous scoring reported, because all of them were arguing about D1.

Go deeper
What should now be retired

Three verdicts, all superseded by not answerable:

  • cycle 104's M1 pass at D1 — obtained with a floor 1.5× wider than its own arm's median and 3.3× wider than the other arm's;
  • cycle 107's F1 (the borrowed floor was 3.3× too generous) — already corrected in cycle 110 for mixing genotypes, and now void for D1 entirely;
  • the GSE299100 beat's conclusion that M1 does not survive — its 2.42% is a real measurement in MV4-11 with MEN1/MLL1 antibodies, imported into a murine MLL-AF9 H2AK119ub comparison.

None of the three was careless. Each measured a floor honestly and applied it. The defect is structural: a single floor was treated as a property of the assay when it is a property of the arm, and no beat could see that until three arms had been measured separately.

What survives untouched is every claim about the TOP stratum, where the effects are large relative to even the widest floor, and every claim of relative ordering between strata, which a share reports faithfully.

Why the D10 result is the one worth keeping

At D10 the marks that break are different in the two arms — H3K79me2 in the non-targeting control, H3K4me3 in the PCGF1-KO clone. That is not what a shared artefact looks like: a floor that was simply too tight would break the same marks everywhere.

It is also the opposite of the stratum everyone argued about. Three cycles of dispute concentrated on D1, the densest bulk decile, and the resolvable signal is in D10, the sparsest — where this project has repeatedly warned that ratios are unstable. The warning and the result now point at the same rows, and separating them needs the per-mark floors Q135 began rather than more argument about D1.

Facts and figures
questiondoes any mark move at D1, scored against a floor from its own arm?
downloads0
recomputed from binsfalse
D1 cells undecided9
D1 cells total10
what survivesevery claim about the TOP stratum, where effects are large against even the widest floor, and every claim of relative ordering between strata
absence classmeasured-not-curated
bite backThis entry tells the project to stop scoring D1 from GSE260742. If the arm floors are wide only because n=2 and n=3, then more conditions would narrow them and D1 would become resolvable after all - in which case this entry will have shut down a line of enquiry on the strength of a small sample, which is the same error it accuses three previous verdicts of making.
same ratios different rulers NT arm
floor_10.00_cycle1040.000floor_10.00_cycle104: 0floor_3.03_cycle1074.000floor_3.03_cycle107: 4floor_2.42_GSE2991004.000floor_2.42_GSE299100: 4
arm floors PCT
{
  "NT": {
    "D1_median": 6.76,
    "D1_spread": [
      3.03,
      10.49
    ],
    "D10_median": 11.09,
    "D10_spread": [
      9.25,
      12.92
    ],
    "n_conditions": 2
  },
  "C1": {
    "D1_median": 1.89,
    "D1_spread": [
      0.4,
      9.06
    ],
    "D10_median": 1.73,
    "D10_spread": [
      1.18,
      13.21
    ],
    "n_conditions": 3
  }
}
D10 breaks
{
  "NT": {
    "mark": "H3K79me2",
    "ratio": 0.8383,
    "dev_pct": 16.17,
    "arm_max_floor": 12.92
  },
  "C1": {
    "mark": "H3K4me3",
    "ratio": 1.1447,
    "dev_pct": 14.47,
    "arm_max_floor": 13.21
  }
}
same ratios different rulers NT arm — the numbers behind the chart
{
  "floor_10.00_cycle104": 0,
  "floor_3.03_cycle107": 4,
  "floor_2.42_GSE299100": 4
}
verdicts superseded
[
  "cycle 104's M1 pass at D1",
  "cycle 107's F1 - the borrowed floor was 3.3x too generous",
  "the GSE299100 beat's conclusion that M1 does not survive"
]
Still open
  1. Measure per-mark, per-arm floors. Q135 measured H3K79me2's and found it 4.1x tighter than H2AK119ub's in the same arm; every verdict here transfers an H2AK119ub floor across marks and that is the single largest remaining defect.
  2. Re-derive cycle 107's per-condition floors with the corrected universe gate. The NT arm here rests on two conditions because the defective gate refused a third; with it restored the NT spread and therefore every NT verdict changes.
  3. Is the D10 signal real or sparse-bin noise? Two marks break in two arms, but they are different marks, and low-coverage instability is the standing explanation. A per-mark floor at D10 would separate them.
  4. How many conditions would it take to make D1 resolvable? The spreads here come from n=2 and n=3; the deposit has 17 same-condition replicate groups, and most have never been scored.
Every source behind this entry (1)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    Deriving arm-specific floors from the per-condition same-condition H2AK119ub replicate floors gives, at D1, a median of 6.76% (spread 3.03-10.49, n=2 conditions) for the NT non-targeting-control arm and 1.89% (spread 0.40-9.06, n=3) for the C1 PCGF1-KO arm; at D10, 11.09% (9.25-12.92) and 1.73% (1.18-13.21). Scoring cycle 104's VTP/DMSO mark ratios against those spreads, 9 of 10 mark-by-arm cells at D1 fall inside the spread and are therefore undecided. At D10, H3K79me2 in NT (ratio 0.8383, |r-1| 16.17%) and H3K4me3 in C1 (ratio 1.1447, 14.47%) exceed the widest condition floor in their own arm. Applying instead a single imported floor to the NT arm gives 0 marks breaking at 10.00%, 4 at 3.03% and 4 at 2.42%.

    What this evidence cannot show

    THE ARM FLOORS REST ON n=2 AND n=3 CONDITIONS and one of them spans 0.40-9.06%, so 'the spread' is barely an estimate - with so few conditions the spread is as likely to reflect which three conditions happened to be scoreable as any property of the arm. THE FLOOR IS AN H2AK119ub FLOOR APPLIED TO H3K27me3, H3K4me3 AND H3K79me2, a transfer across marks that Q135 showed is wrong in a known direction: H3K79me2's own floor is 4.1x tighter, which would make its D10 break MORE significant and says nothing about the others. Cycle 107's per-condition floors, which this rescoring consumes, were computed with a universe gate later found defective; the gate affected WHICH conditions entered, so the NT arm here rests on two conditions where a corrected gate admits three. Every ratio scored is a within-sample share, so all of this concerns relative behaviour and none of it licenses a statement about absolute gain or loss.

    Evidence that argues against this (3)
    • Declaring D1 unresolvable is the most convenient possible outcome for a coordinator who has now published two wrong verdicts about it, and it should be read with that in mind. A sceptic would note that 'inside the spread' is a criterion I chose after seeing that the spreads were wide, and that a defensible alternative - scoring against the arm MEDIAN - gives a clean answer: 2 of 5 marks break in NT and 4 of 5 in C1. That analysis is in the results file and it does not say 'unresolvable'.
    • The D10 result is carried by the sparsest decile, which this project has repeatedly and correctly described as unstable. Two marks exceeding a floor there is exactly what one would expect from noise in low-coverage bins, and the fact that they are different marks in the two arms is equally consistent with independent noise as with genuine arm-specific biology. Nothing here distinguishes those.
    • This entry retires three verdicts and replaces them with none, which leaves the menin-interface thread with less evidence than it had yesterday and no new experiment. That is the honest position but it is also a way of never being wrong again. The constructive version - measure per-mark, per-arm floors and re-score properly - is named in the open questions and was not done here.
Connected to