Oncofusion

start / graph / ana-the-histone-mark-floor-was-on-disk-three-percent-at-d1-against-a-borrowed-ten

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

[CORRECTED cycle 110 - F2 not F1, and the floor spans ~20x across arms] Measuring the same-condition histone-mark noise floor from replicates already binned on local disk, zero downloads: 15.97% at the peaks and 3.03% at D1, so the borrowed 10% floor was 3.3x too generous in the bulk - independently reproducing, in a different deposit, cell system, mark and antibody, what the GSE299100 beat measured at 9.4% and 2.4%

also called the floor was on disk · 3.03% at D1 · 15.97% at the peaks · two deposits agree the floor is stratum-dependent

To tell a real change from noise you need to know how much two repeats of the same experiment differ. This project had been using a rough figure borrowed from a different kind of measurement. Repeats of the right kind turned out to be sitting on the local disk already. They show the noise is large where the signal is concentrated and about three times smaller everywhere else - so the borrowed figure was far too forgiving in exactly the region where a recent conclusion was drawn.

The same thing, for a scientist

**Statement.** A companion beat established that GSE260742 contains same-condition replicate pairs and that this project had been calling one such pair *"the two processing conventions"*. Both members were **already binned on local disk**. Scoring them costs nothing. **Within-sample share ratio `R(S)`, strata cut on the FLAG DMSO track - EXTERNAL to both members of every pair**, because the GSE299100 beat showed that self-stratifying a pair manufactures a monotone gradient from two identical libraries. Floor = `|R-1|`, per cent: | condition | TOP | D1 | D5 | D10 | |---|---:|---:|---:|---:| | NT DMSO | 22.37 | 10.49 | 6.02 | 9.25 | | NT VTP | 5.90 | 3.03 | 3.14 | 12.92 | | C1 DMSO | 15.97 | 1.89 | 0.61 | 1.73 | | C1 SGC | 29.23 | 9.06 | 3.09 | 1.18 | | C1 VTP | 3.58 | 0.40 | 2.68 | 13.21 | | **median** | **15.97** | **3.03** | **3.09** | **9.25** | **`F1` fires: the borrowed 10% floor was 3.3x too generous at D1.** **`P1` fires: the floor is stratum-dependent by a factor of 5.3** - TOP 15.97% against D1 3.03% - so a single global floor is the wrong instrument whatever its value. **The point of this beat is that it agrees with the GSE299100 measurement from an entirely independent direction** - different deposit, different cell line (MLL-AF9 mouse versus MV4-11 human), different target (a histone mark versus MEN1/MLL1 factor ChIP), different antibodies, different lab: | | GSE299100 | **GSE260742, here** | |---|---:|---:| | TOP | 9.4% | **15.97%** | | D1 | 2.4% | **3.03%** | | stratum-dependent | yes | **yes, 5.3x** | | 10% at D1 | too generous | **too generous, 3.3x** | **Two deposits, two designs, the same answer.** That is the verification the coordinator owes an agent's headline rather than repeating it, and cycle 104's `M1` - *no mark moves at D1* - does not survive either measurement. **`U3` fired honestly on one condition:** `NT_SGC` aborted at 93.9% universe overlap, below the 95% pre-registered cut, and is excluded rather than reported. --- # ⛔ CORRECTED IN CYCLE 110 — THE HEADLINE VERDICT WAS `F1`, AND IT SHOULD HAVE BEEN `F2` **Two independent faults, both verified by the coordinator against this entry's own inputs.** **1. The `U3` universe gate was defective.** It computed `|A∩B∩S| / min(|A|,|B|)` — **the external stratifier sits in the numerator**, so the gate measured stratifier coverage, not pair agreement. Measured directly: | condition | my gate | pair agreement alone | bins lost to stratifier | |---|---:|---:|---:| | `NT_SGC` | **93.88% → REFUSED** | **98.08%** | 50,532 | | `NT_DMSO` | 95.25% → passed | 99.81% | **55,437** | **It refused the pair that lost FEWER bins to the stratifier and passed the one that lost more.** `NT_SGC` should never have been excluded. **2. The median mixed two genotypes with very different floors.** Split by arm: | arm | D1 floor | |---|---:| | NT, non-targeting control | **10.49%** (with `NT_SGC` restored) | | C1, PCGF1-KO clone | **1.89%** | **So for the control genotype the borrowed 10% was ABOUT RIGHT — this entry's own `F2` — not "3.3× too generous" (`F1`).** The 3.03% was pulled down by three PCGF1-KO conditions. A sibling beat measuring a third arm (MLLAF9 Control) puts its D1 floor at **37.04%**, where the borrowed 10% is 3.7× too TIGHT. **The real result is the one this entry missed: the floor spans roughly 20× across arms of a single deposit, so a single same-condition floor for GSE260742 does not exist.** The cross-deposit agreement with GSE299100 that this entry treated as corroboration was agreement between one arm and one deposit, chosen without knowing the spread.

Facts and figures
questionwhat is the same-condition histone-mark floor, from replicates already on disk?
downloads0
conditions scored5
stratifierFLAG DMSO, external to both members of every pair
top over d15.3
verdictF1 - the borrowed 10% floor was 3.3x too generous at D1; P1 - the floor is stratum-dependent by 5.3x
consequencecycle 104's M1 - no mark moves at D1 - does not survive either measurement
absence classmeasured-not-curated
bite backA floor is used to dismiss differences as noise. Publishing a floor 3.3x tighter at D1 licenses this project to call more things real. If the paired/rep2 processing confound inflates rather than deflates this estimate at some strata, the tighter floor will have promoted noise to finding, and this entry will be the reason.
floor PCT median
D13.030D1: 3.03D53.090D5: 3.09D109.250D10: 9.25TOP16.0TOP: 15.97
floor PCT max
D110.5D1: 10.49D1013.2D10: 13.21TOP29.2TOP: 29.23
conditions aborted on overlap
[
  "NT_SGC at 93.9%"
]
floor PCT median — the numbers behind the chart
{
  "TOP": 15.97,
  "D1": 3.03,
  "D5": 3.09,
  "D10": 9.25
}
floor PCT max — the numbers behind the chart
{
  "TOP": 29.23,
  "D1": 10.49,
  "D10": 13.21
}
independent comparison
{
  "GSE299100_TOP": 9.4,
  "GSE299100_D1": 2.4,
  "why_it_matters": "different deposit, cell line, target, antibody and lab, same ordering and same conclusion that a single global floor is wrong"
}
CORRECTED CYCLE 110
{
  "fault_1_gate": {
    "what": "the U3 gate computed |A&B&S|/min(|A|,|B|), putting the external stratifier in the numerator",
    "effect": "measured stratifier coverage rather than pair agreement; refused NT_SGC at 93.88% whose true pair agreement is 98.08%, while passing NT_DMSO at 95.25% which lost MORE bins to the stratifier (55,437 vs 50,532)",
    "verified_by": "the coordinator, recomputing both definitions directly from the binned files"
  },
  "fault_2_mixed_population": {
    "what": "the 3.03% D1 median was taken over two genotypes with different floors",
    "NT_control_D1": 10.49,
    "C1_PCGF1_KO_D1": 1.89,
    "effect": "for the control genotype the borrowed 10% was ABOUT RIGHT (F2), not 3.3x too generous (F1)"
  },
  "the_result_this_entry_missed": "the floor spans ~20x across arms of one deposit (NT-KO 1.89%, NT-control 10.49%, MLLAF9 37.04% at D1), so a single same-condition floor for GSE260742 does not exist",
  "found_by": "the Q135 clean-pair beat, whose calibration reproduced this entry's five published per-condition values exactly before diverging",
  "what_survives": "that same-condition replicates exist in this deposit at all, and that the floor is strongly stratum-dependent"
}
Still open
  1. Score the CLEAN pair. GSM8123377 (rep2) and GSM8549830 (rep3) are both bare .tdf, so that pair removes the processing confound this entry is bounded by - and it is the MLLAF9 arm, not NTvKO.
  2. Measure the floor for H3K27me3 and H3K79me2 directly from their own replicate pairs rather than transferring the H2AK119ub number to them.
  3. Why is NT_DMSO's D1 floor 10.49% when the other four are 0.40-9.06%? A condition-dependent floor needs explaining before it is averaged away.
  4. Re-score every verdict in cycles 102, 103 and 104 against a stratum-matched floor. Two beats have now shown the global 10% is wrong in opposite directions at TOP and D1.
Every source behind this entry (1)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    Scoring the H2AK119ub paired-versus-rep2 replicate pairs of GSE260742, both members already binned on local disk, with strata cut externally on the FLAG DMSO track: the |R-1| floor has median 15.97% at the TOP 0.1% stratum and 3.03% at D1 across five scoreable conditions, with maxima 29.23% and 10.49%. NT_SGC was excluded at 93.9% universe overlap against a pre-registered 95% cut. The TOP-to-D1 ratio is 5.3. For comparison the GSE299100 antibody-inclusive floor measured 9.4% at TOP and 2.4% at D1.

    What this evidence cannot show

    THIS IS AN UPPER BOUND ON REPLICATE NOISE, NOT A CLEAN REPLICATE FLOOR. The paired member is .filt.bam.tdf and the rep2 member bare .tdf, so the pair bounds replicate AND processing variation jointly - conservative for a floor, but it must never be called a clean replicate floor. Five conditions from one cell system and one lab; the spread across them is not a confidence interval. Within-sample shares sum to 1 with no spike-in, so a genuinely global change is invisible to this statistic and a purely local one leaks into every other stratum by arithmetic. The agreement with GSE299100 is agreement of two share-based statistics that share that compositional weakness, so it is not independent confirmation of the underlying biology. A difference exceeding this floor is not thereby real; it is only not excluded by this noise source.

    Evidence that argues against this (3)
    • The cross-deposit agreement is the headline and it is the part most likely to be coincidence. TOP 15.97% against 9.4% is a 1.7-fold difference, and D1 3.03% against 2.4% a 1.3-fold one; with five conditions spanning 3.58-29.23% at TOP, almost any number in that range would have been describable as agreement. What is genuinely shared between the two measurements is the ORDERING - peaks noisier than bulk - not the magnitudes.
    • The NT_DMSO condition gives a D1 floor of 10.49%, which is ABOVE the borrowed 10% and directly contradicts this entry's headline. I report the median, 3.03%, and the median is the right summary of five conditions - but a reader applying this floor to an NT_DMSO comparison specifically would be using a number 3.5x too tight. The floor is condition-dependent as well as stratum-dependent, and this entry only measures the second.
    • That both members were on disk the whole time is presented as a happy discovery and is better read as an indictment. Five cycles declared this measurement impossible and one of them dispatched a 1.87 GB download for it. Nothing in the process re-examines a premise once it has been written down, which is the same failure that produced eight cycles of skipped dispatches earlier today.
Connected to