The eviction quartiles are defined in the 'MLLAF9 Control' FLAG culture and the H3K79me2 they stratify is measured in the 'NTvKO Control' and 'NTvKO PCGF1 KO' cultures - zero shared libraries - so the commonest way a stratified analysis fools itself is structurally unavailable here
also called the third-culture argument · the stratifier shares no library with the measurement · why this is not a self-fulfilling split
The groups of DNA stretches were defined using one experiment and the chemical tag was measured in a different experiment on different cells. So a quirk of any single sequencing run cannot have created both the grouping and the pattern inside it.
The same thing, for a scientist
**The failure this rules out.** When a set is stratified by a ratio measured in track A and then scored on a ratio measured in track A', any noise, depth or duplication feature shared by A and A' produces a gradient for free. It is the same family as regression to the mean but worse, because the strata are *defined* by the shared quantity. **Verified from the deposit's own records, not asserted** [E1]: | role | samples | !Sample_title | |---|---|---| | defines `S_FLAG` and its quartiles | GSM8123359, GSM8123361 | *MLLAF9 Control - DMSO FLAG* / *- VTP50469 FLAG* | | the measurement being stratified | GSM8123438, GSM8123440 | *NTvKO Control - DMSO / VTP50469 H3K79me2* | | the second genotype | GSM8123435, GSM8123437 | *NTvKO PCGF1 KO - DMSO / VTP50469 H3K79me2* | | the timecourse | GSM8123452, GSM8123447-450 | *Timecourse Control - ...* | **Zero libraries in common between the stratifier and any measurement**, and Q124 recorded the same fact in its own words when it wrote that `S_FLAG` is *'a set defined in neither arm's culture'* [E1][E2]. **What it does NOT rule out.** The three cultures share a genome, a cell line background, a laboratory and a sequencing facility. Anything that makes the fusion's strongest sites behave distinctively - mappability, copy number, chromatin accessibility - is common to all of them, and this beat measured one such thing: **all three deposited inputs carry their own monotone gradient across the same four quartiles** [E3]. The third-culture argument defeats a shared-*library* artefact, not a shared-*genome* one. The matched null is the control for the latter.
Why this is worth stating separately from the matched null
They defeat different things and this corpus has conflated them. The matched null answers are these bins special? by finding bins elsewhere with the same baseline. The third-culture fact answers did the same sequencing run define the groups and fill them in?, which no null can answer because a null drawn from the same libraries inherits exactly that defect.
A stratified analysis needs both, and the cheap one - reading four !Sample_title lines - is the one nobody runs.
| absence class | measured-not-curated |
|---|---|
| absence class note | The sample titles that settle this were in the GEO records from 2025-11-23 and in this project's own stored copy since Q117 in cycle 68. The structural argument they support had never been written down, even though Q124 had noticed the underlying fact and quoted it for another purpose. |
| absence class assigned | 2026-09-23 (Q130) |
stratifier is a third culture
{
"S_FLAG_and_quartiles_from": [
"GSM8123359 MLLAF9 Control - DMSO FLAG",
"GSM8123362 MLLAF9 Control - VTP50469 FLAG"
],
"h3k79me2_measured_in": [
"GSM8123438/439/440 NTvKO Control",
"GSM8123435/436/437 NTvKO PCGF1 KO"
],
"shared_libraries": 0,
"consequence": "no library is shared between the thing that DEFINES the strata and the thing MEASURED in them, so a shared-library artefact cannot produce the gradient. Verified by reading GSM8123359's own SOFT record, not asserted.",
"verified_from": "C:/Temp/q117/q117_GSM8123359.txt, !Sample_title"
}- Does this argument survive for Q117's own within-culture statistics? Q117's FLAG gain analysis defines and scores in the same libraries by necessity, and whether that is a problem there has not been asked.
- Is there a deposited DMSO-versus-DMSO H3K79me2 contrast anywhere in GSE260742 that would give a direct selection-only baseline at S_FLAG? Cycle 83 flagged the absence of one for the MLL1 channel and it has not been checked for H3K79me2.
the structural control the matched null cannot supply
the two controls defeat different failure modes
the shared-genome artefact the third-culture argument cannot defeat