Cycle 107's headline same-condition floor of 3.03% at D1 is a median over a MIXED population of two genotypes. Restricted to the non-targeting control genotype it is 10.49%, so the borrowed 10% floor cycle 107 declared 'too generous by 3.3x' was in fact about right there - cycle 107's own F2, not its F1.
also called restrict the null with the population · the subgroup reverses F1 · 3.03% is a mixture
A previous analysis averaged the repeat-to-repeat wobble across two different cell lines and got a small number, then used that small number to declare an older, larger estimate too lax. Splitting the average back into its two cell lines shows the older estimate was close to right for the normal cells, and the small number came almost entirely from the gene-knockout cells.
The same thing, for a scientist
**Cycle 107's `verdict_floor` reads:** *"F1 - THE BORROWED 10% FLOOR WAS TOO GENEROUS AT D1: the measured same-condition H2AK119ub floor there is 3.03% ... Cycle 104's M1 excluded effects up to 10% at a stratum whose real floor is 3.3x smaller."* That 3.03% is `median([10.49, 3.03, 1.89, 9.06, 0.40])` over five conditions drawn from **two genotypes** - the non-targeting control and the PCGF1-KO clone C1. Splitting: | population | D1 values | median | |---|---|---:| | **non-targeting control** | 10.49, 10.66*, 3.03 | **10.49%** | | **PCGF1-KO clone C1** | 1.89, 9.06, 0.40 | **1.89%** | | pooled, as published | the five cycle 107 scored | 3.03% | *\* NT_SGC, recovered here; cycle 107's universe gate aborted it.* **Against the control genotype the borrowed 10% is 10.49% - cycle 107's own F2 band ('about right'), not its F1.** The verdict that has propagated is a property of a mixture. **Two independent routes reach the same correction, which is what makes it more than a re-slicing.** (1) The genotype split above. (2) The recovered NT_SGC condition, which cycle 107 never saw: its D1 is **10.66%**, and it lands with the other two control-genotype conditions rather than with the knockout ones. A recovered observation that falls on the side the subgroup analysis predicts is a check on the split, not another slice of it.
Why the mixture pulls down, mechanically
The PCGF1-KO clone is a single CRISPR clone - one guide, no second clone, no rescue. A clonally derived line replicated twice will agree with itself more closely than two independent cultures of a polyclonal control will, and that is what the numbers show: the knockout arm's floor is 5.5x tighter than the control arm's at D1. Pooling them and taking a median puts three knockout conditions into a five-point median, so the median sits in the knockout's range.
This is the shape the project already has a rule for - restrict the null with the population. Here the restriction runs the other way from the usual case: it makes the floor rise, and a floor that rises makes previously-passing claims fail.
What survives of cycle 107, and it is most of it
Survives unchanged: that GSE260742 contains 17 genuine same-condition replicate groups and that the five-cycle claim it had none was false; that the floor is strongly stratum-dependent (its P1, confirmed again here - TOP is 1.9x to 5.3x D1 depending on arm); that the paired/rep2 pair is processing-confounded (its U1, which is why this beat ran).
Does not survive: the F1 verdict as a general statement, and the specific sentence 'a stratum whose real floor is 3.3x smaller'. In the control genotype it is 1.05x, and in the MLLAF9 Control arm the borrowed 10% is 3.7x too TIGHT.
THE RENORMALISATION BOUND, AND IT APPLIES TO EVERY NUMBER IN THIS ENTRY. Within-sample shares sum to 1 and GSE260742 has NO USABLE SPIKE-IN (Q117 verified zero hits for the entire spike-in vocabulary across all 419,824 characters of its 158 sample records, and no scale factor is deposited), so a purely local change leaks into every other stratum by arithmetic. A FLOOR COMPUTED THIS WAY CANNOT DISTINGUISH 'the bulk moved' from 'the peaks moved and the shares renormalised'. A TIGHTER NUMBER IS NOT THEREBY A CLEANER ONE.
| cycle107 published d1 PCT | 3.03 |
|---|---|
| cycle107 published top PCT | 15.97 |
| control genotype d1 PCT | 10.49 |
| knockout genotype d1 PCT | 1.89 |
| borrowed floor cycle104 M1 PCT | 10 |
| verdict reversal | F1 -> F2 within the control genotype |
recovered condition
{
"id": "NT_SGC",
"paired_gsm": "GSM8123405",
"rep2_gsm": "GSM8123421",
"TOP_pct": 19.49,
"D1_pct": 10.66,
"why_cycle107_dropped_it": "its universe gate aborted at 93.88%; the amended gate admits it at 98.08% genuine track agreement"
}- Is the knockout arm's tighter floor a clonality effect? Binning a second independent clone would answer it, and the deposit does not contain one.
- Which of cycles 100-107's verdicts change sign if scored against 10.49% rather than 3.03%? That re-scoring has not been done and is the obvious next beat.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Re-scoring cycle 107's own GSE260742 NTvKO `paired`/`rep2` pairs with an identical statistic reproduces all five of its published per-condition TOP/D1 values exactly and recovers a sixth (NT_SGC, 19.49%/10.66%). Restricted to the non-targeting control genotype the D1 floor is 10.49% (n=3); restricted to the PCGF1-KO clone it is 1.89% (n=3); the published pooled median is 3.03% (n=5).
What this evidence cannot show
SYSTEM: murine MLL-AF9 (KMT2A::MLLT3) myeloid blood cells, Mus musculus, mm10, one lab, one deposit - NOT the infant B-lineage KMT2A::AFF1 disease this graph is about. n = 3 conditions per arm, one cell system; the spread across conditions is reported so stability is visible and IS NOT A CONFIDENCE INTERVAL. THE RENORMALISATION BOUND, AND IT APPLIES TO EVERY NUMBER IN THIS ENTRY. Within-sample shares sum to 1 and GSE260742 has NO USABLE SPIKE-IN (Q117 verified zero hits for the entire spike-in vocabulary across all 419,824 characters of its 158 sample records, and no scale factor is deposited), so a purely local change leaks into every other stratum by arithmetic. A FLOOR COMPUTED THIS WAY CANNOT DISTINGUISH 'the bulk moved' from 'the peaks moved and the shares renormalised'. A TIGHTER NUMBER IS NOT THEREBY A CLEANER ONE. n=3 per genotype. The two genotypes also differ in clonality (the KO arm is one CRISPR clone, the control is not), so 'genotype' here carries a clonality term that this entry does not separate - and clonality is the more likely mechanical cause.
Evidence that argues against this (2)
the arm span is the general statement; this is its effect on the one number that has propagated
the knockout clone whose three conditions pull the pooled median down
the amended gate is what recovers cycle 107's dropped NT_SGC condition