The selection control, run at last: two biologically identical vehicle-treated input libraries produce a MONOTONE decile gradient from 0.9040 to 1.2349 - wider than the real FLAG gradient cycle 103 called a finding - when the strata are cut on the pair itself, and are flat within 0.73% and non-monotone when the strata are cut externally. Monotonicity across deciles is therefore NOT evidence against a coverage artefact. Cycle 103 is safe because it cut externally
also called the selection control · self-stratification manufactures a gradient · monotone is not evidence · the cycle-83 control
A common way to analyse this kind of data is to sort the genome into bins by how much signal they carry, then look for a trend across the sorted bins. This shows that if you do the sorting using the very samples you are comparing, you get a clean, steady trend out of two samples that are biologically identical - bigger than the trend a real biological effect produced. Sorting with a third, independent sample removes it completely.
The same thing, for a scientist
**Cycle 83 first asked for this control, and Q125, Q130 and Q131 all reported they could not run it.** It needs two biologically equivalent samples, which GSE260742 and GSE127507 could not supply. GSE299100 can: `shLuc #1` and `shLuc #2`, both DMSO 72 h, both sequenced 230927, uniform `_hg38_rmdup.tdf`, in MV4;11. **Result, on the input pair (2 kb hg38 bins, blacklist removed, within-sample share):** | stratum | strata cut on **the pair itself** | strata cut on an **external** reference | |---|---:|---:| | TOP 0.1% | 0.9858 | 1.0057 | | D1 | **0.9040** | 0.9933 | | D2 | 0.9164 | 0.9947 | | D3 | 0.9460 | 0.9985 | | D4 | 0.9744 | 0.9978 | | D5 | 0.9997 | 0.9999 | | D6 | 1.0338 | 1.0009 | | D7 | 1.0667 | 1.0032 | | D8 | 1.1073 | 1.0048 | | D9 | 1.1626 | 1.0038 | | D10 | **1.2349** | 1.0073 | | | **MONOTONE, span 0.331, max deviation 23.5%** | **not monotone, max deviation 0.73%** | **Two identical samples, one analysis choice, a 32-fold difference in apparent effect.** **And the manufactured gradient is larger than the real one it could be mistaken for.** Cycle 103's headline was a monotone FLAG gradient from **0.810 to 1.136** across the same ten deciles - span **0.326** - which it correctly called a finding. The artefact produced here spans **0.331**, in the same direction, with the same monotone shape. **The mechanism, and why the pre-registration got it wrong.** This analysis pre-registered a *symmetric* internal cut, on the reasoning that strata cut on the MEAN of the two arms cannot favour either. **That is only true when the two arms have equal measurement noise.** These two input libraries differ **1.74x** in total coverage (9.516e9 against 5.467e9 coverage bases). Selection on the mean preferentially picks bins where the *noisier* (shallower) arm happened to fluctuate up, which pushes the ratio below 1 at the top and above 1 at the bottom - a monotone gradient, from nothing. Recorded as `DEVIATION 1`. **A one-arm cut is worse in the obvious way.** Cutting the top 0.1% on `shLuc-1` alone gives **1.0731** there against the unbiased **1.0057** - a **6.7-point winner's-curse inflation** in favour of whichever arm defined the set. **Cycle 103 passes.** Its strata came from the q117 parental FLAG DMSO track, external to both arms it scored, and on that design the input pair here is flat and non-monotone. **Its `U1` verdict stands and its reasoning was right.** What does not stand is the general inference - *neither input pair is monotone, so monotonicity is evidence of biology* - because monotonicity is exactly what the wrong cut produces.
Does the artefact scale with the depth imbalance? Suggestive at n = 3
| pair | depth ratio sh1/sh2 | max deviation, external cut | max deviation, self cut | monotone when self-cut |
|---|---|---|---|---|
| input | 1.743 | 0.73% | 23.49% | yes |
| MLL1 | 1.188 | 8.83% | 6.68% | no |
| MEN1 | 1.075 | 9.37% | 11.31% | no |
The input pair has both the largest depth imbalance and by far the largest self-cut artefact, and it is the only one that goes monotone. That is a suggestive pattern at n = 3 and not a measured relationship. The two antibody pairs are not clean tests of it, because their external-cut floors are already ~9% at the peaks from IP variability, which swamps a selection term of a few points.
The rule this yields, stated so it can be applied without re-deriving it
Never define a stratum using a sample you then score inside it. Not the sample, and not the average of the two samples being compared. Use a reference track that contains neither arm.
When no external reference exists, the honest options are to report the artefact's size alongside the result (compute both cuts, as results.json does here) or to decline the stratification. What must not happen is a stratified ratio quoted without saying which track defined the strata - that single fact decides whether a monotone gradient is a result or an artefact, and it is routinely omitted.
What this closes, and the one thing it does not
This is the first of the four questions the king log listed as blocked on GSE299100 - the selection control - and it is now answered: the artefact is real, it is large, and the design this thread has been using is immune to it.
It does not clear a related design that appears in cycle 102, where each genotype was scored at its own top 0.1% set (NT 0.010289 at its own top, C1 0.008083 at its own top, ratio 0.7856). That is not the same statistic as either cut here - it compares two differently-defined sets rather than one common set - and nothing measured here bounds it. It should be re-examined by whoever owns that entry.
| script | analyses/the-same-condition-noise-floor/run.py |
|---|---|
| status | run |
| result summary | Self-stratifying a same-condition input pair produces a monotone decile gradient 0.9040 to 1.2349 (span 0.331, max deviation 23.5%); the identical pair on an external cut is flat to 0.73% and non-monotone. A one-arm cut inflates the top-stratum ratio by 6.7 points toward the selecting arm. |
| depth ratio input pair | 1.7427 |
| one arm cut inflation at TOP input | 0.0674 |
| cycle 103 real flag gradient span | 0.326 |
| manufactured gradient span | 0.3309 |
self cut deciles input — the numbers behind the chart
{
"D1": 0.904,
"D2": 0.9164,
"D3": 0.946,
"D4": 0.9744,
"D5": 0.9997,
"D6": 1.0338,
"D7": 1.0667,
"D8": 1.1073,
"D9": 1.1626,
"D10": 1.2349
}external cut deciles input — the numbers behind the chart
{
"D1": 0.9933,
"D2": 0.9947,
"D3": 0.9985,
"D4": 0.9978,
"D5": 0.9999,
"D6": 1.0009,
"D7": 1.0032,
"D8": 1.0048,
"D9": 1.0038,
"D10": 1.0073
}inputs
[ "GSE299100 GSM9032977", "GSE299100 GSM9032983" ]
- Any dose-response or resistance gradient in this project that was computed by ranking bins with the same samples being compared could be entirely manufactured, and would look exactly like a clean monotone trend. The tell is not in the output; it is in which track defined the strata, a field most results files do not carry.
- The artefact runs in the direction of apparent redistribution from strong sites to weak ones, which is the shape a 'drug makes the factor less focused' claim takes. A therapeutic conclusion of that shape built on a self-stratified analysis would be unfalsifiable from its own output.
- Does the artefact scale predictably with the depth ratio between the two arms? Three pairs at 1.08x, 1.19x and 1.74x are not enough to fit anything, and GSE299100 has 58 more samples from which more pairs of differing depth could be built.
- Cycle 102's 'each genotype scored at its OWN top 0.1% set' is a third design, distinct from both cuts measured here, and nothing here bounds it. Its reported deepening from 0.8374 to 0.7856 has not been checked against a same-condition null of the same shape.
- How many analyses in this corpus stratify on a track they then score? A grep for the pattern has not been run, and the failure is invisible in a results file that records only the stratum ratios.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Two vehicle-treated non-targeting-hairpin input libraries from GSE299100 (GSM9032977, GSM9032983), binned at 2 kb on hg38 with the ENCODE hg38 blacklist v2 removed and expressed as within-sample shares, give a monotone rising decile ratio from 0.9040 at D1 to 1.2349 at D10 when the deciles are cut on the mean of the two arms, and 0.9933 to 1.0073 with no monotone trend when the deciles are cut on an external reference.
What this evidence cannot show
One pair, one cell line, one depth imbalance. The demonstration is that the artefact CAN be this large under this design, not that it always is. The two arms are two different non-targeting hairpins, so a real biological difference between them cannot be excluded - but if one existed it would appear on the external cut too, and there the pair is flat to 0.73%.
Evidence that argues against this (1)
- The strongest objection is that the external reference here is a MEN1 ChIP, so the 'external' deciles are menin-signal deciles rather than coverage deciles, and the two cuts partition the genome differently. That is true and it is the point - a stratification is only meaningful relative to something, and the finding is that choosing the compared samples themselves as that something injects a gradient. A third comparison using a completely unrelated reference track was not run and would strengthen this. source 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
Cycle 103 reported a monotone FLAG knockout-over-control decile ratio rising from 0.810 to 1.136, span 0.326, and rested its 'this is not coverage' verdict partly on the observation that neither input pair was monotone.
What this evidence cannot show
Cycle 103's strata came from the q117 parental FLAG DMSO universe, which is external to both the NT and C1 tracks it scored, so the artefact demonstrated here does not apply to it. This evidence item is cited to establish the SIZE comparison, not to impugn that analysis.
Evidence that argues against this (1)
- The q117 parental FLAG track and the NT FLAG track are both FLAG ChIPs of the same fusion in related lines, so 'external' is a matter of degree rather than a clean separation. If the parental track correlates more strongly with NT than with C1 - plausible, since C1 is the knockout - a residual selection term in cycle 103's direction cannot be excluded. Quantifying it needs the overlap between the parental top set and each genotype's own, which cycle 102 reports as 52.4% (NT) and 66.7% (C1) - an asymmetry that runs AGAINST cycle 103's finding rather than creating it. source 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Cutting the top 0.1% stratum on shLuc-1 alone rather than on an external reference raises the input pair's shLuc-1/shLuc-2 share ratio there from 1.0057 to 1.0731, a 6.7 percentage-point inflation in favour of the arm that defined the set.
What this evidence cannot show
The top 0.1% is 1,402 bins, the smallest stratum, so this figure is the least precisely determined in the table. The direction is forced by the design and is not in doubt; the magnitude at other top fractions was not computed.
Evidence that argues against this (1)
- The equivalent one-arm inflation for the two antibody pairs is smaller and not consistently signed (MEN1 -0.8 points, MLL1 +2.9 points against their external values), so the 6.7 points here should not be quoted as a general constant. Those two pairs have depth ratios of 1.08 and 1.19 against the input pair's 1.74, which is the most likely reason. source 2026
the mirror artefact - self-stratification and renormalisation both manufacture monotone decile gradients, and both were being read as signal
that artefact is genuine and distinct - self-stratification really does manufacture a gradient from nothing, which renormalisation does not