Cycle 102 declared before it ran that an input pair 'bounds library and coverage variation but NOT ChIP efficiency' and is therefore a LOWER BOUND - the size of that underestimate is now measured in human KMT2A::AFF1 MV4;11 and it is sixteen-fold at the peaks: 0.6% for an input pair against 9.4% for a menin ChIP pair over the same bins, in the same cells, on the same day
also called an input control cannot bound a ChIP · the antibody is most of the floor · how much does the antibody add
When scientists want to know how noisy a measurement is, they often use a simpler control sample that skips the antibody step. This compares the two directly in identical cells: at the sites that matter most, the simple control looks sixteen times steadier than the real measurement does, so using it as the yardstick makes small changes look much more convincing than they are.
The same thing, for a scientist
**In human MV4;11 (KMT2A::AFF1, but acute MONOCYTIC with FLT3-ITD, not infant B-ALL), two vehicle-treated non-targeting hairpins were compared three ways over identical 2 kb hg38 bins and an identical external peak-set stratification.** The only difference between the three comparisons is whether an antibody was used and which one. | stratum | input pair | MEN1 pair | MLL1 pair | antibody / input | |---|---:|---:|---:|---:| | **TOP 0.1%** | **0.57%** | **9.37%** | **8.83%** | **16.4x** | | D1 | 0.67% | 2.42% | 0.82% | 3.6x | | D2-D10 | 0.01-0.73% | 1.27-2.41% | 0.08-2.55% | 3-25x | **At the peaks the antibody is essentially the entire floor.** Two input libraries from the same experiment agree to 0.6% over the top 0.1% of menin-complex sites; two menin ChIPs over those same bins disagree by 9.4%, and two MLL1 ChIPs by 8.8%. **The two antibodies are different reagents against different proteins and they agree with each other to within 0.6 percentage points**, which argues the ~9% is a property of immunoprecipitation at sharp peaks rather than of one bad antibody. **This is exactly the caveat cycle 102 wrote about itself before it ran** - *"an input pair carries no antibody, so it bounds library and coverage variation but NOT ChIP efficiency; the floor is a LOWER BOUND on a ChIP ratio's variability"*. The caveat was right. **What was missing was its size, and a lower bound with no scale attached was used for three cycles as though it were the quantity itself.** **And the direction of the error is not uniform.** In the bulk deciles the gap narrows to a few percentage points in absolute terms, because both floors are small there. The damage is concentrated where the project does most of its work: the top stratum.
Why this is not the same finding as 'the floor is stratum-dependent'
The companion entry reports that the floor is 9.4% at the peaks and ~2% in the bulk, and that cycles 103/104 applied a peak figure in the bulk. This entry is the orthogonal cut of the same table: holding the stratum fixed and varying only the presence of an antibody.
The two have different consequences. Stratum-dependence says apply the right row. This says an input control is not a control for a ChIP at all at the peaks - no choice of stratum repairs it, because the quantity it measures excludes the dominant term.
What an input pair does still bound
It is not useless. Over the bulk deciles the input pair's deviation stays inside 0.73% and is not monotone, which is a genuine and strong statement that library depth, mappability and copy-number structure are not producing a decile gradient in these cells. That is precisely the use cycle 103 put its input pairs to (U1), and on that use they were the right control and the conclusion held.
The failure is confined to using an input pair to bound a ChIP ratio. Those are different quantities and the difference is a factor of sixteen where it matters.
The practical rule this yields
Any floor quoted for a ChIP contrast must come from a ChIP pair. If only an input pair is available, the honest statement is a lower bound with the word LOWER in the claim, not in the caveat - cycle 102 put it in a field called caveat_declared_before_running, did the right thing, and it still travelled as a floor into three later cycles. A caveat in a field does not protect the number that gets quoted.
| script | analyses/the-same-condition-noise-floor/run.py |
|---|---|
| status | run |
| result summary | Input-only same-condition floor 0.57% at the top 0.1% stratum; MEN1 9.37%, MLL1 8.83% over the same bins in the same cells. The antibody contributes ~16x the input-pair floor at the peaks and 3-25x across the bulk deciles. |
fold antibody over input
{
"TOP": 16.4,
"D1": 3.6
}antibodies
{
"MEN1": "Bethyl A300-105A",
"MLL1": "Bethyl A300-086A"
}inputs
[ "GSE299100 GSM9032977 / GSM9032983 input pair", "GSE299100 GSM9032979 / GSM9032985 MEN1 pair", "GSE299100 GSM9032981 / GSM9032987 MLL1 pair" ]
- If a 9% same-condition floor is generic to sharp-peak ChIP, then published menin-inhibitor target-engagement figures based on a single ChIP pair with no replicate cannot resolve engagement below about 10% at the peaks - and a resistant genotype whose true residual engagement differs by 5% would be scored identical to a sensitive one.
- The reverse risk is worse for a negative: a beat that measures a 7% drug effect at the peaks and calls it real would be inside this floor. Anything in this corpus that reports a single-pair peak-stratum ChIP effect between roughly 1 and 10 percent is unresolved, not negative and not positive.
- Is the ~9% peak floor a property of immunoprecipitation or of menin-complex peak shape? The UBR5, IKAROS, MEIS1, NRF1 and ZNF692 ChIPs in the same deposit discriminate these and none has been binned.
- Does the gap close at higher resolution? The top 0.1% of 2 kb bins is a coarse peak definition; at 200 bp the peak stratum would contain the summits only, where IP efficiency differences might be larger or smaller.
- No entry in this corpus currently states a floor for any other ChIP target. Every bounded 'inside the floor' statement about a non-menin protein is therefore uncalibrated.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Over identical 2 kb hg38 bins, an identical external peak-set stratification and the same two non-targeting hairpins in the same DMSO-treated MV4;11 experiment, the vehicle-versus-vehicle share ratio at the top 0.1% stratum deviates from 1 by 0.57% for the input pair, 9.37% for the MEN1 pair and 8.83% for the MLL1 pair.
What this evidence cannot show
n = 1 pair per assay. The three pairs are not independent of each other in every respect - they are the same two cell populations, so anything that differs between shLuc-1 and shLuc-2 cells is shared across all three rows. That shared term is bounded by the input row at 0.57%, which is why the antibody attribution is defensible, but it is an argument and not a separate measurement.
Evidence that argues against this (1)
- An alternative reading: the input has almost no signal at menin peaks, so its share there is dominated by uniform background, which is intrinsically reproducible. On that reading the 16-fold gap is partly an artefact of comparing a flat quantity with a peaked one rather than of the antibody. Against it: the per-bin spread runs the OTHER way - at the top stratum the input's per-bin central 90% interval is 0.5411-2.1508 and MEN1's is only 0.7796-1.1670, so the input is the noisier of the two per bin and still the steadier in aggregate. That pattern is what a systematic IP-efficiency difference looks like and not what background flatness looks like. source 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
Cycle 102 recorded, in a field named caveat_declared_before_running, that 'an input pair carries no antibody, so it bounds library and coverage variation but NOT ChIP efficiency; the floor is a LOWER BOUND on a ChIP ratio's variability', and the 4-10% figure derived from those input pairs was nevertheless used in cycles 103 and 104 to adjudicate antibody readouts.
What this evidence cannot show
Internal repository sources, not external literature. Cited because the claim is about this project's own reasoning chain.
Evidence that argues against this (1)
- Cycle 104's own pre-registration (U1) restated the limitation again, in stronger terms - 'the inputs have NO DRUG ARM so they bound genotype variation and not drug variation - a weaker control here than it was in cycle 102 and is labelled as such rather than leaned on'. So the project flagged the defect three times and used the number anyway. That makes this less a failure of awareness than of having no alternative, which is the situation GSE299100 changes. source 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The two ChIP arms are different antibodies against different proteins - anti-Menin Bethyl A300-105A for GSM9032979/GSM9032985 and anti-MLL1 Bethyl A300-086A for GSM9032981/GSM9032987 - and their same-condition floors at the top 0.1% stratum agree to within 0.54 percentage points (9.37% and 8.83%).
GEO sample records GSM9032979, GSM9032981, GSM9032985, GSM9032987, chip antibody / antibody vendor / antibody catalog_number fields, fetched individually 2026-09-23 · NCBI Gene Expression Omnibus 2025What this evidence cannot show
Two antibodies is n = 2. Menin and MLL1 occupy largely the same sites as a complex, so the two ChIPs are not independent experiments on independent biology, and a shared property of those sites - peak sharpness - is at least as good an explanation as a shared property of immunoprecipitation.
Evidence that argues against this (1)
- GSE299100 contains ChIP for UBR5, IKAROS, MEIS1, NRF1 and ZNF692 in the same cells. If the ~9% peak floor is a property of IP rather than of menin-complex peaks, those should show it too; if it is a property of these particular peaks, a factor with a different binding-site distribution should not. That experiment is one download away and has not been run. source