IgG passes this project's occupancy gate at TWICE the rate of the fusion antibody - because the gate is normalised to each library's own promoter mean and IgG has almost none
also called q154 C4 · the IgG control that fired and should not have · a unit that is defined by the signal cannot police the signal
The test asks whether a gene has at least twice as much signal as an average gene switch in the same experiment. In a negative-control experiment there is almost no signal at any gene switch, so 'twice the average' is a very low bar and random noise clears it. The negative control therefore passes the test more often than the real antibody does - which tells you the control is broken, not that the real result is.
The same thing, for a scientist
**The pre-registered FALLS condition on the IgG control fired, and it is being overridden with a measurement rather than an argument.** Scored with a one-library form of the same rule (share >= 2.0x that library's own mean promoter, above that library's floor), the 25 IgG libraries pass a **median 1.938% of gene bodies** against **0.948%** for KMT2A-N1 scored the same way - **IgG passes at 2.0 times the rate of the fusion-side antibody** [E1]. The cause is measured, not assumed. The unit is *that library's own mean promoter density*, and **IgG has 8.45-fold lower promoter density than KMT2A-N1** (median 0.00228 against 0.01928 fragments per bp) and a **2.84-fold lower in-promoter fragment fraction** (2.38% against 6.76%) [E1]. So an IgG library's "average promoter" is close to background, and a gene body carrying ordinary background reads clears 2.0x that trivially. **The denominator is defined by the very enrichment the control is supposed to lack, so the control's unit shrinks with its signal and the gate becomes easier exactly where it should be hardest.** This generalises past IgG. **Any within-sample-share gate is uninterpretable in a library whose enrichment differs from the test libraries', and this deposit has no spike-in with which to fix it.**
The measurement
| IgG (25 libraries) | KMT2A-N1 (25 matched samples) | ratio | |
|---|---|---|---|
| median total fragments | 4,570,914 | 10,102,155 | 2.21x |
| median in-promoter fragment fraction | 0.0238 | 0.0676 | 2.84x |
| median promoter density (fragments per bp) | 0.00228 | 0.01928 | 8.45x |
| median one-library gene-body pass rate at 2.0x | 1.938% | 0.948% | 0.49x |
The last row is the control firing; the third row is why. Read together they say the gate is measuring how empty the denominator is 1.
What was pre-registered, what fired, and the decision
The q154 pre-registration lists, under FALLS: "C4 shows IgG passing the rule at a rate within a factor of two of KMT2A-N." IgG passes at 2.0x the rate, so the condition fires on its letter and by a wide margin in the unexpected direction.
It is not being quietly dropped. Three things are recorded instead:
- The condition fired. Written here, in the title, and in the beat's verdict entry.
- The control is invalid for a stated, measured reason - an 8.45-fold denominator difference - not because its answer was unwelcome. The same measurement would have been made had IgG passed at a tenth of the rate.
- The control was mis-specified in two ways I can name. IgG has one library per sample, so the two-antibody rule under test cannot be applied to it at all and the comparison above had to use a one-library variant of the rule - a different rule. And the pre-registration's direction column for C4 said a high IgG rate "means the gate is reading library structure", which is true and is a statement about the control's unit, not about the fusion calls; I wrote it as a FALLS trigger without noticing that the two are different claims.
The right replacement control, for the next beat that needs one, is an absolute one: compare IgG and KMT2A-N gene-body densities in fragments per bp per million fragments, with no within-library normalisation, so the denominator cannot move with the antibody.
Why this was predictable and was predicted
q145 measured IgG libraries in this deposit as 57-94% zero at promoters, and q147 concluded in its own pre-registration that "IgG is nearly worthless here" and replaced it with an eight-locus cannot-be-occupied panel as the floor - while still printing IgG beside every call as information. That decision was right and this beat re-derived its justification from a different direction.
What is new is the mechanism stated as a general rule rather than as a fact about this deposit: a ratio whose denominator is computed from the same library it is scoring cannot be compared across libraries that differ in enrichment. The project already knows the sibling of this: oncoprotein share correlates with raw library coverage at Spearman +0.659 across the 54 KMT2A-N libraries, and the floor-referenced escape route still correlates at +0.631. There is no coverage-invariant measure in this deposit, and the IgG control is the sharpest demonstration of it because IgG is the library where enrichment is lowest.
| script | analyses/q154-gene-scale-occupancy-null/analyze.py |
|---|---|
| status | run |
| result summary | IgG passes the one-library gate at 1.938% against KMT2A-N1's 0.948% because IgG promoter density is 8.45-fold lower; the pre-registered FALLS condition fired on a control whose unit is defined by the signal it is meant to police. |
| prereg condition fired | FALLS condition 4 (C4: IgG within a factor of two of KMT2A-N) - fired, and overridden with the denominator measurement recorded here |
inputs
[ "C:/Temp/q144mag/counts (133 per-library promoter fragment count tables)", "Q154 gene-body occupancy matrix, 19,402 x 133" ]
measurements
[
{
"quantity": "promoter_density_ratio_kmt2a_n1_over_igg",
"subject": "median promoter fragment density, KMT2A-N1 libraries divided by IgG libraries",
"value": 8.45,
"units": "fold",
"n": 25,
"universe": "the 25 GSE252378 samples that have an IgG library, promoter density = in-promoter fragments / (19,404 x 2,000 bp)",
"null": 1,
"note": "IgG median 0.00228 frag/bp, KMT2A-N1 median 0.01928 frag/bp; in-promoter FRACTION differs by 2.84-fold and total fragments by 2.21-fold"
},
{
"quantity": "one_library_gate_pass_rate",
"subject": "fraction of gene bodies clearing a one-library 2.0x-plus-floor gate in IgG",
"value": 1.938,
"units": "% of gene bodies per sample (median)",
"n": 19402,
"universe": "25 GSE252378 IgG libraries, gene-body shares against each library's own mean promoter density",
"null": 0.948,
"note": "the null quoted is the same one-library gate applied to KMT2A-N1; IgG passes at 2.0x the rate of the fusion-side antibody"
}
]- Any claim of the form 'the signal is above background' that uses a within-library normalisation in this deposit is comparing a number to a background it computed from the same library. Where the two libraries differ in enrichment - drug-treated versus untreated is the case that matters clinically - the comparison is not a comparison.
- What does an absolute comparison say - gene-body fragments per bp per million mapped fragments, with no within-library normalisation? That is the control C4 should have been, it needs no new data, and it would say whether the gate is loose as well as whether the IgG unit is wrong.
- How many other entries in this project use a within-sample share to compare libraries of different enrichment? The registry of such comparisons does not exist and the normalisation is the project's default.
- GSE252378 has no spike-in. Is there any deposited KMT2A CUT&RUN with a spike-in against which this deposit's within-sample shares could be calibrated even approximately?
independently confirms the IgG arm is unusable, here by locus length and depth sparsity
the pre-registered IgG FALLS condition fired against this class; it is overridden by an 8.45-fold promoter-density difference that makes the IgG unit incomparable
re-derives that decision from the denominator side: IgG promoter density is 8.45-fold below KMT2A-N1's
the FALLS condition that fired and the measurement that overrides it