Recalibrate the N/C oncoprotein threshold on the fusion-negative CD34+ control at the interval scale it will actually be used at, and report the false-positive count at every scale rather than one
A cutoff that makes almost no mistakes on one kind of genomic region can make hundreds on another. The fix is to measure the mistake rate on a healthy control at exactly the region size you intend to use, and to publish that number beside the results.
The same thing, for a scientist
**The problem this addresses.** The published KMT2A oncoprotein-site criterion (N/C > 1.75, a-value < 1e-5) was calibrated to produce **3 false-positive sites** in a fusion-negative CD34+ HSPC sample, on the authors' own peak intervals. Re-applying the same criterion to **TSS +/- 1 kb refGene promoters** in that same CD34+ sample produces **three orders of magnitude more** false positives. The criterion is not wrong; **a criterion is a property of the intervals it is scored on**, and the number 1.75 does not carry between interval definitions. **Design.** Take the CD34+ HSPC KMT2A N1/N2/C1/C2 AutoCUT&RUN libraries (GSE159608, `GSM5139616`-`GSM5139619`, hg19 fragment BEDs, and the matching `GSM5139613` IgG). For each of a ladder of interval definitions - the authors' called peaks; TSS +/- 250 bp, +/- 500 bp, +/- 1 kb, +/- 2 kb; CpG islands; 5 kb tiles - score every interval with the identical binomial N/C model and record the number of calls. **That curve is the calibration**, and it is what should be published beside any re-use of the threshold. **Then invert it.** For each interval definition, find the ratio threshold that yields a fixed false-discovery target (say 5 false calls) and report that threshold. Applying the resulting per-scale thresholds to the 27 GSE252378 patients gives a between-patient spread that is scale-calibrated rather than scale-inherited.
Controls that must be in the same run
- Same-reagent controls. N1 vs N2 and C1 vs C2 in the same sample call zero sites if the criterion is specific to a protein-species difference. Run them at every interval scale; they separate reagent disagreement from the N-vs-C asymmetry that exists even without a fusion.
- Depth matching. The binomial p-value threshold is depth-dependent, so every library must be thinned to a common fragment total before the call counts are compared, and the thinning must be seeded and recorded.
- IgG. At each scale, the IgG library's promoter structure bounds how much of the signal is antibody-independent.
What makes this cheap
Every input is already public and processed. The five CD34+ libraries are ~150 MB gzipped in total and stream in under a minute each; the scoring is a binomial survival function over at most 20,000 intervals. No wet work, no new sequencing, no tissue. The whole experiment is a re-scoring and it answers a question that currently blocks any quantitative re-use of the published score.
| design | Re-score the fusion-negative CD34+ KMT2A N/C libraries under the published criterion across a ladder of interval definitions (authors' peaks; TSS +/-250bp/500bp/1kb/2kb; CpG islands; 5kb tiles), with same-reagent (N1/N2, C1/C2) and depth-matched controls at every scale. |
|---|---|
| readout | False-positive call count per interval definition, and the ratio threshold that achieves a fixed false-call target at each scale. |
| controls | N1-vs-N2 and C1-vs-C2 same-protein calls; seeded binomial down-sampling to a common fragment total; IgG promoter structure. |
| cost tier | low |
| timeline | under a day of compute; all inputs are public |
inputs
[ "GSE159608 GSM5139613, GSM5139616, GSM5139617, GSM5139618, GSM5139619", "GSE252378 108 KMT2A fragment BEDs" ]
- If oncoprotein-target gene lists are used as a pharmacodynamic readout for a menin inhibitor, a scale-miscalibrated threshold changes which genes are in the list, and therefore which patients are scored as responding. The calibration is not a methodological nicety in that setting.
- At what interval width does the promoter-scale false-positive count fall to the published 3, and is that width narrower than a promoter or simply deeper per interval?
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The published criterion's 3-false-positive calibration was performed on a CD34+ oncoprotein-negative control sample, which is deposited in GSE159608 with a complete KMT2A N1/N2/C1/C2 AutoCUT&RUN set using the same four antibody catalogue numbers as the patient deposit.
MLL oncoprotein levels influence leukemia lineage identities · Janssens DH, Duran M, Otto DJ, Wu W, Xu Y, Kirkey D, Mullighan CG, Yi JS, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Communications 15:9341 2024What this evidence cannot show
The calibration interval definition is the paper's own peak set, which is not reproduced here; the ladder proposed above therefore measures how the number moves with scale rather than reproducing the published 3.
Evidence that argues against this (1)
- If the authors' peak intervals are themselves selected for high coverage, a calibration performed on them is a calibration at high depth per interval, and a promoter ladder will look worse for reasons of statistical power rather than of specificity. The depth-matching control is what separates those two, and it must be run or the ladder is uninterpretable. Janssens DH et al. 2021
the calibration sample this experiment needs, and where it actually lives
the 27 patients the recalibrated threshold would be applied to
- related to →KMT2A / MLL1 — the gene that breaks
N-terminal versus C-terminal KMT2A antibody ratio
the experiment that would settle it
the recalibration this result makes necessary
the per-batch fusion-negative calibrator is the same missing control