Tile chr5:2,750,677-2,752,892 at 100 bp from the GSE252378 bigWigs and ask which TSS the p454 signal sits on
also called IRX2 summit localisation · tiled deconvolution IRX2
The data needed to tell IRX2's signal from its neighbour's already exists; it was thrown away when the signal was averaged over two thousand letters. Reading the same files in hundred-letter steps instead would show where the signal actually sits, at no extra experimental cost.
The same thing, for a scientist
Re-analysis, no wet work, no new deposit. **Design:** for each of the 133 GSE252378 libraries, extract mean coverage in **22 non-overlapping 100 bp bins** spanning `chr5:2,750,677-2,752,892` — the union of the IRX2 and IRX2-DT 2 kb windows. Normalise each library by its own mean over all 19,404 promoters, exactly as the existing matrix is normalised, so the values are on the same within-sample-share scale as every q145/q147/q151 number. **Readout:** the bin index of the maximum, per library, and whether it falls on IRX2's side of the 215 bp inter-TSS segment (bins spanning `2,750,677-2,751,677`), on IRX2-DT's side (`2,751,892-2,752,892`), or in the shared inter-TSS segment itself. **Pre-specified primary comparison:** `p454` KMT2A-N1, the only library where either window exceeds a share of 3.0. **Controls that must be in the same run:** (i) `HOXA9` and `MEIS1` tiled identically, as loci the same rule calls occupied in 21 and 24 of 27 — if their summits do not localise, the tiling has no resolution and the IRX2 answer is uninterpretable; (ii) the IgG library of every sample tiled identically; (iii) an unoccupied negative locus from q147's eight-locus floor set, to give the bin-level noise floor. **Prediction and its falsifier are pre-stated in `data.falsifiers`.** **Cost:** the bins are 22 of 19,404 rows' worth of extraction per library, so this is cheap in compute; the constraint is disk, because the GSE252378 bigWigs must be streamed and this project's disk headroom was ~10 GB on C: and ~13 GB on D: at the time of writing. The existing `extract147.py` already streams rather than downloads whole files, which is the route to use.
Why this and not a new experiment
The attribution failure documented in this beat is not a limitation of the assay. CUT&RUN is explicitly a high-resolution method; the 215 bp question is inside its resolution. The information was lost at the analysis step, when 2,000 bp were averaged into one number. So the correct response is re-analysis of data already public, not a new antibody, a new cohort or a new assay.
That also sets the bar for what counts as success: if the tiling cannot localise a summit at HOXA9, where the same rule fires in 21 of 27 samples, then the deposit's per-library depth is too low for 100 bp bins and the right conclusion is that this question is unanswerable in this deposit — which is a real result and should be recorded as one rather than reported as a null at IRX2.
What each outcome would mean
| outcome at p454 KMT2A-N1 | reading |
|---|---|
| summit on IRX2's side | the promoter-scale near-positive is IRX2's; gene-irx2 gains its first occupancy evidence and the attribution caveat lifts for this locus |
| summit on IRX2-DT's side | the near-positive belongs to an uncharacterised divergent transcript, and every IRX2 promoter sentence in this corpus must be relabelled |
| summit in the 215 bp inter-TSS segment | the shared bidirectional element is the feature and neither gene owns it; the right object is a noncoding_element entry, not a gene claim |
| no localisable summit, controls fine | the signal is broad rather than peaked, which is itself informative about the mechanism at this locus |
| no localisable summit, controls also fail | unanswerable at this depth; record the bound, do not report an IRX2 result |
The last two rows are the ones a pre-registration exists to protect: without them written down, a flat profile would be read as whichever of the first three the reader already believed.
| design | Stream each of the 133 GSE252378 hg38 bigWigs and extract mean coverage in 22 non-overlapping 100 bp bins across chr5:2,750,677-2,752,892, plus identically tiled control loci (HOXA9, MEIS1, and one locus from q147's eight-locus cannot-be-occupied floor set). Normalise each library by its own mean over all 19,404 promoter windows so values are on the existing within-sample-share scale. Machinery: extract147.py / regions147.py, which already stream rather than download. |
|---|---|
| readout | Per library: the bin of maximum share, and whether it falls in IRX2's half (2,750,677-2,751,677), the 215 bp inter-TSS segment (2,751,677-2,751,892), or IRX2-DT's half (2,751,892-2,752,892). Primary comparison pre-specified as p454 KMT2A-N1. |
| controls | HOXA9 and MEIS1 tiled identically as resolution controls - if their summits do not localise, the IRX2 answer is uninterpretable and must be reported as a bound. Every sample's IgG library tiled identically. One floor-set locus for the bin-level noise floor. |
| cost tier | low |
| timeline | one analysis session; the binding constraint is disk headroom (~10 GB on C:, ~13 GB on D: as of 2026-09-24), not compute |
falsifiers
[ "PREDICTION: at p454 KMT2A-N1 the maximum bin falls in IRX2's half of the interval, consistent with the >=19.7% of that window's signal already shown to come from IRX2-only sequence. FALSIFIED IF the maximum bin falls in IRX2-DT's half, or if the profile is flat within the noise floor set by the floor-set control locus.", "RESOLUTION FALSIFIER, which must be scored FIRST: if the HOXA9 and MEIS1 summits do not localise to within 300 bp in the samples where the occupancy rule calls them occupied (21 and 24 of 27), the tiling has insufficient resolution in this deposit and NO IRX2 conclusion may be drawn from the same run. Report the bound instead.", "ATTRIBUTION FALSIFIER: if the maximum bin falls in the 215 bp inter-TSS segment in a majority of informative libraries, then neither gene owns the feature, the correct object is a noncoding_element entry for the shared bidirectional promoter, and every gene-level statement about this locus - including this beat's own - is the wrong shape." ]
measurements
[
{
"quantity": "n_bins_planned",
"subject": "100 bp bins spanning the union of the IRX2 and IRX2-DT 2 kb windows",
"value": 22,
"units": "bins",
"n": 133,
"universe": "chr5:2,750,677-2,752,892 (2,215 bp) across 133 GSE252378 libraries",
"null": null,
"note": "planned, not run; 22 bins x 133 libraries is negligible compute against the 19,404 x 133 extraction the project has already done twice"
}
]- If the summit turns out to sit in the shared 215 bp between the two TSSs, the experiment will have replaced a gene-attribution problem with a harder one: an element that regulates two genes cannot be targeted for one of them, and every therapeutic framing built on 'IRX2 is a fusion target' would have to be rewritten as a statement about a bidirectional element.
- Should the same tiling be run at AFF1/AFF1-AS1 (92.5% shared) in the same session? It is the same code and the same libraries, and AFF1 is the index fusion's own partner, so the marginal cost is close to zero.
- What bin width does this deposit actually support? 100 bp is chosen because it is a fifth of the smallest TSS separation of interest, not because any depth calculation says so; the HOXA9 control will answer it empirically.
- Is there a published base-resolution KMT2A-fusion track at this locus in any other deposit that would answer the question without touching GSE252378 at all? No archive search for one was run in this beat.
the only route that separates two TSSs 215 bp apart in this deposit; would decide whether the p454 near-positive is IRX2's, IRX2-DT's, or the shared element's
tiles chr5:2,750,677-2,752,892 at 100 bp with HOXA9 and MEIS1 as resolution controls
Q153 proposed this tiling at IRX2/IRX2-DT; the same code was run at AFF1/AFF1-AS1 instead and the IRX2 version is still unrun