The third resolution trap in three cycles: an ENRICHED-domain BED has a median domain of 13-17 kb and a per-SAMPLE bin size, so 'the GAPDH promoter is inside an H3K9me3 domain' is true in the file and means nothing about the promoter
also called a domain call is not a track · median domain 15 kb · why the mark-identity control passed and failed at once
Some public datasets do not publish the raw signal, only a list of stretches the software judged enriched, and those stretches average fifteen thousand bases - so a gene's promoter can be 'inside' one while the promoter itself carries none of the tag. Two files from the same experiment can also use different block sizes, and both facts are invisible unless you measure them.
The same thing, for a scientist
Cycle 57 hit a **tooling** limit on GSE210480 (pyBigWig would not build), cycle 58 removed it, cycle 59 found a **resolution** limit underneath (the H3K9 tracks are 1 kb binned; `absence-gse210480-processed-tracks-are-1-kb-binned`), and this cycle finds the **third** kind in a different dataset: GSE71809 deposits **RSEG ENRICHED-domain calls in BED**, not coverage. Three consequences, all measured [E1]: **1. The domains are enormous.** Across the CD34+ samples the modal called-domain width is **2,550-6,800 bp** and the median is **12,750-17,000 bp**; the ChIP-input samples reach a median of **63,600 bp** with only 8,056 domains. A 147 bp exon, a 1,048 bp element and a 7,466 bp BCR are all *sub-domain* objects: each is inside a domain or outside one, full stop. **2. The consequence is a control that passes and fails simultaneously.** In the CD34+ H3K9me3 sample the *ZNF274* body - this project's H3K9me3 positive control - is 100% covered, which is right. **And so are the *GAPDH* and *ACTB* promoters and the whole *MLLT1* and *TCF3* gene bodies**, which is not a statement anyone should make about H3K9me3 at an active housekeeping promoter. The domains simply swallow the neighbourhood. **So a mark-identity control run at this resolution cannot discriminate, and reading a positive control as passing here would be reading the domain size.** **3. The bin size is per SAMPLE, not per series.** Every domain boundary in GSM1846171 is a multiple of 300 bp (100% at moduli 100, 150 and 300). In GSM1846172 and GSM2152593 - the same mark, the same experiment, one the same donor - only **24.8%** of boundaries are multiples of 100, and the modal width 2,550 bp is not a multiple of 100 at all. **RSEG chose a different bin size per library**, presumably from read density, and nothing in the GEO record says so [E1]. This is the same shape as `concept-one-geo-series-two-chrom-tables-the-ctcf-tracks-are-unpadded`: a per-file property that a per-series check cannot see. **What a domain call CAN support** is a domain-scale question - is this locus inside a heterochromatic block, how large is the block, where does it rank among the chromosome's blocks. That is the form the CD34+ result was written in (`ana-q111-cd34-h3k9me2-over-the-kmt2a-locus-measured`), and it is why that entry reports a gap length and a gap rank rather than a BCR/flank ratio.
The detection recipe, and the three things to record
Before quoting anything from a domain-call BED:
- Modal and median domain width, per file. One pass over the BED. If the median is tens of kilobases, every interval narrower than that is a binary membership question.
- The modulus of the boundaries, per file - the fraction of starts and ends that are multiples of 100, 150, 300, 600, 1,000. This recovers the caller's bin size when the deposit does not state it, and it is how the per-sample variation above was found. An all-or-nothing result (100% or ~25%) is the tell.
- Whether any sample calls a domain over your interval at all. If none does, you cannot distinguish "no enrichment" from "unmappable" - and a caller's deadzone file, here
deadzones-k41-hg19.bed, will silently blank a region in every sample. In GSE71809 one of three CD34+ inputs calls the KMT2A locus ENRICHED, which is what establishes it is callable; without that single row the CD34+ H3K9me2 zeros would be uninterpretable.
And record all three in the analysis output, as absence-gse210480-processed-tracks-are-1-kb-binned prescribes for binned tracks. The three resolution traps now found in this project differ in mechanism and share a signature: the file answers a question finer than it supports, returns a well-formed number, and warns about nothing.
| deposit form | the silent failure | the cheap detection |
|---|---|---|
| binned coverage bigWig (1 kb) | a 147 bp query returns one value and equals a 501 bp interval containing it | run-length encode a window; or query a narrow interval and a wider one containing it |
| binned coverage bigWig (200 bp) | an argmax is a bin label, not a summit | use a signal-weighted centroid and quote +/-100 bp |
| ENRICHED-domain BED | a 15 kb domain makes a promoter 'covered'; and the bin size varies per sample | modal and median domain width, and the boundary modulus, per file |
The second-order point, which is the one that keeps costing cycles
All three traps were found only because two quantities were computed that did not need to be. The 1 kb floor surfaced because a design happened to list exon 11 and the summit window separately, so two intervals returned identical numbers. The 200 bp bin label surfaced because a centroid and an argmax were both computed. The domain breadth surfaced here because ZNF274, GAPDH and ACTB were carried through as controls even though the analysis was about KMT2A, and their answers were absurd in a way the KMT2A answer was not.
So the practice worth generalising is not a checklist item; it is carrying redundant controls whose correct answer you already know, through to the output, and reading them. A positive control that must be enriched and a negative control that must not be will expose a resolution problem the moment the resolution is wrong, because a coarse instrument gives them the same answer. This project's existing mark-identity pair - ZNF274 body positive, GAPDH promoter negative - did exactly that here: both came back 100% covered, and that is the whole finding.
And the corollary for reading other people's results: a paper reporting that a locus falls inside an H3K9me2 or H3K9me3 domain has made a domain-scale statement. It is not evidence about a promoter, an exon or a 1 kb element inside that domain, however precisely the coordinates are quoted.
| trap class | silent resolution mismatch in a processed deposit - third distinct mechanism found in three cycles |
|---|---|
| deposit form | RSEG 0.4.8 RC ENRICHED-domain calls in BED, one file per sample, hg19, with deadzones-k41-hg19.bed |
| series | GSE71809 |
| per sample not per series | GSM1846171's boundaries are all multiples of 300 bp; GSM1846172's and GSM2152593's are not multiples of 100 at all for three quarters of domains. Same mark, same experiment, one the same donor. The caller chose a bin per library and the GEO record does not say so. |
| detection recipe | per file: modal and median domain width; the fraction of boundaries that are multiples of 100, 150, 300, 600 and 1,000 bp; and whether ANY sample in the series calls a domain over your interval, which is the only way to separate 'no enrichment' from 'in the caller's deadzone file' |
| what a domain call supports | domain-scale questions only - is this locus inside a block, how large is the block, where does the block rank among the chromosome's blocks. Not a ratio, not a promoter, not an element. |
| how all three were found | by computing a quantity that did not need computing - two intervals of different widths, a centroid alongside an argmax, and redundant positive and negative controls carried through to the output and read. A coarse instrument gives a known-positive and a known-negative control the SAME answer, which is the cheapest resolution alarm available. |
| localization | not applicable - this entry concerns file format and asserts nothing about a cellular location |
| subnuclear localization | not applicable - a domain-call BED has no subnuclear coordinate, though a large H3K9me2 domain is the kind of object often associated with the lamina; no such position is claimed here |
| cell type | the files measured are from normal human CD34-positive bone-marrow haematopoietic progenitors, granulocytes, K562 and primary AML myeloblasts |
| cell state | resting primary cells and asynchronous K562 culture with and without UNC0638 |
| epigenetic readout | H3K9me2, H3K9me3, H3K27me3, H3K27ac and H3K4me2 ENRICHED-domain calls; the finding is about the resolution of the deposit rather than about any mark |
| shown for fusion | Not applicable to any fusion. The finding concerns file format; none of the cells involved carries a known KMT2A fusion. |
measured domain widths — the numbers behind the chart
{
"CD34+ H3K9me2 GSM1846171": {
"n_domains": 45778,
"modal_bp": 6300,
"median_bp": 15300
},
"CD34+ H3K9me2 GSM1846172": {
"n_domains": 46839,
"modal_bp": 2550,
"median_bp": 12750
},
"CD34+ H3K9me2 GSM2152593": {
"n_domains": 39515,
"modal_bp": 6800,
"median_bp": 17000
},
"CD34+ input GSM1846168": {
"n_domains": 8274,
"modal_bp": 16800,
"median_bp": 63600
},
"CD34+ input GSM1846173": {
"n_domains": 8056,
"modal_bp": 19200,
"median_bp": 63600
},
"CD34+ input GSM2152598": {
"n_domains": 19686,
"modal_bp": 8100,
"median_bp": 23400
}
}boundary modulus fraction — the numbers behind the chart
{
"GSM1846171": {
"100": 1,
"150": 1,
"300": 1,
"600": 0.253,
"1000": 0.009
},
"GSM1846172": {
"100": 0.248,
"150": 0.113,
"300": 0.027,
"600": 0.006,
"1000": 0.002
},
"GSM2152593": {
"100": 0.248,
"150": 0.113,
"300": 0.028,
"600": 0.006,
"1000": 0.002
}
}the control that passes and fails
{
"sample": "GSM2152597, CD34+ H3K9me3",
"ZNF274_body_covered_fraction": 1,
"GAPDH_promoter_pm1kb_covered_fraction": 1,
"ACTB_promoter_pm1kb_covered_fraction": 1,
"MLLT1_gene_covered_fraction": 1,
"TCF3_gene_covered_fraction": 1,
"KMT2A_locus_covered_fraction": 0,
"reading": "the H3K9me3 positive control is covered, which is right, and so are two active housekeeping promoters, which is not a promoter-scale statement. At a 15-27 kb median domain the mark-identity pair cannot discriminate, and only the KMT2A zero is informative because a zero at domain scale IS a domain-scale statement."
}the three traps so far
{
"1 kb binned bigWig": "a 147 bp query returns one value and equals a 501 bp interval containing it - GSE210480 H3K9 tracks, cycle 59",
"200 bp binned bigWig": "an argmax is a bin label, not a summit - GSE210480 CTCF tracks, cycles 59-60",
"ENRICHED-domain BED": "a 13-17 kb median domain makes a promoter 'covered', and the bin size varies per sample - GSE71809, cycle 60"
}- Applied too broadly this would discard domain-level evidence altogether, which would be wrong and would cost this graph the one haematopoietic H3K9me2 measurement it has. A 15 kb domain call is a perfectly good instrument for a 200 kb question, and the CD34+ result - the KMT2A locus inside the 7th-longest H3K9me2-free gap on chr11 - is exactly the kind of statement it supports.
- The per-sample bin-size finding could be used to dismiss the six-of-ten AML samples that show no domain at the KMT2A locus as low-resolution artefacts, and that would be motivated reasoning in the direction this graph's chain prefers. The AML split is 6 with no domain and 4 with one; if bin size tracks depth, the honest move is to test it rather than to explain away whichever half is inconvenient.
- How many processed-track re-analyses in this corpus quote an interval narrower than the file's resolution? Named as an open question in cycle 59 and still unrun. The sweep is one resolution check per dataset cited by an analysis entry, and there are fewer than thirty such datasets.
- What bin size did RSEG actually use per library in GSE71809, and does it track read depth? The boundary modulus infers a bin and does not read one. The SRA run metadata carries spot counts, so testing bin size against depth is one join, and it would say whether coarser samples are simply shallower - which would matter for the six of ten AML samples that show no domain at the KMT2A locus.
- Does this graph hold any claim that rests on a locus being INSIDE a published heterochromatin domain? That is a domain-scale statement being used at locus scale whenever the interval is much smaller than the domain, and the H3K9me3 and LAD entries are the obvious places to look.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
Measured from GSE71809's 31 deposited RSEG 0.4.8 RC ENRICHED-domain BED files: modal called-domain width is 2,550-6,800 bp and median 12,750-17,000 bp in the CD34+ H3K9me2 samples, rising to a median of 63,600 bp with only 8,056 domains in a CD34+ ChIP input; 100% of domain boundaries in GSM1846171 are multiples of 300 bp while only 24.8% of those in GSM1846172 and GSM2152593 are multiples of 100 bp, so the caller's bin size differs between samples of the same mark and experiment; no domain in any file exceeds its hg19 chromosome length; and in the CD34+ H3K9me3 sample the ZNF274 body, the GAPDH promoter, the ACTB promoter and the MLLT1 and TCF3 gene bodies are all 100% covered by called domains while the KMT2A locus is 0% covered.
GSE71809 RSEG ENRICHED-domain BED files for 31 samples, parsed for per-file domain-width distributions, boundary moduli and per-interval coverage, 2026-09-22; code and full output at analyses/q111-cd34-h3k9me2-over-the-kmt2a-bcr/run.py and results.json · Salzberg AC, Harris-Becker A, Popova EY, Keasey N, Loughran TP, Claxton DF, Grigoryev SA · NCBI GEO (data accompanying PLoS One 12:e0173723) 2017What this evidence cannot show
The boundary-modulus test infers a bin size and does not read one: a caller could produce boundaries at multiples of 300 for a reason other than a 300 bp bin, and the 24.8%-of-100 result in two files means their bin is not a round number rather than telling us what it is. RSEG's own parameters are recorded in the GEO data-processing field as '-s 10000 -i 20 -duplicates', and -s is not a bin size, so the deposit states nothing that contradicts or confirms this. Domain widths were tabulated genome-wide per file; the interval coverages are for the intervals this analysis happened to carry. The GAPDH and ACTB observations are from ONE CD34+ H3K9me3 sample, so 'the mark-identity control cannot discriminate at this resolution' rests on one file, and the same check on the H3K9me2 files was not tabulated in this entry.
Evidence that argues against this (1)
- A broad domain call covering an active promoter may be correct rather than coarse. H3K9me3 and H3K9me2 form genuinely large domains, and a promoter inside such a domain is a real biological configuration - active promoters do sit in H3K9me2 LOCKs, and the paper's own framing is about blocks over 100 kb. On that reading the file is not failing; the reader is asking a question the mark does not have an answer to at promoter scale. That is the more charitable and probably more accurate diagnosis, and it leads to the same prescription: use a domain call for domain-scale questions only. Salzberg AC et al. 2017
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
This project has now found three distinct silent resolution limits in deposited processed files in three consecutive cycles - 1,000 bp binning in GSE210480's H3K9 bigWigs, 200 bp binning in the same series' CTCF bigWigs which makes an argmax a bin label, and 13-17 kb ENRICHED-domain calls in GSE71809 - and in every case the file returned a well-formed, plausible number for a query finer than it supported, with no error, no NaN and no warning.
absence-gse210480-processed-tracks-are-1-kb-binned and data/symbol-collisions.json row 'A BINNED coverage track answers a finer query without error, and the answer is plausible', read 2026-09-22, together with this cycle's domain-width measurement · Iqbal Lab oncofusion knowledge graph 2026What this evidence cannot show
Three instances in three cycles is a pattern in this project's recent work, not a measured rate over the corpus: no systematic sweep has asked how many of this graph's processed-track re-analyses quote an interval narrower than the file's resolution. That sweep is named as an open question in the cycle-59 entry and is still unrun. So the claim is that the failure recurs and is cheap to detect, not that it is prevalent at any particular rate.
Evidence that argues against this (1)
- The recurrence may be selection rather than prevalence: this project has spent three cycles on two datasets that happen to deposit processed files at coarse resolution, and a project working mostly with base-resolution bigWigs would meet this once. The ENCODE K562 H3K9me1 tracks read by an earlier beat are base-resolution (modal run 1 bp), which is the counter-example from inside this graph's own work. Iqbal Lab oncofusion knowledge graph 2026
the third distinct silent resolution limit in three cycles, with the same signature and a different mechanism
the resolution bound that decides what that measurement can and cannot say
same shape: a per-file property that a per-series check cannot see - here the caller's bin size varies between samples of one mark