Two processing batches inside one GEO series: the FLAG and KDM2B tracks put 1.7-1.9% of their coverage in ENCODE blacklist regions and the BCOR, RING1B and input tracks put 14-46%, with a single 2 kb bin at chr2:98,666,000 carrying 18.1% of the BCOR SGC0946 track on its own
also called Q117 blacklist asymmetry in GSE260742 · trimmed_ versus .filt.bam in one series
Half the public files in this dataset were cleaned of artefacts before deposit and half were not, and nothing in the record says so. In the uncleaned ones, a single two-kilobase stretch of repetitive DNA holds up to a fifth of the entire experiment's signal, which quietly distorts every comparison that divides by the total.
The same thing, for a scientist
GSE260742's processed files fall into two naming conventions that turn out to be two processing states. The first-batch files are named `*.filt.bam.tdf` and the second-batch files `trimmed_*.tdf`. Measured against the **ENCODE mm10 blacklist v2** (3,435 intervals, 239 Mb, 8.9% of the genome; 79,267 of 1,316,897 2 kb bins = 6.02% of the grid) [E2]: | track | batch / filename form | coverage in blacklist regions | in the single bin chr2:98,666,000-98,668,000 | |---|---|---:|---:| | FLAG DMSO / SGC / VTP | 1, `*.filt.bam.tdf` | **1.65 / 1.68 / 1.73%** | 0.00% | | KDM2B DMSO / SGC / VTP | 1, `*.tdf` | **1.85 / 1.86 / 1.87%** | 0.00% | | BCOR DMSO / SGC / VTP | 2, `trimmed_*.tdf` | **36.2 / 45.6 / 38.3%** | **13.7 / 18.1 / 14.8%** | | RING1B DMSO / SGC / VTP | 2, `trimmed_*.tdf` | **32.7 / 26.6 / 27.3%** | 11.5 / 8.8 / 9.1% | | input | 2, `trimmed_*.tdf` | **14.4%** | 4.5% | **The worst single bin holds 229,022,150 coverage-bases in BCOR DMSO** - a mean depth of 114,511 over 2 kb - against a track median of 360 [E1]. chr2:98,659,400-98,668,200 is an ENCODE *High Signal Region*; the other large offenders, chr12:3,102,800-3,111,000, chr14:19,414,800-19,633,500, chr9:0-3,053,100 and chr17:39,842,000-39,849,700 (rDNA), are all blacklisted too [E2]. **Two consequences, and the second is the dangerous one.** First, any total-referenced statistic on a batch-2 track is computed against a denominator that is a third to a half artefact. Second, and worse, **the artefact fraction differs between the conditions being compared** - BCOR is 36.2% in DMSO, 45.6% under SGC0946 and 38.3% under VTP50469 - so the distortion does not cancel. Removing the blacklist changed this beat's BCOR in-fill ratio from 1.208 to 1.327 and its RING1B ratio from 1.020 to 0.935, and it moved a pre-registered gain criterion from fail to pass [E1].
What can and cannot be concluded about the paper's own figures
The paper's Methods state that duplicate reads were filtered with Picard and that quantitative figures were made from RPGC-normalised bigWigs. This entry says nothing about whether those figures are affected, because the bigWigs are not deposited and the filtering state of the BAMs behind them cannot be read off a .tdf filename. What is established is narrower and is enough:
- the deposited files differ in artefact load by a factor of twenty between two halves of one series;
- nothing in the 158 sample records distinguishes them;
- a re-analyst joining them into one comparison, which is exactly what this beat set out to do, will produce cross-antibody statements distorted by an amount that varies by condition.
The general form, which is not about this series
A deposit can be internally inconsistent in its processing and give no signal of it except a filename. The checks that would have caught it, in increasing cost:
- Print the maximum bin and its fraction of the track total for every track before using any of them. One bin at 13.7% of a track is not subtle and takes one line to find. This beat found it only because the BCOR maximum looked physically impossible.
- Report the blacklist fraction per track, not per series, and compare it across the conditions you intend to contrast. A blacklist fraction that differs by ten points between arms is a warning about the denominator, not about the regions.
- Treat filename conventions inside a series as a hypothesis about batches and test it.
This is the same family as the project's existing rule that a filter which silently removes the hardest cases makes the remaining data look well-behaved - here the filter was applied to half the data and the halves were then compared.
| script | analyses/q117-gained-fusion-occupancy/run117.py |
|---|---|
| status | run |
| result summary | Two processing batches inside one GEO series with a twenty-fold difference in blacklist artefact load, and the load differs between the conditions being compared (BCOR 36.2 / 45.6 / 38.3% for DMSO / SGC0946 / VTP50469). |
| absence class | measured-not-reported |
| absence class note | ONE gap, and the token describes it: the processing state of each deposited file. The submitters know which BAMs were duplicate-filtered - the paper's Methods say Picard was used - and the per-sample records do not say which files reflect it. It is not a statement about the data being absent; the files are all there. |
| absence class assigned | 2026-09-23 (Q117) |
blacklist fraction of track total
{
"FLAG_DMSO": 1.65,
"FLAG_SGC": 1.68,
"FLAG_VTP": 1.73,
"KDM2B_DMSO": 1.85,
"KDM2B_SGC": 1.86,
"KDM2B_VTP": 1.87,
"BCOR_DMSO": 36.22,
"BCOR_SGC": 45.6,
"BCOR_VTP": 38.29,
"RING1B_DMSO": 32.67,
"RING1B_SGC": 26.57,
"RING1B_VTP": 27.31,
"INPUT_B2": 14.42,
"unit": "percent of the track's total coverage-bases on chr1-19 and chrX"
}worst single bin
{
"locus": "chr2:98666000-98668000 (mm10)",
"encode_interval": "chr2:98659400-98668200 High Signal Region",
"BCOR_DMSO_coverage_bases": 229022150,
"BCOR_DMSO_pct_of_track": 13.69,
"BCOR_SGC_pct_of_track": 18.09,
"INPUT_pct_of_track": 4.54,
"track_median_bin_BCOR_DMSO": 360
}effect on this beat — the numbers behind the chart
{
"BCOR_infill_ratio_before_blacklist": 1.2084,
"BCOR_infill_ratio_after": 1.3272,
"RING1B_infill_ratio_before": 1.0201,
"RING1B_infill_ratio_after": 0.9349,
"FLAG_held_share_ratio_before": 0.2113,
"FLAG_held_share_ratio_after": 0.1793,
"gain_criterion_G1_before": 42,
"gain_criterion_G1_after": 52,
"G1_threshold": 50
}batches
{
"batch_1": "GSM8123xxx, filenames *.filt.bam.tdf or *_F3_*.tdf - FLAG, KDM2B, MLL1, histone marks",
"batch_2": "GSM8549xxx, filenames trimmed_*.tdf - BCOR, RING1B rep2, H2AK119ub rep3, input"
}- A correlative ChIP assay run in a clinical laboratory will not have an ENCODE blacklist step unless somebody writes one into the pipeline, and the effect measured here - a single repeat carrying a fifth of a library - is large enough to move a reported drug effect across a decision threshold. The failure is invisible in any summary statistic that divides by a total.
- Do the paper's own RPGC bigWigs carry this artefact load? They are not deposited, so the only way to know is to ask the authors or to re-align from SRA - and the second of those would answer it for free as a by-product of any proper re-analysis.
- How many of this project's other re-analysis targets mix processing batches inside one accession? Nothing in this graph currently records per-track artefact load for any deposit it relies on.
- The rDNA region chr17:39,842,000-39,849,700 is blacklisted and this project has a standing interest in KDM2B at rDNA. What does the BCOR and RING1B signal there look like once the unfiltered batch is accounted for, and is the batch-2 pileup there real or duplicate-driven?
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
In GSE260742, the BCOR and RING1B and input tracks place 14.4-45.6% of their total coverage in ENCODE mm10 blacklist regions while the FLAG and KDM2B tracks place 1.65-1.87%, and the single 2 kb bin chr2:98,666,000 carries 13.7-18.1% of each BCOR track.
NCBI GEO GSE260742 - deposited .tdf files for GSM8123359-61, GSM8123392-94, GSM8549823-29, re-binned to 2 kb · NCBI Gene Expression Omnibus 2026What this evidence cannot show
Computed 2026-09-23 from coverage-bases (igvtools mean coverage times window width) summed into 2 kb bins over chr1-19 and chrX. Coverage-bases are proportional to reads but are not a read count, so 'a mean depth of 114,511' assumes the deposited window values are coverage and not some other summary; the TDF header records trackType COVERAGE. The batch attribution rests on filename form plus submission batch (GSM8123xxx versus GSM8549xxx), not on a statement by the submitters.
Evidence that argues against this (1)
- Duplicate filtering is not the only thing that could produce this difference. Different sonication, different library complexity, a different sequencing run or simply a lower-complexity ChIP would all concentrate reads in high-signal repeats. The measurement is the artefact load; the explanation offered - that one batch was deposited unfiltered - is an inference from the filenames and could be wrong while the numbers stand. Nature Cell Biology 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The ENCODE mm10 blacklist v2 comprises 3,435 intervals covering 239.0 Mb, and it contains every one of the extreme pileup regions found here, including chr2:98,659,400-98,668,200, chr12:3,102,800-3,111,000, chr14:19,414,800-19,633,500, chr9:0-3,053,100 and the rDNA region chr17:39,842,000-39,849,700.
The ENCODE Blacklist: Identification of Problematic Regions of the Genome. · Amemiya HM, Kundaje A, Boyle AP · Scientific Reports 2019What this evidence cannot show
PMID, title, journal, year and DOI fetched from a live PubMed esummary on 2026-09-23; the BED file was downloaded the same day from the Boyle Lab repository and the interval count and total span were computed from it rather than quoted. 3,416 of the 3,435 intervals lie on the chromosomes analysed here.
Evidence that argues against this (1)
- A blacklist is a blunt instrument and removes 8.9% of the mouse genome, including regions that carry real biology - the rDNA arrays among them, which this project has an active interest in. Blacklisting is the right default for a cross-condition comparison and the wrong default for a question about repetitive loci, and this beat's use of it makes its results structurally blind to rDNA. Scientific Reports 2019
the BCOR and RING1B tracks are the unfiltered batch; the in-fill ratios in that entry are computed after removing the blacklist
a second way the deposited processed files cannot carry a cross-condition comparison on their own
- related to →BCOR — the other side of the CBX8 conflict, and a target you would be inhibiting in the direction of disease
13.7-18.1 percent of each deposited BCOR track sits in one blacklisted 2 kb bin
the deviation was the blacklist added for the artefact that entry documents
the batch-1 classification holds for the MLL1 files: 1.74-1.89% of coverage in blacklist intervals, against 14-46% for the batch-2 trimmed_* files
a third processing defect in the same series