Oncofusion

start / graph / ana-q117-instrument-check-the-fusions-top-bins-recovered-from-a-tdf-are-the-hoxa-cluster-meis1-jmjd1c-and-six1

Analysisestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.Q117-gained-fusion-occupancy

Instrument check before any conclusion: reading a browser-track .tdf with a hand-written binner and ranking the result blind returns the Hoxa cluster in five of the top twenty bins, plus Meis1, Jmjd1c, Six1, Arid2, Ncor1 and Tsc22d2 - the canonical MLL-AF9 programme, which is what the track claims to be

also called Q117 positive control · the tdf reader is reading what it says it is

Before trusting a home-made tool that reads an unusual file format, we ranked the strongest signals it found and looked at which genes they sit on. They are the exact genes this leukaemia fusion is famous for occupying, which means the tool and the file are both doing what they claim.

The same thing, for a scientist

This beat's entire analysis rests on a **hand-written reader** for IGV's `.tdf` browser-track format, because 152 of GSE260742's 158 samples ship nothing else. A reader that silently mis-assigns coordinates, mis-scales values or reads the wrong zoom level would produce well-formed numbers and wrong biology, and nothing downstream would notice. So the reader was validated the only way that is convincing: **rank the bins, annotate the top of the ranking against a live gene track, and read the names.** Top twenty 2 kb bins of FLAG-MLLAF9 in DMSO, mm10, annotated from the UCSC mm10 refGene track on 2026-09-23 [E1]: | rank | bin | nearest refGene(s) | |---|---|---| | 1 | chr13:103,920,000 | *Erbin* | | 2, 3 | chr3:58,414,000 / 58,416,000 | ***Tsc22d2*** | | 4, 5 | chr4:58,910,000 / 58,912,000 | *Ecpas* | | 6, 8, 13, 15, 16 | chr6:52,216,000-52,234,000 | ***Hoxa5 / Hoxa7 / Hoxa9 / Hoxa10 / Mir196b*** | | 7 | chr4:28,814,000 | *Epha7* | | 9 | chr12:73,044,000 | ***Six1*** | | 10 | chr10:67,126,000 | ***Jmjd1c*** | | 11, 12 | chr11:62,456,000 / 62,538,000 | *Ncor1*, *Cenpv* | | 14 | chr15:96,286,000 | *Arid2* | | 17 | chr11:19,014,000 | ***Meis1*** | | 18-20 | chr3:60,500,000; chr15:8,444,000; chr5:65,696,000 | *Mbnl1*, *Nipbl*, *Pds5a* | **Five of the top twenty are the Hoxa cluster.** *Meis1*, *Jmjd1c* and *Six1* are the three loci the depositing paper itself names as its example PRC1.1-rescued genes, and *Tsc22d2* is one of the two loci it shows as a browser snapshot in Extended Data Fig. 4e [E2]. **None of that was known to the ranking**, which used no gene annotation at all until after the bins were chosen. Separately, the genome build was verified from **each file's own coordinates** rather than from its metadata: the chromosome extents written inside all thirteen `.tdf` files are within 64 bp of live UCSC mm10 sizes on every chromosome and about 318 kb from mm39 on chr1 [E3]. The files' internal `genomeId` field also reads `mm10`, but the verdict does not depend on it.

Go deeper
What this check does and does not license

It licenses: that the reader assigns coordinates correctly, that values are proportional to coverage, that the correct dataset (/<chr>/raw) is being read rather than a zoom level, that NaN-for-uncovered is handled, and that the file is mm10.

It does not license anything about normalisation, about replicate stability, or about whether a 2 kb bin is the right unit for any particular question. A positive control on the strongest signal in a track says the instrument points the right way; it says nothing about its behaviour in the tail, which is exactly where this beat's gained-site question lives.

The corresponding negative control was also run and is worth recording: the reader returns zero coverage-bases in 6.8% of the mm10 grid and NaN rather than zero for uncovered windows, and the FLAG tracks place only 1.65-1.73% of their coverage in ENCODE blacklist regions - which is the behaviour of a normally-processed ChIP and not of a mis-parsed file.

Facts and figures
scriptanalyses/q117-gained-fusion-occupancy/run117.py, Q117Bin.java
statusrun
result summaryThe tdf reader is validated: the top twenty FLAG-MLLAF9 DMSO bins are the canonical MLL-AF9 programme, and the build is mm10 by the files' own coordinates.
hoxa bins in top205
readerQ117Bin.java over IGVTools 2.3.98 TDFReader; /<chr>/raw dataset, tiles streamed by genomic position, NaN skipped, coverage-bases accumulated into 2 kb bins
top20 loci
[
  "Erbin",
  "Tsc22d2",
  "Ecpas",
  "Hoxa5-Hoxa11 (5 bins)",
  "Epha7",
  "Six1",
  "Jmjd1c",
  "Ncor1",
  "Cenpv",
  "Arid2",
  "Meis1",
  "Mbnl1",
  "Nipbl",
  "Pds5a"
]
build check
{
  "method": "chromosome extent written inside each .tdf (the /<chr>/z0/mean tileWidth) versus live UCSC chromosome sizes",
  "max_abs_diff_vs_mm10_over_all_13_files": "< 64 bp on every chromosome",
  "diff_vs_mm39_chr1": 317692,
  "verdict": "mm10"
}
negative controls
{
  "fraction_of_mm10_grid_with_zero_coverage": 0.068,
  "FLAG_blacklist_coverage_fraction": "1.65-1.73%"
}
How this could bite us back
  • A positive control that recovers the famous loci is the easiest kind to pass and the easiest kind to over-read. This one licenses the reader and the build; it does not license the tail, and this beat's headline question was entirely about the tail.
Still open
  1. Erbin and Ecpas rank above the Hoxa cluster in this track. Are they documented MLL-AF9 targets, or is this an artefact of 2 kb binning over a narrow very-high peak? Neither gene appears anywhere in this graph and neither was followed up.
  2. What window size did igvtools count use for these deposits? The record does not say, and it sets the finest resolution any re-analysis of these files can reach.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    Ranking the deposited FLAG-MLLAF9 DMSO .tdf by 2 kb bin and annotating the top twenty against the live UCSC mm10 refGene track returns the Hoxa cluster in five bins plus Meis1, Jmjd1c, Six1, Tsc22d2, Arid2, Ncor1, Epha7, Erbin, Mbnl1, Nipbl and Pds5a.

    What this evidence cannot show

    Gene assignment is any refGene transcript within 20 kb of the bin, which is generous and will attribute an intergenic bin to a neighbour. The Hoxa cluster is gene-dense, so five bins there resolve to overlapping gene lists rather than to five distinct genes - the correct statement is that five of the top twenty bins lie within the Hoxa5-Hoxa11 interval, not that five distinct Hox genes were recovered. Erbin and Ecpas topping the list is not something this beat can independently corroborate as MLL-AF9 biology.

    Evidence that argues against this (1)
    • A positive control drawn from the same file it is validating cannot detect an error that is uniform across the file - a constant coordinate offset smaller than a bin, or a global scale error, would leave the ranking and its gene assignment intact. The build check partly covers the first; nothing here covers the second, and nothing needs to, because every statistic in this beat is scale-invariant. Broad Institute 2017
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    The depositing paper names Meis1, Jmjd1c and Six1 as example PRC1.1-rescued genes and shows Tsc22d2 as a browser snapshot, all of which appear independently in the top twenty of this beat's blind ranking.

    DOT1L provides transcriptional memory through PRC1.1 antagonism. · Neville D, Ferguson DT, Heikamp EB, et al. · Nature Cell Biology 2026
    PMID 41634341DOI
    What this evidence cannot show

    Read from the retrieved fullTextXML on 2026-09-23. 'Blind' means the ranking used no gene annotation; it does not mean the analyst was unaware of the paper, which had been read before the ranking was produced. That is a real limitation of this control and is why the build check in E3 is reported alongside it.

    Evidence that argues against this (1)
    • Recovering famous loci is a weak control precisely because they are famous: any reasonable processing of an MLL-AF9 ChIP will find Hoxa and Meis1, so this check would pass for a reader with several kinds of bug that do not disturb the strongest peaks. It is a necessary check, not a sufficient one. NCBI Gene Expression Omnibus 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    The genome build of all thirteen downloaded .tdf files was verified from the chromosome extents written inside each file against live UCSC chromosome sizes: every file is within 64 bp of mm10 on every one of chr1-19 and chrX, and about 318 kb from mm39 on chr1.

    What this evidence cannot show

    The extent written in a .tdf is the chromosome length the writer was given, which is one step removed from the coordinates of the reads themselves; a file aligned to mm39 and then written with an mm10 chromosome table would pass this check. The check is a strong discriminator between assemblies with different chromosome lengths and is not a proof of alignment reference.

    Evidence that argues against this (1)
    • This project has recorded a corpus entry asserting a genome build that was wrong in both directions two cycles ago, so a build check that relies on a single field, however internal, should not be the only one. The independent corroboration here is biological: the Hoxa cluster bins fall at chr6:52.2 Mb, which is the mm10 position and not the mm39 one. UCSC Genome Browser 2026
Connected to