start / graph / ana-q118-the-deposited-tss-table-is-blind-to-83-percent-of-the-gained-set

AnalysisstrongQ118-menin-at-the-human-index-fusion

The convenient deposited promoter table finds 3 gained windows where the genome-wide re-binning finds 53, because 83% of the gained set is not at a RefGene TSS - so the processed file that makes this dataset cheap to re-use is structurally blind to the answer

also called 3 versus 53 · the processed table cannot see the result

This dataset comes with a ready-made table of signal at gene starts, which is what anybody reusing it would open first. Running the same question against the raw files instead finds seventeen times more changed places, because most of them are not at gene starts at all.

The same thing, for a scientist

**The same pre-registered question, asked twice with two instruments on the same samples** [E1]: | instrument | universe | gained set | |---|---|---:| | A - the deposited ROSE TSS table | 21,461 RefGene 6 kb TSS windows | **3** windows (RNF149, U2AF2, TNFSF9) | | B - genome-wide re-binning of the same `.tdf` files | 1,389,514 2 kb bins, input > 0, blacklist-filtered | **53** bins | **And the reason, measured** [E1]: of the 53 gained bins, **9 (17.0%)** fall inside one of the deposited table's own 21,461 TSS windows, against **76.1%** of the 1,390 held bins and **6.4%** of the universe (Fisher, gained versus held, **p = 1.2e-18**). The gained set is 2.7-fold enriched over the genome for being near a promoter and 4.5-fold **depleted** relative to the held set. **So a beat that had used only the processed file would have reported a clean null**, and the null would have been a fact about the file. This project has documented the general form - *an index that covers one part of a document cannot bound a claim about the whole document* - for PubMed abstracts; **this is the genomic version, and the missing part is 83% of the answer.** **It cuts the other way too, and that is why both instruments are kept.** Instrument A is the only route to the DOT1L arm and the ML-2 line in this beat, because those `.tdf` files were not downloaded, and it is the only instrument that carries all three lines and all three antibodies at once. The rule is not *do not use the processed table*; it is **state what the processed table's universe excludes, and measure it if the claim depends on it.**

Facts and figures

script	analyses/q118-menin-at-the-human-index-fusion/binsD.py (D9 block)
status	run
result summary	3 gained windows on the deposited promoter table versus 53 gained bins genome-wide; 17.0% of the 53 are at a RefGene TSS against 76.1% of the held set and 6.4% of the universe, Fisher p = 1.2e-18.
cell line	RS4;11
genome build	hg19
absence class	measured-not-searched
absence class note	The gap this token describes is the 44 gained bins that are outside every window of the deposited processed table: the signal is in the archive and was unfindable because the convenient search space excluded it. It does not describe the H3K79me2 arm, which is measured-not-analysed in this beat.

inputs

[
  "GSE127507 ROSE TSS signal table",
  "GSM3635695/96 RS4;11 MLL1n",
  "GSM3635699 RS4;11 input"
]

Still open

How many of this project's GEO re-analyses rest on a deposited processed file whose universe excludes the compartment the question is about? No sweep of the analyses directory for that pattern was run and it is the obvious follow-up to this measurement.
Does the reverse error occur - a gained set that instrument A sees and instrument B does not? The 3 instrument-A windows (RNF149, U2AF2, TNFSF9) were not checked for membership of the 53, and one of them (U2AF2) also appears in the RS4;11 MENIN gained window list, which is worth a look.

Every source behind this entry (1)

E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The pre-registered gain question returns 3 windows on the deposited GSE127507 ROSE TSS table and 53 bins on a genome-wide 2 kb re-binning of the same RS4;11 samples, and 9 of the 53 (17.0%) fall inside one of the table's own 21,461 TSS windows against 76.1% of the 1,390 held bins and 6.4% of the universe (Fisher gained-versus-held p = 1.2e-18).
Q118 re-analysis of GSE127507, instrument B: 2 kb genome-wide bins from the nine deposited .deduped.tdf files; analyses/q118-menin-at-the-human-index-fusion/results-instrumentB.json and results-instrumentB-tss-and-molm13.json, run 2026-09-23 · Iqbal Lab oncofusion knowledge graph 2026
What this evidence cannot show
The two instruments also differ in region size (6 kb versus 2 kb), in filtering (the table is unfiltered and input-subtracted by the depositors; the bins are blacklist- and input-filtered by this beat) and in the top fraction used (1% of 21,461 versus 0.1% of 1,389,514). So the 3-versus-53 gap is the combined effect of all of those, of which the TSS restriction is the one measured here. The universe figure is a 1-in-37 systematic sample. RefGene is not a complete transcript set, so 'not at a TSS' means 'not at a RefGene TSS in this table'.
Evidence that argues against this (1)
- Not all of the 3-versus-53 gap is the TSS restriction: the genome-wide instrument also uses a ten-fold smaller top fraction of a 65-fold larger universe, so it has far more regions in its candidate set to begin with. The measured part of the explanation is the 17% versus 76% TSS rate; the rest is not decomposed here. Iqbal Lab oncofusion knowledge graph 2026

Connected to

← related toFifty-three 2 kb bins gain MLL1-N-reactive signal under VTP-50469 in RS4;11, and the same 53 coordinates gain 1.77-fold in MOLM13 against a class-matched null of 1.01 with the MOLM13 input flat at 1.01 - so there IS a small, reproducible, input-controlled redistribution, and it carries no CpG-island signature
the deposited promoter table finds 3 of these and the genome-wide re-binning finds 53, because 83% are not at a RefGene TSS
← related toThe deposited GSE127507 processed signal table is an Excel export in which every value at or above 1,000 is a quoted comma-separated string, so a plain numeric parse silently zeroes 34,955 of 386,298 cells - and those cells are the entire top of all eighteen distributions
two independent ways the convenient deposited processed file misleads a re-user: 9% of its cells are unparseable at the top of every distribution, and its universe excludes 83% of the gained set
related to →The deposited GSE127507 processed signal table is an Excel export in which every value at or above 1,000 is a quoted comma-separated string, so a plain numeric parse silently zeroes 34,955 of 386,298 cells - and those cells are the entire top of all eighteen distributions
both are properties of the same deposited processed file
← supportsAt the 692 RS4;11 bins where an MLL1 N-terminus and menin co-occupy, VTP-50469 takes menin to 0.20 and DOT1L to 0.20, drops H3K79me2 to 0.66, and leaves MLL1-N-reactive material at 0.83 - four factors, one drug, one chromatin, and a clean ordering
the H3K79me2 arm is excluded from the depositors' own processed table by their own processing note, so this measurement exists only because the raw .tdf files were re-binned
← supportsTwo instruments on the same cell line give opposite answers about DOT1L under a menin inhibitor: the depositors' own TSS-window table puts DOT1L at 0.6638 beside menin's 0.2513, and 2 kb genome-wide bins put it at 0.2219 beside menin's 0.1657 - and only the genome-wide instrument can see the 86% of the co-occupied set that is not at a TSS
a second, independent measurement of the same blindness: 86.3% of the MOLM13 co-occupied bins fall outside every deposited ROSE TSS window, and the consequence is an opposite verdict about DOT1L