Oncofusion

start / graph / concept-a-promoter-window-is-not-a-gene

Explainerstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q153-can-an-irx2-promoter-measurement-be-attributed

A promoter window is not a gene: what it means to say a protein 'sits on gene X's switch'

also called promoter window · divergent promoter · bidirectional promoter · attribution of a promoter measurement

Scientists cannot point at a gene's on-switch directly, so they draw a box of about two thousand DNA letters around where the gene starts and average the signal inside it. If two genes start close together facing opposite ways, one box covers both, and the number belongs to both genes at once.

The same thing, for a scientist

Almost every genome-wide occupancy statement of the form *"protein P occupies gene G's promoter"* is in fact a statement about a **fixed-width interval centred on G's annotated transcription start site** — 2 kb in this project, `TSS ± 1000`. Three things follow that a reader should hold at all times. **First, the window is a convention, not a biological object**: a real promoter may be 300 bp or 5 kb, and the number reported is a mean, so a narrow tall peak and a broad low one are indistinguishable. **Second, the window can contain other genes.** Human genes are frequently arranged head to head sharing one bidirectional promoter: Trinklein and colleagues found **more than 10% of 23,752 human genes** in divergently transcribed pairs with TSSs less than 1 kb apart, co-expressed in many cases, with the intervening segment initiating transcription in both directions and carrying elements that regulate both genes [E1]. In this project's own window set the figure is **21.6% of 19,404 windows** containing another gene's TSS within 1 kb [E2]. **Third — and this is the part that is easy to get wrong — the resulting ambiguity is asymmetric.** *No signal* in the window means no signal at the gene's promoter, because the promoter is a subset of what was measured. *Signal* in the window does not say which gene's promoter it is on [E2]. **And strand does not help**: chromatin occupancy assays such as CUT&RUN sequence DNA fragments, and a DNA fragment has no transcriptional strand, so the two genes of a divergent pair are read by the same fragments however deeply they are sequenced [E3]. The only routes out are a **narrower or base-resolution readout**, or a measurement over a region the two genes do not share — for a head-to-head pair, the **gene bodies are disjoint** and are therefore attributable [E2].

Go deeper
The four things a promoter number does not tell you
  1. Which gene. If another TSS is in the window, the gene name on the row is an annotation choice, not a measurement.
  2. Where in the window. A 2 kb mean of 4.0 could be a single 100 bp spike of 80 or a flat 4.0 across the whole interval.
  3. Whether the promoter is the right interval at all. A distal element 3 kb upstream is outside every claim built on a 2 kb window.
  4. Whether the signal is the protein you think. Antibody specificity, background and library depth are separate problems and this concept is about none of them.

What it does tell you, reliably: this interval of this chromosome carries, or does not carry, the signal. That is a real and useful statement. The discipline is to write it that way when the window is shared, and to name the gene only when the window is the gene's alone.

How to tell which case you are in, in one step

Take your gene's TSS. Ask whether any transcript of any other gene in the annotation starts within half your window width. Two rules learned the hard way in this project:

  • Ask the full annotation, not your own table. A one-row-per-gene, coding-transcripts-only table structurally cannot see an antisense or microRNA neighbour. AFF1's 2 kb window is 92.5% AFF1-AS1's, and AFF1-AS1 is not in this project's promoter table at all, because it has no NM_ transcript. Checking against the table returns a clean, confident no neighbour.
  • Use the strand-correct TSS. For a minus-strand gene the TSS is the higher coordinate of the gene body. Using the lower one puts the window at the far end of the gene — 16 kb away for IRX2 — and the resulting number is about unrelated sequence while looking completely normal.
Why divergent pairs exist at all

This is not a nuisance of annotation; it is how a large slice of the mammalian genome is built. A single nucleosome-depleted, CpG-rich segment can load polymerase in both directions, so one regulatory element serves two genes. Trinklein and colleagues showed the arrangement is often conserved between human and mouse orthologs, which is the argument that it is selected for rather than incidental 1.

The consequence for interpretation is genuinely double-edged. If two genes share a promoter and are co-regulated, then "which gene does the signal belong to" may be the wrong question — the element regulates both, and an intervention at it will affect both whatever the answer. But Trinklein also reports pairs that are anti-regulated, so co-location does not license assuming co-regulation either. Both readings have to stay open until something measures the two transcripts.

Facts and figures
window convention in this projectTSS +/- 1000, width 2000, strand-aware (minus-strand TSS = txEnd)
asymmetry rulea null over a superset entails a null over its subset; a positive over a superset does not entail a positive over its subset. Trust a negative at a shared window; do not attribute a positive.
routes that separate two nearby tsss
[
  "base-resolution coverage from the same bigWigs",
  "a window over the disjoint gene bodies",
  "a strand-specific TRANSCRIPT assay (which measures expression, not occupancy)"
]
routes that do not
[
  "any fixed-width promoter mean wider than the TSS separation",
  "strand information from a chromatin occupancy assay, which has none",
  "deeper sequencing of the same 2 kb mean"
]
measurements
[
  {
    "quantity": "fraction_of_promoter_windows_with_another_genes_tss",
    "subject": "this project's 19,404 promoter windows, any refGene gene's TSS within 1 kb",
    "value": 21.6,
    "units": "percent",
    "n": 19404,
    "universe": "19,404 rows of the project promoter table against 78,362 refGene transcripts on chr1-22,X,Y",
    "null": null,
    "note": "published comparator: >10% of 23,752 genes in divergent pairs under 1 kb (Trinklein 2004, PMID 14707170)"
  }
]
How this could bite us back
  • The asymmetry means a promoter-scale screen produces false gene attributions and never false clears, so every error it makes points toward starting a programme rather than stopping one. That is the worst possible direction for an error in target nomination, and it is invisible unless the neighbour census is run.
  • For a shared promoter, no DNA-level intervention can be gene-selective: the sequence is the same sequence. A therapeutic hypothesis that requires silencing one member of a divergent pair and not the other has no route at the DNA level and must move to the transcript or the protein.
Still open
  1. What window width minimises the trade between attribution and noise for this deposit? Narrowing to 500 bp roughly halves the fraction of ambiguous windows but makes every mean noisier, and the trade has not been measured here.
  2. Does fragment-end asymmetry carry enough orientation information to separate a divergent pair in CUT&RUN? The founding paper reports directional contact mapping at CTCF, so the question is not obviously closed, and nobody in this project has tried it.
  3. For how many of this graph's promoter-scale statements is the gene name determined by the data rather than by the annotation? The census gives the exposure; the per-claim audit has never been run.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    More than 10% of 23,752 human genes are arranged as divergently transcribed pairs with transcription start sites less than 1,000 bp apart; many such pairs are co-expressed and some are anti-regulated, the intervening promoter segment often initiates transcription in both directions and contains elements shared by both genes, and the arrangement is frequently conserved in mouse orthologs.

    An abundance of bidirectional promoters in the human genome · Trinklein ND, Aldred SF, Hartman SJ, Schroeder DI, Otillar RP, Myers RM · Genome Research 2004
    PMID 14707170DOI
    What this evidence cannot show

    Abstract read in full via efetch XML on 2026-09-24. A 2004 analysis on full-length cDNA alignments to a then-current assembly, so the gene set is smaller and more conservative than a modern annotation; the >10% is a lower bound. The co-expression and bidirectional-initiation results are reported in the abstract for 'many' and 'many of the promoter segments' without a fraction, so this entry does not quote one. No claim is made here about IRX2, AFF1 or any leukaemia gene from this paper.

    Evidence that argues against this (1)
    • This is a 2004 genome-wide architecture paper and says nothing about whether a divergent pair is co-regulated in any particular cell type, let alone in KMT2A-rearranged infant leukaemia. It establishes that the arrangement is common and selected for; it cannot be cited for what any specific pair does. NCBI Gene 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    In this project's own 19,404 promoter windows, 21.60% contain another gene's TSS within 1 kb, the nearest neighbour is on the opposite strand in 82.3% of those cases, and 43.87% of all occupancy calls the project's rule makes fall on such windows; for a head-to-head pair the gene bodies are disjoint and therefore attributable.

    q153 neighbour census and shared-window enrichment over the 19,404-promoter window table · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026
    What this evidence cannot show

    All figures are for a 2 kb window convention against the UCSC hg38 refGene snapshot with dataTime 2020-08-17; a narrower window or a different annotation gives different numbers, and the tiers 12.66% (200 bp) to 24.02% (2 kb) are reported so the dependence is visible. The 43.87% is a property of one occupancy rule on one deposit and does not transfer to other rules or datasets.

    Evidence that argues against this (1)
    • The 21.60% is inflated relative to a strict reading by 534 -AS/-DT and 402 LINC/LOC neighbours, which a sceptic may regard as annotations of the same locus. Discounting all of them leaves 15.25%, which is the figure this concept should be read with if the reader is unwilling to count non-coding annotations as genes. Iqbal Lab oncofusion knowledge graph et al. 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    CUT&RUN measures protein-DNA binding by antibody-targeted MNase cleavage and paired-end sequencing of released DNA fragments, so the readout is DNA and carries no transcriptional strand; the method is nonetheless explicitly high-resolution, which is why base-resolution re-analysis rather than a new assay is the route to separating two nearby start sites.

    PMID 28079019DOI
    What this evidence cannot show

    Abstract read in full via efetch XML on 2026-09-24. `strand-specific` is in expected_absent_terms deliberately: the paper makes no strand-specificity claim because the assay has none, and that absence is the support. The inference that a DNA-fragment readout cannot assign a fragment to a transcriptional strand is the author's, not the paper's.

    Evidence that argues against this (1)
    • The same abstract reports that CUT&RUN combined with native ChIP-seq mapped DIRECTIONAL long-range contact sites at CTCF, so 'no strand information' is not the same as 'no directional information' - fragment-end asymmetry can carry orientation for some factors. That does not rescue attribution between two TSSs 215 bp apart, but it means the blanket phrase 'CUT&RUN has no directionality' would be wrong. Skene PJ et al. 2017
Connected to