21.6% of this project's 19,404 promoter windows contain another gene's TSS, so an attribution caveat is corpus-wide and not an IRX2 peculiarity
also called promoter neighbour census · shared promoter windows hg38 · 19404 promoter attribution
This project measures each gene by taking the average signal in a 2,000-letter stretch around its start. For one gene in five, a different gene starts inside that same stretch, so the number belongs to two genes at once. IRX2 is not a special case - it is an ordinary member of a large group.
The same thing, for a scientist
Census over the exact window table this project scores (`C:\Temp\q145irx1\promoters.tsv`, 19,404 rows, each window asserted at load to be TSS±1000 with `end-start` = 2,000 and every symbol unique). Against full hg38 refGene, **4,191 of 19,404 = 21.60%** of these windows contain the TSS of a **different gene symbol** within 1 kb, i.e. they share **at least half their sequence** with another gene's window [E1]. The tiers are 12.66% at 200 bp, 18.64% at 500 bp, 21.60% at 1 kb and 24.02% at 2 kb [E1]. Restricted to neighbours that are also rows of the project's own one-row-per-symbol table, the figure is **14.37%** — the gap between 14.37% and 21.60% is the contaminating genes the project's table never sees [E1]. The nearest neighbour is on the **opposite strand in 3,451 of 4,191 (82.3%)**, i.e. this is overwhelmingly the divergent-promoter architecture Trinklein and colleagues described in 2004, where they found **>10% of genes** in head-to-head pairs with TSSs under 1 kb apart across 23,752 genes [E2]. Our figure is larger for a reason that must be stated: refGene 2020 carries `-AS1`, `-DT` and `LINC` annotations that a 2004 full-length-cDNA set did not. **Discounting all of them still leaves the problem intact:** the nearest neighbour is an *ordinary named gene* — not antisense, not a lincRNA, not a microRNA, not a readthrough — in **2,959 of the 4,191**, which is **15.25% of all 19,404 windows** [E1]. **The pre-registered threshold was 10%; every reading clears it.** Within this graph specifically, of 572 `gene-*` and `protein-*` entries, 515 resolve to a promoter row and **141 (27.4%) share ≥50% of their window with another gene, 116 (22.5%) share ≥75% and 70 (13.6%) share ≥90%** [E3]. The graph's own entries are **more** promoter-crowded than an average promoter (27.4% against a 21.60% base rate, one-sided binomial p = 0.0011), which is a real enrichment and most likely reflects that well-studied genes sit in gene-dense CpG-island regions rather than anything about this disease [E3].
The census, with its denominator and what it counted
Denominator: 19,404, the exact rows this project extracts bigWig means over. Neighbour = any hg38 refGene transcript TSS whose name2 differs from the row's symbol (so alternative promoters of the same gene never count).
| distance | (a) neighbour also in the 19,404 table | (b) any refGene gene |
|---|---|---|
| within 200 bp (windows share ≥90%) | — | 12.66% |
| within 500 bp (share ≥75%) | 12.35% | 18.64% |
| within 1 kb (share ≥50%) | 14.37% | 21.60% |
| within 2 kb (share anything) | 15.71% | 24.02% |
At 1 kb, census (b): nearest neighbour divergent (opposite strand) 3,451, tandem (same strand) 740.
What the neighbour is (nearest, 1 kb, census b), because a census that does not say what it counted is a ruler and not a measurement:
| class of the nearest neighbour | n | of 19,404 |
|---|---|---|
| ordinary named gene | 2,959 | 15.25% |
-AS/-DT antisense or divergent-transcript symbol | 534 | 2.75% |
LINC/LOC | 402 | 2.07% |
small RNA (MIR, SNOR, RNU, …) | 150 | 0.77% |
| readthrough / other hyphenated symbol | 143 | 0.74% |
intronic transcript (-IT) | 3 | 0.02% |
The conservative reading is the one that matters. Throw away every antisense, lincRNA, microRNA and readthrough annotation — the categories a sceptic would call artefacts — and 15.25% of this project's promoter windows still contain an ordinary named gene's start site.
Where this graph's own entries fall
572 entries are typed gene or protein with a gene-/protein- id; 515 resolve to a row of the promoter table. The 57 that do not are complex entries (FACT-COMPLEX, NELF-COMPLEX), *-DOMAINS protein-architecture entries, six co-mutation entries, and five genuine symbol misses (AFG2A, AFG2B, BOD1L, SKIC8, KDM3C — all recent HGNC renames, and IRX2-DT, which the project's table carries under its previous symbol C5orf38).
| shared fraction of the 2 kb window | entries | share | genome base rate | one-sided p |
|---|---|---|---|---|
| any sharing | 162 | 31.5% | 24.02% | — |
| ≥50% | 141 | 27.4% | 21.60% | 0.0011 |
| ≥75% | 116 | 22.5% | 18.64% | 0.015 |
| ≥90% | 70 | 13.6% | 12.66% | n.s. |
The enrichment is real at the ≥50% tier and fades at ≥90%. The honest interpretation is a confound, not a discovery: this graph's genes are well-studied genes, and well-studied genes are concentrated in gene-dense, CpG-island-rich regions. The number is recorded because it sets the size of the caveat, not because it says anything about leukaemia.
The 25 worst in this graph, by shared fraction — read them rather than counting them:
LEPR 100.0% (LEPROT, 0 bp, tandem) · ELP5 99.9% (CTDNEP1) · U2AF1 99.6% (U2AF1L5) · SUPT6H 99.2% (SDF2) · CSNK2B 99.2% (GPANK1) · FTSJ3 99.0% (PSMC5) · RBBP8 99.0% (MIR4741) · FOXM1 98.8% (RHNO1) · RAD21 98.8% (MIR3610) · KDM5A 98.7% (CCDC77) · LARP7 98.4% (ZGRF1) · TOP2B 98.4% (MIR4442) · ICE2 98.2% (RORA-AS1) · SEPTIN2 97.8% (HDLBP) · EAF1 97.5% (METTL6) · METTL14 97.5% (METTL14-DT) · UBAP2L 97.5% (C1orf43) · TOPBP1 97.2% (TF) · KMT2E 96.8% (KMT2E-AS1) · RAD51 96.6% (RAD51-AS1) · PES1 96.5% (TCN2) · PRMT7 96.4% (SLC7A6OS) · NIPBL 96.3% (NIPBL-DT) · EP300 96.0% (MIR1281) · BRD9 95.9% (TRIP13).
LEPR/LEPROT share a TSS to the base. SUPT6H, EAF1, ICE2, LARP7, KMT2E and EP300 are elongation and chromatin machinery this project reasons about constantly.
What is clean, which is the part that protects the corpus's main claims
The entries this graph's central arguments rest on have no other TSS within 5 kb and are therefore fully attributable at promoter scale: KMT2A, MEN1, MLLT1, MLLT3, MLLT10, ELL, AFF4, DOT1L, BRD4, WDR5, KAT7, PSIP1, CBX8, FLT3, BCL2 and IRX1.
IRX1 is clean. The brief that commissioned this beat suggested the same problem might affect IRX1; it does not — the nearest other refGene TSS is more than 5 kb away, and the q145/q147 IRX1 occupancy results carry no attribution caveat from this direction.
HOXA9's nearest neighbour is HOXA10-AS at 3,361 bp, outside the window, so the HOXA positive control is clean too.
The contaminated ones among the load-bearing entries are AFF1 (92.5%, AFF1-AS1), CDK9 (94.2%, MIR2861) and PAF1 (86.5%, MED29), each of which has its own entry in this beat or is named there.
| denominator | 19404 |
|---|---|
| denominator source | C:\Temp\q145irx1\promoters.tsv, the exact window table q145/q147/q151 extract bigWig means over |
| annotation | hg38 refGene, UCSC snapshot dataTime 2020-08-17, 78,362 transcripts on chr1-22,X,Y |
| prereg threshold | if the 1 kb full-annotation figure exceeds 10% of 19,404, the corpus needs the number and every promoter-scale claim inherits an attribution caveat |
| prereg outcome | THRESHOLD CLEARED at 21.60%, and still cleared at 15.25% after discounting every antisense, lincRNA, microRNA and readthrough neighbour |
| members file | analyses/q153-can-an-irx2-promoter-measurement-be-attributed/q153_census_1kb_members.tsv |
clean load bearing entries
[ "KMT2A", "MEN1", "MLLT1", "MLLT3", "MLLT10", "ELL", "AFF4", "DOT1L", "BRD4", "WDR5", "KAT7", "PSIP1", "CBX8", "FLT3", "BCL2", "IRX1", "HOXA9" ]
contaminated load bearing entries — the numbers behind the chart
{
"AFF1": 92.45,
"CDK9": 94.2,
"PAF1": 86.5,
"IRX2": 89.2,
"MEIS1": 17.3
}measurements — the numbers behind the chart
[
{
"quantity": "fraction_of_promoter_windows_with_another_genes_tss",
"subject": "this project's 19,404 promoter windows, neighbour = any hg38 refGene TSS of a different gene symbol within 1 kb",
"value": 21.6,
"units": "percent",
"n": 19404,
"universe": "19,404 rows of C:\\Temp\\q145irx1\\promoters.tsv, each TSS+/-1000; neighbour set = 78,362 refGene transcripts on chr1-22,X,Y",
"null": null,
"note": "4,191 of 19,404; tiers 12.66% at 200 bp, 18.64% at 500 bp, 24.02% at 2 kb. Published comparator: Trinklein 2004 found >10% of 23,752 genes in divergent pairs under 1 kb"
},
{
"quantity": "fraction_of_promoter_windows_with_another_genes_tss",
"subject": "same, but the neighbour must be an ORDINARY NAMED GENE (not -AS/-DT, not LINC/LOC, not small RNA, not a readthrough symbol)",
"value": 15.25,
"units": "percent",
"n": 19404,
"universe": "same 19,404 windows; 2,959 of the 4,191 nearest neighbours at 1 kb are ordinary named genes",
"null": null,
"note": "the figure that survives the annotation-artefact objection; it is the number to quote to a sceptic"
},
{
"quantity": "fraction_of_promoter_windows_with_another_genes_tss",
"subject": "same census restricted to neighbours that are themselves rows of the 19,404-row table",
"value": 14.37,
"units": "percent",
"n": 19404,
"universe": "2,788 of 19,404; the gap to 21.60% is contaminating genes the project's one-row-per-symbol table cannot see",
"null": null,
"note": "the within-set census is the one an analysis using only this table would notice; it understates the problem by 7.2 points"
},
{
"quantity": "fraction_of_graph_entries_with_a_ge50pct_shared_promoter_window",
"subject": "this graph's gene- and protein-typed entries",
"value": 27.38,
"units": "percent",
"n": 515,
"universe": "515 of 572 gene/protein entries that resolve to a promoter row; 141 share >=50% of their 2 kb window with another gene's window",
"null": 21.6,
"note": "one-sided binomial p = 0.0011 against the genome-wide base rate; interpreted as a study-bias confound (well-studied genes are in gene-dense regions), not as biology. The enrichment is absent at the >=90% tier"
},
{
"quantity": "fraction_divergent_among_shared_windows",
"subject": "orientation of the nearest neighbour for the 4,191 windows with a TSS within 1 kb",
"value": 82.3,
"units": "percent",
"n": 4191,
"universe": "3,451 opposite-strand against 740 same-strand nearest neighbours",
"null": null,
"note": "head-to-head divergent architecture dominates, as Trinklein 2004 reported; a divergent pair has DISJOINT gene bodies, which is what makes the attribution answer scale-dependent"
}
]- One entry in five in a promoter-scale screen is a two-gene readout, so a target nominated from such a screen has roughly a one-in-five chance of being the wrong member of a pair. A programme that goes straight from 'occupied promoter' to 'knock this gene down' will, at that rate, eventually knock down the neighbour's partner and get a phenotype it attributes to the wrong gene.
- The same applies in reverse to safety: a CRISPRi or enhancer-deletion strategy aimed at a gene in the 21.60% silences two genes, and the off-target one is by construction the gene whose promoter overlaps - i.e. the one least likely to be in the safety panel.
- How many of this corpus's promoter-scale OCCUPANCY CLAIMS are on a window in the 21.60%? The census names the windows; nobody has yet crossed it against the entries that assert a promoter-scale result, and that cross-tab is the number a reader actually needs.
- Would a narrower window fix it? At 500 bp the fraction of windows containing another TSS falls to 12.66% (200 bp tier), but a 500 bp mean is noisier and the q147 rule's floor was calibrated at 2 kb. Nobody has measured the trade.
- Does the enrichment of this graph's entries (27.4% against 21.60%) survive matching on CpG-island status and local gene density? It is treated here as a confound on those grounds without the matched analysis having been run.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
Of the 19,404 promoter windows in this project's window table, 4,191 (21.60%) contain the TSS of a different gene symbol within 1 kb against full hg38 refGene; the nearest such neighbour is on the opposite strand in 3,451 cases (82.3%) and is an ordinary named gene rather than an antisense, lincRNA, microRNA or readthrough annotation in 2,959 cases (15.25% of all 19,404).
q153 neighbour census over the 19,404-promoter window table, hg38 refGene · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026What this evidence cannot show
The denominator is this project's table, not 'all human promoters': it is one row per gene symbol, 19,404 symbols, built from refGene, and it reproduces from refGene under an 'earliest NM_ TSS on the gene's own strand' rule for 18,525 of 19,404 rows (95.5%) - the 879 that do not reproduce are multi-locus symbols and tie-breaks, and the census uses each row's OWN stored TSS as the query point, so that mismatch does not enter the result. The neighbour set is hg38 refGene as snapshotted by UCSC with dataTime 2020-08-17, downloaded by an earlier beat; a newer annotation with more lncRNAs would give a LARGER figure, so 21.60% is a floor. Genes whose symbol appears at two loci are handled by excluding the symbol, never by position, so an alternative promoter of the same gene is never counted as a neighbour. The identities of all 4,191 are persisted in q153_census_1kb_members.tsv, not only the count.
Evidence that argues against this (1)
- A large part of the census is annotation-dependent: 534 of the 4,191 nearest neighbours are -AS/-DT symbols and 402 are LINC/LOC, which some would regard as annotations of the same locus rather than distinct genes. The figure that survives that objection is 2,959 (15.25%), and that number is reported beside the 21.60% for exactly this reason. Trinklein ND et al. 2004
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Published independently of this project: analysing distances between 23,752 human genes, more than 10% are divergently transcribed pairs whose transcription start sites are separated by less than 1,000 base pairs, many of them co-expressed and some anti-regulated, with the intervening segment often initiating transcription in both directions.
An abundance of bidirectional promoters in the human genome · Trinklein ND, Aldred SF, Hartman SJ, Schroeder DI, Otillar RP, Myers RM · Genome Research 2004What this evidence cannot show
Abstract read in full via efetch XML on 2026-09-24. This is a 2004 analysis built on full-length cDNA alignments to a then-current human genome assembly, so its gene set is smaller and cleaner than hg38 refGene and its >10% is a lower bound by construction. It counts divergent pairs only; this beat's 21.60% counts any orientation, and the divergent-only subset here is 17.8% (3,451/19,404), which is the number directly comparable to Trinklein's.
Evidence that argues against this (1)
- Trinklein and colleagues report that many bidirectional pairs are CO-EXPRESSED and share regulatory elements. If that holds at IRX2/IRX2-DT then the two genes are not merely confounded in the measurement, they may be genuinely co-regulated - in which case separating them is not only technically hard, it may not be the biologically meaningful question. That possibility is not tested anywhere in this corpus. NCBI Gene 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Of 572 gene- and protein-typed entries in this graph, 515 resolve to a row of the promoter table, and 141 (27.4%) have a window sharing at least half its sequence with another gene's window against a genome base rate of 21.60% (one-sided binomial p = 0.0011); KMT2A, MEN1, MLLT1, MLLT3, MLLT10, ELL, AFF4, DOT1L, BRD4, WDR5, KAT7, PSIP1, CBX8, FLT3, BCL2 and IRX1 all have no other refGene TSS within 5 kb.
q153 per-entry promoter-window sharing census over this graph's gene and protein entries · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026What this evidence cannot show
Symbols are taken from data.gene_primary where present and otherwise derived from the node id by stripping the type prefix and uppercasing, which is why 57 entries do not resolve - complex entries, *-DOMAINS architecture entries, co-mutation entries and five HGNC renames. The first version of this census counted only type=='gene' and therefore MISSED gene-aff1, gene-kmt2a, gene-men1 and gene-hoxa9, which are typed 'protein' in this graph: the highest-degree entries were excluded by a type filter, and the numbers here are from the corrected run over both types. The enrichment against the base rate is almost certainly a study-bias confound (well-studied genes sit in gene-dense CpG-island regions) and is not interpreted as biology.
Evidence that argues against this (1)
- The enrichment disappears at the tier that matters most for attribution: at >=90% sharing the graph's entries are at 13.6% against a 12.66% base rate, which is not distinguishable. So the claim 'this graph is unusually exposed' holds only at the moderate tier, and the worst-affected entries are no more common here than anywhere in the genome. Iqbal Lab oncofusion knowledge graph et al. 2026
the census that gives the concept its size: 4,191 of 19,404 windows, 15.25% even after discounting every non-coding annotation
- refutes →IRX1 — Iroquois homeobox 1, the gene that names the HOXA-negative half of infant t(4;11) ALL
refutes the dispatch premise that IRX1 may share this problem: IRX1 has no other refGene TSS within 5 kb and its promoter window is fully attributable
the IRX1 and HOXA9 windows both carry no other gene's TSS, so this beat's attribution caveat does not touch the q145 axis result
the worked case for why the within-table census (14.37%) understates the full-annotation census (21.60%): the contaminating neighbour has no row
the 21.60% base rate this 43.87% is measured against
the 21.6% census was run against a 2020 snapshot, so it is a floor; this locus gains a third TSS under the current release
would re-run Q153's census against the current annotation rather than the 2020 snapshot it was computed on
independently reproduces Q153's census on a different release with a different implementation: 14.19% / 22.63% against 14.37% / 21.60%