22 of the 37 promoters in the published B-ALL-specific peak class are the clustered protocadherin gamma locus - one 23.8 kb interval at chr5:141,473,710-141,497,491 - and it reads below an average promoter in all 27 samples
A gene list that looks like 37 genes is really 16 places in the genome, because 22 of the entries are neighbouring copies of one repeated gene family sitting under a single peak. That one place is quiet in every patient, so it drags the whole list down.
The same thing, for a scientist
The authors' `ALL_specific` class resolves to 37 promoters in the 19,404-promoter set. **22 of them (59.5%) are `PCDHGA1-12`, `PCDHGB1-7` and `PCDHGC3-5`** - the clustered protocadherin gamma locus at 5q31, where tandem variable exons share three constant exons. **All 22 come from a single author interval, chr5:141,473,710-141,497,491, 23.8 kb wide.** So the class is **16 independent genomic neighbourhoods, not 37 genes**, and one of those neighbourhoods supplies 59.5% of the promoters. It is also the quiet one. The PCDHG-only share runs **0.062x to 0.898x** - **never above an average promoter in any of the 27 samples** - while the other 15 promoters run **0.658x to 7.627x**. Removing PCDHG raises the B-lineage class from 1.039-2.474 to 1.998-5.119 in the B-lineage samples and the lineage contrast survives: AML median **1.828** against B-lineage **0.607** (Mann-Whitney two-sided **p = 0.0200**, against p = 0.0080 with PCDHG included), and **A73051 stays inside the HOXA-positive AML range under both versions**. So the reagent defect **does not change any verdict in this beat**. It is recorded because a 37-gene set that is really 16 loci will be quoted as n = 37 by the next person who uses it, and because the 15 non-PCDHG members - **AK5, BCL7A, BLNK, CD34, CDCA7, CEP78, CLEC14A, CLEC2B, DNTT, ETS2, GPR174, KLRF2, MPZL1, PRDM8, TM9SF3** - are the biologically legible part: `DNTT` (TdT), `BLNK` and `CD34` are B-precursor genes.
Where the inflation comes from
It is not an error in the authors' pipeline. Their construction intersects peak intervals with an hg38 gene list using bedtools intersect -wao and then explodes the result one row per overlapping gene. A 23.8 kb interval over a tandem gene cluster therefore produces 22 rows, all correct. The inflation appears only when the exploded gene column is read as a gene set, which is exactly what any downstream reuse does - including this one.
The same mechanism affects the other classes far less: 0 PCDHG members in AML_specific, and the 284 intervals with no annotated gene are lost rather than inflated. The general rule: after converting intervals to a gene set, count the distinct intervals, not the genes.
Both versions of every headline number
| with PCDHG (37 promoters) | without PCDHG (15 promoters) | |
|---|---|---|
| AML samples, AML/ALL median (n = 6) | 3.031 | 1.828 |
| AML range | 1.263 - 7.188 | 0.569 - 3.602 |
| B-lineage, AML/ALL median (n = 8) | 1.268 | 0.607 |
| B-lineage range | 0.874 - 1.987 | 0.422 - 1.033 |
| Mann-Whitney two-sided p | 0.0080 | 0.0200 |
| A73051 AML/ALL | 3.493 | 2.270 |
| A73051 inside the HOXA-positive AML range | yes | yes |
Denominator: depth-ADEQUATE, untreated, non-circular samples. The smallest attainable two-sided p for 6 against 8 is 2/3003 = 0.00067, so neither p is at its floor.
| class | ALL_specific (the authors' B-ALL-specific peak class) |
|---|---|
| n joined promoters | 37 |
| n pcdhg | 22 |
| pcdhg share | 0.5946 |
| pcdhg interval | chr5:141473710-141497491 (23781 bp), one author interval |
| effective loci | 16 |
| script | analyses/q148-what-the-hoxa-negative-blast-occupies/pcdhg148.py |
| status | run |
| result summary | 59.5% of the B-lineage class is one 23.8 kb protocadherin interval; no verdict changes when it is removed |
non pcdhg members
[ "AK5", "BCL7A", "BLNK", "CD34", "CDCA7", "CEP78", "CLEC14A", "CLEC2B", "DNTT", "ETS2", "GPR174", "KLRF2", "MPZL1", "PRDM8", "TM9SF3" ]
pcdhg share range across 27 samples
[ 0.062, 0.898 ]
measurements — the numbers behind the chart
[
{
"quantity": "fraction_of_class_from_one_interval",
"subject": "authors' ALL_specific peak class, promoter-joined",
"value": 0.595,
"units": "fraction",
"n": 37,
"universe": "37 joined promoters of the ALL_specific class, 19,404-promoter hg38 set",
"null": null,
"note": "22 PCDHG promoters from chr5:141473710-141497491; the class is 16 loci, not 37"
},
{
"quantity": "within_sample_promoter_share",
"subject": "the 22 PCDHG promoters, KMT2A-N",
"value": 0.898,
"units": "x an average promoter (maximum across 27 samples)",
"n": 27,
"universe": "GSE252378 KMT2A-fusion samples",
"null": 1,
"note": "minimum 0.062; never exceeds 1.0 in any sample, so it uniformly deflates the class mean"
},
{
"quantity": "mannwhitney_p_two_sided",
"subject": "AML/ALL without PCDHG, AML (n=6) vs B-lineage (n=8)",
"value": 0.02,
"units": null,
"n": 14,
"universe": "depth-ADEQUATE untreated non-circular GSE252378 samples",
"null": "smallest attainable two-sided p at 6 vs 8 is 0.00067",
"note": "0.0080 with PCDHG included; the verdict is unchanged"
}
]- A signature whose size is inflated by a tandem cluster will look well-powered and behave like a single locus, so any patient stratifier built on it inherits the failure modes of one 23.8 kb region.
- Is the protocadherin gamma cluster genuinely a KMT2A-fusion target in B-ALL, or does it enter this class because a 23.8 kb interval over a tandem array is easy to call as differential? Its share never exceeds 1.0x, which argues for the latter.
- How many other entries in this corpus quote a gene-set size that came from an exploded interval-to-gene join? The rule 'count intervals, not genes' has not been applied to any of them.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
22 of the 37 joined promoters in the authors' ALL_specific class are PCDHG* genes from one 23.8 kb interval at chr5:141,473,710-141,497,491, and that interval's share never exceeds 1.0x an average promoter in any of the 27 samples.
analyses/q148-what-the-hoxa-negative-blast-occupies - results148_pcdhg.json; author interval list from p148752_152985_Menin_Signal.csv · Q148 (Oncofusion Knowledge Graph) · this repository 2026What this evidence cannot show
The promoter set is RefSeq canonical TSS, so the 22 PCDHG promoters are 22 real annotated TSSs inside the cluster, not a duplication artefact in my join - the cluster genuinely has that many transcription start sites. The point is that they are not 22 independent tests.
Evidence that argues against this (1)
- Each PCDHG variable exon does have its own promoter and its own regulation, so treating them as one locus understates the cluster's complexity in the other direction. The defensible statement is that they are 22 non-independent promoters within one peak, not that they are one promoter. Janssens DH 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Removing the PCDHG promoters changes every headline magnitude in this beat but reverses no verdict: the AML-versus-B-lineage contrast survives at Mann-Whitney two-sided p = 0.0200 and A73051 remains inside the HOXA-positive AML range.
analyses/q148-what-the-hoxa-negative-blast-occupies - results148_pcdhg.json with_pcdhg / without_pcdhg blocks · Q148 (Oncofusion Knowledge Graph) · this repository 2026What this evidence cannot show
n = 6 against n = 8. Institution and lineage are the same split in this subset, so the contrast is a lineage OR institution contrast and cannot be attributed to either.
Evidence that argues against this (1)
- A result that survives removal of 59.5% of one class's members might also survive removal of the other 40.5%, in which case the class carries little information either way. That test was not run. Q148 (Oncofusion Knowledge Graph) 2026
the interval-scale reagent is smaller again: 20 ALL_specific intervals against 37 promoters
the volatile denominator is itself 59.5% one 23.8 kb interval
the reagent audit for that measurement's B-lineage class