Oncofusion

start / graph / ana-q148-the-gmp-over-b-lineage-contrast-tracks-the-samples-own-lineage-and-institution-not-its-hoxa-status

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q148-what-the-hoxa-negative-blast-occupies

The GMP-over-B-lineage occupancy contrast tracks the sample's own lineage - and institution, which is the identical split in this deposit - and not its HOXA status

The myeloid-versus-lymphoid signal ratio separates the myeloid patients from the lymphoid ones, which is what you would expect and tells you nothing new. It does not separate the patients whose fusion has left the HOXA genes from the ones whose fusion is still there.

The same thing, for a scientist

On the 15 depth-ADEQUATE, untreated, non-circular GSE252378 samples: | split | n | AML/ALL median | range | Mann-Whitney two-sided p | |---|---:|---:|---|---:| | lineage AML | 6 | **3.031** | 1.263 - 7.189 | **0.0080** vs B-lineage | | lineage B | 8 | **1.268** | 0.874 - 1.987 | | | Fred Hutch | 6 | **3.031** | 1.263 - 7.189 | **0.0087** vs Texas | | Texas Children's | 6 | **1.182** | 0.874 - 1.791 | | | **HOXA-negative (< 1.0x)** | **2** | 2.216 | **0.940 - 3.493** | **no statistic** | | HOXA-positive (>= 2.0x) | 13 | 1.709 | 0.874 - 7.189 | | **The lineage split and the institution split are the SAME six samples.** All six depth-adequate Fred Hutch samples are paediatric AML and all six depth-adequate Texas Children's samples in this set are B-lineage. **They cannot be separated in this deposit**, so neither p licenses a claim about institution or about lineage individually - only about that stratum. The HOXA split does nothing. Its two members are **p454 at 0.940** (KMT2A::AFF1, infant B-ALL, Texas Children's) and **A73051 at 3.493** (KMT2A-AF10, paediatric AML, Fred Hutch) - one at the bottom of the whole cohort and one in the middle, each sitting with its own lineage. **n = 2 against n = 13 has a smallest attainable exact two-sided p of 0.019 and no statistic is computed for it.**

Go deeper
Why this is the confound that decides the beat

The GMP-like class is an AML programme, and the one depth-adequate HOXA-negative non-AFF1 blast, A73051, is an AML. So a high GMP score in A73051 is doubly explained before any redistribution hypothesis is needed. The pre-registered discriminator was whether A73051's ratio exceeds the other Fred Hutch paediatric AMLs, all of which are HOXA-POSITIVE. It does not: rank 3 of 6, inside 1.263-7.189, and the same is true with the protocadherin cluster removed (2.270 inside 0.569-3.602).

Put the other way round: the two HOXA-negative samples behave like their lineages, not like each other. p454 is a HOXA-negative B-ALL with a low GMP ratio (0.940) and A73051 is a HOXA-negative AML with a mid GMP ratio. Whatever HOXA-negativity is, it does not come with a GMP-like address.

The institution batch effect this sits on top of

GSE252378 carries a measured institution batch effect - unrelated-pair median excess 1.456 within institutions (n = 83 pairs) against 1.372 across (n = 236 pairs), Mann-Whitney U = 11,930.5, ONE-SIDED p = 0.0016 - the corpus entry that computed it states the tail explicitly and this beat carries it as one-sided, not two-sided - so institution is not a nuisance label in this deposit, it is a measured source of between-sample similarity. With institution and lineage perfectly aligned in the depth-adequate subset, the honest report is a single stratum name: Fred Hutch paediatric AML versus Texas Children's infant/paediatric B-ALL.

The two strata that break the alignment are both too small to help: Seattle Children's (R3595, R3652, n = 2, both B-lineage, AML/ALL 1.545 and 1.987) and St. Jude (TB971992, n = 1, T-lineage, 1.709). Both sit between the two big strata, which is weakly reassuring and carries no statistic.

Facts and figures
n15
populationdepth-ADEQUATE, untreated, non-circular GSE252378 samples
institution and lineage are the same splittrue
scriptanalyses/q148-what-the-hoxa-negative-blast-occupies/controls148.py
statusrun
result summarythe contrast is a lineage/institution contrast; HOXA status does nothing
by lineage
T1.000T: 1AML6.000AML: 6B8.000B: 8
by institution
St. Jude Children's Hospi…1.000St. Jude Children's Hospital: 1Seattle Children's Hospit…2.000Seattle Children's Hospital: 2Fred Hutch Cancer Center6.000Fred Hutch Cancer Center: 6Texas Children's Hospital6.000Texas Children's Hospital: 6
measurements
AML_over_ALL2.000AML_over_ALL: 2mannwhitney_p_two_sided12.0mannwhitney_p_two_sided: 12mannwhitney_p_two_sided14.0mannwhitney_p_two_sided: 14
by lineage — the numbers behind the chart
{
  "AML": {
    "n": 6,
    "median": 3.031,
    "range": [
      1.263,
      7.189
    ]
  },
  "B": {
    "n": 8,
    "median": 1.268,
    "range": [
      0.874,
      1.987
    ]
  },
  "T": {
    "n": 1,
    "value": 1.709
  }
}
by institution — the numbers behind the chart
{
  "Fred Hutch Cancer Center": {
    "n": 6,
    "median": 3.031
  },
  "Texas Children's Hospital": {
    "n": 6,
    "median": 1.182
  },
  "Seattle Children's Hospital": {
    "n": 2,
    "median": 1.766
  },
  "St. Jude Children's Hospital": {
    "n": 1,
    "median": 1.709
  }
}
hoxa negative members
{
  "p454": 0.94,
  "A73051": 3.493
}
measurements — the numbers behind the chart
[
  {
    "quantity": "mannwhitney_p_two_sided",
    "subject": "AML/ALL, lineage AML (n=6) vs B-lineage (n=8)",
    "value": 0.008,
    "units": null,
    "n": 14,
    "universe": "depth-ADEQUATE untreated non-circular GSE252378 samples, 40/37 promoters, KMT2A-N",
    "null": "attainable floor 0.00067",
    "note": "identical to the institution split"
  },
  {
    "quantity": "mannwhitney_p_two_sided",
    "subject": "AML/ALL, Fred Hutch (n=6) vs Texas Children's (n=6)",
    "value": 0.0087,
    "units": null,
    "n": 12,
    "universe": "same samples",
    "null": "attainable floor 0.0022",
    "note": "the same six samples as the AML arm; institution and lineage are not separable here"
  },
  {
    "quantity": "AML_over_ALL",
    "subject": "the two HOXA-negative samples p454 and A73051",
    "value": [
      0.94,
      3.493
    ],
    "units": "ratio",
    "n": 2,
    "universe": "same 15 samples; HOXA share below 1.0x carried from q147",
    "null": "HOXA-positive median 1.709, range 0.874-7.189",
    "note": "n=2 vs n=13, smallest attainable exact two-sided p 0.019; NO statistic computed"
  }
]
How this could bite us back
  • A stratifier that fires on 'GMP-like occupancy' will select AML patients, which is a lineage selection wearing a mechanism's name, and in this deposit those are also the patients from one institution.
Still open
  1. Is there a deposit in which KMT2A-rearranged AML and B-ALL samples were profiled at the SAME institution with the same protocol? Without one, every lineage claim in this deposit is an institution claim as well.
  2. Would a HOXA-negative KMT2A-r B-ALL other than p454 behave like p454? n = 1 per lineage on the HOXA-negative side is the binding constraint on this whole question.
Every source behind this entry (2)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Across 15 depth-adequate untreated non-circular GSE252378 samples the AML/ALL occupancy ratio separates AML from B-lineage at Mann-Whitney two-sided p = 0.0080 and Fred Hutch from Texas Children's at p = 0.0087, and those are the same six samples on each side.

    What this evidence cannot show

    n = 6 against n = 6 and n = 6 against n = 8. Both p values are above their attainable floors (0.0022 and 0.00067). Lineage and institution are the identical split, so neither variable is controlled for the other and no claim about either alone is made.

    Evidence that argues against this (1)
    • The same ratio is 93.5% denominator-driven, so 'AML samples have a higher AML/ALL ratio' is largely 'AML samples have less B-lineage-class occupancy', which is unsurprising and is not evidence that the GMP-like programme is occupied in them. Q148 (Oncofusion Knowledge Graph) 2026
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    The two HOXA-negative samples in that set sit at 0.940 (p454, B-ALL) and 3.493 (A73051, AML), each with its own lineage, and the HOXA split cannot carry a statistic at n = 2 against n = 13.

    What this evidence cannot show

    HOXA-negativity is defined here as the q147 HOXA share below 1.0x, carried unchanged. Two other samples are HOXA-negative but depth-MARGINAL or INADEQUATE (TB920488, TB986292) and one more is drug-exposed (152985); including them would not create a statistic either.

    Evidence that argues against this (1)
    • The published subgroup model is an EXPRESSION dichotomy in infant t(4;11) only, so extending it to a six-partner occupancy cohort was never the authors' claim, and a null here is a null about the extension. Symeonidou V et al. 2021
Connected to