start / graph / ana-q147-eight-of-twenty-seven-gse252378-samples-fail-a-preregistered-depth-or-readability-gate

AnalysisestablishedQ147-hoxa-irx1-across-partners

Eight of 27 GSE252378 samples fail a pre-registered depth or readability gate, and two of them are the only HOXA-negative non-AFF1 candidates

Some samples in this public dataset were not sequenced deeply enough to say that a gene is unoccupied - a blank there means 'not measured', not 'not there'. Eight of the 27 samples are in that category, and they were set aside before the main question was asked.

The same thing, for a scientist

Two gates were fixed in the pre-registration before any signal was read. **Depth.** `p_dz` = fraction of the 19,404 promoters reading zero in KMT2A-N1, times the same fraction in KMT2A-N2. Cuts: ADEQUATE <= 0.10, MARGINAL <= 0.20, INADEQUATE above. Outcome: **19 ADEQUATE, 3 MARGINAL (A93949, TB051939, TB920488), 5 INADEQUATE (TB000694, TB040576, A80956, TB986292, TB995291)** [E1]. The range is 0.0000 to 0.4934 - a 3,000-fold spread inside one deposit and one protocol. **Readability.** MEIS1 or CDK6 must clear 2.0x an average promoter in both KMT2A-N libraries, or the library is not reporting fusion occupancy and no locus call is made from it. Outcome: **2 failures, TB000694 and 152985** [E1]. PAX5 was declared in advance as B-lineage-only and NOT a gate, which matters because 11 of the 27 samples are AML - A70498's PAX5 is 0.21-0.56x while its MEIS1 is 10.6x. **The consequence for the headline.** Four of the 18 non-AFF1 samples have IRX1 UNMEASURED (exact zero in both N antibodies at inadequate depth): A80956, A93949, TB000694, TB995291. The 0-of-18 is therefore **12 fully adequate samples plus 6 weaker ones**, and computed on the 12 alone the rule-of-three bound is **25.0%** rather than 16.7% [E1]. That is the honest version of the bound and it is stated in both forms.

Go deeper

The gate table

sample	partner	zero N1	zero N2	p_dz	depth	readable
148752	ENL	0.000	0.003	0.0000	ADEQUATE	yes
A89691	AFDN	0.082	0.133	0.0108	ADEQUATE	yes
A73051	AF10	0.125	0.096	0.0120	ADEQUATE	yes (MEIS1 only)
R3595	EPS15	0.075	0.160	0.0120	ADEQUATE	yes
p279	AFF1	0.315	0.060	0.0191	ADEQUATE	yes
...
A93949	AF9	0.229	0.695	0.1590	MARGINAL	yes
TB040576	AF9	0.382	0.536	0.2051	INADEQUATE	yes
TB995291	ENL	0.417	0.571	0.2381	INADEQUATE	yes
A80956	AFDN	0.368	0.699	0.2572	INADEQUATE	yes
TB986292	AFF1	0.303	0.858	0.2602	INADEQUATE	yes
TB000694	AF10	0.929	0.531	0.4934	INADEQUATE	NO
152985	ENL	0.001	0.002	0.0000	ADEQUATE	NO

Full table in results147.json. Note the two failure modes are orthogonal: 152985 has the best depth in the deposit and still reports nothing, because the oncoprotein is off chromatin under revumenib. Depth and readability are different questions and a single quality score would merge them.

TB000694 fails both, and its HOXA value is the one to distrust

TB000694 (KMT2A-AF10) has 92.9% of promoters zero in N1 and fails both positive controls there (MEIS1 0.00, GAPDH 0.00) while N2 looks normal (MEIS1 31.1x). Its HOXA9 reads 43.4x and it is scored HOXA-OCCUPIED - on one antibody effectively. It is the one sample in the deposit whose two N libraries disagree about whether anything was sequenced, and its PAX5 reads 68.9x in an AF10 sample, which is not a B-lineage assay result to take at face value. It contributes to no claim in this beat.

Facts and figures

n samples	27
depth adequate	19
depth marginal	3
depth inadequate	5
rule of three all18 PCT	16.7
rule of three adequate12 PCT	25

p dz range

[
  0,
  0.4934
]

marginal

[
  "A93949",
  "TB051939",
  "TB920488"
]

inadequate

[
  "TB000694",
  "TB040576",
  "A80956",
  "TB986292",
  "TB995291"
]

readability failures

[
  "TB000694",
  "152985"
]

irx1 unmeasured nonaff1

[
  "A80956",
  "A93949",
  "TB000694",
  "TB995291"
]

analyses

[
  "ana-q147-irx1-fusion-occupancy-is-confined-to-kmt2a-aff1"
]

Still open

Would deeper sequencing of the five INADEQUATE samples change the 0-of-18? Three of them are St. Jude samples and all three have IRX1 at exact zero, which is the state that cannot be distinguished from unmeasured.

Every source behind this entry (1)

E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
In GSE252378, p_dz (the chance a promoter reads zero in both KMT2A-N antibodies) ranges 0.0000-0.4934 across 27 samples: 19 ADEQUATE (<=0.10), 3 MARGINAL (A93949 0.159, TB051939 0.140, TB920488 0.135), 5 INADEQUATE (TB000694 0.493, TB986292 0.260, A80956 0.257, TB995291 0.238, TB040576 0.205). Two samples fail the MEIS1/CDK6 readability gate: TB000694 and 152985. IRX1 is UNMEASURED in 4 of 18 non-AFF1 samples, so the rule-of-three bound on non-AFF1 IRX1 occupancy is 16.7% over all 18 and 25.0% over the 12 that are adequate and readable.
q147 re-analysis of GSE252378: HOXA and IRX1 oncoprotein occupancy across all six KMT2A fusion partners, 27 samples, 133 libraries (analyses/q147-hoxa-irx1-across-partners) · Iqbal Lab oncofusion knowledge graph, beat q147 · internal analysis over deposited GSE252378 bigWigs 2026
What this evidence cannot show
The depth cuts were pre-declared but the specific thresholds (0.10, 0.20) are a judgement calibrated on q145's measured range (0.019-0.061 deep, 0.135-0.260 shallow), not derived from first principles; a reader preferring different cuts can recompute from the p_dz column in `results147.json`, which is why it is stored per sample. p_dz treats the two N antibodies as independent, which they are not - they share the library prep - so it is an optimistic estimate of the double-zero probability. Every number is a WITHIN-SAMPLE share: an interval's mean coverage divided by that same library's mean coverage over all 19,404 hg38 RefSeq promoters, so 1.0 = an average promoter in that library. GSE252378 has no spike-in, so no between-library absolute statement is licensed; the only cross-sample quantity used is a ratio of two loci measured inside one library, where the unknown per-library scale cancels.
Evidence that argues against this (1)
- The deposit paper's own quality criterion is an N-over-C ratio at called peaks rather than a promoter zero-rate, and by that criterion some of the samples failed here may be usable; q144 measured the per-sample oncoprotein-target promoter counts under the paper's own criterion and found them spanning 0 to 2,282, which is the same heterogeneity seen through a different instrument. Janssens DH et al. 2024

Connected to