IRX1 carries the KMT2A::AFF1 oncoprotein in exactly the patients where HOXA does not — seven patients, GSE252378

analyses/q145-irx1-in-hoxa-negative-infants/PREREGISTRATION.md was written before any IRX-locus value was opened, with an explicit A1 positive, A2 negative, a four-branch A3 'neither', and nine uninformativeness clauses. Two of its registered predictions were wrong and are reported as wrong:

• Q2/B1 did not fire. I registered that the seven patients would split bimodally on HOXA occupancy with a >=3-fold gap. The largest consecutive gap in the ranking is 6.56x, and it is between the two lowest samples — a tail artefact. The biologically meaningful gap, between TB051939 (7.00x) and p279 (2.61x), is 2.68x, below my threshold. So B2 fired: HOXA occupancy is a continuum, not a switch, and every downstream statement is made on the continuous rank rather than on a dichotomy.
• Q4 was refuted. I registered blind that if IRX1 is the reciprocal of HOXA, IRX1 should be higher in p318 than in p279 across the lineage switch. It is lower (1.58x vs 3.52x at the gene). HOXA collapsed ~28-fold across the switch while IRX1 fell only ~2.2-fold, so the ratio moved reciprocally while the absolute did not. With no spike-in only the ratio is interpretable, which is why the ratio was the registered primary readout — but the registered directional prediction still failed and is recorded as failed.

KMT2A-N share (mean of the two N antibodies), Every number is a WITHIN-SAMPLE share: an interval's mean coverage divided by that same library's mean promoter coverage, so 1.0 = an average promoter in that library. GSE252378 has no spike-in and no drug arm, so no between-sample 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.

Italicised P values mark the three libraries shallow enough that a zero is weak evidence — see finding-a-shallow-library-cannot-support-an-absence.

Three of three HOXA-high patients have zero IRX1. Two of two HOXA-dead deep libraries have high IRX1. But:

• p279 is DOUBLE-POSITIVE. HOXA9/HOXA10 at 2.92x and IRX1 at 6.08x, both well clear of floor. The published claim is mutually exclusive expression; this is co-occupancy in one patient, and it is the clearest single refutation of reading the expression dichotomy as an occupancy dichotomy.
• TB920488 is DOUBLE-NEGATIVE (HOXA 0.97x, IRX1 0.59x). So HOXA-negative does not imply IRX1-positive. Of the five HOXA-low sample records, only three are IRX1-positive.

So the axis is a real, significant anti-correlation across patients and NOT a binary switch within a patient.

• IgG floor (U1). IgG has zero reads at the IRX1 promoter in all eight sample records, so the floor is 0 and any non-zero factor signal clears it. Because a zero IgG in a sparse library is itself weak, the zero-rate control below carries the argument instead.
• Mappability (U3) does not fire. 42 of 66 libraries have non-zero coverage at the same IRX1 window, so the locus is mappable under repeat-masked hg38 + bowtie2 and a zero elsewhere is a property of the sample, not the reference.
• It is IRX1, not chr5p15 (A3d does not fire). Of the eight promoters within 2 Mb, MRPL36, NDUFS6 and ICE1 are flat across all patients and IRX4/ADAMTS16 are zero in all; IRX1 is the only one that goes 3.5x -> 0. IRX2 and its bidirectional neighbour C5orf38 rise only in p454.
• It is IRX1, not the IRX family. Spearman against HOXA: IRX1 -0.571, IRX2 -0.071, IRX3 +0.452, IRX5 +0.476. IRX3/IRX5 move with HOXA, so no cluster-wide effect.
• A control FAILED and is reported: EGR1 does not clear its own IgG. EGR1 KMT2A-N / EGR1 IgG is 1.11, 3.90, 1.16, 1.13 and 0.50 across the samples where IgG is non-zero. The EGR1 promoter is a background hotspot in this assay and no EGR1 claim is made in this beat, although Kühn's IRX1-induces-EGR1 result made it an attractive one.

The comparison was run at 2 kb, 8 kb, gene body, 20 kb and cluster scale. The IRX1-vs-HOXA ordering is identical at every gene-level scale and the correlation strengthens from -0.571 (2 kb promoter) to -0.714 (gene body). It collapses only at the 1.72 Mb IRXA-cluster window, where the interval is ~99.7% intergenic and reads 0.02-0.29x in every patient.

That collapse is itself a finding rather than an artefact: HOXA is activated as a cluster and IRX1 as a single gene. In the HOXA-high patients the whole 107 kb HOXA cluster carries 2.65-3.23x, whereas neither IRX cluster ever exceeds 0.29x in any patient. See finding-hoxa-is-occupied-as-a-cluster-and-irx1-as-a-single-gene.

GSE252378 has no RNA-seq, so no patient here can be assigned to the published expression subgroup; chromatin HOXA occupancy is a proxy and its use is registered as an assumption. It has no menin track in any KMT2A::AFF1 sample (the only menin tracks are in the menin-inhibitor-treated KMT2A::ENL pair), so nothing here tests the menin interface directly. Occupancy cannot establish direction of causation: IRX1 may be occupied because the fusion drives it, or the fusion may be there because IRX1 was already open in the cell of origin.