Almost every gene set reads higher in the index arm: a size-matched empirical null for small rare-tumour expression series, and what it leaves of the EWSR1::NR4A3 direct-target catalogue

Abstract

On a small rare-tumour expression series, almost every gene set anyone scores comes back higher in the index arm. In a 10-versus-6 extraskeletal myxoid chondrosarcoma (EMC) series, PPARγ targets, hypoxia metagenes and adipogenesis all move alike, because a set's per-sample score is one draw from a distribution whose width depends on the set's size and on the platform, not on its biology: an arbitrary 19-gene set can print t = 3.16 and be indistinguishable from a random one. We supply the calibration that refuses such a read — a size-matched empirical null drawn from the platform's own genes — and apply it to this disease's best-warranted set. Across 2,276 retrieved documents, the set with a DNA-binding assay against an NR4A3 chimera is three genes: SEMA3C, PPARG, ENO3. In three cohorts on three platforms that aggregate reaches 39% and 88% of its null threshold and does not clear, while the published EMC phenotype clears it 11.9-fold and 4.2-fold in the same run — the instrument reads this disease, not this set. Exact label permutation, every comparator stratum separately, a matrix covariate and a muscle control then separate three genes usually treated alike: SEMA3C survives nothing and reverses sign with the comparator; PPARG's strongest reading is circular, scored on the cohort that first published it; ENO3 survives everything, but was the pre-designated positive control and is not an independent finding here. The binding constraint is not sample size. No experiment has measured where an NR4A3 fusion binds, or what chromatin does, in EMC material — the one genome-wide chromatin readout that exists for these fusions reads accessibility in HEK293T (GSE243553), and the 110 NR4A peak sets that exist are the wrong protein or the wrong disease, with no class-A gene carrying unusual occupancy against a background panel. Until a fusion cistrome in EMC chromatin exists, "elevated in EMC" and "driven by the fusion" are inseparable. The null is not EMC-specific: any series with a small index arm and heterogeneous comparators fails the same way. ---

Keywords