What is shared across 33 cancer types is the tissue, not the tumour: a non-malignant axis of the microenvironment

Abstract What a pan-cancer causal network reports is the composition of the samples, not the regulation of the tumour. Comparing 33 cancer types from The Can- cer Genome Atlas (TCGA) (10,459 samples), each fitted by NOTEARS on a common 100-gene panel, the reproducible structure sits on non-malignant axes: a co-induced, interferon-γ-inducible CXCR3-ligand triplet—CXCL9, CXCL10 and CXCL11—the only three-node module whose edges keep one orienta- tion, together with neutrophil/CXCR2 chemokines, blood content, plasma- cell/immunoglobulin, stroma/complement and myeloid/cytokine signalling. The dominant axis is non-malignant content (first principal component, PC1; λ1 = 16.04, against the Marchenko–Pastur edge; r = 0.80 to markers outside the panel, 33/33 cohorts). Canonical drivers are absent structurally, not by chance: their dispersion lies at the 50.8th percentile whereas the panel selects the 0.487th, so drivers enter only at k ≳ 1000. These findings are bounded. Only 14 of 9,900 directed pairs recur in ≥ 10 cohorts—above the gene-wise permutation null but not against covariance-matched nulls in count, so the recurring set is identified by composition, not by its size—and the comparison is defined only after alignment: misaligned cohorts give a zero median adjacency matrix (0.048) against 0.731 when aligned. Shuffing sample labels abolishes the module (14 → 0 shared edges) while preserving the sample-size dependence, locating an identifiability floor: every cohort still returning edges on shuffed labels has n ≲ 1.7d, so the floor is empirical, not universal. What recurs is the tissue, not the tumour: edge orientations are estimator-selected and unidentified from Gaussian observational data; all results are bulk HiSeqV2 tumours.