When diagnostic signatures detect tissue architecture: quantifying the adipose–stromal component of breast tumor–normal transcriptomic separation

Bulk RNA-seq classifiers distinguish breast tumors from normal tissue with near-perfect accuracy and minimal gene signatures, yet the selected genes are rarely interrogated for what they actually measure. Using TCGA-BRCA [1] (518 tumors, 113 adjacent normals; 23,214 expressed genes), GTEx v10 breast (514 donors), and METABRIC (1,980 tumors; independent cohort and platform), we show that a major component of this separation is captured by an adipose-associated transcriptional axis. Under strictly nested, leakage-free cross-validation, five genes selected within the training data alone reach 98.2% accuracy (full transcriptome: 98.9%), and a pre-specified panel of twelve canonical adipocyte-associated genes reaches 94.7%. A uniformly random ten-gene panel achieves a median accuracy of 92.7% (IQR 90.1–94.9%; B = 500), showing that the tumor–normal transcriptomic displacement is sufficiently global that even uniformly random gene panels retain substantial discriminative information; exceeding this null required effect-size matching (98.1%) and mixed adipose–stromal composition (98.7%). Residualizing the adipose axis from the transcriptome reduces nested classification from 98.9% to 90.9% (AUC 0.999 → 0.950), leaving a considerable residual discriminative signal. The axis replicates across cohorts and platforms: TCGA-trained models call 99.6% of GTEx normal breast as normal, and 93.5% of METABRIC tumors as tumor. Histology showed by far the strongest clinical gradient: capture is markedly lower in invasive lobular than ductal carcinoma (BH ≤ 1.7 × 10⁻⁴), consistent with preserved adipose architecture in diffuse lobular infiltration. Among diagnostic-signature studies examined, none explicitly modeled adipose-tissue composition as a potential confounder of the normal–tumor task. We provide a leakage-free, publicly reproducible evaluation framework that decomposes bulk tumor–normal discrimination into a compositional component and a residual non-adipose-associated component.