CADE: a reference-free workflow for coefficient-sensitivity analysis in bulk transcriptomic differential expression

Bulk differential-expression (DE) coefficients can change when disease alters cell-type abundance, but matched single-cell references are often unavailable in small or rare-disease studies. Existing marker-covariate adjustments can produce an adjusted DE table, yet they do not directly report which gene-level coefficients are stable, attenuated, amplified or representation-sensitive after composition adjustment. Coefficient-sensitivity analysis deliberately asks a narrower question than deconvolution: for each gene, how much does the reported group effect change once marker-derived composition weights are added to the model, in which direction does it change, and is that change stable to marker choice and to the coordinate representation of the weights? The output is a per-gene ranking of coefficient stability, not an estimate of cell-type fractions. We developed CADE (Composition-Aware Differential Expression), an R workflow that estimates marker-derived relative composition weights, fits paired unadjusted and composition-adjusted limma models, and computes a per-gene Composition Confounding Index (CCI). CADE reports signed coefficient-change labels, marker-dropout bootstrap rank stability (repeatedly omitting markers and re-ranking genes, to show which rankings persist), label-permutation calibration, denominator-threshold diagnostics, stabilised-denominator output and isometric log-ratio coordinate-sensitivity analysis. Genes whose classification depends on the coordinate representation of the composition weights are explicitly identified. In the controlled synthetic composition-bias benchmark, CADE preserved recovery of cell-intrinsic DE genes under composition bias, including AUROC 0.996 at 100% synthetic bias versus 0.902 for unadjusted limma. Marker-derived weights recovered PBMC ground-truth relative composition with mean r = 0.993, and a macrophage activation syndrome pseudobulk evaluation against scRNA-seq-derived composition showed mean r = 0.779, with all seven mappings that had a usable scRNA-seq reference showing p < 0.001 (of eight preset categories; Neutrophils had no reference in this ground truth). In familial hemophagocytic lymphohistiocytosis PBMCs (GSE26050; n = 11 versus 33 controls), CADE illustrated the separation of low-CCI from model-sensitive coefficients within a targeted immune, iron and ferroptosis panel. CADE provides a reference-free, reproducible workflow for characterising how sensitive bulk transcriptomic DE coefficients are to marker-derived composition adjustment. It reports coefficient stability and does not estimate absolute cell fractions. The FHL analysis is used as a small-cohort worked example, and disease-specific findings remain hypothesis-generating.