DeconRel: Measurement Reliability Limits Compositional Biomarkers From Single-Cell-Informed Deconvolution in Triple-Negative Breast Cancer

DeconRel is a small open-source framework for testing whether a composition estimated by reference-based deconvolution is reliable enough to use as a biomarker. It does not introduce a new deconvolution estimator. Instead, it wraps a standard non-negative least-squares (NNLS) mixing model with three checks that should be run before any composition-derived quantity is tested against clinical outcome: Conditioning guard. Detects signatures made rank-deficient by per-gene centering, a silent failure that confines every solution to a face of the simplex. Split-half reliability. Measures within-method reliability of program fractions, compositional entropy and the dominant-program call under random partition of the signature genes (Spearman-Brown corrected). It is compared against marker-set scores computed from the same expression data. Attenuation-aware inference. Converts the measured reliability into the minimum effect a downstream model could have detected, so that a null association can be read as either biological or a measurement ceiling.