CSsingle: a unified tool for robust decomposition of bulk and spatial transcriptomic data across diverse single-cell references

Accurate deconvolution of bulk and spatial transcriptomes is essential for studying tissue architecture and disease, yet remains challenged by unmodeled differences in cellular RNA content and cross-source heterogeneity. We introduce CSsingle, a unified deconvolution framework that explicitly corrects for cell-type-specific RNA content differences using either External RNA Controls Consortium (ERCC) spike-ins or a computational estimator, while robustly harmonizing data across platforms. CSsingle employs an iteratively reweighted least-squares model initialized by marker-gene sectional linearity, enabling accurate inference of cell-type proportions from diverse single-cell references. In bulk data, CSsingle outperforms existing methods by correcting systematic errors, including neutrophil underestimation in blood and tumor purity underestimation in breast tumor. Applied to spatial transcriptomics, CSsingle enables fine-grained mapping of cellular organization in the developing human pancreas and reveals functionally distinct niches in colon cancer. By integrating cell size awareness with cross-platform robustness, CSsingle advances the integrative analysis of complex tissues.