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

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Accurate deconvolution of bulk and spatial transcriptomes is crucial for understanding tissue architecture and disease mechanisms.
• Current deconvolution methods struggle with unmodeled cellular RNA content variations and cross-platform data heterogeneity.

Purpose of the Study:

• To introduce CSsingle, a unified deconvolution framework designed to accurately infer cell-type proportions from transcriptomic data.
• To address challenges in deconvolution, including cell-type-specific RNA content differences and data harmonization across diverse sources.

Main Methods:

• CSsingle employs an iteratively reweighted least-squares model, initialized by marker-gene sectional linearity.
• The framework explicitly corrects for cell-type-specific RNA content using External RNA Controls Consortium (ERCC) spike-ins or a computational estimator.
• It robustly harmonizes data from various platforms, integrating cell size awareness.

Main Results:

• CSsingle outperforms existing methods in bulk data deconvolution, correcting systematic errors like neutrophil and tumor purity underestimation.
• Application to spatial transcriptomics enables fine-grained mapping of cellular organization in the developing human pancreas.
• CSsingle reveals functionally distinct niches within colon cancer tissue.

Conclusions:

• CSsingle provides a robust and accurate solution for transcriptomic deconvolution, essential for complex tissue analysis.
• The framework's ability to harmonize data and correct for RNA content differences enhances the integrative analysis of bulk and spatial transcriptomic data.
• CSsingle advances the study of tissue architecture and disease by enabling more precise cellular composition inference.