Deciphering cell-type composition in disease-relevant tissues is essential for identifying cellular drivers of pathology. We first developed MuSiC, which leverages single-cell RNA sequencing (scRNA-seq) data to deconvolve bulk RNA-seq and characterize cellular heterogeneity. By weighting genes with cross-subject and cross-cell consistency, MuSiC enabled accurate transfer of cell type-specific information across datasets and species.
To address mismatches between bulk RNA-seq and scRNA-seq reference conditions, we introduced MuSiC2, which uses an iterative estimation procedure to improve condition-specific proportion estimates. Applications to pancreatic islets and retina demonstrated its accuracy in disease settings where matched scRNA-seq references are limited.
Most recently, we developed InteRD (Integrated and Robust Deconvolution), which integrates multiple scRNA-seq references, incorporates prior biological knowledge, and remains robust to imperfect external data. Through penalized regression with a novel evaluation criterion, InteRD achieved more accurate and biologically meaningful estimates. Together, these advances systematically improve deconvolution accuracy, robustness, and generalizability.