The DeMixSC deconvolution framework uses single-cell sequencing plus a small benchmark dataset for improved analysis of cell-type ratios in complex tissue samples.

We introduce a novel deconvolution framework, DeMixSC, to resolve technological discrepancies between bulk and single-cell/nucleus RNA-seq data, a critical issue unaddressed by existing single-cell-based deconvolution methods. Built upon the weighted non-negative least squares framework, DeMixSC introduces two key improvements: it leverages a small benchmark dataset to identify and rescale genes affected by technological discrepancies; it employs a novel weight function to account for variations across subjects and cells. The advanced utility of DeMixSC is demonstrated by its superior deconvolution accuracy on a benchmark dataset of healthy retinas and its broad applicability to a large aged-macular degeneration (AMD) cohort. Our work is the first to systematically evaluate the impact of technological discrepancies on deconvolution performance and underscores the importance of using a benchmark dataset to counteract these discrepancies. Our study positions DeMixSC as a transferable tool for accurate deconvolution of large bulk RNA-seq cohorts, necessitating only a tissue-type match between the benchmark and targeted datasets.