DECONVersation: Single Cell Foundation Model-Derived Embeddings for Robust Cell Type Deconvolution of Bulk RNA-seq

Abstract

AI and deep learning have transformed single-cell transcriptomics, empowering unsupervised annotation, trajectory inference, and cross-dataset integration at scale. Here, we propose bulk RNA-seq deconvolution as a novel task for single-cell foundation models. Deconvolution estimates cell type proportions from bulk RNA-seq using single-cell reference data, with applications spanning tumor microenvironment profiling to the detection of disease-associated shifts in tissue composition. However, existing methods remain vulnerable to batch effects, subject-level variation, and marker gene selection sensitivity. We introduce DECONVersation, which leverages embeddings from pretrained and fine-tuned single-cell foundation models, coupled with a non-negative least squares solver, to estimate cell type proportions. Foundation model embeddings are robust to batch effects and implicitly encode complex gene relationships, sidestepping explicit marker gene selection. Benchmarked against MuSiC, DWLS, and BayesPrism, DECONVersation achieved comparable or superior performance across diverse tissues. Performance further improved with tissue-specific fine-tuning.