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

Abstract

Motivation: Bulk RNA-seq deconvolution uses single-cell references to estimate cell type proportions in bulk tissue. Existing methods face challenges with batch effects, subject-level variation, and marker gene selection. We propose a new approach leveraging general-purpose single-cell foundation models. Results: DECONVersation combines pretrained or fine-tuned single-cell foundation model embeddings with non-negative least squares for deconvolution. The embeddings eliminate the need for explicit marker selection, and reduce batch effects. Compared against MuSiC, DWLS, and BayesPrism, DECONVersation with Geneformer or Cell2Sentence-2B achieved competitive performance across six bulk datasets with ground truth, with performance improvements after reference-specific fine-tuning. Availability and Implementation: DECONVersation is available through the Python Package Index. The source code is available at https://github.com/nygenome/DECONVersation, and the datasets and Apptainer/Singularity images are archived on https://zenodo.org/records/22949405.