RETROFIT: Reference-free deconvolution of cell-type mixtures in spatial transcriptomics

Abstract Spatial transcriptomics (ST) enables genome-wide measurement of gene expression in intact tissues, but typically captures mixtures of multiple cell types at each spatial location. Deconvolving these mixtures is essential for resolving cell-type-specific spatial organization and transcriptional programs. Existing approaches often rely on matched single-cell references or curated marker genes, which may be unavailable, incomplete, or difficult to integrate across platforms. We present RETROFIT, a Bayesian framework for reference-free deconvolution of spatial transcriptomics data that operates directly on sequencing measurements and incorporates external information only at a post hoc annotation stage when available. Across extensive simulations and multiple real datasets, RETROFIT demonstrates robust performance, outperforming existing reference-free methods and matching or exceeding reference-based approaches when references are imperfect. Notably, RETROFIT remains effective at near–single-cell resolution, as demonstrated on Visium HD data, recovering fine-grained spatial patterns without requiring single-cell references or marker genes. These results establish RETROFIT as a broadly applicable approach for reference-free spatial transcriptomics analysis across platforms and resolutions. RETROFIT is available at https://bioconductor.org/packages/retrofit/ .