CoexpressDeconvolve enables reference-free single-cell-resolution deconvolution from spot-based spatial transcriptomics.

Spot-based spatial transcriptomics captures the transcriptome of multiple adjacent cells per spot, obscuring cell type-specific signals. Most deconvolution tools, therefore, depend on external single-cell references and return cell-type fractions rather than the number of cells, and only some output full expression profiles. Here we present CoexpressDeconvolve, a reference-free framework that combines a hybrid housekeeping-library-size calibration with topic modeling on a spatial gene co-expression manifold to recover integer cell counts and cell type-specific transcriptomes. Benchmarking synthetic Visium data against Tangram, cell2location, and STdeconvolve shows that CoexpressDeconvolve attains competitive expression-reconstruction fidelity, the lowest cell-count error, and the highest per-slide cell-type concordance. Our framework outputs a feature-barcode matrix that mimics standard Space Ranger output and loads directly into the standard single-cell downstream analytical stack. We applied it to human breast cancer and tongue squamous cell carcinoma, where it resolved tumor microenvironment composition and identified malignant progression axes.