Benchmarking cell type deconvolution in spatial transcriptomics and application to cancer immunotherapy

Abstract

Cell type deconvolution is essential for resolving multicellular composition in spatial transcriptomics data, yet the accuracy and consistency of current methods across biological contexts remain uncertain. We introduce a benchmarking framework using realistic simulations that incorporate spatial and transcriptional complexity. We find that a simple marker gene signature scoring approach performs competitively, often outperforming more complex models, particularly for rare cell types. We further apply this method to a new spatial transcriptomics dataset profiling the early response to a single dose of anti-PD1 checkpoint blockade immunotherapy in a mouse cancer model, revealing pronounced spatially localized immune and microenvironmental changes in both tumors and draining lymph nodes. Our findings have implications for the interpretation of many published spatial transcriptomics studies that rely on cell type deconvolution, and they provide a robust, interpretable strategy for future analyses, even in the absence of matched single-cell references.