Abstract
Spatial transcriptomics technologies have significantly enhanced the analysis of gene expression profiles by retaining the spatial information of intact tissue sections and enabling the possibility of a more profound comprehension of tissue structures and cellular relationships. Despite this, most platforms have limited resolution, and at numerous capture spots, multiple signals from various cells are present, requiring deconvolution, a set of computational steps to deduce the underlying cellular composition. Over the last few years, a range of algorithms has been proposed to address this problem, each employing distinct computational principles and processing paradigms. The present review seeks to present a comprehensive analysis of twenty such algorithms, focusing on their methodological foundations. We contrast the underlying computational algorithms, modeling methods, and data processing pipelines that underlie them, and how they deal with external references, noise and sparsity in the data. By drawing out the conceptual as well as technical foundations of each algorithm, we aim to provide researchers a complete and hands-on grasp of the computational landscape of spatial transcriptomics deconvolution. This review is a methodological handbook to enable deep understanding of current deconvolution methods to develop novel strategies and help in selecting or applying these existing tools for different biological contexts.