Characterizing spatial dynamics in the tumour microenvironment (TME), with its diverse cellular organization and crosstalk, could provide valuable insights into tumour formation, progression and therapeutic development. However, sequencing-based spatial transcriptomics fail to disentangle single-cell gene expression, despite transcriptome-wide profiling, limiting their capability to analyse tumour tissue from multiple patients.

We developed Starfysh to address these limitations by providing an end-to-end framework for large-scale spatial transcriptomic data analysis. Unlike existing methods, Starfysh does not rely on single-cell RNA-sequencing data as a reference; it leverages statistical machine learning approaches and histologically intrinsic information through haematoxylin and eosin images, known cell type markers, and archetypes to achieve cell type-specific and refined cell state delineations of the TME from large cohorts. We identified spatial hubs as groups of capture locations (spots) with similar compositions and assessed the role of immune cell states in altering tumour phenotypes.