Extended Data Fig. 2: Framework for spatial ecotype discovery.
From: Non-invasive profiling of the tumour microenvironment with spatial ecotypes
a,b, Graphical depiction of the Spatial EcoTyper discovery module, illustrating its application to a single sample (steps 1–3; panel a) and to multiple samples (steps 1–9; panel b) profiled by single-cell ST. In step 1, spatial neighbourhoods (SNs) are defined over a grid and cell-type-specific gene expression profiles (GEPs) are constructed for each SN. Cell types within the same SN have the same matrix index and are co-registered. In step 2, the similarity of SNs is assessed and the covariance structures of their cell-type-specific GEPs are fused using similarity network fusion (SNF)32. In steps 3 and 4, Louvain clustering is applied to the fused similarity matrix (step 3), and cell-type-specific average expression is computed within each SN cluster (step 4), resulting in a gene by SN cluster expression matrix for each cell type. In step 5, steps 1 to 4 are repeated for multiple samples (e.g., tumour specimens). The resulting cell-type-specific expression matrices from all specimens are concatenated column-wise into a single GEP matrix per cell type (step 6), with rows representing common genes and columns representing the union of SN clusters across all samples. In steps 7 and 8, the covariance of cell-type-specific GEPs across all SN clusters is assessed and SNF is used to integrate these covariance matrices across cell types, resulting in a fused similarity matrix among SN clusters. In step 9, non-negative matrix factorization is applied to define robust spatial clusters termed spatial ecotypes82. Full details are provided in Methods. c, Inputs and outputs of Spatial EcoTyper, here showing three samples from the MERSCOPE discovery cohort used in this work. Cell icons in c were created using BioRender; Steen, C. (2026) https://BioRender.com/xxbpjx0.
