Extended Data Fig. 3: Benchmarking, MERSCOPE samples, and analysis of spatial neighbourhoods from Spatial EcoTyper embeddings.

a, Spatial ecotype clusters defined by Spatial EcoTyper and previous methods in a representative melanoma specimen profiled by MERSCOPE (Melanoma 1, Supplementary Table 8; Methods). b, Comparison of methods for identifying spatial ecotype clusters in tumour samples profiled MERSCOPE (n = 9; Supplementary Table 8), applied to each tumour sample individually. Left: Dot plot showing the relative performance of each method for identifying spatial ecotype clusters using three metrics: (1) “spatial colocalization”, which captures the local contiguity of cells within each cluster, (2) “cell type mixing”, which captures the diversity of cell types per cluster (higher scores denote more cell types), and (3) “mean silhouette width”, which captures, for each cell type, the degree of gene expression profile (GEP) separation between clusters and the degree of GEP similarity within the same cluster (higher scores denote greater cluster separation and compactness) (Supplementary Methods). Dot sizes reflect the relative ranking of methods based on each metric (larger=better). Each quantity was first averaged across clusters within each sample, converted to rank space, and then averaged across samples. Right: Box plot aggregating the three metrics from each tumour sample by geometric mean of their ranks. The box centre lines, bounds of the box, and whiskers denote medians, 1^st and 3^rd quartiles, and minimum and maximum values within 1.5 × IQR (interquartile range) of the box limits, respectively. c, UMAP embeddings showing integration of two single-cell ST samples of the same cancer type (Melanoma 1 and Melanoma 2, Supplementary Table 8) for selected methods, with cells coloured by samples (top), tumour and adjacent stroma (centre), and cell types (bottom). In the Spatial EcoTyper UMAP, a small amount of jitter was applied to display individual cells within each spatial cluster. d, Scatter plot comparing Spatial EcoTyper against all methods in the benchmarking analysis that enable sample integration, showing their relative performance for identifying spatial ecotypes conserved across Melanoma 1 and Melanoma 2 (Supplementary Table 8). Performance was assessed using metrics that quantify the degree of cell type mixing (x-axis) and sample mixing (y-axis), averaged across identified spatial ecotype clusters (Supplementary Methods). e, Scatter plot summarizing the performance of each method for identifying spatial ecotype clusters, combining single-sample analysis (panel b) and integrative analysis (panel d). Integration performance was computed as the geometric mean of cell type mixing and sample mixing metrics (panel d) in rank space. For additional details, see Supplementary Methods. f, Composition of MERSCOPE specimens used for SE discovery and validation. Only samples with more than 5% TME cells derived from tumour or adjacent stroma regions were included (Supplementary Table 8). g, Robustness of spatial embeddings to spatial neighbourhoods of diverse radii, related to Fig. 2b. Here, Spatial EcoTyper was applied to a melanoma specimen profiled by MERSCOPE (Melanoma 1, Supplementary Table 8) (left) to create spatial embeddings as illustrated in Fig. 2a and detailed in Methods, but where the spatial neighbourhood radius was varied (centre). Each point in the embedding denotes an individual spatial neighbourhood. Right: Concordance between spatial neighbourhoods in the Spatial EcoTyper embedding and physical distance to the tumour margin, determined using Spearman correlation as described in Supplementary Methods. A radius of 50 µm, reflecting a balance between higher concordance and smaller radius, was selected for subsequent analysis. h, Analysis of the relationship between the organization of sample-specific spatial embeddings in Fig. 2b and physical distance to the tumour margin. Concordance was quantified as described in Supplementary Methods. Significance was calculated with two-sided t tests. i, Same as Fig. 2b, but colouring individual spatial neighbourhoods by the expression of an immune-hot signature (average log₂ expression of 13 signature genes³⁴; Supplementary Methods).