Benchmarking sketching methods on spatial transcriptomics data

High-throughput spatial transcriptomics (ST) now profiles hundreds of thousands of cells or locations per section, creating computational bottlenecks for routine analysis. Sketching, or intelligent sub-sampling, addresses scale by selecting small, representative subsets. While used in scRNA-seq, most sketching methods optimize coverage in expression space alone and ignore physical location, risking spatial bias. We systematically benchmark uniform sampling, leverage-score sampling, Geosketch (minimax/Hausdorff), and scSampler (maximin) across multiple real ST datasets (mouse ovary, MERFISH brain, human breast cancer, lung) and simulations, using three input representations: PCA embeddings, spatial coordinates, and spatially smoothed embeddings. We show that expression-only designs capture global transcriptomic heterogeneity but distort tissue architecture by over-sampling high-variability regions and under-sampling homogeneous areas. Coordinate-only sampling restores tissue coverage but misses transcriptional extremes. A simple spatially aware extension, computing leverage scores from a randomized SVD basis smoothed by a spatial weights matrix, strikes a favorable balance, recovering rare cell states while maintaining uniform tissue coverage and avoiding edge effects. Across robust Hausdorff distances, clustering stability (ARI), PCA loading drift, and local cell-type MSE, spatially smoothed leverage scores match or outperform alternatives. These results motivate joint spatial-transcriptomic sketching objectives to enable fast, unbiased analyses of increasingly large ST datasets.