Randomized Spatial PCA (RASP): a computationally efficient method for dimensionality reduction of high-resolution spatial transcriptomics data

Spatial transcriptomics (ST) provides critical insights into the spatial organization of gene expression, enabling researchers to unravel the intricate relationship between cellular environments and biological function. Identifying spatial domains within tissues is key to understanding tissue architecture and mechanisms underlying development and disease progression. Here, we present Randomized Spatial PCA (RASP), a novel spatially-aware dimensionality reduction method for ST data. RASP is designed to be orders-of-magnitude faster than existing techniques, scale to datasets with 100,000+ locations, support flexible integration of non-transcriptomic covariates, and reconstruct de-noised, spatially-smoothed gene expression values. It employs a randomized two-stage PCA framework with sparse matrix operations and configurable spatial smoothing. RASP was compared to BASS, GraphST, SEDR, SpatialPCA, and STAGATE using diverse ST datasets (10x Visium, Stereo-Seq, MERFISH, 10x Xenium) on human and mouse tissues. RASP demonstrates comparable or superior tissue domain detection with substantial improvements in computational speed, enhancing exploration of high-resolution subcellular datasets.