FAST: a fast and scalable factor analysis for spatially aware dimension reduction of multi-section spatial transcriptomics data

Biological techniques for spatially resolved transcriptomics (SRT) have advanced rapidly in both throughput and spatial resolution for a single spatial location. This progress necessitates the development of efficient and scalable spatial dimension reduction methods that can handle large-scale SRT data from multiple sections. Here, we developed FAST as a fast and efficient generalized probabilistic factor analysis for spatially aware dimension reduction, which simultaneously accounts for the count nature of SRT data and extracts a low-dimensional representation of SRT data across multiple sections, while preserving biological effects with consideration of spatial smoothness among nearby locations. Compared with existing methods, FAST uniquely models the count data across multiple sections while using a local spatial dependence with scalable computational complexity. Using both simulated and real datasets, we demonstrated the improved correlation between FAST estimated embeddings and annotated cell/domain types. Furthermore, FAST exhibits remarkable speed, with only FAST being applicable to analyze a mouse embryo Stereo-seq dataset with >2.3 million locations in only 2 hours. More importantly, FAST identified the differential activities of immune-related transcription factors between tumor and non-tumor clusters and also predicted a carcinogenesis factor CCNH as the upstream regulator of differentially expressed genes in a breast cancer Xenium dataset.