STADiffuser: high-fidelity simulation and full-view 3D modeling of spatial transcriptomics

Spatial transcriptomics technologies have provided invaluable insights by profiling gene expression alongside precise spatial information. However, they encounter high costs, data sparsity, and limited resolution, hindering their broader adoption and utility. Moreover, the lack of flexible simulators capable of generating high-fidelity simulated data has impeded the development of computational tools for spatial transcriptomic data analysis. To this end, we introduce STADiffuser, a versatile deep generative model that leverages diffusion modeling for accurate simulation of spatial transcriptomic data. STADiffuser employs a two-stage architecture: an autoencoder with a graph attention mechanism for learning spot embeddings, followed by a latent diffusion model integrated with a spatial denoising network for data generation. STADiffuser is the first simulator designed for spatial transcriptomic data, capable of handling multiple samples and 3D coordinates while supporting user-defined conditions. STADiffuser facilitates various downstream analyses, including accurate imputation, super-resolution, and full-slice generation. Furthermore, its generative scheme enables in silico experiments, thereby enhancing the statistical power in detecting differentially expressed genes and identifying cell-type-specific genes while effectively controlling confounding factors. Notably, STADiffuser scales to millions of spots and supports the full-view 3D modeling of a marmoset cerebellum atlas with over 20 million spots, enabling detailed investigations from arbitrary viewing angles.