Spatial transcriptomic data denoising and domain identification by a community strength-augmented graph autoencoder.

The rapid development of spatial sequencing technologies has generated large amounts of spatial transcriptomic data, which provide an opportunity to explore complex tissue structures and functional domains. However, such data often suffer from high noise and sparsity, which bring a big challenge for deciphering spatial domains and further understanding the structural and functional organization of biological tissues. In this study, we propose a novel method named Community Strength-Augmented (CSA) that incorporates community strength-augmented graph autoencoder by considering spatially heterogenous structures. Moreover, attention mechanism is designed in CSA to take full advantage of both spatial transcriptomic data and corresponding histology image information. We applied CSA to several spatial transcriptomic datasets derived from various platforms. Compared with the state-of-the-art methods, CSA exhibits superiority in revealing spatially functional domains. Moreover, CSA is able to denoise the data, enabling the identification of biologically meaningful marker genes.