Spatial Domain Identifying: Graph Attention Network with Two Different Decoders

Exploring spatial domains to investigate tissue structures is a chance provided by spatial transcriptome technology, while also a significant challenge in spatial transcriptomics research. Current approaches only focus on spatial gene expression and cannot simultaneously incorporate spatial location information. Graph deep learning models can simultaneously encode node features and positional information. However, during decoding, most of the models still only focus on reconstructing feature information, ignoring positional information. Here, we propose a new method, DeepDomain, which aims to improve the latent representation of nodes by jointly reconstructing gene expression profiles and spatial neighborhood networks using a deep graph attention network with two distinct decoders. Utilizing enhanced spatial latent representations to identify spatial domains in three datasets, DeepDomain achieved higher accuracy in evaluation metrics and a better description of organizational structure when compared to existing methods.