stDGAC: A novel identifying spatial domains method via graph attention contrastive network for spatial transcriptomics data

Complex biological tissues are composed of many cells in a highly coordinated manner and play a variety of biological functions. In recent years, many methods for spatial clustering have been developed. However, it remains a major challenge to effectively utilize these high-dimensional and noisy spatial transcriptomics data, with similar gene expression and histological features, to more accurately identify spatial domains. Here, a novel method of spatial domain identification, named stDGAC, was proposed by jointly using a denoising autoencoder and a graph attention contrastive network for subsequent analysis of spatial transcriptomics data. The pre-trained denoising autoencoder performed dimensionality reduction and denoising, and then the low-dimensional latent representation was learned through a graph attention contrastive network to aggregate the neighborhood information of the spatial context and acquire a more robust representation. Finally, the proposed method stDGAC was experimentally demonstrated to outperform other existing methods in downstream analysis, such as identifying spatial domains, trajectory inference, and gene expression data denoising.