Contrastive Masked Graph Autoencoders for Spatial Transcriptomics Data Analysis

Spatial transcriptomics is an emerging genomics technology aimed at revealing the spatial distribution of gene expression at the tissue or cellular level. This technique enables the acquisition of gene expression profiles for each spot, constructing a spatial gene expression map. While numerous methods have been developed to integrate expression profiles and spatial information for spatial domain detection, accurate identification remains a challenging task. To address this issue, we propose a novel self-supervised learning model, the Contrastive Masked Graph Autoencoder (stCMGAE) for spatial transcriptomics data analysis. By ingeniously incorporating the masked mechanism and contrastive learning into a graph neural network, the model leverages their respective strengths to learn highly informative gene embedding representations for spatial domain recognition. We evaluate the performance of the stCMGAE method on three spatial transcriptomics datasets, achieving the highest ARI indices in all cases. Additionally, we obtain clearer boundaries in spatial recognition. Our source code is available at https://github.com/donghaifang/stCMGAE .