DiffuScope: A diffusion-regularized autoencoder for spatial transcriptomic clustering

In recent years, the rapid advancement of spatial transcriptomics technologies has led to the public availability of a large and diverse collection of datasets spanning multiple species, organs, and tissue types. These datasets exhibit substantial biological and technical heterogeneity, highlighting the urgent need for a generalizable clustering algorithm capable of adapting to such diversity. To address this challenge, we propose DiffuScope, a clustering framework based on Graph Convolutional Variational Autoencoders (GC-VAE). DiffuScope leverages self-supervised learning to extract informative latent representations from spatial transcriptomics data and incorporates two complementary loss functions - Reconstruction Loss and Diffusion Consistency Loss - to enhance feature learning, thereby improving clustering accuracy and cross-dataset generalization. We systematically and fairly benchmark DiffuScope against several state-of-the-art spatial transcriptomics clustering methods across 12 publicly available datasets. In addition, we assess the robustness of the proposed model by introducing random dropout noise into the spatial expression data. Finally, we apply DiffuScope to breast cancer and gastric cancer spatial transcriptomics datasets, demonstrating its ability to effectively delineate spatial domains and uncover biologically meaningful tissue structures.