VARGG: a deep learning framework advancing precise spatial domain identification and cellular heterogeneity analysis in spatial transcriptomics.

Spatial transcriptomics has revolutionized our ability to measure gene expression while preserving spatial information, thus facilitating detailed analysis of tissue structure and function. Identifying spatial domains accurately is key for understanding tissue microenvironments and biological progression. To overcome the challenge of integrating gene expression data with spatial information, we introduce the VARGG deep learning framework. VARGG combines a pretrained Vision Transformer (ViT) with a graph neural network autoencoder, utilizing ViT's self-attention mechanism to capture global contextual information and enhance understanding of spatial relationships. This framework is further enhanced by multi-layer gated residual graph neural networks and Gaussian noise, which improve feature representation and model generalizability across different data sources. The robustness and scalability of VARGG have been verified on different platforms (10x Visium, Slide-seqV2, Stereo-seq, and MERFISH) and datasets of different sizes (human glioblastoma, mouse embryo, breast cancer). Our results demonstrate that VARGG's ability to accurately delineate spatial domains can provide a deeper understanding of tissue structure and help identify key molecular markers and potential therapeutic targets, thereby improving our understanding of disease mechanisms and providing opportunities for personalization to inform the development of treatment strategies.