Flexible integration of spatial and expression information for precise spot embedding via ZINB-based graph-enhanced autoencoder

Domain identification is a critical problem in spatially resolved transcriptomics data analysis, which aims to identify distinct spatial domains within a tissue that maintain both spatial continuity and expression consistency. The degree of coupling between expression data and spatial information in different datasets often varies significantly. Some regions have intact and clear boundaries, while others exhibit blurred boundaries with high intra-domain expression similarity. However, most domain identification methods do not adequately integrate expression and spatial information to flexibly identify different types of domains. To address these issues, we introduce Spot2vector, a computational framework that leverages a graph-enhanced autoencoder integrating zero-inflated negative binomial distribution modeling, combining both graph convolutional networks and graph attention networks to extract the latent embeddings of spots. Spot2vector encodes and integrates spatial and expression information, enabling effective identification of domains with diverse spatial patterns across spatially resolved transcriptomics data generated by different platforms. The decoders enable us to decipher the distribution and generation mechanisms of data while improving expression quality through denoising. Extensive validation and analyses demonstrate that Spot2vector excels in enhancing domain identification accuracy, effectively reducing data dimensionality, improving expression recovery and denoising, and precisely capturing spatial gene expression patterns.