A neighborhood contrast famework for cell clustering for spatial transcriptomics data

Spatial transcriptomics data can provide high-throughput gene expression profiling and the spatial structure of tissues simultaneously. Most studies have relied on only the gene expression information but cannot utilize the spatial information efficiently. We combine the strengths of spatial transcriptomics and graph neural networks to introduce graph neural networks into cellular clustering of spatial transcriptomics data, and propose a neighborhood comparison framework for clustering, STAGC, to overcome the above limitations. All the modules of STAGC are optimized in a unified framework, and thus the learned node representations contain clustering-oriented messages. Efficiently embedding high-dimensional histological features and complex graph structures to generate robust low-dimensional cellular representations. Through join tanalysis of transcriptomics and epigenetic data, the model enables spatial transcriptomics cell clustering, which significantly outperforms traditional clustering methods. Through validation on different real data sets, the experimental results show that the clustering accuracy is significantly improved and robustness is enhanced compared with other traditional methods. This study provides a generalized tool for cellular analysis.