Abstract
Recent improvements in spatial transcriptomics technologies have enabled the characterization of complex cellular mechanisms within tissue context through unbiased profiling of genome-wide transcriptomes in conjunction with spatial coordinates. These technologies require a systematic analysis approach to deciphering the complex tissue architecture. Here, we develop SpaSEG, an unsupervised convolutional neural network-based method towards this end by jointly learning gene expression similarity of spots and their spatial contiguousness via adopting a loss function for spatial boundary continuity. Using several spatial transcriptomics datasets generated by different platforms with varying resolutions and assayed tissue sizes, we extensively demonstrate that not only can SpaSEG better identify spatial domains, but also be much more computationally and memory efficient than existing methods. In addition, SpaSEG is able to effectively detect genes with spatial expression patterns and infer spot-wise intercellular interactions as well as cell-type colocalization within the tissue section by utilizing the identified domains. Taken together, our results have indicated the flexibility of SpaSEG in multiple analysis tasks in spatial transcriptomics, making it as a desirable tool in facilitating the exploration of tissue architecture and the knowledge of underlying biology.
Competing Interest Statement
The authors have declared no competing interest.