Efficient reliability analysis of spatially resolved transcriptomics at varying resolutions using SpaSEG

Abstract

Spatially resolved transcriptomics (SRT) for characterizing cellular heterogeneities and activities requires systematic analysis approaches to decipher gene expression variations in physiological contexts. Here we develop SpaSEG, an unsupervised convolutional neural network-based model for multiple SRT analysis tasks by jointly learning the transcriptional similarity of spots and their spatial dependence. SpaSEG adopts an edge strength constraint to encourage spatial domain coherence and allows integrative analysis by automatically aligning the spatial domains across multiple adjacent sections. It also enables the detection of domain-specific gene expression patterns and the inference of intercellular interactions and colocalizations within a tissue. In an invasive ductal carcinoma sample analysis, SpaSEG facilitates the unraveling of intratumor heterogeneity and the understanding of immunoregulatory mechanisms. Through comprehensive evaluation over a collection of SRT datasets generated by different platforms at various resolutions, SpaSEG shows superior reliability and computational efficiency over existing methods, endowing it with a great potential for the exploration of tissue architectures and pathological biology.