DiST: disentangling structure via self-supervised tensors for heterophilous spatial transcriptomics

Abstract

Spatially resolved transcriptomics (SRT) serves as a transformative technology for examining gene expression within tissue architecture. However, accurate spatial domain identification remains challenging due to the inherent data sparsity and the tendency of graph-based methods to over-smooth representations at tissue boundaries. To overcome these limitations, we propose DiST, a novel self-supervised tensor learning framework. Distinct from conventional approaches that rely on neighborhood propagation, DiST introduces a Semantically Aligned Ego-Network Tensor construction strategy. By stacking semantically similar ego-networks from the global context, we transform flat graph data into high-order tensor structures that preserve local heterogeneity. Furthermore, we employ a robust low-rank and sparse decomposition mechanism to separate a coherent component used for representation learning from sparse entrywise deviations associated with sequencing dropouts and technical artifacts. This decomposition yields sparsity-resilient representations that recover sharp tissue structures. Extensive experiments on five benchmark datasets demonstrate that DiST outperforms representative methods in spatial clustering. Moreover, downstream biological validation on human breast cancer data confirms DiST’s capability to uncover biologically meaningful molecular signatures and functional heterogeneity.