Multi-directional State Space Modeling for Deciphering Spatial Domains From Spatially Resolved Transcriptomics

Spatially resolved transcriptomics (SRT) enables the investigation of tissue architecture by mapping gene expression within its spatial context. However, current computational methods for spatial domain identification mainly depend on spatial adjacency, limiting the capture of long-range dependencies and often failing to identify identical cell states distributed across distant and non-contiguous regions. To address these challenges, we propose MambaST, a hybrid deep-learning framework that integrates Mamba with self-supervised learning. MambaST utilizes a Context-Aware Contrastive Learning (CACL) strategy to enhance local biological consistency. Furthermore, we introduce Six-Directional (SS6D) and Four-Directional (SS4D) Selective Scan algorithms for different spatial patterns, which transform non-sequential graph structures into pseudo-sequences that maintain the topology. These sequences are subsequently processed by the spaMamba Block (SMB) to capture long-range semantic patterns while reducing technical noise. Extensive evaluations across distinct SRT datasets demonstrate that MambaST outperforms existing methods, achieving a mean Adjusted Rand Index (ARI) of 0.58 on the DLPFC dataset, surpassing state-of-the-art methods by 2.7%. Beyond spatial domain identification, MambaST excels in downstream applications including gene expression denoising differential expression analysis, and the dissection of cancer heterogeneity.