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

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatially resolved transcriptomics (SRT) maps gene expression within tissue context.
• Current methods struggle with long-range dependencies and identifying distant, identical cell states.

Purpose of the Study:

• To develop MambaST, a hybrid deep learning framework for enhanced spatial domain identification in SRT.
• To overcome limitations of existing methods in capturing long-range spatial relationships.

Main Methods:

• MambaST integrates Mamba with self-supervised learning, using Context-Aware Contrastive Learning (CACL).
• Introduces Six-Directional (SS6D) and Four-Directional (SS4D) Selective Scan algorithms to process graph structures as pseudo-sequences.
• Employs the spaMamba Block (SMB) for capturing long-range semantic patterns and reducing noise.

Main Results:

• MambaST achieved a mean Adjusted Rand Index (ARI) of 0.58 on the DLPFC dataset, outperforming state-of-the-art methods by 2.7%.
• Demonstrated superior performance across diverse SRT datasets.
• Showcased effectiveness in downstream tasks like gene expression denoising and cancer heterogeneity analysis.

Conclusions:

• MambaST offers a powerful new approach for spatial domain identification in SRT.
• The framework effectively captures long-range dependencies and improves biological consistency.
• MambaST has broad applicability in analyzing complex biological tissues and diseases.