SpatialMamba: a graph-based model for spatial transcriptomics with feature- and degree-informed spot prioritization for spatial domain annotation

Spatial transcriptomics (ST) enables high-resolution mapping of tissue architecture, yet automated spatial domain annotation remains challenging because existing methods often overemphasize local neighborhoods, underuse biological priors, and fail to preserve long-range tissue continuity. Here we present SpatialMamba, a biologically informed framework for spatial domain annotation in high-resolution ST data. SpatialMamba integrates spatial, transcriptomic, and protein-protein interaction views within a unified graph representation and combines a Mamba-based global branch with a graph convolutional local branch to capture both long-range anatomical dependencies and fine-scale topology. A graph-guided fusion strategy incorporates gene-level biological priors, and joint optimization with reconstruction and contrastive objectives improves structural consistency and boundary discrimination. Across multiple tissues and evaluation settings, SpatialMamba improves whole-slice annotation, generalizes from partially labeled regions, resolves ambiguous domains, and preserves continuous biological gradients in the latent space, providing an accurate and scalable framework for automated tissue mapping.