Answers are grounded in this paper and cite their sources.
Transferring tissue domain labels across spatial transcriptomics sections remains a challenge. A key problem is how to jointly address technical dropout and inter-section batch effects while preserving spatial tissue organization. Existing methods usually address only part of these challenges and often fail to maintain robust annotation across sections. In this study, we introduce AnnST, a robust framework for cross-section label transfer in spatial transcriptomics. AnnST constructs clean, dropout-perturbed, and batch effect-perturbed expression views. Contrastive learning improves robustness to dropout noise, while optimal transport reduces variation caused by batch effects. Label-aware spatial regularization is further used during supervised fine-tuning to preserve local tissue organization. We evaluate AnnST on the DLPFC Visium, SEA-AD MERFISH, and human breast cancer Visium datasets. AnnST achieves competitive annotation performance across these datasets, while showing stable results across random seeds. It also remains robust under multiple technical perturbations. Downstream analyses further show that the predicted annotations preserve cortical laminar organization and support biologically meaningful functional analysis. AnnST provides a robust and biologically interpretable framework for cross-section annotation in spatial transcriptomics and offers a practical solution for downstream spatial analysis in datasets generated using different ST platforms.
No takes yet. Share an insight, caveat, or question.
Guo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: