Abstract
The growing availability of spatial transcriptomics data offers key resources for annotating query datasets using reference datasets. However, batch effects, unbalanced reference annotations, and tissue heterogeneity pose significant challenges to alignment analysis. Here, we present stGuide, an attention-based supervised graph learning model designed for cross-slice alignment and efficient label transfer from reference to query datasets. stGuide leverages supervised representations guided by reference annotations to map query slices into a shared embedding space using an attention-based mechanism. It then assigns spot-level labels by incorporating information from the nearest neighbors in the learned representation. Using human dorsolateral prefrontal cortex and breast cancer datasets, stGuide demonstrates its capabilities by (i) producing category-guided, low-dimensional features with well-mixed slices; (ii) transferring labels effectively across heterogeneous tissues; and (iii) uncovering relationships between clusters. Comparisons with state-of-the-art methods demonstrate that stGuide consistently outperforms existing approaches, positioning it as a robust and versatile tool for spatial transcriptomics analysis.