Abstract
Spatial mapping of multi-slice multi-omics data enables the identification of shared and slice-specific cellular components across spatiotemporal axes. However, conventional graph neural networks assume uniform contributions from neighboring cells, neglecting directional and angular influences that shape central cell states and limiting their ability to dissect complex spatial structures. Here, we present stLVG, a vector-guided lightweight graph model for spatial mapping, label transfer, and niche identification across multi-slice multi-omics datasets. Specifically, stLVG (1) learns two distinct shared feature spaces across slices by aggregating neighbor information through adversarial learning with distance- and direction-informed weights and (2) integrates these features via a multi-view contrastive learning framework. Compared to existing methods, stLVG achieves superior performance across technologies, modalities, and resolutions; it accurately delineates tumor edge regions in breast cancer samples. Notably, it uses pre-computed weights and can be efficiently executed on a standard laptop within minutes, ensuring scalability to large-scale spatial omics analyses.