HarveST uses a heterogeneous graph learning framework to reveal spatial transcriptomics patterns.

Spatial transcriptomics enables in situ gene expression profiling, yet precise spatial domain identification and marker gene detection remain challenging. We present HarveST, a heterogeneous graph-based framework that integrates spatial, transcriptomic, and gene-gene interaction data through a unified computational model. HarveST employs dual learning strategies: self-supervised learning for feature extraction and partially supervised refinement for domain delineation. Additionally, it implements a Random Walk with Restart algorithm for identifying spatial domain-marker spatially variable genes (SVGs). Applied to human cortical tissue, mouse olfactory bulb, and tumor microenvironments across multiple platforms, HarveST demonstrates superior performance in detecting biologically meaningful spatial domains and associated marker genes. HarveST further supports joint analysis across consecutive spatial transcriptomics sections, enabling consistent reconstruction of functional domains across slices. By capturing both spatial topology and molecular relationships in a single graph-theoretical framework, HarveST advances spatial transcriptomics analysis beyond conventional clustering approaches, offering deeper insights into tissue architecture and cellular interactions in normal and pathological contexts.