CauST: Causal Gene Intervention for Robust Spatial Domain Identification

Spatial transcriptomics methods identify tissue domains by clustering spatial gene expression patterns, but current approaches select genes using variance-based heuristics such as highly variable genes (HVGs). Many HVGs reflect donor-specific noise rather than genes that truly drive spatial organization, which leads to poor generalization when models are applied across tissue slices from different donors. This project addresses this limitation by introducing CauST, a framework for identifying causally important genes for spatial domain identification. CauST evaluates each gene through in-silico gene knockout experiments, measuring how much removing the gene changes the learned spatial embedding produced by a spatial graph neural network. To ensure robustness, the method then applies a cross-slice invariance criterion, selecting genes whose effects are both strong and stable across multiple donors. The resulting causally invariant gene set is used to retrain spatial domain identification models, enabling improved cross-donor performance without modifying the underlying model architecture. The project will deliver a production-quality open-source CauST Python package, integration with multiple spatial transcriptomics backbones (STAGATE, GraphST, and SpaGCN), a multi-dataset benchmark evaluating cross-slice robustness, visualization tools for interpreting causal gene importance, and tutorial notebooks and documentation to support adoption by the spatial transcriptomics community.