AI Insight
Researchers have developed SpCAST, a computational framework that uses Kolmogorov-Arnold networks to identify cell types in spatial transcriptomics data by comparing them to reference datasets. The tool addresses key limitations in current spatial transcriptomics methods, which are either restricted to analyzing predetermined genes or suffer from incomplete data capture. Testing on 53 datasets containing over 413,000 spatial cells across five different technologies showed SpCAST achieved accurate cell-type identification faster than existing methods while also reconstructing gene expression patterns and identifying marker genes.
Why it matters
This tool could accelerate research in disease biology and drug development by enabling scientists to more efficiently and accurately map which cell types are present in tissue samples and what genes they express. The method's ability to work across species and handle incomplete data makes it particularly valuable for analyzing diverse biological samples.
arXiv:2605.26904v1 Announce Type: new
Abstract: Single-cell-resolution spatial transcriptomics profiles gene expression at cellular locations in native tissues, yet accurate cell-type annotation remains challenging: imaging-based platforms are constrained by targeted gene panels, whereas sequencing-based platforms often suffer from sparse molecular capture and dropout. Reliable transfer of cell-type labels from single-cell RNA sequencing references is therefore critical for interpreting targeted and sparse spatial datasets. Here, we present SpCAST, a Kolmogorov–Arnold network-based framework for reference-guided spatial transcriptomics analysis. SpCAST captures nonlinear mappings between reference and spatial expression profiles and uses feature attribution to prioritize genes supporting cell-type predictions. Within a unified framework, SpCAST performs cell-type label transfer, spatial gene-expression reconstruction and marker-gene candidate prioritization. We benchmarked SpCAST on 53 datasets comprising 413,376 spatial cells across five technologies and diverse tissue contexts. SpCAST achieved competitive annotation performance with reduced runtime relative to representative existing methods. Case studies demonstrated that SpCAST supports cross-species label transfer and candidate assignment of originally unlabeled cells. It also reconstructs marker-gene expression patterns with improved spatial concordance and prioritizes cell-type-associated marker genes. Together, these results support SpCAST as an efficient and interpretable framework for extracting cell-type and gene-level information from targeted and sparse single-cell-resolution spatial transcriptomics data.