SpCAST enables scalable and interpretable integration of single-cell RNA sequencing and single-cell-resolved spatial transcriptomics
Single-cell-resolution spatial transcriptomics (scST) preserves tissue architecture but often provides targeted or sparse transcriptomic measurements, whereas scRNA-seq offers broader coverage without spatial context. We present SpCAST, a scalable and interpretable framework that uses scRNA-seq references to transfer cell identity, reconstruct expression and expose gene-level decision evidence in scST. SpCAST jointly learns reference-cell classification, reference--query alignment and query reconstruction in mini-batches, avoiding the need for a global reference-by-query correspondence matrix. Spatially Aware Gene Attribution (SAGA) approximates the learned decision function with a sparse additive Kolmogorov--Arnold network. Across 53 sections comprising 413,404 spatial cells from five technologies, SpCAST achieved the highest aggregate annotation rank among seven methods and scaled to ten million simulated cells. Controlled masking recovered cell-type-associated expression signals and improved spatial marker concordance. SAGA further resolved expression-dependent gene evidence and distinguished evidence retained or attenuated across intra- and cross-species reference settings.