Toward simultaneous pseudo-space reconstruction and cell-type deconvolution of single-cell spatial transcriptome using SpaDicer.

Single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) offer complementary insights into tissue heterogeneity. Each modality is characterized by inherent trade-offs: scRNA-seq offers high resolution but lacks spatial context, while ST retains spatial information but compromises resolution. Here, we introduce SpaDicer, an end-to-end deep-learning framework designed to bridge these gaps by unifying pseudo-spatial reconstruction of single cells and cell-type deconvolution of ST spots. SpaDicer employs cross-domain feature disentanglement to extract domain-invariant feature encoding both cellular identity and spatial localization. An adaptive loss weighting strategy harmonizes multiple objectives across data reconstruction, cell-type deconvolution, and spatial regression tasks. We evaluated SpaDicer on two simulated datasets and on five real-world datasets covering a wide range of human tissue contexts, demonstrating consistent improvement over state-of-the-art methods in reconstructing cellular spatial localization and resolving cell-type composition. Through a unified multi-task learning framework, SpaDicer provides a powerful tool for advancing spatially informed biological discovery.