stSCI: A multi-task learning framework for integrative analysis of single-cell and spatial transcriptomics data

Spatial transcriptomics (ST) preserves spatial context in gene expression analysis yet faces limitations like low resolution and RNA capture inefficiency. To address these, we present stSCI, a computational method integrating single-cell (SC) and ST data into a unified, batch-corrected embedding space. stSCI employs a fusion module with three specialized optimization tasks to generate biologically preserved joint latent representations, enabling five key analyses: spatial domain identification in single/multi-slice ST data, ST deconvolution predicting cell type proportions in low-resolution spots, SC spatial coordinate reconstruction using ST references, and crossmodality batch correction. Evaluated on 13 different ST datasets spanning sequencing- and imaging-based platforms, and benchmarked against 27 state-of-the-art methods, stSCI improves spatial domain identification, maps cell type proportions in ST data, accurately reconstructs tissue architecture and regional structures, and integrates SC/ST datasets by removing batch effects without compromising biological signals. In a key application, stSCI successfully resolves the dynamic spatiotemporal response of a lymphatic niche during <i>Salmonella</i> infection, demonstrating its power to generate novel biological insights from complex disease models. stSCI's robustness and versatility make it a powerful tool for uncovering tissue organization and molecular functions.