Linking large-scale pan-cancer single-cell and spatial transcriptomics profiles using SDI 9110

Description Single-cell and spatial transcriptomics have become pivotal tools for elucidating the phenotypic heterogeneity of immune cells and identifying spatially recurrent immune niche states. These technologies have produced extensive datasets containing whole-transcriptome profiles of dissociated single cells and spatially resolved tissues. However, both approaches are subject to significant technical and biological batch effects, which complicates the assignment of cell types from single-cell data to spatial transcriptomics datasets and hinders the comparative analysis of spatial niche and region labels across tissue slices. To overcome these limitations, we developed Synchronized Deconvolution and Integration (SDI), a deep learning model designed to integrate dissociated single-cell and spatial transcriptomics data across diverse platforms and spatial resolutions. Pretrained on over 5 million dissociated cells and more than 1 million spatial spots from pan-cancer tissues, SDI demonstrates the ability to accurately project cell type labels from single-cell to spatial datasets, with adaptability to novel datasets. We show that SDI facilitates batch-unbiased deconvolution of immune cell composition and provides consistent predictions of tertiary lymphoid structures (TLS) across over 400 tissue slices from more than 10 cancer types. In conclusion, SDI enables integrated analyses linking immune cell types from both single-cell and spatial transcriptomics.