SMURF: soft-segmentation for single-cell reconstruction and topological analysis of spatial transcriptomic data

High-resolution spatial transcriptomics requires computational methods to accurately assign transcripts to individual cells. We present SMURF (Segmentation and Manifold UnRolling Framework), a cross-platform soft-segmentation algorithm that uses deep learning to map mRNAs from capture spots to nearby nuclei. SMURF also unrolls complex tissue architectures by projecting cells onto Cartesian coordinates, enabling analyses of cell-type organization and gene expression gradients. We show that SMURF assigns mRNAs to single cells more accurately than existing approaches and robustly unrolls complex tissues to reveal zonated transcriptional programs and cell-type organization across multiple tissues and spatial transcriptomic technologies. To showcase the biological insights enabled by SMURF, we segment over 400,000 cells from the mouse ileum using Visium HD data. We identify zonated gene expression programs along the maturing intestinal villus and the transcription factors that regulate them. Importantly, we show that gene expression gradients along the proximal-distal axis of the intestine accumulate in the upper villus and that upper villus gene expression is reprogrammed by environmental signals in the lumen, suggesting that environmental inputs are major determinants of regional transcriptional identity. Together, these results establish SMURF as a powerful framework for analyzing gene expression of cells within native tissue contexts.