SpaDiff denoises sequence-based spatial transcriptomics via diffusion process

Spot-swapping in sequence-based spatial transcriptomics redistributes RNA molecules across neighboring spots, distorting spatial gene expression and biasing downstream analysis. We introduce SpaDiff, a mass-conserving denoising framework that corrects this artifact through score-guided reverse diffusion. Across simulated and real spatial transcriptomics datasets, SpaDiff consistently improves the recovery of spatial gene-expression distributions. In particle-level simulations with known ground truth, SpaDiff robustly restores gene-level spatial patterns across diverse contamination regimes. In real tissues, SpaDiff relocalizes misplaced transcripts in human-mouse chimeric samples, improves spatial-domain recovery in human dorsolateral prefrontal cortex, and preserves local tumor heterogeneity while sharpening regional structure in colorectal cancer. These results show that SpaDiff improves the spatial accuracy of gene expression and supports more accurate downstream analyses such as clustering and spatial domain identification.