Abstract Deciphering the cellular composition of spatial spots in spatial transcriptomics (ST) data is fundamental for elucidating the heterogeneity of tissue spatial structures. However, existing models often require retraining for each new deconvolution task, reflecting limitations in both generalization performance and computational efficiency. To address this problem, we design a foundational diffusion model to deconvoluting spatial transcriptomics based on combinatorial optimization, termed DiSCO. DiSCO formulates the deconvolution of ST data as a task-specific deconvolutional combinatorial optimization (CO) problem, wherein single cells (SCs) are assigned to spatial spots to optimally preserve the gene expression profiles of each spot. DiSCO introduces a bipartite graph diffusion model as an optimization solver, specifically designed to be generalizable to any new deconvolutional CO problem. Pretrained on a large number of deconvolution tasks using gene expression profiles of both SCs and spatial spots as inputs, DiSCO learns the distribution of true solutions and generates approximate solutions through sampling, thereby enabling the determination of the cellular composition for each spot. As a generalizable deconvolution solver, the DiSCO is evaluated by experiments on both simulated datasets and real datasets, demonstrating that the pretrained DiSCO model performs effectively and efficiently on datasets with varying resolutions and different numbers of genes, thus highlighting its capacity to effectively generalize to diverse datasets.
Liu et al. (2026) studied this question.