Leverage multi-scale dependencies of single cells through a physics-informed deep diffusion model

Single-cell transcriptomics is typically analyzed by modeling gene expression within individual cells and hypothetical cell adjacencies. However, existing computational methods often struggle to capture the full scope of multi-scale dependencies among single cells in feature space, limiting their effectiveness and robustness in downstream applications, such as cell type identification and cell fate inference. To address this limitation, we introduce scDiffusion, a graph-based diffusion model that incorporates long-range information propagation across cells to uncover intercellular dependencies from transcriptomic data. Inspired by diffusion in physics, scDiffusion integrates both local and global diffusion processes to reveal fine-grained structures and large-scale patterns. This approach models the hierarchical organization of cell graphs, capturing biological relationships among individual cells, sub-cell types, and cell types across multiple scales. scDiffusion achieves a high level perception of inherent cell types and potential lineages, and preserves cell identities in batch-imbalanced datasets. It enhances various downstream tasks, including batch correction, cell type annotation, and cell fate inference. We evaluated scDiffusion through extensive benchmarks, demonstrating its effectiveness compared to existing methods.