CellDiffusion: Enhancing scRNA-seq data integration with accurate inference of cell-cell association.

UNLABELLED: Integrating single-cell transcriptomic datasets across laboratories, conditions, and technologies remains challenging due to cell-type-specific batch effects and heterogeneous biological compositions. Current methods often fail to separate genuine biological variation from technical noise, misaligning distinct cell populations or leaving batch effects uncorrected. To address these challenges, we present CellDiffusion, an unsupervised graph-based model that propagates cellular features through a diffusion process to strengthen meaningful cell-cell associations in data integration. By learning a cell-cell attention network, CellDiffusion adaptively identifies corresponding populations across batches. Our approach applies cell-specific correction in data integration, generating cell embeddings that align cells across batches while preserving intrinsic cell types and developmental trajectories. We validate CellDiffusion across various datasets, including the human endometrium, one of the most dynamic and complex tissues featuring extensive cellular and molecular changes with the menstrual cycle. This framework also improves other essential analyses such as label transfer, clustering, and trajectory inference. CellDiffusion represents a powerful paradigm for accurate single-cell transcriptomic integrative analysis.

TEASER: A diffusion-based framework removes technical noise and preserves biological variation in single-cell datasets.