scCADI: Cell-Aware Diffusion Model for Single-Cell RNA-seq Imputation

Single-cell RNA sequencing (scRNA-seq) enables high-throughput transcriptomic analysis at single-cell resolution, revealing cell identities and underlying gene regulatory mechanisms. This technology provides critical insights into cellular differentiation and disease mechanisms. However, due to technical limitations, scRNA-seq data suffer from dropout events that may frustrate the analyses. Existing imputation methods mainly rely on scRNA-seq data without effectively integrating additional biological information, and still have limited ability to model complex data structures, making it difficult to capture intrinsic structure. To solve these limitations, we propose scCADI, a conditional diffusion-based imputation framework for scRNA-seq data. This framework aims to achieve robust imputation across multiple datasets and application scenarios while facilitating downstream analyses. In scCADI, an autoencoder first projects each cell into the latent space, where the observed data and metadata are integrated as conditional priors to guide the diffusion process in learning biologically meaningful expression patterns. With this design, scCADI achieves accurate gene expression recovery while preserving intrinsic biological variability among cells. We evaluated the imputation performance on multiple datasets, results show that scCADI outperforms other competing methods across various analytical tasks, including gene expression recovery, cell clustering, cellular trajectory inference and differential expression analysis.