<p>Advances in single-cell RNA sequencing (scRNA-seq) technology have facilitated the analysis of genome-wide transcriptional profiles and enabled researchers to address biological problems at the single-cell level. Clustering methods are the primary data analysis approach in single-cell studies. However, clustering scRNA-seq data remains statistically and computationally challenging due to noise caused by substantial dropout events and the high dimensionality resulting from the large number of genes. In this study, based on a single-cell model, we developed a deep embedded subspace clustering method, scDESC. This approach combines a denoising autoencoder with a zero-inflated negative binomial (ZINB) distribution for deep embedded subspace clustering. Specifically, our method simulates the scRNA-seq data generation process through a denoising ZINB model-based autoencoder. In this process, scRNA-seq data is mapped into a low-dimensional space to learn subspace bases. Subsequently, by applying three constraint conditions to the subspace, the learned bases are refined to enhance the representation learning capability of the deep neural network. Simultaneously, scDESC effectively integrates feature representation learning and clustering while reducing data noise. The experimental results on four representative real datasets of single-cell sequencing demonstrate that scDESC has promising applications in the field of single-cell clustering.</p>