Abstract The rapid advancement of single-cell RNA sequencing (scRNA-seq) has significantly propelled biological research at the cellular level, particularly in uncovering cellular heterogeneity. However, the high dimensionality, sparsity, and frequent zero counts caused by dropout events present challenges for clustering, a critical step in scRNA-seq data analysis. To address these challenges, we introduce scDeGAEsa, a deep graph learning framework for unsupervised cell-type clustering. scDeGAEsa employs a self-adaptive cell graph and a doubly enhanced graph autoencoder(GAE) to jointly learn cell-cell relationships and cluster assignments. The adaptive graph is constructed from low-dimensional representations produced by a deep autoencoder coupled with a zero-inflated negative binomial (ZINB) model. The doubly enhanced autoencoder integrates these representations into a joint embedding, from which clusters are derived via a self-optimized module. During training, the adjacency matrix is dynamically updated to reflect refined relationships. Extensive experiments on 12 scRNA-seq datasets demonstrate that scDeGAEsa consistently outperforms state-of-the-art clustering methods.
Liu et al. (Thu,) studied this question.