scSCDT: Self-contrastive neural network with deep topology mining for scRNA-seq data clustering

Abstract

Advancements in single-cell sequencing technologies have enabled researchers to better identify cells based on gene-level information. Cell clustering is a key task in single-cell analysis and plays an important role in distinguishing cell types. However, due to the high dimensionality and sparsity of scRNA-seq data, single-cell clustering remains a major challenge. Although many methods based on deep learning and machine learning have been developed for single-cell clustering, they often fail to capture the deep topological structure between cells, which limits clustering precision. In addition, most existing clustering approaches cannot effectively construct suitable sample pairs to optimize clustering models. To address these issues, we propose a topology-aware deep contrastive clustering model for single-cell data, named scSCDT. First, scSCDT employs a ZINB-based autoencoder to simultaneously learn cell embeddings and topological information, effectively handling the challenges posed by the high-dimensional and sparse nature of the data. Then, we introduce a dual clustering-guided loss to supervise the clustering task, combining a probabilistic soft assignment strategy and a hard pseudo-labeling strategy for optimization. Finally, based on the topological structure in the low-dimensional embedding space, we construct negative pairs within a single view and design a self-contrastive learning method to further improve clustering performance. We conduct extensive experiments on ten real scRNA-seq datasets and evaluate performance using four clustering metrics. The results indicate that scSCDT achieves strong clustering performance across multiple datasets, thereby facilitating more accurate cell type identification in single-cell transcriptomic analysis.