scSCC: A swapped contrastive learning-based clustering method for single-cell gene expression data

Cell clustering plays a pivotal role in deciphering the intricacies of cell types, facilitating subsequent cell annotation endeavors within scRNA-seq data analysis. In this paper, we propose a novel swapped contrastive clustering algorithm for scRNA-seq data called scSCC. scSCC combines two contrastive learning modules, namely the instance contrastive learning module and the swapped prediction module, to extract clustering-friendly cell representations. Through the combination of swapped prediction module and instance contrastive learning module, scSCC can retrieve disentangled cell representations and amplify the clustering signals in the latent space, leading to satisfactory clustering performance. Different from existing contrastive-learning-based scRNA-seq data clustering algorithms, the swapped prediction module of scSCC injects clustering signals to the latent space through some clustering prototypes. The swapped prediction module encourages cells of the same cluster to gravitate toward the common clustering prototype and naturally stay away from other prototypes in the latent space, hence cell representations obtained by scSCC are more clustering-friendly compared to other algorithms. Experimental results on real scRNA-seq datasets show that scSCC achieves improved clustering performance compared with the benchmark methods. The ablation study on two contrastive modules exhibits the promotion by the combination of instance learning module and swapped prediction module. The source codes are available at the GitHub website (EnchantedJoy/scSCC).