Abstract
Single-cell clustering is a pivotal technique in deciphering single-cell omics data, enabling us to gain insights into cell function and heterogeneity. However, the presence of batch effects in single-cell omics data poses significant challenges to the accuracy and reliability of single-cell clustering. To address these challenges, we develop a Batch Effect-Removed Clustering method(scBERC) that incorporates a novel single-cell augmentation strategy and contrastive learning. By augmenting cells with artificially induced variations, our approach captures the full range of cell states, accounting for both biological heterogeneity and batch-specific variations. This augmentation strategy enriches the training data while the contrastive learning strategy better extracts real biological features, allowing the deep subspace clustering algorithm to learn robust and generalized patterns, thereby mitigating the biases introduced by batch effects. Experimental results on twelve real single-cell datasets demonstrate that scBERC outperforms existing single-cell clustering methods, significantly improving clustering performance and facilitating downstream analysis.