SDMCC: Sample-wise Debiased Multilevel Contrastive Clustering for Single-cell Gene Expression Data

Single-cell gene expression profiling has emerged as a powerful technology for dissecting complex tissues at unprecedented resolution. Accurate cell clustering is a fundamental computational prerequisite for cell type identification. In recent years, numerous single-cell contrastive clustering algorithms have been developed to more effectively identify individual cells and characterize cellular heterogeneity. Despite these advances, the presence of noise and data sparsity in single-cell datasets continues to impede algorithmic performance. Mainstream clustering methods face three principal limitations: (i) most clustering methods process data in mini-batches, and batch-related value shifts can destabilize feature extraction; (ii) existing contrastive learning strategies fail to achieve multilevel, coordinated information fusion, making them prone to collapsed solutions; and (iii) noise and sparsity can cause centroid shifts, yet existing methods lack effective mechanisms for dynamic centroid updating. To address these challenges, we propose sample-wise debiased multilevel contrastive clustering with shrinkage risk regularization (SDMCC). Instead of directly learning from the original sparse data, SDMCC applies a sample-wise correction module in conjunction with a batch-size adaptation module to mitigate distortion. Furthermore, we introduce a multilevel contrastive learning strategy that integrates instance-level and cluster-level information, enabling the capture of fine-grained cellular variations while maintaining coarse-grained semantic consistency across cell subpopulations. In addition, a shrinkage risk regularization term prevents convergence to trivial solutions. Extensive experiments conducted on multiple single-cell datasets demonstrate that SDMCC not only achieves superior clustering accuracy but also effectively uncovers biologically meaningful patterns in gene expression, thereby highlighting its potential for broader biomedical applications.