Two-Stage Multi-View Graph Spectral Clustering for Single-Cell RNA-Seq Data.

Introduction: The appearance of single-cell RNA sequencing (scRNA-seq) data has brought a distinctive perspective to studying gene expression at the cell level. However, it faces challenges such as large data volume, sparsity, heterogeneity, and the curse of dimensionality. Current clustering methods still face many challenges in studying cell type distribution and have not utilized the structural relationship information between cells. Methods: To avoid the insufficiency of the single characteristic space of scRNA-seq data in characterizing cell function, this paper constructs multiple view characteristic spaces and utilizes multi-view learning to characterize scRNA-seq information from distinctive perspectives comprehensively. In multi-view learning, the similarity graph is divided into weighted learning and structural learning stages. Through weighting the multi-view similarity graphs, the significance of diverse views and features is underscored. During the structural stage, the emphasis is placed on uncovering potential relationships among different views by preserving common edges in the multi-view similarity graphs. The optimization of the attribute and structure graphs was conducted separately by the alternating direction multiplier method. Results: The performance of the MVGSC was validated using eight different scales of real scRNA-seq datasets, and the experimental results showed that the proposed multi-view clustering method significantly surpasses other single-view clustering methods and multi-view clustering methods. Discussion: When the features of scRNA-seq data are complex and there are significant differences between views, the two-stage multi-view graph method can better capture the complex relationships in the data, demonstrating superior performance compared to a single framework. Conclusion: Two-stage multi-view learning can more accurately capture complex relationships in the data, thereby improving the accuracy of the model. It can also better capture consistency and complementary information in multi-view data, thereby enhancing the generalization ability of the model.