scMSDA: A Novel Multi-View Fusion Framework for Single-Cell RNA-seq Data Clustering with Semantic and Distribution...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high resolution for cellular heterogeneity but faces analytical challenges.
• Existing clustering methods often overlook local data structures, impacting the capture of semantic relationships.
• Technical noise and high dimensionality in scRNA-seq data complicate accurate downstream analysis.

Purpose of the Study:

• To develop a novel multi-view fusion framework, scMSDA, for enhanced scRNA-seq data clustering.
• To learn robust cell representations by enforcing semantic consistency and distribution alignment.
• To improve the accuracy and reliability of scRNA-seq data clustering for biological insights.

Main Methods:

• scMSDA employs data augmentation via dropout regularization and global feature aggregation.
• A distance-guided adaptive-negative contrastive learning strategy dynamically adjusts negative sample contributions.
• Iterative centroid refinement and optimal transport (OT)-based cross-view alignment enforce distribution alignment and cluster separation.

Main Results:

• scMSDA demonstrates superior performance across 17 public scRNA-seq datasets.
• The proposed method outperforms 10 baseline clustering approaches based on multiple metrics.
• Experimental results validate the effectiveness of scMSDA in learning robust representations for scRNA-seq data.

Conclusions:

• scMSDA provides an effective multi-view fusion framework for scRNA-seq data clustering.
• The method successfully addresses challenges of sparsity, dimensionality, and noise in scRNA-seq analysis.
• scMSDA offers a significant advancement in computational biology for understanding cellular heterogeneity.