Clustering single-cell multi-omics data via multi-subspace contrastive learning with structural smoothness.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell multi-omics data integration is crucial for understanding cell heterogeneity.
• Existing clustering methods struggle with noise, sparsity, and intercellular heterogeneity in single-cell data.
• Current multi-omics approaches often overlook diverse feature extraction and post-fusion smoothing.

Purpose of the Study:

• To propose a novel method, scMUSCLE, for robust single-cell multi-omics data clustering.
• To address limitations in feature extraction and smoothing consistency in existing integration methods.
• To improve the accuracy and robustness of clustering for diverse cell types and states.

Main Methods:

• Leveraging degree structure to enhance structural diversity within each omics modality.
• Employing multi-subspace contrastive learning for improved cross-modal feature exploration.
• Utilizing an adaptive graph convolution clustering module with intra-cluster smoothness feedback.

Main Results:

• Demonstrated effectiveness and robustness of scMUSCLE on four benchmark multi-omics datasets.
• scMUSCLE successfully addresses challenges in single-cell multi-omics data integration.
• The method shows superior performance in clustering complex cellular data.

Conclusions:

• scMUSCLE offers a significant advancement in single-cell multi-omics data analysis.
• The proposed method enhances feature extraction and smoothing for more accurate clustering.
• This approach provides a robust framework for uncovering regulatory mechanisms in diverse cell populations.