PLNMFG: Pseudo-label guided non-negative matrix factorization model with graph constraint for single-cell multi-omics...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell multi-omics sequencing allows simultaneous analysis of diverse molecular data within individual cells.
• Accurate cell clustering is vital for interpreting complex biological functions from multi-omics data.
• Existing integration methods struggle to incorporate prior biological knowledge and capture cross-omic interactions effectively.

Purpose of the Study:

• To develop a novel computational model, PLNMFG, for robust and accurate clustering of single-cell multi-omics data.
• To address limitations in current methods by integrating prior biological knowledge and unified latent representation learning.
• To improve the capture of complementary information and cross-platform interactions within multi-omics datasets.

Main Methods:

• Developed PLNMFG, a non-negative matrix factorization model combining latent representation and cluster structure learning.
• Implemented adaptive imputation for handling data dropouts and used prior pseudo-labels as constraints.
• Incorporated Graph Laplacian constraints for preserving multi-omics data structure and adaptively learned omic weights.

Main Results:

• PLNMFG achieved superior clustering accuracy compared to existing methods across 8 benchmark datasets.
• The model demonstrated computational efficiency in processing single-cell multi-omics data.
• PLNMFG successfully preserved double similarity information and captured intrinsic data structures.

Conclusions:

• PLNMFG offers a robust framework for single-cell multi-omics data integration and clustering.
• The model's ability to incorporate prior knowledge and capture cross-omic relationships enhances biological insights.
• PLNMFG represents a significant advancement in analyzing complex single-cell multi-omics data.