BANMF-S: a blockwise accelerated non-negative matrix factorization framework with structural network constraints for...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high-resolution transcriptome profiling but is limited by technical 'dropout' events.
• These dropouts represent zero counts and can obscure true biological variation and hinder downstream analysis.

Purpose of the Study:

• To develop and validate a novel computational framework, BANMF-S, for accurate imputation of dropout events in scRNA-seq data.
• To improve the reliability of cell heterogeneity studies and trajectory inference using imputed data.

Main Methods:

• Proposed Blockwise Accelerated Non-negative Matrix Factorization with Structural network constraints (BANMF-S).
• Constructed gene-gene similarity networks using PPI data and cell-cell similarity networks via Minimum-Spanning Trees.
• Integrated these networks as regularizations within a matrix factorization framework.
• Employed a blocklization strategy with distributed Stochastic Gradient Descent for accelerated optimization.

Main Results:

• BANMF-S effectively imputes technical zeros in scRNA-seq data by leveraging gene and cell similarity networks.
• The method demonstrated improved accuracy in downstream clustering and pseudo-trajectory inference tasks.
• Performance of BANMF-S surpassed seven existing state-of-the-art imputation algorithms in simulations and real datasets.

Conclusions:

• BANMF-S provides a robust and efficient solution for addressing dropout events in scRNA-seq data.
• The framework enhances the biological insights obtainable from scRNA-seq by improving data quality.
• The developed method offers superior performance compared to current imputation techniques.