CoGAPS 3: Bayesian non-negative matrix factorization for single-cell analysis with asynchronous updates and sparse...

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Bayesian factorization methods like Coordinated Gene Activity in Pattern Sets (CoGAPS) are valuable for single-cell data analysis.
• High computational costs limit their application to large single-cell datasets.
• Existing parallelization strategies are constrained by CoGAPS' prior distributions.

Purpose of the Study:

• To overcome computational limitations of Bayesian matrix factorization for single-cell data analysis.
• To enhance the efficiency of the CoGAPS algorithm for large-scale single-cell datasets.

Main Methods:

• Developed a new software framework for parallel matrix factorization within the CoGAPS R/Bioconductor package.
• Implemented asynchronous updates for sequential steps to boost computational efficiency.
• Introduced new software architecture and sparse data structures to minimize memory usage.

Main Results:

• The enhanced CoGAPS software framework significantly improves computational efficiency for Bayesian matrix factorization.
• Algorithmic and software architecture improvements reduce memory overhead for single-cell data.
• The new framework enables analysis of datasets up to 1000 times larger than previously possible.

Conclusions:

• The updated CoGAPS software dramatically enhances the efficiency of Bayesian matrix factorization.
• This advancement allows for the analysis of significantly larger single-cell datasets.
• The improved computational performance makes CoGAPS more accessible for large-scale single-cell research.