Stochastic gradient descent estimation of generalized matrix factorization models with application to single-cell RNA...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables gene expression quantification at the single-cell level, crucial for studying cellular heterogeneity.
• Dimensionality reduction is essential for visualizing and analyzing complex scRNA-seq data but is computationally challenging due to data size.
• Existing methods like principal component analysis struggle with large-scale scRNA-seq datasets.

Purpose of the Study:

• To develop a generalized matrix factorization model for scRNA-seq data analysis.
• To propose a scalable algorithm for efficient dimensionality reduction on large single-cell datasets.
• To provide a robust and accurate method for analyzing cellular heterogeneity and gene expression dynamics.

Main Methods:

• A generalized matrix factorization model based on exponential dispersion family distributions was developed.
• A scalable adaptive stochastic gradient descent algorithm was designed for efficient model estimation.
• The proposed method was benchmarked against state-of-the-art techniques using numerical experiments and real-world biological data.

Main Results:

• The generalized matrix factorization model encompasses many existing dimensionality reduction approaches.
• The stochastic gradient descent algorithm enables efficient analysis of datasets containing millions of cells.
• The proposed method demonstrated superior performance in execution time, memory usage, and reconstruction accuracy compared to existing methods.
• The method successfully scaled to analyze large single-cell datasets, facilitating dimensionality reduction.

Conclusions:

• The novel generalized matrix factorization model and scalable algorithm offer significant advancements in single-cell data analysis.
• The method provides a powerful tool for researchers to explore cellular heterogeneity and gene expression dynamics in large scRNA-seq datasets.
• An open-source R package, sgdGMF, is available for implementing the discussed methods.