Probabilistic count matrix factorization for single cell expression data analysis

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• High-throughput single-cell sequencing reveals cellular transcriptome diversity and variability.
• Statistical challenges arise in summarizing and visualizing complex single-cell expression data.
• Traditional methods like Principal Component Analysis (PCA) are limited by Euclidean distance for over-dispersed count data with dropouts.

Purpose of the Study:

• To introduce a novel probabilistic Count Matrix Factorization (pCMF) approach for single-cell expression data analysis.
• To develop a method that can jointly represent cells and genes in a low-dimensional space.
• To provide a statistically robust framework for visualizing and clustering single-cell data.

Main Methods:

• Developed a sparse Gamma-Poisson factor model for probabilistic Count Matrix Factorization (pCMF).
• Inferred the hierarchical model using a variational Expectation-Maximization (EM) algorithm.
• Evaluated pCMF against standard representation methods like t-SNE for single-cell data.

Main Results:

• pCMF jointly builds low-dimensional representations of cells and genes.
• The probabilistic framework provides a suitable geometry for single-cell data visualization.
• pCMF demonstrates powerful data compression for clustering purposes, outperforming existing methods.

Conclusions:

• pCMF offers an effective probabilistic approach for analyzing single-cell expression data.
• The method enhances data visualization and clustering capabilities.
• The pCMF R-package is available for broader research use.