Finite mixtures of matrix variate Poisson-log normal distributions for three-way count data.

Area of Science:

• Bioinformatics
• Statistical Genomics
• Computational Biology

Background:

• Three-way data structures are common in biological studies, particularly in RNA sequencing, involving genes, conditions, and time points.
• Clustering gene expression data is crucial for discovering gene co-expression networks.

Purpose of the Study:

• To propose a novel statistical model for clustering RNA sequencing data using mixtures of matrix variate distributions.
• To leverage the matrix variate structure to simultaneously consider conditions and time points, reducing parameter estimation complexity.

Main Methods:

• A mixture of matrix variate Poisson-log normal distributions is proposed for clustering RNA sequencing read counts.
• Three parameter estimation frameworks are presented: Markov chain Monte Carlo, variational Gaussian approximation, and a hybrid approach.
• Model selection is performed using various information criteria.

Main Results:

• The proposed models were applied to both real and simulated RNA sequencing data.
• The approaches demonstrated the ability to successfully recover underlying cluster structures.
• Simulation studies showed good parameter recovery when true model parameters were known.

Conclusions:

• The mixture of matrix variate Poisson-log normal distribution provides an effective method for clustering RNA sequencing data.
• The developed methods offer robust parameter estimation and cluster recovery for gene expression analysis.
• An R package is available for implementing the proposed clustering approaches.