Identifying Subpopulations of Cells in Single-Cell Transcriptomic Data: A Bayesian Mixture Modeling Approach to Zero...

Area of Science:

• Computational Biology
• Genomics
• Statistical Modeling

Background:

• Identifying cell subpopulations is crucial for analyzing single-cell RNA sequencing (scRNA-Seq) data.
• Existing machine learning methods often lack uncertainty estimation for cluster assignments.
• scRNA-Seq data presents challenges due to a high proportion of zero counts (dropout), complicating probabilistic modeling.

Purpose of the Study:

• To develop a novel probabilistic model for scRNA-Seq data analysis that addresses dropout and provides uncertainty estimates.
• To improve the identification of cell subpopulations and gene expression patterns within these subpopulations.
• To offer a more robust alternative to existing methods for modeling scRNA-Seq count data.

Main Methods:

• Development of a Dirichlet process mixture model incorporating cell-level and transcript-level mixtures.
• Utilizing a zero-inflated negative binomial distribution to model transcript counts, accounting for dropout.
• Employing a Bayesian approach to model gene expression and quantify uncertainty in cluster assignments.

Main Results:

• The proposed model outperforms multinomial and non-zero-inflated negative binomial models for scRNA-Seq counts.
• Demonstrated ability to distinguish cell subpopulations in mouse cortex and hippocampus scRNA-Seq data.
• Successfully identified gene sets characteristic of specific cell subpopulations.

Conclusions:

• The novel Bayesian Dirichlet process mixture model effectively handles dropout in scRNA-Seq data.
• This approach provides reliable cell subpopulation identification with quantified uncertainty.
• The method facilitates the discovery of marker genes for distinct cell types, advancing single-cell data analysis.