Bayesian model selection reveals biological origins of zero inflation in single-cell transcriptomics

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for studying cellular heterogeneity.
• High variability and excessive zero counts in scRNA-seq data pose analytical challenges.
• The origin and treatment of zero counts in scRNA-seq are debated, with no consensus on zero-inflated models.

Purpose of the Study:

• To rigorously assess the presence and origin of zero inflation in scRNA-seq data.
• To evaluate the utility of zero-inflated models versus standard count distributions.
• To provide guidance on appropriate statistical modeling for scRNA-seq analysis.

Main Methods:

• Application of a Bayesian model selection framework.
• Analysis of multiple biologically realistic scRNA-seq datasets.
• Comparison of parameter estimates from zero-inflated and non-zero-inflated models.

Main Results:

• Unambiguous demonstration of zero inflation in scRNA-seq datasets.
• Identification of biological factors, rather than technical artifacts, as the primary cause of zero inflation.
• Finding that parameter estimates from zero-inflated negative binomial models are unreliable indicators of zero inflation.

Conclusions:

• Zero inflation is a genuine feature of scRNA-seq data, stemming from biological processes.
• Despite zero inflation, a generalized linear model with a non-zero-inflated negative binomial distribution is proposed as a robust reference model.
• This recommendation simplifies scRNA-seq analysis by avoiding complex, potentially unreliable, zero-inflated models.