Comparison and evaluation of statistical error models for scRNA-seq

Summary

This study evaluates statistical error models for scRNA-seq data, finding that while Poisson models suffice for sparse data, a negative binomial model is necessary for genes with sufficient depth due to observed overdispersion that varies across datasets and systems.

Key findings

• Poisson model is appropriate for sparse scRNA-seq data.
• Overdispersion is a common phenomenon in scRNA-seq data with sufficient sequencing depth.
• Negative binomial model is superior to Poisson for modeling overdispersed data.
• Data-driven parameter estimation is crucial due to dataset-specific overdispersion.

Methods

• Analysis of 59 scRNA-seq datasets covering diverse technologies, systems, and sequencing depths.
• Evaluation of different statistical error models, including Poisson and negative binomial distributions.
• Assessment of parameter settings for error models.
• Comparison of model performance based on observed overdispersion.

Limitations

• The study focuses on statistical error models and may not encompass all sources of technical variation.
• The recommendations are based on the analyzed datasets and may require further validation for novel technologies or systems.
• The performance of generalized linear models and likelihood-based approaches is discussed in the context of preprocessing and downstream analysis, but specific implementations are not detailed.

Applications

• Preprocessing of scRNA-seq data
• Downstream analysis of scRNA-seq data
• Development of robust computational tools for scRNA-seq analysis
• Interpretation of cellular heterogeneity and biological variation in scRNA-seq studies