Quantifying stability via count splitting to guide model selection in RNA velocity analyses.

Area of Science:

• Computational biology
• Single-cell genomics
• Systems biology

Background:

• RNA velocity (RV) predicts cell fate from single-cell RNA sequencing (scRNA-seq) data.
• Quantifying uncertainty and stability in RV predictions remains a challenge.
• Existing methods lack standardized evaluation metrics for RV tool performance.

Purpose of the Study:

• To develop a generalizable framework for evaluating the stability and uncertainty of RNA velocity predictions.
• To introduce novel metrics for assessing RV method performance and guiding model selection.
• To compare the robustness and biological relevance of different RNA velocity tools.

Main Methods:

• Developed a novel framework using negative binomial count splitting for data replication.
• Introduced 'replicate coherence' as a metric for RNA velocity stability.
• Implemented a signal-to-random coherence metric for model selection.

Main Results:

• Identified significant performance differences and inconsistencies among five tested RNA velocity methods.
• Demonstrated the framework's robustness across diverse datasets (mouse erythroid, pancreatic, human brain).
• Showed that selecting models with high replicate coherence enhances the discovery of biologically informative gene pathways.

Conclusions:

• The proposed framework provides a rigorous approach to assess and compare RNA velocity methods.
• This tool enhances the reliability of cell state predictions and biological interpretation from scRNA-seq data.
• The findings facilitate more accurate understanding of dynamic biological processes using RNA velocity analysis.