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

Area of Science:

• Single-cell genomics
• Computational biology
• Developmental biology

Background:

• RNA velocity analysis predicts future cell states from single-cell RNA sequencing (scRNA-seq) data.
• Existing methods lack robust ways to quantify prediction uncertainty and stability.
• This limits the reliable application of RNA velocity in understanding dynamic biological processes.

Purpose of the Study:

• To introduce a novel framework for evaluating the stability and reliability of RNA velocity predictions.
• To provide a metric for comparing the performance of different RNA velocity computational methods.
• To guide the selection of RNA velocity models that yield more biologically meaningful insights.

Main Methods:

• Developed a framework using negative binomial count splitting to create independent data replicates.
• Introduced 'replicate coherence' as a metric to quantify RNA velocity prediction stability.
• Proposed a signal-to-random coherence metric to aid in model selection.

Main Results:

• Tested five RNA velocity methods across mouse and human developmental datasets.
• Identified significant performance differences and inconsistencies among tested methods.
• Demonstrated that high replicate coherence selects models uncovering more biologically informative gene pathways.
• Showcased framework robustness even with missing intermediary cell states.

Conclusions:

• The proposed framework offers a rigorous approach to assess and compare RNA velocity methods.
• This tool enhances the reliability of scRNA-seq based cell state predictions.
• Selecting models based on replicate coherence improves the biological interpretability of RNA velocity analyses.