Single-cell RNA sequencing has become one of the most powerful tools in modern biology. By measuring gene activity in individual cells, researchers can identify different cell types, study how cells change over time, and better understand development and disease. However, a single-cell RNA sequencing experiment captures only a snapshot of a cell at one moment in time. Scientists often want to know not only where a cell is now, but also where it is heading.

To address this challenge, researchers at the University of California, San Francisco have developed a computational approach called RNA velocity. RNA velocity estimates the future state of a cell by examining the balance between unspliced RNA, newly produced RNA molecules, and spliced RNA, which have already undergone processing. By comparing these RNA populations, researchers can infer the direction of cellular change and predict developmental trajectories.

Different RNA velocity methods vary in directionality predictions

a. Overview of the RNA velocity workflow, with the original Velocyto [6] model as an example. The rates of transcription, splicing and degradation are notated as , , and  respectively. All RNA velocity models use spliced (s) and unspliced (u) mRNA counts as model inputs and predict the directionality of a cell in transcriptomic space. b. Summary of the five RNA velocity models studied in this work and the methodologies implemented in each model. c. UMAP embeddings with RNA velocity predictions of mouse pancreas (n = 3696 cells) across the five different methods, highlighting different directionality predictions in the ‘Ductal’ cell type (bottom panel). d. A schematic describing examples of inconsistency and disagreement in the velocity predictions between different methods: (1) the consistency within the single cells in a neighborhood (the colored dots represent single cells), and (2) the agreement in directionality between methods (black arrows between boxes). The DNA icon is from BioRender.com.

RNA velocity has gained popularity because it can help researchers study processes such as embryonic development, cellular differentiation, aging, and disease progression. However, multiple RNA velocity algorithms are available, and they do not always produce the same results. Understanding the strengths and limitations of each method is therefore important for scientists who rely on these predictions.

The researchers created a benchmarking pipeline to compare five widely used RNA velocity methods using three developmental datasets, including mouse pancreas development and zebrafish embryo datasets. They assessed several key performance metrics, including whether nearby cells produced consistent trajectory predictions, how closely different methods agreed with one another, how each method identified genes that drive biological processes, and how well the methods performed when sequencing depth was reduced.

The analysis showed that all five methods were capable of recovering known developmental trajectories under certain conditions. However, performance varied depending on the complexity of the biological system and the amount of sequencing data available. In some cases, different methods generated substantially different predictions about where cells were heading.

One important finding was that the identification of driver genes often depended on the RNA velocity method used. Since these genes are frequently used to generate biological hypotheses, researchers should interpret such results carefully and consider validating findings using multiple approaches.

The study also found that sequencing depth can significantly influence RNA velocity performance. When sequencing data were reduced, some methods maintained accuracy better than others. This highlights the importance of considering data quality when selecting an RNA velocity workflow.

The authors emphasize that RNA velocity should be viewed as a hypothesis-generating tool rather than definitive proof of cellular behavior. They recommend that researchers compare results from multiple methods and validate important conclusions with additional experimental evidence whenever possible.

As single-cell RNA sequencing continues to expand across biology and medicine, benchmarking efforts like this provide valuable guidance for selecting appropriate computational tools. By helping researchers understand where different RNA velocity methods perform well and where they may struggle, these comparisons can improve the reliability and interpretation of future single-cell studies.