Comprehensive benchmarking of RNA velocity methods across single-cell datasets

BACKGROUND: RNA velocity provides a powerful framework for inferring cellular dynamics from single-cell RNA sequencing data. The rapid proliferation of computational methods within this field has prompted a need for systematic evaluation. However, existing comparisons often suffer from limited scope or incomplete task design, leaving users without clear guidance. Consequently, there is a lack of a comprehensive and standardized benchmark that evaluates methods across diverse biological and technical scenarios using appropriate, context-specific metrics. RESULTS: In this study, we present a comprehensive benchmark of 19 computational RNA velocity tools covering 30 distinct methods. We systematically evaluate 25 RNA-only methods across eight evaluation tasks, designating directional consistency, temporal precision, negative control robustness, and sequencing depth stability as core tasks, while assessing five multimodal-enhanced methods specifically on the multimodal integration task. These assessments utilize 34 datasets spanning 26 real-world and eight simulated scenarios. Our results reveal a clear trade-off between directional consistency and negative control robustness, distinct group-wise behaviors across temporal modeling strategies, and variability driven by sequencing depth and quantification choices. This study also identifies several methodological gaps, including the need for improved modeling of gene dependence, more accurate temporal inference strategies, and better-designed multimodal architectures. CONCLUSIONS: This benchmark establishes a unified framework for evaluating RNA velocity methods. Crucially, we provide task-aware guidance to facilitate method selection based on specific biological contexts and technical constraints, rather than relying on a single overall ranking.