Benchmarking algorithms for RNA velocity inference.

Area of Science:

• Computational biology
• Genomics
• Single-cell analysis

Background:

• RNA velocity (scRNA-seq) infers cell state transitions.
• Numerous methods exist with varying assumptions and performance.
• Lack of consensus on evaluation and reliability hinders adoption.

Purpose of the Study:

• To systematically compare RNA velocity algorithms.
• To evaluate methods across simulated and real scRNA-seq data, including spatial and multi-omics.
• To provide guidance for selecting appropriate RNA velocity tools.

Main Methods:

• Benchmarked 29 RNA velocity inference algorithms.
• Utilized 114 simulated and 62 real scRNA-seq datasets.
• Extended evaluation to spatial and multi-omics data.
• Assessed performance using accuracy, scalability, stability, and usability.

Main Results:

• Performance rankings varied significantly across metrics and datasets.
• No single RNA velocity method demonstrated uniform optimality.
• Feasibility and robustness are critical constraints for practical deployment.
• Guidance for tool selection based on data modality and computational resources was developed.

Conclusions:

• Method selection for RNA velocity analysis requires careful consideration of specific project needs.
• Scalability, gene selection sensitivity, and lack of multimodal/spatially explicit models are key limitations.
• Further development is needed for robust and scalable RNA velocity tools, especially for large datasets and emerging modalities.