Comparison between a conventional tool and deep learning models for RNA velocity analysis of scRNA-Seq data

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-Seq) provides gene expression data at single-cell resolution.
• RNA velocity analysis infers transcriptional dynamics using spliced and unspliced mRNA ratios from scRNA-Seq.
• Classical methods like scVelo use gene-specific kinetic modeling.

Purpose of the Study:

• To systematically evaluate deep learning RNA velocity tools against classical dynamical models.
• To assess the advantages of variational autoencoder (VAE)-based methods in RNA velocity analysis.
• To compare the performance of VAE-based tools (DeepVelo, VeloVI, LatentVelo, SymVelo, scTour) with scVelo.

Main Methods:

• Utilized public scRNA-Seq datasets (GSE149689, GSE203233) processed with a standard pipeline.
• Employed cosine similarity of velocity vectors to evaluate directional concordance.
• Applied mean squared error analysis to assess trajectory continuity for deep learning models.

Main Results:

• VAE-based deep learning methods generated more coherent, consistent, and directionally accurate velocity fields compared to the classical model.
• Deep learning models demonstrated superior performance in predicting cell-state trajectories.
• Findings highlight the potential of VAE frameworks for advancing RNA velocity analysis.

Conclusions:

• Deep learning RNA velocity tools, especially VAE-based ones, offer significant improvements in inferring cell-state trajectories.
• These methods provide richer and more biologically plausible dynamics but require robust splicing quantification and higher computational resources.
• Careful data preprocessing is crucial for optimal performance of VAE-based RNA velocity analysis.