DeepVelo: Single-cell transcriptomic deep velocity field learning with neural ordinary differential equations

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Single-cell sequencing offers insights into gene expression and RNA velocity.
• Modeling transcriptional dynamics is complex due to high-dimensional, sparse data and nonlinear relationships.

Purpose of the Study:

• To introduce DeepVelo, a neural network-based ordinary differential equation model.
• To model complex transcriptome dynamics and continuous-time gene expression changes in individual cells.

Main Methods:

• Applied DeepVelo to public single-cell sequencing datasets.
• Utilized neural networks to solve ordinary differential equations for gene expression.
• Performed perturbation analysis to identify driver genes.

Main Results:

• DeepVelo accurately represents the velocity field, outperforming existing methods.
• Modeled transcriptome dynamics across various time scales.
• Quantified cell state instability and identified developmental driver genes.
• Revealed potential chaotic properties in single-cell dynamical systems.

Conclusions:

• DeepVelo enables data-driven discovery of differential equations for single-cell transcriptome dynamics.
• The model advances the analysis of complex cellular processes using single-cell data.