A data-driven method to learn a jump diffusion process from aggregate biological gene expression data.

Area of Science:

• Computational Biology
• Systems Biology
• Genetics

Background:

• Dynamic models of gene expression are crucial for understanding cellular processes.
• Current models often struggle to capture the inherent stochasticity and complexity of gene expression.
• A need exists for methods that can model the time evolution of gene expression from aggregate data.

Purpose of the Study:

• To develop a novel dynamic model for gene expression using a jump diffusion process.
• To enable prediction of population distributions and individual cell trajectories.
• To provide a new computational approach for determining RNA velocity.

Main Methods:

• Learning a jump diffusion process directly from aggregate gene expression data.
• Utilizing Wasserstein distance to minimize the difference between empirical and predicted distributions.
• Treating gene expression data at a time point as an empirical marginal distribution.

Main Results:

• The learned jump diffusion process accurately predicts population distributions and long-time cell trajectories.
• The model offers a novel method for computing RNA velocity based on instantaneous rates of state change.
• Demonstrated superior performance in recovering nonlinear dynamics compared to existing parametric and Brownian motion diffusion models on synthetic and real datasets.

Conclusions:

• The jump diffusion process provides a powerful stochastic dynamic perspective for studying biological systems.
• The method effectively models cell development, heterogeneity, and RNA velocity.
• The approach is robust to data perturbations due to its reliance on population expectations.