Multiscale learning of gene network-driven phenotypic dynamics of single cells

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Gene regulatory networks (GRNs) dynamically control cell fate determination.
• Reconstructing these complex dynamics from temporal single-cell RNA-seq data is challenging.

Purpose of the Study:

• To develop a computational framework, GRNvelo, for reconstructing multiscale cell fate dynamics.
• To integrate GRNs with phenotypic dynamics from temporal single-cell RNA-seq data.
• To predict cell fate alterations in response to genetic perturbations.

Main Methods:

• GRNvelo employs a multiscale model coupling GRN-driven single-cell velocity with population dynamics.
• A two-phase optimization algorithm using physics-informed neural networks (PINNs) is utilized.
• TC-PINN infers GRN velocity and latent time; MP-PINN refines GRN velocity within population dynamics.

Main Results:

• GRNvelo demonstrated superior performance on synthetic and real datasets, including branching development.
• The framework accurately infers GRN-driven cell fate dynamics.
• GRNvelo successfully predicts altered cell fates under various genetic perturbations.

Conclusions:

• GRNvelo establishes a new computational paradigm for predicting and modulating cell fate outcomes.
• The framework provides biologically interpretable and mathematically rigorous insights into cell fate dynamics.
• GRNvelo advances the understanding of gene regulatory mechanisms in cell development and response to perturbations.