Biologically informed NeuralODEs for genome-wide regulatory dynamics.

Area of Science:

• Computational biology
• Systems biology
• Bioinformatics

Background:

• Ordinary differential equations (ODEs) model gene expression dynamics for insights into cellular processes and disease.
• Existing ODE estimation methods lack biological interpretability and scalability.
• Learning causal gene-regulatory networks (GRNs) requires accurate modeling of nonlinear gene expression dynamics.

Approach:

• Developed PHOENIX, a framework combining neural ordinary differential equations (NeuralODEs) with Hill-Langmuir kinetics.
• Incorporates prior domain knowledge and biological constraints for interpretable ODE representations.
• Benchmarked PHOENIX against existing ODE estimation tools using in silico experiments.

Key Points:

• PHOENIX promotes sparse, biologically interpretable ODE models of gene expression.
• Demonstrated flexibility in modeling yeast cell oscillating expression data.
• Assessed scalability by modeling genome-scale breast cancer expression data.

Conclusions:

• PHOENIX integrates prior biological knowledge to encode GRN properties effectively.
• Enables biologically explainable predictions of gene expression patterns.
• Offers a scalable and interpretable approach for learning gene regulatory dynamics.