Biologically informed NeuralODEs for genome-wide regulatory dynamics.

Area of Science:

• Computational Biology
• Systems Biology
• Bioinformatics

Background:

• Gene regulatory network (GRN) models using ordinary differential equations (ODEs) explain gene expression dynamics but face challenges in accurately encoding GRNs and nonlinear relationships.
• Existing ODE estimation methods often lack biological insight or impose restrictive assumptions, limiting scalability and explainability.

Purpose of the Study:

• To develop a novel modeling framework, PHOENIX, for learning gene regulatory ODEs that are biologically interpretable and scalable.
• To overcome limitations of current methods by incorporating prior biological knowledge and constraints.

Main Methods:

• PHOENIX utilizes neural ordinary differential equations (NeuralODEs) combined with Hill-Langmuir kinetics.
• The framework incorporates user-defined prior knowledge and systems biology functional forms as soft constraints.
• It promotes sparse and biologically interpretable representations of GRN ODEs.

Main Results:

• PHOENIX demonstrated accuracy in silico, outperforming several existing tools.
• The framework successfully modeled oscillating gene expression profiles from synchronized yeast cells.
• PHOENIX showed scalability by modeling genome-scale GRNs for breast cancer and B cells.

Conclusions:

• PHOENIX effectively integrates biological 'first principles' as soft constraints for predicting gene expression patterns.
• The framework provides a biologically explainable approach to GRN modeling.
• PHOENIX offers a flexible, accurate, and scalable solution for complex GRN analysis.