Jaxkineticmodel: Neural ordinary differential equations inspired parameterization of kinetic models

Area of Science:

• Systems Biology
• Computational Biology
• Biophysics

Background:

• Metabolic kinetic models, described by Ordinary Differential Equations (ODEs), are crucial for understanding biological systems.
• Parameterizing large-scale ODE models presents significant computational challenges.
• Neural Ordinary Differential Equations (NODE) offer promising approaches for time-series data modeling.

Purpose of the Study:

• To introduce jaxkineticmodel, a novel simulation and training framework for Systems Biology Markup Language (SBML) models.
• To leverage JAX/Diffrax for accelerated kinetic model parameterization and enable hybrid modeling.
• To provide a robust platform for efficient fitting of large metabolic kinetic models.

Main Methods:

• Implementation of a simulation and training framework using JAX/Diffrax.
• Utilizing JAX's automatic differentiation and just-in-time compilation for performance enhancement.
• Developing capabilities for hybridizing kinetic models with neural networks.

Main Results:

• Demonstrated robust training of kinetic models across diverse SBML models with varying parameter initialization.
• Successfully applied the framework to a complex glycolysis model.
• Showcased hybridization of kinetic models with neural networks for unknown reaction mechanisms.

Conclusions:

• The jaxkineticmodel framework efficiently fits large metabolic kinetic models.
• It offers a powerful platform for advancing systems biology research and modeling.
• The framework facilitates the integration of mechanistic and data-driven modeling approaches.