Distilling dynamical knowledge from stochastic reaction networks.

Area of Science:

• Computational Chemistry
• Systems Biology
• Machine Learning

Background:

• Stochastic reaction networks model complex systems in biology, chemistry, physics, and ecology.
• Understanding their dynamics is challenging due to exponential state space growth.

Purpose of the Study:

• To develop a knowledge distillation method using reinforcement learning to compress dynamical information from stochastic reaction networks.
• To create a predictive neural network model for these systems.

Main Methods:

• Employed reinforcement learning principles for knowledge distillation.
• Trained a singular neural network to capture the dynamics of stochastic reaction networks.
• The network predicts state conditional joint probability distributions.

Main Results:

• The trained neural network accurately predicts system dynamics and probabilities.
• Enables direct estimation of state and trajectory probabilities without integrating over the state space.
• Demonstrated high accuracy in multimodal and high-dimensional systems.

Conclusions:

• The knowledge distillation approach effectively compresses dynamical knowledge from stochastic reaction networks.
• The neural network serves as a foundational model for parameter inference and trajectory generation.
• Highlights the potential of large-scale pretrained models for diverse stochastic dynamical systems.