DeepCME

DeepCME approximates solution statistics of high-dimensional chemical master equations (CMEs) for stochastic models of biomolecular reaction networks.

Key Features:

• Deep Learning Integration: DeepCME employs deep neural networks (DNNs) to approximate CME solution statistics and estimate expectations for user-defined functions of the state-vector.
• Reformulation Using Kolmogorov's Backward Equation: The stochastic dynamics are reformulated using Kolmogorov's backward equation to enable more efficient computation than forward-equation (CME) approaches.
• Reinforcement Learning-Based Algorithm: A reinforcement learning framework is used to train a policy function from a moderate number of stochastic simulations, reducing simulation demands.
• Sensitivity Analysis: DeepCME computes sensitivities of CME solutions with respect to all reaction network parameters, such as rate constants, for parameter estimation and model refinement.

Scientific Applications:

• Systems biology: Analysis of stochastic fluctuations in biomolecular reaction networks relevant to systems biology.
• Synthetic biology: Estimation of expectations and sensitivities for design and analysis of synthetic biological circuits under stochastic dynamics.
• Parameter sensitivity analysis: Conducting parameter sensitivity analyses to inform parameter estimation and improve model accuracy.

Methodology:

Reformulate stochastic dynamics using Kolmogorov's backward equation; train deep neural networks (DNNs) using reinforcement learning principles to learn a policy function from a moderate number of stochastic simulations; approximate expectations for user-defined functions of the state-vector and compute sensitivities with respect to reaction network parameters.