Deep Reinforcement Learning for Intervention of Partially Observable Regulatory Networks.

Area of Science:

• Computational Biology
• Systems Biology
• Bioinformatics

Background:

• Gene regulatory networks (GRNs) control cellular functions.
• Real-world GRN analysis is challenged by partial observability and noisy data.
• Existing intervention methods often assume complete system state information.

Purpose of the Study:

• To develop a deep reinforcement learning framework for optimal intervention policies in partially observable GRNs.
• To address limitations of existing methods that assume full observability.
• To manage uncertainties in gene expression data and gene activity stochasticity.

Main Methods:

• Extension of Boolean network models to incorporate partial observability.
• Formulation of optimal intervention policies in the belief space, utilizing belief states to represent state posterior distributions.
• Application of deep Q-network (DQN) for scalable approximation of optimal policies.
• Analytical demonstration of convergence to optimal dynamic programming solutions under reduced uncertainty.

Main Results:

• The proposed deep reinforcement learning framework effectively handles partial observability in GRNs.
• The belief state successfully captures data uncertainty and gene activity stochasticity.
• Numerical experiments on a melanoma GRN show improved performance in maintaining desired states and reducing cancer-related gene activation compared to existing methods.

Conclusions:

• The developed framework provides a robust approach for designing intervention strategies in complex, partially observable biological systems.
• Deep reinforcement learning offers a scalable solution for optimizing interventions in GRNs.
• The method holds promise for applications in precision medicine and disease control by targeting specific gene regulatory pathways.