On approximate stochastic control in genetic regulatory networks

Area of Science:

• Computational Biology
• Systems Biology
• Control Theory

Background:

• Probabilistic Boolean networks (PBNs) model genetic regulatory networks (GRNs).
• Optimal stochastic control using dynamic programming is computationally infeasible for large PBNs due to dimensionality issues.
• Exponential complexity arises in both control calculations and probability distribution estimation.

Purpose of the Study:

• To develop an approximate stochastic control method for PBNs that overcomes computational limitations.
• To mitigate the "curses of dimensionality" inherent in traditional dynamic programming approaches.
• To enable control of larger and more complex genetic regulatory networks.

Main Methods:

• An approximate stochastic control method based on reinforcement learning (RL) was proposed.
• The RL method is model-free, eliminating the need for explicit probability distribution estimation.
• A simulator was used to implement and test the RL-based control strategy.

Main Results:

• The proposed RL method achieves polynomial time complexity, significantly reducing computational demands.
• The method effectively bypasses the need for complex probability distribution estimation.
• Experimental results show performance comparable to optimal stochastic control, even for large networks.

Conclusions:

• Reinforcement learning provides a viable and computationally efficient alternative for controlling large-scale probabilistic Boolean networks.
• This model-free approach expands the applicability of optimal control techniques to complex genetic regulatory network models.
• The method successfully addresses the computational bottlenecks that limit traditional dynamic programming solutions.