Research on predicting 2D-HP protein folding using reinforcement learning with full state space

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Folding

Background:

• Protein structure prediction is a critical bioinformatics challenge.
• The hydrophobic-polarity model for 2D protein structure prediction is an NP-hard problem.
• Existing optimization methods often fail to find optimal protein conformations.

Purpose of the Study:

• To develop a novel reinforcement learning (RL) method for optimizing the hydrophobic-polarity model.
• To enhance the accuracy and robustness of 2D protein structure prediction.

Main Methods:

• Structured the full state space for reinforcement learning.
• Designed an energy-based reward function.
• Implemented a rigid overlap detection rule.

Main Results:

• Reinforcement learning with full states achieved convergence to lowest energy conformations for all tested sequences.
• Full state RL outperformed partial state RL (50% convergence) and greedy algorithms.
• Full state RL achieved an average of 5 lowest energy hits, significantly higher than alternatives.

Conclusions:

• Reinforcement learning with full states is a powerful and competitive method for 2D hydrophobic-polarity protein structure prediction.
• This approach offers significant advantages over traditional greedy algorithms and partial state RL methods.