A reinforcement learning approach for protein-ligand binding pose prediction

Area of Science:

• Computational Chemistry
• Structural Biology
• Artificial Intelligence in Drug Discovery

Background:

• Protein ligand docking is crucial for predicting protein functions and screening drug candidates.
• Current challenges include limited understanding of protein-ligand energetics and vast conformational search spaces.
• Existing methods struggle with accuracy and efficiency in identifying binding sites.

Purpose of the Study:

• To develop a novel reinforcement learning (RL) approach for enhanced protein ligand docking.
• To address the limitations of current docking methods using the asynchronous advantage actor-critic (A3C) model.
• To improve the accuracy of binding site prediction in computational drug discovery.

Main Methods:

• Developed a two-model framework utilizing the asynchronous advantage actor-critic (A3C) reinforcement learning algorithm.
• The actor model selects actions based on the current location during the search process.
• The critic model evaluates actions and predicts the distance to the true binding site.

Main Results:

• The A3C model demonstrated substantial improvements in binding site prediction compared to a naive model for both single- and multi-atom ligands.
• For single-atom ligands (e.g., copper ion), a median root-mean-square deviation (RMSD) of 2.39 Å was achieved.
• For multi-atom ligands (e.g., sulfate ion), a median RMSD of 3.82 Å was achieved, indicating high accuracy.

Conclusions:

• The developed RL framework offers a significant advancement in protein ligand docking accuracy.
• Ligand-specific models are valuable for solvent mapping studies and can be extended to diverse ligands.
• This approach provides a scalable and effective tool for computational drug screening and protein function prediction.