Steering semi-flexible molecular diffusion model for structure-based drug design with reinforcement learning.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence

Background:

• Traditional molecular generation methods often use static ligand modeling, which doesn't reflect the dynamic nature of real molecular interactions.
• Semi-flexible docking acknowledges ligand conformational changes but is computationally intensive.

Purpose of the Study:

• To develop a semi-flexible molecular generation framework using reinforcement learning (RL) and diffusion models.
• To improve the efficiency and accuracy of generating drug-like molecules for protein targets.

Main Methods:

• A reinforcement learning (RL)-steered diffusion framework was developed, treating the denoising process as a Markov decision process.
• Multiple molecular properties were used as conditions to guide the denoising policy towards drug-like molecules.
• Self-supervised rigid training was performed on both target-free and target-specific molecules.
• A fast sampling strategy was implemented, accelerating sampling by 20 times.

Main Results:

• The proposed method achieved a Vina score of -7.23 kcal/mol and an 11.53% success rate, outperforming state-of-the-art methods.
• Generated molecules successfully preserved canonical interaction patterns with unseen proteins.
• The framework discovered novel binding chemotypes for real-world protein targets.

Conclusions:

• The RL-steered diffusion framework enables efficient and accurate semi-flexible molecular generation.
• This approach advances structure-based drug design by capturing dynamic molecular interactions.
• The method holds promise for discovering new drug candidates with improved binding affinities and novel chemotypes.