SHARP: Generating Synthesizable Molecules via Fragment-Based Hierarchical Action-Space Reinforcement Learning for...

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Artificial Intelligence in Drug Discovery

Background:

• Designing molecules with multiple properties like high binding affinity and synthesizability is a complex optimization challenge.
• Current deep learning models often fail to generate valid, synthetically feasible molecules due to limitations in optimization feedback.

Purpose of the Study:

• To introduce a novel molecular generator, SHARP, that addresses limitations in AI-driven drug discovery.
• To improve the generation of drug-like molecules with high affinity and synthetic accessibility.

Main Methods:

• Developed Synthesizable Hierarchical Action-space Reinforcement learning for Pareto optimization (SHARP), a fragment-based hierarchical reinforcement learning model.
• Integrated a Synthesizability Estimation Model (SEM) for action masking to ensure synthetic accessibility.
• Employed a composite reward function including docking scores, pharmacophore matching, and solvent accessibility for RL policy training.

Main Results:

• SHARP consistently outperformed existing methods in generating high-affinity molecules with good synthesizability across four lead optimization tasks.
• The model successfully addressed issues of invalid structures, local optima, and synthetic infeasibility.
• Demonstrated effectiveness in fragment growing, linker design, scaffold hopping, and side chain decoration.

Conclusions:

• Reinforcement learning with a chemically intuitive action space is a powerful approach for AI-driven drug discovery.
• SHARP offers a robust framework for rational molecular design in structure-based applications.
• The method enhances the generation of functionally relevant and experimentally tractable drug candidates.