Multi-objective molecular generation via clustered Pareto-based reinforcement learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• De novo molecular design aims to generate novel chemical structures with desired properties.
• Existing methods often struggle to balance multiple drug targets in polypharmacology, limiting molecule validity and desirability.
• Drug resistance necessitates molecules with multiple pharmacological activities.

Purpose of the Study:

• To propose a novel approach, clustered Pareto-based reinforcement learning (CPRL), for de novo molecular design.
• To address the limitations of existing methods in balancing multiple molecular properties and targets.
• To improve the validity and desirability of generated molecules for drug discovery.

Main Methods:

• CPRL integrates a pre-trained model for molecular knowledge acquisition via supervised learning.
• A clustered Pareto optimization algorithm, using aggregation-based molecular clustering, identifies optimal solutions across objectives.
• A reinforcement learning agent balances multiple properties, guided by Pareto frontier ranking, with a fixed-parameter exploration model for diversity.

Main Results:

• CPRL effectively balances multiple molecular properties, crucial for polypharmacology.
• The method achieved high desirability (0.9551) and validity (0.9923) in generated molecules.
• Experimental results validate CPRL's capability in exploring chemical space and generating effective drug candidates.

Conclusions:

• CPRL offers a significant advancement in de novo molecular design for multi-target drug discovery.
• The approach enhances the balance of multiple pharmacological properties, leading to more effective and desirable drug candidates.
• CPRL demonstrates superior performance in generating molecules with improved validity and desirability compared to existing methods.