A Reinforcement Learning-Guided Genetic Algorithm Integrating Medicinal Chemistry-Inspired Molecular Transformations.

Area of Science:

• Computational chemistry and drug discovery.
• Artificial intelligence in medicinal chemistry.
• Molecular modeling and simulation.

Background:

• Drug discovery faces challenges in optimizing target activity, synthetic accessibility, and drug-likeness.
• Existing de novo generative models often produce chemically invalid or synthetically intractable molecules.
• Simultaneous optimization of multiple objectives in molecular design remains a significant hurdle.

Purpose of the Study:

• To introduce ALCHIMIA, an interpretable hybrid framework for de novo molecular design.
• To address limitations of current generative models in producing valid, synthesizable, and drug-like molecules.
• To enable efficient optimization of multiple objectives in computational drug discovery.

Main Methods:

• ALCHIMIA combines reinforcement learning (RL) and a genetic algorithm (GA) using 33 medicinal chemistry-inspired molecular transformations.
• The RL component trains a policy network to prioritize transformations improving synthetic accessibility (SA) and quantitative estimate of drug-likeness (QED).
• The GA component uses the learned policy for population-based optimization guided by molecular docking.

Main Results:

• ALCHIMIA generated chemically valid molecules with SA and QED scores comparable to or exceeding baseline methods.
• The framework was successfully applied to human Cannabinoid Receptor 2 (CB2R) and Sigma nonopioid intracellular Receptor 1 (S1R).
• Scenarios included unconstrained hit identification, scaffold-constrained lead optimization, and dual modulator design.

Conclusions:

• ALCHIMIA provides a practical, interpretable, and scalable framework for molecular de novo design.
• The hybrid RL-GA approach effectively integrates medicinal chemistry actions into molecular generation.
• This framework advances the design of high-affinity ligands with desirable drug-like properties.