Integrating Reaction Schemes, Reagent Databases, and Virtual Libraries into Fragment-Based Design by Reinforcement...

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• AI-driven generative models are crucial for lead optimization in drug design.
• Key success factors include reagent availability, novelty, and multi-property optimization.
• Directed fragment-replacement strategies mimic medicinal chemistry approaches.

Purpose of the Study:

• To present novel variations of fragment-based reinforcement learning for enhanced drug design.
• To improve synthesizability and balance novelty with diversity in generated molecules.
• To explore large chemical spaces efficiently with synthesizable drug candidates.

Main Methods:

• Utilized an actor-critic model for fragment-based reinforcement learning.
• Introduced novel features: freezing fragments and using reagents as fragment sources.
• Incorporated reaction schemes for molecular splitting to improve synthesizability.
• Tuned network output probabilities to balance novelty and diversity.
• Combined fragment-based optimization with virtual library encodings.

Main Results:

• Achieved design of high-quality molecules with favorable profiles.
• A validation study across 15 pharmaceutically relevant targets yielded novel structures.
• Obtained molecules were identical or related to independent validation sets for most targets.
• Demonstrated significant improvements in the value of fragment-based reinforcement learning.

Conclusions:

• The presented modifications substantially enhance fragment-based reinforcement learning for drug design.
• The approach effectively generates novel, synthesizable, and high-quality drug candidates.
• This AI-driven methodology offers a powerful tool for exploring chemical space and optimizing molecular properties.