MoleGuLAR: Molecule Generation Using Reinforcement Learning with Alternating Rewards.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Developing novel drug inhibitors is critical, especially given global health challenges like COVID-19.
• Conventional methods like virtual screening are often time-consuming and data-intensive.
• De novo molecular generation offers a promising alternative for identifying new drug candidates.

Purpose of the Study:

• To develop a computational strategy for the de novo generation of molecules with high binding affinities.
• To optimize molecular generation for desirable drug-like properties using reinforcement learning.
• To improve the efficiency and success rate of identifying potential drug candidates.

Main Methods:

• Utilized a deep generative model based on a stack-augmented recurrent neural network.
• Employed reinforcement learning to optimize molecule generation for specific properties.
• Implemented a novel multi-objective optimization strategy by periodically changing the reward-calculating property.

Main Results:

• The reinforcement learning approach successfully optimized the generative model for drug-like properties.
• The novel periodic reward strategy enhanced the generation of molecules with desired characteristics.
• Significantly higher numbers of molecules with desirable properties were generated compared to conventional weighted sum approaches.

Conclusions:

• The proposed computational strategy offers an effective approach for de novo drug design.
• Reinforcement learning, coupled with a novel optimization technique, significantly improves the generation of drug-like molecules.
• This method presents a powerful tool for accelerating the discovery of new therapeutic agents.