Memory-assisted reinforcement learning for diverse molecular de novo design

Area of Science:

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

Background:

• Recurrent neural networks (RNNs) are effective for de novo molecular design, generating novel chemical structures.
• Reinforcement learning (RL) optimizes RNNs to generate molecules with desired properties using scoring functions.
• Current RL methods often produce low-diversity or duplicate molecular structures, limiting exploration of chemical space.

Purpose of the Study:

• To address the issue of low structural diversity in reinforcement learning-based molecular design.
• To introduce a novel memory-assisted reinforcement learning (RL) method for enhanced molecular generation.
• To evaluate the performance of memory-assisted RL in generating diverse and active molecules for specific targets.

Main Methods:

• Developed a memory-assisted reinforcement learning (RL) framework by incorporating a memory unit into the standard RL process.
• Applied the method to generate molecules with a specific AlogP value.
• Utilized the method to design ligands for dopamine D2 and serotonin 5-HT1A receptors, employing machine learning models for activity prediction.

Main Results:

• Memory-assisted RL successfully generated molecules with a desired AlogP value.
• For both dopamine D2 and 5-HT1A receptor targets, the method produced a higher number of predicted active compounds.
• Generated molecules exhibited significantly greater chemical diversity compared to established RL methods, improving chemical space coverage.

Conclusions:

• Memory-assisted RL is an effective strategy to overcome the limitations of low diversity in current RL-based molecular design.
• The proposed method enhances the exploration of chemical space, leading to a broader range of novel and potentially active molecular structures.
• This approach holds promise for improving the efficiency and scope of de novo drug design and discovery.