Predicting chemical structure using reinforcement learning with a stack-augmented conditional variational...

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Developing novel molecules with high binding affinity to target proteins is crucial for drug discovery.
• Existing generative models often struggle to produce valid, unique compounds with desired properties and high affinity.

Purpose of the Study:

• To propose a reinforcement learning model that maximizes predicted binding affinity between generated molecules and target proteins.
• To utilize the Stacked Conditional Variation AutoEncoder (Stack-CVAE) as an agent for molecule generation, optimizing for chemical properties and binding affinity.

Main Methods:

• Employed a Stacked Conditional Variation AutoEncoder (Stack-CVAE) integrated with reinforcement learning.
• Generated 1000 chemical formulas based on sorafenib's properties and its target kinases.
• Evaluated generated molecules for validity, uniqueness, chemical properties, and predicted binding affinity compared to other models.

Main Results:

• The Stack-CVAE model generated valid and unique chemical compounds with desired properties and superior predicted binding affinity.
• Analysis of top-scoring molecules revealed novelty, not present in existing databases.
• Generated molecules demonstrated significantly higher predicted binding affinity for Raf kinases and exhibited high druggability and synthesizability.

Conclusions:

• The proposed Stack-CVAE reinforcement learning model effectively generates novel, druggable, and synthesizable molecules with high predicted binding affinity.
• This approach shows promise for accelerating the discovery of targeted therapeutics, particularly for kinases like Raf.