DrugEx v3: scaffold-constrained drug design with graph transformer-based reinforcement learning.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• Rational drug design relies on modifying chemical scaffolds to discover novel molecules.
• Deep learning methods, including reinforcement learning, have advanced de novo drug design.
• Previous DrugEx versions lacked scaffold input flexibility and user-defined prior information.

Purpose of the Study:

• To enhance the DrugEx method for scaffold-based drug molecule generation.
• To incorporate user-provided scaffolds and molecular fragments into the design process.
• To improve the general applicability of deep learning models in drug discovery.

Main Methods:

• Utilized a Transformer deep learning model with an encoder-decoder architecture.
• Developed a novel positional encoding for molecular graph representations using adjacency matrices.
• Implemented a graph Transformer with fragment-based growing and connecting procedures.
• Trained the molecule generator using a reinforcement learning framework.

Main Results:

• The updated DrugEx successfully designed drug molecules based on user-defined scaffolds.
• Generated molecules demonstrated 100% validity.
• A significant portion of generated molecules exhibited high predicted affinity for the adenosine A2A receptor.
• Performance was validated against traditional SMILES-based methods.

Conclusions:

• The scaffold-based graph Transformer approach enhances de novo drug design capabilities.
• The method offers flexibility by accepting user-defined scaffolds and fragments.
• This advancement holds promise for accelerating the discovery of targeted therapeutic agents.