Geometric Deep Learning for Structure-Based Ligand Design.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Expanding ligands with chemical fragments is crucial for improving drug properties in lead optimization and fragment-based drug design.
• Current methods for ligand expansion often lack efficiency and drug-like properties in generated molecules.

Purpose of the Study:

• To develop a comprehensive machine learning framework (FRAME) for intelligent ligand expansion.
• To improve predicted affinity and selectivity of drug candidates.
• To generate molecules with enhanced drug-like properties.

Main Methods:

• Utilized machine learning and 3D protein-ligand structures.
• Developed an iterative approach to determine fragment addition sites, select fragments, and predict their geometry.
• Benchmarked FRAME against existing docking-based methods.

Main Results:

• FRAME consistently improved predicted affinity and selectivity compared to initial ligands.
• Generated molecules exhibited superior drug-like chemical properties.
• The method accurately describes molecular interactions without prior information.

Conclusions:

• FRAME offers a powerful framework for molecular hypothesis generation in drug design.
• The framework can be integrated into workflows for lead optimization, fragment-based drug discovery, and de novo drug design.
• FRAME advances the field by providing an accurate and efficient method for ligand expansion.