FragGen: towards 3D geometry reliable fragment-based molecular generation

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• 3D structure-based molecular generation using AI shows promise in drug discovery.
• Atom-wise generative models often yield molecules with poor drug-like properties, such as low synthesizability.
• Fragment-wise generation offers an alternative but struggles with co-designing chemical and geometric structures.

Purpose of the Study:

• To develop a novel fragment-wise molecular generation method that addresses limitations in geometric reliability and molecular properties.
• To introduce a new protocol for handling complex 3D geometry in generative models.
• To establish a validated 3D fragment-based drug design algorithm.

Main Methods:

• Introduced the Deep Geometry Handling protocol to decompose 3D geometry into manageable variables.
• Developed FragGen, a hybrid strategy employing a six-category taxonomy for fragment-wise generation.
• Focused on co-designing plausible chemical structures and reliable 3D geometries.

Main Results:

• FragGen significantly enhances both the geometric quality and synthesizability of generated molecules.
• The method overcomes major limitations of previous atom-wise and fragment-wise generative models.
• Achieved successful application in designing type II kinase inhibitors with nanomolar potency.

Conclusions:

• FragGen is the first validated 3D fragment-based drug design algorithm, demonstrating improved molecular properties.
• The Deep Geometry Handling protocol and FragGen offer a robust framework for geometry-centric AI tasks.
• This work provides a practical example of customizing generative AI for effective drug design.