GRELinker: A Graph-Based Generative Model for Molecular Linker Design with Reinforcement and Curriculum Learning

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Fragment-based drug discovery (FBDD) is a key strategy in modern drug design.
• Designing effective linkers is crucial for optimizing molecular properties in FBDD.
• Existing methods for linker generation have limitations in efficiency and control over molecular attributes.

Purpose of the Study:

• To introduce GRELinker, a novel generative model for fragment linking in FBDD.
• To demonstrate GRELinker's capability in generating molecules with desired properties.
• To showcase GRELinker's superiority over existing methods in benchmark tasks.

Main Methods:

• Utilized a gated-graph neural network architecture.
• Integrated reinforcement learning (RL) and curriculum learning for molecular generation.
• Employed docking scores as a scoring function within the RL framework for affinity optimization.

Main Results:

• GRELinker efficiently controls molecular properties such as log P, synthesizability, and predicted bioactivity.
• The model generated molecules with high 3D similarity and low 2D similarity to lead compounds.
• GRELinker outperformed the DRlinker method on benchmark tasks and successfully optimized molecular affinity in a real FBDD case.
• Curriculum learning facilitated the efficient generation of structurally complex linkers.

Conclusions:

• GRELinker is a powerful and versatile tool for linker design in FBDD.
• The model offers significant advantages in molecular optimization and drug discovery.
• GRELinker demonstrates the feasibility and benefits of integrating advanced machine learning techniques in drug design.