DRlinker: Deep Reinforcement Learning for Optimization in Fragment Linking Design.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Fragment-based drug discovery (FBDD) is a key strategy in drug design.
• Current deep generative models struggle to generate linkers with specific attributes for FBDD.

Purpose of the Study:

• To present DRlinker, a novel framework for controlled fragment linking in drug design.
• To enable the generation of drug compounds with specified attributes using reinforcement learning.

Main Methods:

• Developed a novel framework, DRlinker, utilizing reinforcement learning.
• Applied DRlinker to control linker length, log P, and optimize bioactivity.
• Conducted a quasi-scaffold-hopping study to assess 3D and 2D similarity.

Main Results:

• DRlinker successfully generated 91.0% and 93.9% of compounds with desired linker length and log P.
• Achieved a 7.5 pChEMBL value improvement in bioactivity optimization.
• Generated nearly 30% of molecules with high 3D similarity but low 2D similarity to lead inhibitors.

Conclusions:

• DRlinker effectively controls fragment linking for drug design.
• The framework demonstrates applicability in generating novel drug candidates with desired properties.
• DRlinker offers significant benefits for practical fragment-based drug discovery.