Exploring the macrocyclic chemical space for heuristic drug design with deep learning models

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Macrocyclic compounds show therapeutic potential but face challenges in structural optimization due to limited bioactive candidates.
• Exploring structure-activity relationships for macrocycles is hindered by data scarcity.

Purpose of the Study:

• Introduce CycleGPT, a generative chemical language model to address challenges in macrocycle drug design.
• Improve the generation of novel and adaptable macrocyclic structures.
• Demonstrate the utility of CycleGPT in identifying potential drug candidates.

Main Methods:

• Developed CycleGPT using a progressive transfer learning paradigm for specialized macrocycle generation.
• Implemented a probabilistic sampling strategy to enhance structural novelty and domain adaptability.
• Integrated CycleGPT with a JAK2 activity prediction model for prospective drug design.

Main Results:

• CycleGPT effectively overcomes data shortage issues in macrocycle generation.
• Generated macrocycles exhibit improved structural novelty and domain-specific adaptability.
• Successfully identified a novel JAK2 drug candidate with high selectivity (inhibiting 17 wild-type kinases).

Conclusions:

• CycleGPT demonstrates practical utility in deep learning-based macrocyclic drug design.
• The identified JAK2 inhibitor shows promise for treating polycythemia in vivo.
• This generative model facilitates systematic exploration of macrocycle structure-activity relationships.