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

Abstract

Macrocyclic compounds hold great promise as therapeutic agents. However, their structural optimization remains constrained by the limited availability of bioactive candidates, which in turn hampers the systematic exploration of structure-activity relationships. Here we introduce CycleGPT, a generative chemical language model designed specifically to address these challenges. CycleGPT is characterized by a progressive transfer learning paradigm that incrementally transfers knowledge from pre-trained chemical language models to specialized macrocycle generation, thereby overcoming the data shortage issue. Meanwhile, it adopts an innovative probabilistic sampling strategy that effectively improves the structural novelty of generated macrocycles while ensuring domain-specific adaptability. In a prospective drug design based on CycleGPT and a JAK2 activity prediction model, we successfully developed a new JAK2 drug candidate with a good selectivity profile (inhibiting 17 wild-type kinases) and promising potential for treating polycythemia in vivo, demonstrating the practicality of deep learning methods in macrocyclic drug design.