Abstract
Heterocycles are important scaffolds in medicinal chemistry that can be used to modulate the binding mode as well as the pharmacokinetic properties of drugs. The importance of heterocycles has been exemplified by the publication of numerous data sets containing heterocyclic rings and their properties. However, those data sets lack synthetic routes toward the published heterocycles. Consequently, novel and uncommon heterocycles are not easily synthetically accessible. While retrosynthetic prediction models could usually be used to assist synthetic chemists, their performance is poor for heterocycle formation reactions due to low data availability. In this work, we compare the use of four different transfer learning methods to overcome the low data availability problem and improve the performance of retrosynthesis prediction models for ring-breaking disconnections. The mixed fine-tuned model achieves top-1 accuracy of 36.5%, and, moreover, 62.1% of its predictions are chemically valid and ring-breaking. Furthermore, we demonstrate the applicability of the mixed fine-tuned model in drug discovery by recreating synthetic routes toward two drug-like targets published in 2023. Finally, we introduce a method for further fine-tuning the model as new reaction data becomes available.