A computational Approach to Assess the Effect of Training Set on Generated Molecules Using Deep Learning-driven Scaffold Hopping

Discovery and development of a new drug is a risky, time-consuming and costly process. To speed up the process and increase the success rate, scaffold hopping, an important technique for small-molecule drug design, has received extensive attention. In recent years, some skeleton transition deep learning models have been developed. However, most of these scaffold hopping models use the same training set and do not consider the impact of the training set on the model-generated molecules. As a result, this project aims to develop a novel computational approach to assess the effects of training set using deep learningdriven scaffold hopping. We apply our approach to a dataset on four key drug-like properties, commonly shared by approved market drugs: hydrogen bond donor, hydrogen bond acceptor, lipophilicity, and molecular weight. In our results, we show the significant impact of training sets on generated molecules by intentionally selecting the training set.