Universal multi target drug design with generative model

Abstract

The deep learning-based de novo drug design is one of the promising approaches to help drug discovery. However, most of the compound generation studies are focused on the single target drug. Recently, some deep learning-based multi-target compound design studies have been suggested, but those studies fixed the target pair. In this research, we proposed a generalizable fine-tuning framework of the multi-target compound generative model using the LOGICS framework and GraphATT-DTA model. The random forest affinity prediction model in LOGICS was replaced with GraphATT-DTA. GraphATT-DTA can learn the pattern of protein and compound structures. So, the DTA model can predict the affinity of a protein whose affinity data is few with stable performance. Also, the LOGICS framework can explore diverse regions in the chemical space. These two models are appropriate for building the generalizable fine-tuning framework of the multi-target compound generation model. We checked the performance of the affinity prediction model and the fine-tuned compound generation model. Then, we calculated the similarity of the generated dual-target compounds to the approved multi-target drugs. We also applied a docking method to the generated compounds from the fine-tuned model and the pre-trained model. The results showed that the fine-tuned compound generation model based on LOGICS and GraphATT-DTA can generate diverse compounds with desirable structures against various multi-target protein pairs.