Automated Design vf Drug Candidate Molecules with Deep Graph Learning

Abstract

The discovery of new drug candidate molecules is an important step in the process of drug development. Deep generative learning, a frequently used approach in the field of artificial intelligence in recent years, has emerged as a promising method for generating realistic synthetic data within a defined theme. Additionally, the utility of these models in the drug development process depends on their ability to generate molecules specific to the biological target. In this study, a new generative system called "DrugGEN" has been developed specifically for the de novo design of drug candidate molecules that will interact with selected target proteins. The system represents compounds and protein structures as graphs and processes them using two sequentially connected generative adversarial networks (GANs) incorporating graph transformers. The training dataset of the system was created from a large collection of drug-like compound records and target-specific bioactive molecules obtained from the ChEMBL database. The developed model was trained with the aim of designing new molecules targeting the AKT1 protein, which plays a critical role in various cancer types. The performance of the DrugGEN model was evaluated comparatively with other methods in the literature using fundamental criteria. In addition, explanatory data analysis was performed on the generated results. The results demonstrated the novelty of molecules designed de novo by DrugGEN. Furthermore, it was shown that the outputs were comparable to the known ligands of the AKT1 protein both in terms of physicochemical properties and structure. Consequently, in this study, an artificial intelligence model was developed using deep learning algorithms and extensive chemical and biological data to automatically design completely novel molecules with the ability to target selected proteins.