De Novo Drug Property Prediction using Graph Convolutional Neural Networks

Abstract

Drug property prediction poses a complex task in the healthcare domain, because the drug-like molecules greatly vary in chemical structures. Inhibitory concentration (IC50) prediction is essential to reduce the cost and labor of a drug's preclinical and clinical trials, because a lot of drug evaluations (cellular, animal, and clinical) can be spared due to the predicted IC50. In the era of Artificial Intelligence, Drug Discovery processes can benefit from deep learning, which has been widely used in computational chemistry and bioinformatics with state-of-the-art performance. In this paper, we propose a novel (in other words, de novo) graph convolutional network approach, cross-validated by traditional methods like Lipinski's rule of five and PaDEL-described one-hot encoding. The experiment has been carried out on the ChEMBL bioactivity dataset of Acetylcholinesterase protein, achieving an R2 score of 0.52.