Molecular graph Self attention and graph convolution for drug discovery

Abstract

Drug development is an important, but complicated and expensive process. By utilizing the power of deep learning, we aim to improve the current process of drug development. We model molecules as undirected graphs and use graph convolutions and self-attention to predict molecular properties. With a series of ablation studies, we demonstrate the added value of several key components in our network. We analyze two standard datasets: BBBP, which includes classication data on whether molecules pass the blood-brain barrier, and ClinTox, which includes toxicity information. Using our architecture, we are able to match state of the art performance on the BBBP prediction task.