Development of machine learning systems for drug discovery

Abstract

Deep learning have been widely utilized in computer vision, natural language processing, playing games, robotics, and other numerous fields. Molecular science is not an exception, prediction systems and generative models have paved a way of predicting molecular properties, generating novel compounds with desired properties, and planning chemical synthesis. For successful molecular applications, we consider a number of key learning principles that should be applied for neural molecular models. Firstly, we propose improved graph convolutional networks (GCNs) with attention and gate mechanisms. Our augmented-GCNs outperforms the baseline GCN in predicting several molecular properties. They also show better representation learning results, providing post-hoc interpretations on how our model make decisions in given tasks. Secondly, we address reliability issue in using deep neural networks, which is usually considered as black-box models. For this purpose, we incorporate Bayesian inference for property prediction tasks. Bayesian inference with approximate variational posterior enables us to quantify predictive uncertainty and demonstrates that it greatly helps us tackle over-confident predictions. As a result, we can verify more effective virtual screening can be possible by demonstrating on bio-activity predictions. Furthermore, since predictive uncertainty is related to the mutual information between a model parameter and a predictive outcome, we extend our Bayesian GCN for active learning on HIV-inhibitors discovery. Our demonstrations have great potentials in experimental designs with communication between computational and experimental chemists, being expected to reduce a heavy cost problem in current drug discovery industry. Lastly, we improve previous molecular generative models with adversarially regularized autoencoders. Also, our model can effectively perform conditional molecular generation with variational information disentanglement. Our algorithm has advantages in de novo molecular design, as shown in the experiment generating EGFR-active compounds. Beyond our previous works, we propose future directions, especially focus on generalization issues as key research topics that should be addressed in near future.