Deep learning for novel drug development

Abstract

During the 2010s, numerous developments in the fields of artificial intelligence (AI) and statistics led to the current boom around deep learning or inference, prediction, and decision support with deep neural networks (DNNs). Such developments include the availability of graphics processing unit (GPU) kernels that facilitated considerably faster NN training; the access to massive annotated datasets in several domains, which prevented overfitting; new mathematical advances and architectural designs that mitigated convergence issues, for example, mitigating the vanishing gradient problem; and, finally, the provision of automatic differentiation libraries, such as TensorFlow, Torch, or Caffe. As other chapters in this volume expose, introducing a new drug to the market follows a costly process that typically spans over several years with a high attrition rate. Thus, accelerating this process with innovative technologies would be very beneficial. While traditional computational and statistical techniques over the last decades greatly speed up drug development, the application of recent AI tools and methods, such as deep learning, has entailed a major disruption in the drug discovery development. This chapter overviews outstanding recent advances in DNNs and their application in drug development.