5 - Big data and deep learning: extracting and revising chemical knowledge from data

Abstract

The recent advancements in machine learning and deep learning (DL) methods are making it possible to create systems that automatically mine patterns and learn from data. Applications of those methods in chemistry, in particular quantitative structure–activity relationships (QSAR) and drug discovery, are already available. While DL can be applied as a conventional way of learning from chemical descriptors, the potentialities of the method are far more. In particular, the capabilities of DL to autonomously extract, through multiple transformations, the structural elements that are correlated with the property under investigation can help in discovering the link between a chemical and its biological/physical effects. After presenting the principal DL methods developed for chemical problems, the focus is on a study case in mutagenicity prediction that uses directly the chemical graph, either as SMILES, graphs, or images, and applies convolutional and recurrent networks. The knowledge extracted from the networks is analyzed and compared with the accepted structural alerts for mutagenicity. The next challenges and the future of DL for QSAR are finally discussed.