On the synthesis of machine learning and automated reasoning for an artificial synthetic organic chemist

Abstract

This perspective outlines current capabilities and limitations of state-of-the-art artificial intelligence methods as applied to automating the planning of synthetic routes in organic chemistry. Synthetic organic chemistry is viewed from the perspective of two prominent approaches: deep neural networks and SAT-solver based automated reasoning. After introducing these concepts to non-computer scientists, the expected performance of these approaches is estimated by surveying comparable problems in artificial intelligence. A truly artificial synthetic organic chemist that automatically constructs viable synthetic routes is clearly a challenging artificial intelligence problem and not directly amenable to existing approaches but chemistry could encourage new combinations of machine learning methods with automated reasoning to realize this goal. The importance of objective and open competitions with standardized problems and evaluations is also detailed as critical to realizing tangible computer programs that automate the planning of plausible synthetic routes.