Button, A. et al. Nat. Machine Intell. 1, 307–315 (2019).

Discovery of drug-like compounds is often limited by the associated synthetic challenges. Button et al. described a machine-intelligence-based method termed DINGOS, which is trained on known reaction data from chemical patent literature as well as structural information of reactants and products. To generate compounds bearing desired physiochemical properties and structural similarity to an input structure, the algorithm chooses appropriate building blocks and optimizes pre-encoded rules for synthesis as well as modification of the intermediate compounds. The method was tested for de novo design of four approved drugs, and the suggested synthetic routes were found to be appropriate for the desired structures. Successful experimental validation of these compounds suggests that machine-learning models are capable of capturing chemical knowledge and suggesting feasible synthetic reactions.