Significance Machine learning has achieved great success in retrosynthesis planning. We introduced chemical information, including NMR chemical shifts, bond energies, catalysts, and solvents into the descriptor of molecules and reactions… Click to show full abstract
Significance Machine learning has achieved great success in retrosynthesis planning. We introduced chemical information, including NMR chemical shifts, bond energies, catalysts, and solvents into the descriptor of molecules and reactions and into molecular graphs to represent molecules and reactions, and constructed a retrosynthesis planning model. It was developed using five molecular graph–based neural networks and Monte Carlo tree search. Our model, trained with a dataset of 1.4 million reaction data, achieved a top 50 accuracy of 0.94 for reaction template selection, a top 10 accuracy of 0.93 for catalyst prediction, and a top 10 accuracy of 0.89 for solvent prediction. The introduction of chemical information greatly enhances the accuracy, reliability, and efficiency of both single-step and multistep path planning.
               
Click one of the above tabs to view related content.