Transformer: Linking Atom Mapping and Neural Machine Translation

Atom mapping reveals the corresponding relationship between reactant and product atoms in chemical reactions, which is important for drug design, exploration for underlying chemical mechanism, reaction classification and so on. Here, we present a new method that links atom mapping and neural machine translation using the transformer model. In contrast to the previous algorithms, our method runs reaction prediction and captures the information of corresponding atoms in parallel. Meanwhile, we use a set of approximately 360K reactions without atom mapping information for obtaining general chemical knowledge and transfer it to atom mapping task on another dataset which contains 50K atom-mapped reactions. With manual evaluation, the top-1 accuracy of the transformer model in atom mapping reaches 91.4%. we hope our work can provide an important step toward solving the challenge problem of atom mapping in a linguistic perspective.