Predicting Chemical Reaction Product by Graph Transformer

The substantial cost and prolonged duration associ-ated with drug discovery present significant challenges within the pharmaceutical industry. A potential solution to alleviate these issues is molecular translation, a process where an input molecule is converted into another molecule following specific rules or conditions. In this study, we introduce a novel machine learning technique for predicting chemical reaction outcomes: the molecular graph translation method, employing a transformer architecture designed for efficient chemical reaction product prediction. We assessed the performance of our proposed method using the USPTO dataset for chemical reaction product prediction. Our results demonstrated an accuracy of 86.5%. Although this accuracy was slightly lower than that of the Molecular Transformer (a string-based method using a similar architecture), our approach surpassed both the Seq2Seq and other graph-based methods in terms of performance.