Translating the InChI: adapting neural machine translation to predict IUPAC names from a chemical identifier

Abstract We present a sequence-to-sequence machine learning model for predicting the IUPAC name of chemical from its standard International Chemical Identifier (InChI). The uses two stacks transformers in an encoder-decoder architecture, setup similar to neural networks used state-of-the-art translation. Unlike translation, which usually tokenizes input and output into words or sub-words, our processes InChI predicts character by character. was trained on dataset 10 million InChI/IUPAC pairs freely downloaded National Library Medicine’s online PubChem service. Training took seven days Tesla K80 GPU, achieved test set accuracy 91%. performed particularly well organics, with exception macrocycles, comparable commercial generation software. predictions were less accurate inorganic organometallic compounds. This can be explained inherent limitations representing inorganics, as low coverage training data.