RetroTRAE: retrosynthetic translation of atomic environments with Transformer

We present a new single-step retrosynthesis prediction method, viz. RetroTRAE, using fragment-based tokenization and the Transformer architecture. RetroTRAE mimics chemical reasoning, and predicts reactant candidates by learning the changes of atom environments (AEs) associated with the chemical reaction. AEs are the ideal stand-alone chemically meaningful building blocks providing a high-resolution molecular representation. Describing a molecule with a set of AEs establishes a clear relationship between translated product-reactant pairs due to the conservation of atoms in the reactions. Our model achieved a top-1 accuracy of 58.3% on the USPTO test dataset. When highly similar analogs were considered the accuracy increased to 61.6%. These results outperform other state-of-the-art neural machine translation-based methods. Besides yielding a high level of overall accuracy, the proposed method does not suffer from the SMILES-based translation issues such as invalid SMILES. Additionally, the attention matrices of RetroTRAE are shown to capture chemical changes around reaction sites successfully. Through careful inspection of reactant candidates, we demonstrated that AEs are promising descriptors for studying reaction route prediction and discovery, which has been underexplored yet. Our methodology offers a novel way of devising a retrosynthetic planning model using fragmental and topological descriptors as natural inputs for chemical translation tasks, and opens new possibilities for developing other sequence-based machine-learning methods in chemistry.