Predicting Retrosynthetic Reactions Using Self-Corrected Transformer Neural Networks

Abstract

Synthesis planning is the process of recursively decomposing target molecules into available precursors. Computer-aided retrosynthesis can potentially assist chemists in designing synthetic routes; however, at present, it is cumbersome and cannot provide satisfactory results. In this study, we have developed a template-free self-corrected retrosynthesis predictor (SCROP) to predict retrosynthesis using transformer neural networks. In the method, the retrosynthesis planning was converted to a machine translation problem from the products to molecular linear notations of the reactants. By coupling with a neural network-based syntax corrector, our method achieved an accuracy of 59.0% on a standard benchmark data set, which outperformed other deep learning methods by >21% and template-based methods by >6%. More importantly, our method was 1.7 times more accurate than other state-of-the-art methods for compounds not appearing in the training set.

RetroXpert: Decompose Retrosynthesis Prediction Like A Chemist

“…Motivated by the success of the Transformer [20] in natural machine translation, the Transformer is also adapted for reaction outcome prediction [21] and retrosynthesis [14], in which both the product and reactant are represented in SMILES. Transformer is a sequence-to-sequence model equipped with attention mechanism, both the self-attention and the "encoder-decoder" attention [20].…”

“…which reactant to be generated first if there are two) can be determined by aligning reactants in the target with the synthons in the source, thanks to the intermediate synthons which are associated with reactants one to one. The generation order of reactants is undetermined in previous methods [7] [14], which naively treats the sequence-to-sequence model as a black box. This uncertainty makes the model training more difficult since there may be multiple orders for the similar training samples.…”

“…Secondly, once the synthons are obtained by splitting the target molecule according to the identified disconnection bonds, our Reactant Prediction Network (RPN) will try to predict associated reactants given the target and synthons accordingly. Unlike in previous methods [7] [14], the reactants generation order is undetermined since there is no such supervised correspondence information in the training data, which makes the model training converge much slower and inferior, the reactants generation order is unique in our method. Besides, we find the robustness of the RPN plays an important role in the retrosynthesis.…”

“…On the other hand, some research [7] [14] also attempts to use a SMILES sequence representation for the reactants and products, and they simply treat the retrosynthesis as a machine translation problem. Although simple and expressive, these approaches ignore rich chemistry knowledge within the chemical reaction and lack of generalization.…”

“…To overcome these challenges, inspired by the expert experience from chemists, we devise a two-step framework to automate the procedure that chemists used to do for retrosynthesis prediction. Note that in this work we focus on the "single step " version (predict possible reactants given the target product) of retrosynthesis problem following previous methods [7] [9] [14]. Our model tackles the "single step " retrosynthesis prediction problem in two steps.…”