Predicting Retrosynthetic Reaction 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, but at present it is cumbersome and provides results of dissatisfactory quality. In this study, we develop a template-free self-corrected retrosynthesis predictor (SCROP) to perform a retrosynthesis prediction task trained by using the Transformer neural network architecture. In the method, the retrosynthesis planning is converted as a machine translation problem between molecular linear notations of reactants and the products. Coupled with a neural network-based syntax corrector, our method achieves an accuracy of 59.0% on a standard benchmark dataset, which increases >21% over other deep learning methods, and >6% over template-based methods. More importantly, our method shows an accuracy 1.7 times higher than other state-of-the-art methods for compounds not appearing in the training set.

Predicting Retrosynthetic Pathways Using a Combined Linguistic Model and Hyper-Graph Exploration Strategy

“…Schwaller et al [22] recently proposed to ignore reactant and reagent roles for the reaction prediction task. In contrast to previous works [32,33,35,36], the single-step retrosynthetic model presented here predicts reactants and reagents. In an effort to simplify the prediction task, the most common precursors with a length of more than 50 tokens were replaced by molecule tokens.…”

“…Duan et al [37] increased the batch size and the training time for their Transformer model and were able to achieve a top-1 accuracy of 54.1% on the 50k USPTO data set [44]. Later on, the same architecture was reported to have a top-1 accuracy of 43.8% [36], in line with the three previous transformer-based approaches [32,33,35] but significantly lower than the accuracy previously reported by Duan et al [37]. Interestingly, the transformer model was also trained on a proprietary data set [36], including only reactions with two reactants with a Tanimoto similarity distribution peaked at 0.75, characteristic of an excessive degree of similarity (roughly 2 times higher than the USPTO).…”

“…Here, we dispute the previous use of top-N accuracy [6,9,12,16,[32][33][34][35][36][37] and to introduce four different metrics, namely, round-trip accuracy, coverage, class diversity and Jensen-Shannon divergence [50], as seen in Figure 3, to evaluate single step retrosynthetic models and through them retrosynthetic tools as a whole. All these four metrics have been critically designed and assessed with the help of human domain experts (see Section 4.2 for a detailed description).…”

“…While this extensive production of AI models for Organic chemistry was made possible by the availability of public data [28,29], the noise contained in this data and generated by the text-mining extraction process is heavily reducing their potential. In fact, while rule-based systems [30] demonstrated, through wet-lab experiments, the capability to design target molecules with less purification steps and hence, leading to savings in time and cost [31], the AI approaches [6,9,12,16,[32][33][34][35][36][37][38] still have a long way to go.…”

“…Inspired by the success of the Molecular Transformer [22,42,43] for forward reaction prediction, a few retrosynthetic models based on the same architecture were reported shortly after [32,33,[35][36][37]. Zheng et al [32] proposed a template-free self-corrected retrosynthesis predictor built on the Transformer architecture. The model achieves 43.7% top-1 accuracy on a small standardized (50k reactions) data set [44].…”