RetroXpert: Decompose Retrosynthesis Prediction Like A Chemist

Abstract

<div>Retrosynthesis is the process of recursively decomposing target molecules into available building blocks. It plays an important role in solving problems in organic synthesis planning. To automate the retrosynthesis analysis, many retrosynthesis prediction methods have been proposed.</div><div>However, most of them are cumbersome and lack interpretability about their predictions.</div><div>In this paper, we devise a novel template-free algorithm, RetroXpert, for automatic retrosynthetic expansion by automating the procedure that chemists used to do.</div><div>Our method disassembles retrosynthesis into two steps: i) we identify the potential reaction center within the target molecule through a graph neural network and generate intermediate synthons; and ii) we predict the associated reactants based on the obtained synthons via a reactant generation model. </div><div>While outperforming the state-of-the-art baselines by a significant margin, our model also provides chemically reasonable interpretation.</div>

Molecule Edit Graph Attention Network: Modeling Chemical Reactions as Sequences of Graph Edits

“…G2G 61 is a template-free model that, similarly to MEGAN, generates reactions by modifying molecular graphs but with a separate module for predicting reaction center. GraphRETRO 63 employs two separate graph models for predicting reaction center and synthon completion. Retro-Prime 74 uses transformer-based models for both the reaction center and leaving group prediction.…”

“…In Table 5, we compare the performance of MEGAN and molecular transformer on USPTO-MIT in top K predictions. We use the best non-ensemble models provided by Schwaller et 12 Somnath et al, 62 Yan et al, 63 Wang et al, 74 and Tetko et al 77 † denotes concurrent work. We also note that AT averages predictions over 100 different augmentations, which significantly increases inference time.…”

“…Concurrently to our work, the graph–edit reaction generation approach has been employed to retrosynthesis in G2Gs and GraphRETRO models. − The G2Gs model presented in Shi et al and Yan et al consists of two separate modules for predicting the reaction center and generating final substrates from the disconnected synthons. GraphRETRO employs a similar framework but completes the synthons with substructures which are selected from a predefined set found on the training data. In contrast to these concurrent works, we propose an end-to-end model for generating reaction through a sequence of edits and apply it both to retro and forward synthesis prediction.…”

“…The highest accuracy is achieved in the ordering called BFS Rand-augmented transformer (AT). In this ordering, for retrosynthesis, we give bond deletion the highest priority, as it is usually the step that determines the reaction center, which is the first step of reaction prediction in other methods. − , Analogously, adding a bond has the highest priority for forward synthesis, as this usually determines the reaction center in forward prediction. Priorities of other types of actions were determined experimentally; we did not observe a significant change in validation error when modifying these priorities.…”

NeuralTPL: a deep learning approach for efficient reaction space exploration

“…Another group of methods adopt either sequence-based or graph-based deep learning architectures to directly model the entire reactions. These template-free methods [9,10,11,12,13,14,15,16,17,18] first learn the chemical knowledge from the visible reaction space and perform synthetic routes generation directly from the entire reaction space. They hold the potential to be able to infer novel chemical transformations beyond the training scope.…”