最新NLP论文解析--附RetroXpert复现教程

编者按

RetroXpert--Decompose Retrosynthesis Prediction Like A Chemist是NeurIPS2020上的一篇关于逆合成的文章。其中，RetroXpert模型在目前基于模版的逆合成方向做端到端预测中Top-n accuracy 最高的模型。本文总结分析了文献的主要实验方法，预测结果以及预测可视化成果，以供读者更好的理解文献。

1.引言

This article was led by AI Lab in November 2020, and was published in NeurIPS 2020 with Texas University, Tsinghua University and Sun Yat-sen University. The design problem is mainly divided into two steps: firstly, a new graph neural network is used to identify the potential reaction center of the target molecule, and an intermediate synthon is generated, and then a robust reactant generation model is used to generate the synthon related The reactants. The difference between it and G2Gs is that G2Gs can only predict a special key disconnection at most. In addition, G2Gs independently produce multiple reactants, ignoring the relationship between multiple reactants. In order to overcome two difficulties, three metho