Abstract:

Chemical reaction selection and design plays a key role in the field of drug synthesis and material synthesis. The selection of a proper chemical reaction could greatly optimize the synthesis reaction conditions, reduce time and improve synthesis yield of the product. In order to better distinguish the advantage degree of chemical reactions, a graphical neural network model was proposed to distinguish reaction superiority by using a reaction graph descriptor based on reaction atomic mapping relationships. Firstly, a reaction dataset for distinguishing the superiority of chemical reactions was established by using USPTO data. Then, the graphical neural network modeling method was used to establish the mapping relationship between the chemical reactions molecular and atomic features and the reaction superiority probability values. Finally, the aspirin different chemical synthesis reactions were taken as examples to prove and verify the feasibility and superiority of the reaction indicator by using the relevant experimental information.

Key words: chemical reaction, neural networks, model, reaction graph descriptor, reaction indicator design

摘要：

化学反应的选择与设计在药物合成、材料合成领域中起到了关键性的作用，选择一条合适的化学反应能极大地优化反应条件，降低产品的合成时间，提升目标产物的合成产率。为能更好地分辨出化学反应的优劣程度，本文提出了一种用于评价化学反应优劣的图神经网络模型，通过采用基于反应前后原子映射关系构建的反应图描述符提取化学反应特征评估化学反应的优劣程度。首先，利用USPTO中反应数据建立包含反应映射关系的化学反应优劣数据集。然后，利用图神经网络方法建立起化学反应中各分子、原子特征与反应难易程度概率值的映射关系。最后，本文以阿司匹林合成反应为例，利用该模型对药品不同合成反应进行对比评估，并通过相关实验信息进行验证，证明当前评价指标的可行性与优越性。

关键词: 化学反应, 神经网络, 模型, 反应图描述符, 反应评价指标设计