基于深度学习的神经性毒剂及其相关化合物特征基团结构分类模型研究  

Research on Classification Models of Functional Groups of Nerve Agents and Their Related Compounds Based on Deep Learning

作　　者：李雪萌 徐斌 徐亚平 谢剑炜 赵东升 LI Xuemeng;XU Bin;XU Yaping;XIE Jianwei;ZHAO Dongsheng(Academy of Military Medical Sciences,Academy of Military Sciences,Beijing 100850,China)

机构地区：[1]军事科学院军事医学研究院,北京100850

出　　处：《防化研究》2025年第1期69-74,共6页CBRN DEFENSE

摘　　要：在神经性毒剂的识别中,相似性谱图搜索是公认的可靠谱图注释方法,然而该方法受限于参考标准谱图库,无法实现库外谱图注释,因而对库外未知化合物的鉴定耗时长且人工判定误差大。本文提出了一种基于深度卷积神经网络的神经性毒剂及其相关化合物的特征基团结构分类模型,并在中央分析数据库(Official Central Analytical Database,OCAD)数据集上完成了模型的训练和测试。构建平衡训练数据集后,模型在测试集上的精确率、召回率和F1分数分别为96.89%、96.88%和96.88%。测试结果表明,其在OCAD数据集上精确率为98.78%,在美国国家标准与技术研究院(National Institute of Standards and Technology,NIST)数据集上精确率为95.73%。该方法在预测神经性毒剂特征基团结构时不依赖于标准谱图库,不借助预先建立的规则、程序或峰值匹配,可为实验人员提供辅助决策。In the identification of nerve agents,similarity search of spectra are widely recognized as a reliable method for spectral annotation.However,this approach is constrained by reference standard spectral libraries and cannot annotate out-of-library spectra.Consequently,the identification of unknown compounds outside existing libraries proves to be time-consuming and prone to significant human judgment errors.The time-consuming nature of manual analysis and the high propensity for human error further exacerbate these challenges.In this paper,a classification model based on deep convolutional neural network specifically designed for the classification of the functional groups of nerve agents and related compounds was introduced.The model was trained and tested using Official Central Analytical Database(OCAD)dataset.After constructing a balanced training dataset,the model achieved precision,recall and F_1-score on the test set were 96.89%,96.88%and 96.88%,respectively.Testing on the OCAD and National Institute of Standards and Technology(NIST)datasets showed a precision of 98.78%and 95.73%,respectively.This method offered a significant advantage by predicting functional groups without reliance on standard reference libraries or the need for pre-established rules,procedures or peak matching,thus providing valuable assistance in decision-making for experimental personnels.

关 键 词：深度卷积神经网络 神经性毒剂及其相关化合物 特征基团 结构分类 决策支持 

分 类 号：E929[军事—军事装备学] TP181[兵器科学与技术—武器系统与运用工程]