基于集成深度学习框架的新型冠状病毒感染治疗药物活性预测  

Prediction of Drug Activity for COVID-19 Based on Ensemble Deep Learning Framework

作　　者：许强 罗杰斯 杨明[1] 张永林 XU Qiang;LUO Jiesi;YANG Ming;ZHANG Yonglin(Affiliated Hospital of North Sichuan Medical College,Nanchong,Sichuan,China 637000;Southwest Medical University,Luzhou,Sichuan,China 646000)

机构地区：[1]川北医学院附属医院,四川南充637000 [2]西南医科大学,四川泸州646000

出　　处：《中国药业》2024年第14期47-53,共7页China Pharmaceuticals

基　　金：西南特色中药资源国家重点实验室开放基金[SKLTCM2022028];川北医学院校级科研发展计划项目[CBY22-QNA38]。

摘　　要：目的 建立预测新型冠状病毒感染治疗药物活性的集成深度学习框架。方法 采用卷积神经网络(CNN)和递归神经网络(RNN)从简化分子线性输入规范(SMILES)字符串序列信息中筛选出代表性的特征标识,以深度神经网络(DNN)从离散特征信息中提取更高级别的抽象特征,均以网格筛选法生成1个主框架模型和7个离散特征模型的最优结构,构成8种架构的127种可能组合。通过准确率(ACC)、F、召回率(Recall)、精确度(PRE)和马修斯相关系数(MCC)5个标准指标评估模型的预测性能。建立和维护最终框架。结果 最终建立了1个以BiLSTM为集成深度学习框架的核心架构和4个不同的离散特征模型组成的集成深度学习模型,训练集ACC为72.84%,F为69.70,Recall为72.21%,PRE为68.03,MCC为0.456 9;测试集中成功预测了23种可能对新型冠状病毒感染有治疗作用的药物。结论 集成深度学习框架相较于单个模型具有更强的预测能力,该研究为新型冠状病毒感染治疗药物的筛选提供了新的选择。Objective To establish an ensemble deep learning framework for predicting the activity of drugs for Corona Virus Disease 2019(COVID-19).Methods Convolutional neural network(CNN)and recursive neural network(RNN)were used to screen the representative feature identifiers from the simplified molecular input line entry system(SMILES)sequence.Deep neural network(DNN)was used to extract higher-level abstractfeatures from discrete feature information.The optimal structure of one main framework model and seven discrete feature models was generated by the grid search method,forming 127 possible combinations of eight architectures.The predictive performance of model was evaluated by the accuracy(ACC),F,Recall,precision(PRE)and Matthews correlation coefficient(MCC).The final framework was established and maintained.Results An ensemble deep learning model with BiLSTM as the core architecture and consisting of four different discrete feature models was ultimately established.The ACC of the training set was 72.84%,the F was 69.70,the Recall was 72.21%,the PRE was 68.03,and the MCC was 0.4569.Twenty-three drugs that might be effective against COVID-19 were successfully predicted in the test set.Conclusion The ensemble deep learning framework has better predictive performance than a singular model,this study provides a new choice for the screening of the drugs for COVID-19.

关 键 词：集成深度学习框架 新型冠状病毒感染 药物活性 神经网络 自动生物序列 

分 类 号：R97[医药卫生—药品] TQ311.5[医药卫生—药学]