基于古今特征融合与图卷积网络的活血药对配伍预测  

Study on Compatibility and Efficacy of Blood-activating Herb Pairs Based on Graph Convolution Network

作　　者：王晶爱 牛琪锴 宗文静 曾子玲 田思玮 张思琪 赵钰文 张华敏[2] 霍炳杰[3] 李兵[1] WANG Jingai;NIU Qikai;ZONG Wenjing;ZENG Ziling;TIAN Siwei;ZHANG Siqi;ZHAO Yuwen;ZHANG Huamin;HUO Bingjie;LI Bing(Institute of Chinese Materia Medica,China Academy of Chinese Medical Sciences,State Key Laboratory for Quality Ensurance and Sustainable Use of Dao-di Herbs,Institute of Chinese Materia Medica,China Academy of Chinese Medical Sciences,Beijing 100700,China;Institute of Basic Theory for Chinese Medicine,China Academy of Chinese Medicine Science,Beijing 100700,China;The Fourth Hospital of Hebei Medical University,Key Laboratory of Traditional Chinese Medicine Treatment of Digestive Tract Tumors in Hebei Province,Shijiazhuang 050010,China)

机构地区：[1]中国中医科学院中药研究所,道地药材与品质保障全国重点实验室,北京100700 [2]中国中医科学院中医基础理论研究所,北京100700 [3]河北医科大学第四医院河北省消化道肿瘤中医辨治重点研究室,石家庄050010

出　　处：《中国实验方剂学杂志》2025年第8期228-234,共7页Chinese Journal of Experimental Traditional Medical Formulae

基　　金：国家重点研发计划项目(2023YFC3502900,2022YFC3500900);国家自然科学基金项目(82474376);中国中医科学院科技创新工程项目(CI2023C063YLL,CI2021B015,CI2023E002);中央级公益性科研院所基本科研业务费专项(ZXKT21024)。

摘　　要：目的:该研究旨在构建一种基于图卷积网络(GCN)的中药药对配伍预测模型(HC-GCN),该模型融合了传统中药药性和现代药理机制特征以预测具有特定功效的药对配伍,并以活血类中药为示范进行了预测应用和效能验证。方法:通过系统收集常用中药药对及其所包含的中药性味、归经、靶点基因等特征数据,构建药对预测训练数据集。结合中药的传统特质和现代生物信息建立图卷积网络,通过药对配伍复杂关系的机器学习和功效特征的加权方法,构建相应功效导向的药对配伍预测HC-GCN模型。采用准确率(ACC)、召回率(Recall)、精确度(Precision)、F1分数(F1-score)和ROC曲线下面积(AUC)指标评估HC-GCN模型的性能,并与极端梯度提升(XGBoost)、逻辑回归(LR)、朴素贝叶斯(Naive Bayes)、K最近邻(KNN)、支持向量机(SVM)5种机器学习模型的预测效果进行比较分析。结果:以具有活血功效的药对为示范,基于46个活血药对及其中药性味、归经、靶点基因特征的基础数据集构建了预测模型。HC-GCN模型在ACC、Recall、Precision、F1-score和AUC等关键性能指标上均优于其他常用机器学习模型。通过HC-GCN模型的预测分析,成功预测了60个可能具有活血功效的中药药对。在所预测药对配伍中,有44个药对中至少含有1味具有活血功效的中药。结论:该研究通过结合传统中药特质和现代药理机制,建立了基于图卷积网络的功效导向药对配伍预测模型,展现出较高的预测性能,为中药组方智能筛选优化和及其临床应用提供新的方法。Objective:This study aims to develop a prediction model for the compatibility of Chinese medicinal pairs based on Graph Convolutional Networks(GCN),named HC-GCN.The model integrates the properties of herbs with modern pharmacological mechanisms to predict pairs with specific therapeutic effects.It serves as a demonstration by applying the model to predict and validate the efficacy of blood-activating herb pairs.Methods:The training dataset for herb pair prediction was constructed by systematically collecting commonly used herb pairs along with their characteristic data,including Qi,flavor,meridian tropism,and target genes.Integrating traditional characteristics of herb with modern bioinformatics,we developed an efficacy-oriented herb pair compatibility prediction model(HC-GCN)using graph convolutional networks(GCN).This model leverages machine learning to capture the complex relationships in herb pair compatibility,weighted by efficacy features.The performance of the HC-GCN model was evaluated using accuracy(ACC),recall,precision,F1 score(F1),and area under the ROC curve(AUC).Its predictive effectiveness was then compared to five other machine learning models:eXtreme Gradient Boosting(XGBoost),logistic regression(LR),Naive Bayes,K-nearest neighbor(KNN),and support vector machine(SVM).Results:Using herb pairs with blood-activating effects as a demonstration,a prediction model was constructed based on a foundational dataset of 46 blood-activating herb pairs,incorporating their Qi,flavor,meridian tropism,and target gene characteristics.The HC-GCN model outperforms other commonly used machine learning models in key performance metrics,including ACC,recall,precision,F1 score,and AUC.Through the predictive analysis of the HC-GCN model,60 herb pairs with blood-activating effects were successfully identified.Among of these potential herb pairs,44 include at least one herb with blood activating effects.Conclusion:In this study,we established an efficacy-oriented compatibility prediction model for herb pairs based on GCN b

关 键 词：中药配伍 图卷积网络 活血功效 预测模型 临床决策 

分 类 号：R259[医药卫生—中西医结合] R289[医药卫生—中医内科学] R287[医药卫生—中医学]