Objective To construct a deep learning model(SGRN-Trans) based on knowledge graph and attention mechanism for predicting the interaction between pharmacodynamic components and targets in classic traditional Chinese medicine(TCM) prescriptions with Wendan Decoction as an example,and to assess its predictive performance. Methods The SGRN-Trans predictive model was proposed for the first time. Multiple biological data sources were used to construct the knowledge graph of Wendan Decoction(WDKG),and graph neural networks were used to learn the low-dimensional embedding representation of each entity in the knowledge graph. The respective structural features of TCM components and targets were introduced,and the Transformer model based on attention mechanism was used to predict the interaction between pharmacodynamic components and targets. Molecular docking and literature review were used for validation. Results WDKG contained 10 types of entities,with 14292 entities in total,which could be used for the research on deep learning models. The SGRN-Trans predictive model showed the best performance compared with other knowledge graph embedding models such as TransE,TransR,ComplEx,DistMult,and ConvKB. Molecular docking and visualized presentation were performed for the top 20 groups of pharmacodynamic components and targets,among which 8 combinations suggested the potential interaction between pharmacodynamic components and targets. With the interaction between soya-cerebroside (an effective constituent of Pinellia ternata in Wendan Decoction) and low-density lipoprotein receptor as an example,the literature review showed that it might be one of the mechanisms for Wendan Decoction in the treatment of atherosclerosis. Conclusion The SGRN-Trans model based on knowledge graph and attention mechanism proposed in this study can be widely used to predict the interaction between components and targets in the complex network system of classic TCM prescriptions,which provides a new tool for clarifying the pharmacodynamic material basis of classic TCM prescriptions and related mechanisms of action.
本研究以中草药数据库TCMSP(Traditional Chinese Medicine Systems Pharmacology Database and Analysis Platform,TCMSP)和生物医药知识图谱DRKG(Drug Repurposing Knowledge Graph,DRKG)为数据基础,构建以“Head-Relation-Tail”(头-关系-尾)形式的三元组组成的WDKG。首先,从TCMSP中获取温胆汤六味中药(Herb)所包含的药物分子(compound)和各药物分子关联的靶点或基因,去重并滤除无法在PubChem(https://pubchem.ncbi.nlm.nih.gov/)数据库中查找到的药物分子和无法在DrugBank(https://go.drugbank.com/)数据库中查询得到或者非人类靶点的基因(Gene)。保留下来的药物和靶点之间的相互作用关系作为一类三元组(Compound:Gene),同时也是本研究的DTI数据集。将获得的基因实体在DRKG提取出包含这些基因实体在内的所有三元组。分析统计WDKG中不同类实体的数量、来源、占比。将WDKG中的实体关系通过 Neo4j (https://neo4j.com/)图数据库进行存储和可视化。
在评估知识图谱嵌入的性能时,本课题组计算每个真实三元组的得分并将它们降序排列,得到每个三元组在其中的排名,并计算它们的平均倒数排名(mean reciprocal rank,MRR)以及排名前N的三元组所占的比例(Hits@N,N=1,3,10,50,100)。这些评估指标适用于评估知识图谱嵌入的整体性能,却并不适用于针对知识图谱中特定关系类型的下游任务。在DTI预测任务中,将其处理为一个二分类任务,为了评估其效果,引入了AUROC(area under the receiver operating characteristic curve,AUROC)和AUPR(area under precision-recall curve,AUPR)这2个在分类任务中应用广泛的评价指标,同时还引入了F1-score,其能够评估分类模型在不平衡数据集(在本研究中正样本和负样本的比例为1∶4)上的综合性能。
1.6 模型参数设置
模型使用Adam(Adaptive Moment Estimation)算法作为训练所有模型时的优化算法,在分别训练SGRN与Transformer时预先对以下超参数进行调整并最终设置见表1。
Notice of the Office of the State Administration of Traditional Chinese Medicine and the Comprehensive and Planning Finance Department of the State Medical Products Administration on Issuing the“Key Information Research Principles for Ancient Classic Famous Prescriptions”and“Key Information Table for Ancient Classic Famous Prescriptions( 7 Prescriptions)”[EB/OL]. (2020-11-11)[2024-05-26].
YangSY, DongLP, TanHS,et al. Quality value transmitting of Wendan Decoction reference samples[J]. Chin Tradit Pat Med,2023,45(6):1788-1794.
[11]
YangYT, ChenR, LiCX,et al. A synthetic external control study comparing the clinical efficacy of Wendan Decoction and 19 antidepressants[J]. Int J Neuropsychopharmacol,2023,26(10):739-746.
[12]
WangN, LiP, HuXC,et al. Herb target prediction based on representation learning of symptom related heterogeneous network[J]. Comput Struct Biotechnol J,2019,17:282-290.
[13]
XuZG. Modernization:one step at a time[J]. Nature,2011,480(7378):90-92.
[14]
BagherianM, SabetiE, WangK,et al. Machine learning approaches and databases for prediction of drug-target interaction:a survey paper[J]. Brief Bioinform,2021,22(1):247-269.
[15]
RuXQ, YeXC, SakuraiT,et al. Current status and future prospects of drug-target interaction prediction[J]. Brief Funct Genomics,2021,20(5):312-322.
WanFP, HongLX, XiaoA,et al. NeoDTI:neural integration of neighbor information from a heterogeneous network for discovering new drug-target interactions[J]. Bioinformatics,2019,35(1):104-111.
[18]
ChengFX, LiuC, JiangJ,et al. Prediction of drug-target interactions and drug repositioning via network-based inference[J]. PLoS Comput Biol,2012,8(5):e1002503.
[19]
ChenX, LiuMX, YanGY. Drug-target interaction prediction by random walk on the heterogeneous network[J]. Mol Biosyst,2012,8(7):1970-1978.
[20]
MorganHL. The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service[J]. J Chem Doc,1965,5:107-113.
[21]
DongJ, YaoZJ, ZhangL,et al. PyBioMed:a python library for various molecular representations of chemicals,proteins and DNAs and their interactions[J]. J Cheminform,2018,10(1):16.
[22]
DubchakI, MuchnikI, HolbrookSR,et al. Prediction of protein folding class using global description of amino acid sequence[J]. Proc Natl Acad Sci U S A,1995,92(19):8700-8704.
YeQ, HsiehCY, YangZY,et al. A unified drug-target interaction prediction framework based on knowledge graph and recommendation system[J]. Nat Commun,2021,12(1):6775.
ChenXJ, XuDP, ChenXG,et al. Overview of the relationship between phlegm syndrome and objective indexes in coronary heart disease[J]. China J Tradit Chin Med Pharm,2017,32(7):3089-3092.
[27]
XuJH, HuangYM, LingW,et al. Wen Dan Decoction for hemorrhagic stroke and ischemic stroke[J]. Complement Ther Med,2015,23(2):298-308.
HeHT. Effects of the Wendan Decoction on angina of incomplete revascularization patients with coronary heart disease of the Qixu Tanyu type[J]. Clin J Chin Med,2019,11(10):73-75.
ChenXJ, ZhongBY, WuHL,et al. Inhibitory effect of Wendan Decoction on formation of foam cells induced by ox-LDL[J]. Chin J Pathophysiol,2020,36(11):1952-1959.
ParkYM, FebbraioM, SilversteinRL. CD36 modulates migration of mouse and human macrophages in response to oxidized LDL and may contribute to macrophage trapping in the arterial intima[J]. J Clin Invest,2009,119(1):136-145.