| Background:Currently,the metabolic processes are still unclear for most chemical constituents from traditional Chinese medicine(TCM),which seriously blocked the discovery of effective material basic and mechanism of TCM.Study on the metabolites is contributed to reveal the dynamic change process of TCM and provide valuable information for the discovery of effective material basis and mechanism of TCM.Compared to detect metabolites experimentally,to predict metabolites using in silico methods is more time-saving and low-cost.Currently,in silico methods have been considered as another effective approaches to identification of metabolites beside experimental methods.During the past decades,there have been continuous attempts in the prediction of drug metabolism.However,most of these works can only provide signal about site of metabolism(SOM),and the reliability these models is not so optimistic.Therefore,it is compelled to construct a reliable metabolite prediction model.Objective:To compensate for the lack of information on metabolites or drug-metabolizing enzymes of prior works,an efficient and effective metabolite prediction model for the chemical constituents from TCM will be developed,which will be further applied to the construction of metabolic pathway prediction model.Methods:(1)Investigation of metabolism data.We assessed the accuracy,completeness,and applicability of 11 metabolism databases by investigating their original data source,number of metabolic reactions,and data format,respectively.The availability of the metabolism data was also taken into account.Eventually,the database of which data derived from the literature was regarded as the optimal data source,with maximum number of metabolic reactions and high applicability.(2)Construction of metabolite prediction model for chemical constituents from TCM.Based on the metabolism data collected in(1),each microsomal metabolic reaction system(MMRS)was marked as positive or unlabeled,which depended upon whether the MMRS undergone biotransformation.According to the available literature reports,if an SOM was experimentally determined as a positive SOM,we marked the corresponding MMRS as a positive MMRS.In contrast,if an SOM has not been experimentally detected as a positive SOM,the corresponding MMRS was marked as an unlabeled MMRS.Then a voting method was used to select negative MMRSs from the unlabeled MMRSs.Based on the negative and positive MMRSs,four feature selection methods(CHI,GR,IG,Relief)were utilized to select feature subsets.The feature subsets were used as input to train eight classifier procedures(BayesLibSVM?Kstar?IBK?AdaBoost?Boosting?J48?RandomForest),then a series models were obtained.Eventually,comparisons among the models were conducted,and the optimal model was attained.Then,we evaluated the reliability of the model in both "SOM-scale" and"molecule-scale",and comparisons between prior works and the model developed in this work were also performed.(3)Construction of metabolic pathway prediction model for chemical constituents from TCM.The metabolic pathway prediction model was constructed based on two sets metabolite prediction models developed in(2).Then drug metabolic pathways reported in the literature were utilized to estimate the reliability of the metabolic pathway prediction model.Results:(1)Construction of metabolite prediction model for CYP450 3A4,2D6,and 2C9.In this step,we developed metabolite prediction model based on metabolism data reported in the literature.Five major biotransformations,including aliphatic C-hydroxylation,aromatic C-hydroxylation,N-dealkylation,O-dealkylation and S-oxidation,were involved.Accuracy of the internal test ranged from 0.940 to 0.987.Sensitivity(SE)and specificity(SP)were generally in excess of 0.856 and 0.968,respectively.For the independent test,accuracy was 0.947 to 0.994,SE and SP exceeded 0.864 and 0.933,respectively.In addition,for receiver operating characteristic(ROC)analysis,each of these models gave a significant area under curve(AUC)value>0.953.For the external test set gathered from the literature,SE and SP were 0.821 and 0.956,respectively.Compared to the results provided by previous works(SE:0.701,SP:0.963),the ability of our model to correctly identify true positives has risen by 12%.For the ability to correctly identify true negatives,there was no significant difference between our model and prior works.In "molecule-scale",17 out of 34 molecules were predicted perfectly by our models.However,only 15 molecules were predicted perfectly by prior works.In summary,we may speculate that our models are better than prior works.(2)Construction of metabolite prediction model for oxidoreductases.11 metabolism databases were investigated,and BKM database was chosen as the optimal database.Based on the metabolism data included in BKM database,seven metabolite prediction models were established,including Alcohol Oxidation to Ketone,Alcohol Oxidation to Aldehyde,Carbon Double bond Formation,N-Dealkylation,aliphatic C-hydroxylation,aromatic C-hydroxylation,and Oxidative Deamination.Altogether 655 metabolic enzymes were involved.For the training set,ACC ranged from 0.901 to 0.995.Almost each model gave SE value ranged from 0.875 to 0.988 except for Alcohol Oxidation to Ketone and aromatic C-hydroxylation.SE of Alcohol Oxidation to Ketone and aromatic C-hydroxylation was 0.777 and 0.765,respectively.SP and AUC were generally in excess of 0.963 and 0.916,respectively.For the independent test set,AUC was 0.885 to 0.969.SE was commonly greater than 0.777 except for aromatic C-hydroxylation.SP and AUC were generally exceeded 0.963 and 0.916,respectively.The balanced accuracy of training set and test set was higher than 0.858 and 0.813,respectively.For the external test set collected from the literature,31 out of 34 positive SOMs was correctly identified.SE was 0.912.In "molecule-scale",we found that 23 out of the 31 molecules were in accordance with reports in the literature.(3)Construction of metabolic pathway prediction model for chemical constituents from TCM.Based on the metabolite prediction models developed in(1)and(2),a metabolic pathway prediction model was constructed.This model was aimed to predict the metabolic pathways of chemical constituents from TCM.The scope of this model was demonstrated as follows:Five biotransformations mediated by CYP450 3A4,2D6,and 2C9,including aromatic C-hydroxylation,N-dealkylation,O-dealkylation,S-oxidation and aliphatic C-hydroxylation.Seven biotransformations mediated by oxidoreductases,including Alcohol Oxidation to Ketone,Alcohol Oxidation to Aldehyde,Carbon Double bond Formation,N-Dealkylation,aliphatic C-hydroxylation,aromatic C-hydroxylation,and Oxidative Deamination.Then the metabolic pathways of citalopram and aconitine reported in the literature were utilized to evaluate the reliability of our metabolic pathway prediction model.Consequently,the prediction results can reveal the metabolic pathways of citalopram and aconitine to certain extent.Conclusions:In the current study,a novel approach to construction of metabolite prediction model was proposed.On the basis of this method,five metabolite prediction models for CYP450 3A4,2D6,2C9,and seven metabolite prediction models for oxidoreductases were established.All of these models were reasonably successful.Then a metabolic pathway prediction model was developed based on the metabolite prediction models developed.Then some drug metabolic pathways reported in the literature were utilized to evaluate the reliability of metabolic pathway prediction model developed.It turned out that the metabolic pathway prediction model can reveal the metabolic pathways of chemical constituents from TCM partly.In summary,the metabolite prediction models and metabolic prediction model developed in the present study will provide valuable information for the discovery of effective material basis and mechanism of TCM.Furthermore,our work will also assist in the process of drug design and optimization. |
PDF Full Text Request