Computer-aided Prediciton Of ADME Properties Of Drugs

Since the late 90's,drug development research has shown that the main reason leading to the failure of the expensive drug development in the late stage is poor pharmacokinetics and toxicity(ADMET) properties.ADMET is an acronym in pharmacokinetics and pharmacology for absorption(A),distribution(D),metabolism (M),excretion(E) and toxicity(T).ADMET properties in drugs early research mainly use human tissue function proteins as "drug target",combined with the in vitro techniques and the computation simulation methods to explore how the drugs are treated in the biological body.The evaluation of ADMET in the early drug research stage can significantly improve the success rate and reduce the costs,toxicity,side effects in drug development,also can provide rational clinical guidance.It can be seen that the ADMET theoretical prediction is of great significance.Pharmacokinetics of drugs is determined by the chemical structures and the physico-chemical properties.The changes in the chemical structure of drugs and their physical and chemical properties can lead to the changes in biological activity. Quantitative structure-activity/property relationship(QSAR/QSPR) establishes a quantitative relationship between the physiological activity of compounds and their molecular structure using a number of mathematical statistics methods.Through QSAR/QSPR models,we can predict the physiological activity of compounds or certain other properties,and then provide useful information on design and synthesis the compounds with higher activity.QSAR is one of the most applicable drug design methods.The first chapter of this thesis is the introduction on the ADMET properties of drug and the principle of quantitative structure activity/property relationship(QSAR /QSPR).Here the QSPR/QSAR theoretical method was used to perform three research works on ADMET.Just as following respectively:In the second chapter,QSPR method was use to study the relationship between the intrinsic aqueous solubility of 60 druglike compounds and their structure.Genetic algorithm(GA) selected out five parameters which were used to establish the genetic algorithm-multiple linear regression method(GA-MLR) model and the partial least squares support vector machine(LS-SVM) model.The derived models were evaluated by a variety of validation methods,including the leave-one-out(LOO) cross-validation(Q~2),Y-randomization test and external validation.RMSR of 0.51, coefficient of determination R~2 of 0.88,Q~2 of 0.84 and Y-randomization R~2 of 0.07 respectively in training set of GA-MLR model and the results of test set are R~2 of 0.82, RMSE of 0.84.While in the training set of LS-SVM model,Q~2 is 0.86,R~2 is 0.90 and RMSE is 0.47,and in the test set R~2 is 0.83,RMSE is 0.49 respectively.It can be seen that the results of the GA-MLR and LS-SVM models are close to each other and both of them are robust to accurately predict the intrinsic aqueous solubility of druglike compounds.In the third chapter,the GA-MLR method was used to predict the microemulsion electrokinetic chromatography(MEEKC) capacity factor of the central nervous system(CNS) drugs,which was related closely to the blood-brain permeability,then research the blood-brain barrier indirectly.A correlation between molecular structure and capacity factor was analyzed.Leave-one-out cross-validation(Q~2),Y-random test and external validation were employed to evaluate the GA-MLR model.Index of evaluation for the training set are RMSE of 0.25,R~2 of 0.87,Q~2 of 0.75 and Y-random R~2 of 0.10 respectively,while for test set are R~2 of 0.91 and RMSE of 0.23.As can be seen,the GA-MLR model has good predictive capability.In the fourth chapter,the prediction on the free energies of salvation of 223 druglike compounds in the olive oil was discussed.Five variables selected by GA were used to construct GA-MLR and LS-SVM models.Here various validation methods also were employed to evaluate these models,such as leave-one-out cross-validation(Q~2),Y-random test and external validation.The R~2 of Y-random for GA-MLR is 0.03.For training set,Q2,R2,RMSE,AARD,F values of GA-MLR model vs LS-SVM model are 0.88 vs 0.87,0.89 vs 0.89,0.67 vs 0.67,20.08%vs 19.80%,271.86 vs 275.60 respectively,which are very close to each other;For the tests set,R2,RMSE,AARD.F of GA-MLR model vs LS-SVM model are 0.91 vs 0.92.0.73 vs 0.74.15.26%vs 14.62%,82.86 vs 84.92,which close too.That is to say the predictive ability of GA-MLR model and LS-SVM model are similar,so use only linear model can well predict the free energy of salvation of druglike compounds in olive oil.In the fifth chapter,the main content is about QSPR study on the diamagnetic susceptibility of 1080 organic compounds.GA used to select out three most important descriptors to build GA-MLR model,and then the GA-MLR was estimated to have good predictive ability by,leave-one-out(LOO) cross validation(Q~2),Y-random test as well as external validation.The statistics indicators of the training set of GA-MLR model are list as follows:AARD of 4.51%,RMSE of 5.63,R~2 of 0.98,Q~2 was 0.98 and the Y-random R~2 of-0.02.Those of test set are as follows:R~2 of 0.98,RMSE of 7.14 and AARD of 5.24%respectively.Comparatively,the statistics indicators of the training set of the model form original literature are R~2 of 0.98,RMSE of 5.41 and AARD of 5.07%.Those of test set are R~2 of 0.98,RMSE of 6.05 and AARD of 5.52%. Because the GA-MLR model can give a smaller value of AARD while the model in original literature can give a smaller RMSE,the prediction accuracy of these two models is similar.