QSAR Model Based On PLS Variable Selection Procedure For Rapid Evaluation On The Performance Of Drug ADME

It is a trend in the drug QSAR research and development that straightforwardmathematical models could be established quickly and easily. One major objective ofthis paper is to build a reliable variable selection method by PLS. Optimizationdescriptor set is quickly extracted from a large number of molecular descriptorsaccording to drug pharmacokinetics parameter, especially for drug’s absorption,distribution, and excretion. Then PLS regression model is built based on optimizationdescriptor set for predicting drug pharmacokinetics parameter and verified by interiorand exterior calibration methods. The PLS variable selection method is constructed by134molecules’ penetration coefficient on the Caco-2cell layers--log (P). Firstly,3764descriptors of these molecules are quickly calculated by DRAGON software. Thesedescriptors are pretreated according to the variables variance, the number of non-zerosamples and the correlation coefficient on variables. Then the PLS method is repeated todetermine the number of latent variables A and to calculate the corresponding descriptorVIP. An optimization descriptor set with58descriptors is derived from a large numberof descriptors according to descriptor VIP values. Finally, the PLS regression model isbuilt based on optimization descriptor set for predicting penetration coefficient andverified by interior and exterior calibration methods (n=96, A=5, r=0.8548, RMSE=0.45, q=0.7576, RMSV=0.59, n_p=38, r_p=0.7432,RMSP=0.60). This model has good fittingability, robustness and high predictive ability. It determined that a high quality andutility model could be built with optimization descriptor set selected by PLS variableselection method. The same time, this method has been effectively applied to predictdrug ’s activity parameters including human intestinal absorption, percutaneousabsorption, and penetration coefficient on placental barrier. PLS variable selectionmethod is a fast and effective variable screening method in QSAR research.The PLSR models evaluating the BBB permeability were derived and validated byusing the MEDV descriptors of70compounds. The PLSR models (M1: n=57, A=4,m=39, r=0.9202, RMSE=0.28, F=71.87, q=0.7956, RMSV=0.44, n_p=13, Rp=0.6649,RMSP=0.78) have a good estimation ability, high stability, and proper predictive power.The MEDV descriptors were directly and rapidly computed from two-dimensiontopological structure of a molecule by VSMP. The MEDV descriptors could berepresented as an interaction between two atomic types. Some important descriptors affecting BBB permeability were picked up by the VIP ordering analysis. The higher theVIP value of a variable is, the higher its contribution to the model is. It has been foundthat main structural factors influencing the BBB permeability of compounds are themolecular substructures,–CH3,–CH2–,=CH–,=C=,≡C–,–CH<,=C<,=N–,–NH–,=O, and–OH. Intuitive and simple QSAR model derived from the moleculartwo-dimensional structure is so an ideal QSAR model. Then MEDV descriptors mayalso be applied in the QSAR/QSPR studies on many complicated molecular systems toevaluate drug ADME parameters.Finally,103drugs whose6pharmacokinetics parameters (including humanintestinal absorption, oral bioavailability, plasma protein binding rate, volume ofdistribution, renal clearance and half-life) were entirely determined according to thereference are chose for systemic QSAR analysis. Six different quality QSAR models areobtained after that optimization descriptor set were chosen from a large number ofDRAGON and MEDV descriptors by the PLS variable selection methods. Afterrigorous validation, oral bioavailability models and half-life models failed to predictcorresponding parameters. And other four kinds of models can be used to reliablypredict drug molecules’ corresponding parameters in the defined application domainwith high quality, respectively. Drug four ADME activities parameters can be quickly,systematic forecasted by PLSR models. DRAGON descriptors and MEDV descriptorscould be effectively applied in drug molecular structure characterization and biologicalactivity QSAR studies.As a fast and effective variable screening method, PLS variable selection method isavailable in QSAR research. Most constructed QSAR models in drug ADME activityhave good fitting ability, robustness and high predictive ability. Then those models aresuitable for systematic reviews in drug ADME activity instead of animal experiments.To some extent, those models, ultimately, can support far more utility data to guideclinical medication and shorten the drug development cycle in drug design.