| Selectivity refers to drug has specific affinity to one target comparing to the others, and compound with high selectivity is generally considered to be of higher safety and less side effects, which can reduce economic lost caused by clinical stage development failure through determining the selectivity of candidate compounds in drug discovery and development early phase. However, drug selectivity experiments spend huge time and cost simultaneously experimental measurement is not feasible for the theory design molecule that has not been synthesized, therefore, there is of great significance to examining theory drug selectivity prediction.Selectivity forecasting methods can be divided into receptor structure-based methods, such as molecule docking and free energy calculation which protein structure is needed, and ligand-based methods most cases need superimposing molecules, such as QSAR and pharmacophore, both cannot calculate or filter batch data automatically. BRS-3D is a multidimensional descriptor for three dimensional molecule structure characterization proposed by our laboratory, and the purpose of this paper is to verify its possibility in predicting selectivity of enormous compounds. The molecule structure and activity data for dopamine receptor(DR) subtypes and serotonin receptor(5-HTR) subtypes is from Ch EMBL.The models for 5 DR subtypes:1) 5 DR subtypes activity fitting models built by SVM show that, q2 and r2 of BRS-3D is higher than 2D descriptor for activity fitting models of DR1, DR2, DR3 and DR4, for 5 fitting models based on BRS-3D,q2 mean = 0.512 as q2 highest = 0.588 and r2 mean = 0.549 as r2 highest = 0.572,meanwhile the 2D+3D descriptor is more effective as q2 highest and r2 highest equal to 0.618 and 0.636;2) 10 selectivity binary data combined by 5 subtypes was applied to bulid SVM model with relative binding ability as selectivity, the result indicates BRS-3D is better than 2D descriptor with q2 mean = 0.600 as r2 mean = 0.655, and q2 highest = 0.766 as r2 highest = 0.899, meanwhile the 2D+3D descriptor is more effective as q2 highest and r2 highest equal to 0.800 and 0.884.The models we built and results for 14 5-HTR subtypes:1) 12 BRS-3D based activity fitting SVM models have 9 models q2> 0.4 and 8 models r2> 0.5, as the best is q25-HTR1 D = 0.621 and r25-HTR1 D = 0.747;2) 17 selectivity binary data with more than 100 molecules combined by 14 subtypes were applied to build model with relative binding ability as selectivity, the result demonstrates SVM and k NN arithmetic has barely discrepancy, 17 SVM models has 5 q2 and r2 > 0.5 and the highest is q22A-2C = 0.744 and r22A-2C = 0.801;3) two-classifation model 1A-2A, 1A-2A and 2A-6 were built and the best prediction of test set accuracy is 99.346%, and prediction accuracy of the four-classifation model(1A, 2A, 6 and MDDR) achieves 91.329%, 84.188%, 92.193% and 93.646%;4) agonist-antagonist discrimination model was built by 34 agonists and 38 antagonists of 5-HTR2 C, ROC score = 1 and 12 molecules of 13 test set were predicted accuracy.The research results demonstrates BRS-3D can build activity or selectivity prediction model of GPCR receptor subtype ligands effectively, simultaneously indicate that BRS-3D is of molecule structure characterization ability, and the information carrying capacity is equal to traditional two-dimensional descriptors, meanwhile the BRS-3D and 2D descriptors combination can improve the prediction ability. Compared to 2D descriptor, BRS-3D is more advantageous in compound characteristics representation when structures are variously different, and can be used in scaffold jumping research. |
PDF Full Text Request