The Study Of Relationships Between Drug Clearance And Their Structure Parameters By Neural Network

Not only the pharmadynamic factors but also the pharmacokinetic ones should be considered in the rational drug design. Previously, the researchers took much attention on the former and the latter was overlooked. In this paper, a correlative analysis is given between the pharmacokinetic parameter-total drug clearance and the drug structure parameters by use of neural network. One hundred drugs are selected and their clearances have been determined. We calculate the physical and chemical parameters of those drugs: volume of molecule (V), mole weight of molecule (M), refractive index of molecule (MR), parameter of distant water (logP), heat of formulation (HF), polarizational rate (a ), total energy (Ex), and quantum chemical parameters: EHOMO and ELUMO, and analysissitus parameters: 0~6 connective indexes, the total parameters are eighteen. Firstly, a correlative analysis is given for all swatchs and three parameters are selected. Secondly, the three parameters are given an analysis of Partial Least Squares and three main components are singled out. Secondly, two neural network models including two hidden layers are set up by 3 structural parameters and by 3 main components. As comparation, the one including 18parameters is set up too. The network's structural parameters are described as: the maximal cycles are 3,000, the nodes of the first-hidden layer are 18 and the ones of second-hidden layer are 34, original learning rate is 0.1 and its increasing ratio is 1.05 and its decreasing ratio is 0.7, the momentum factor is 0.9.By this network, 90 swatchs act as the training gather and the 10 ones are selected stochasticly as the predicting gather for validating the our network. The result of final calculation shows: for those drugs owning lower total drug clearance (CL<1.0ml/min/kg), the correct ratio of being learned by the first network is 26.3%, the correct ratio of being predicted is 50%. For the main components network, the two numbers are 15.8% and 0%. For the third network, they are 5.3% and 0%. On the other hand, for those drugs of higher total drug clearance (CL>1.0ml/min/kg), the correct ratio of being learned by the first network are 80.3%, the correct ratio of being predicted is 87.5%. For the main components network, the two numbers are 76.1% and 87.5%. For the third network, the two numbers are 54.9% and 12.5%. From the results, in the third network, both learning correct ratio and predicting correct ratio are most discontent, and we don't find the clear distinction exist between the first and second models. From the learning and predicting results, the first network is beyond the second one slightly. Perhaps there are more complicated nonlinear relationships between the total drug clearance and the structure parameters of drug molecules. The main components of being calculated by linear Partial Least Squares Method maybe don't embody their effects properly in the model. All results show the model between the total drug clearance and the structure parameters of drug molecules is reasonable and neural network will become an effective instrument in quantitative structure pharmacokinetic relationships.