Prediction of the soil organic carbon to water distribution coefficients parameter for some neutral organic compounds from the theoretical derived molecular descriptors and QSPR modeling

The term of sorption is used frequently in environmental studies to denote uptake of a solute by a solid without referring to specific mechanism. Sorption play an important role in determining environmental fate and impact of organic chemicals. Soil organic carbon to water distribution coefficient, quantitatively describe extent to which an organic chemical is distributed between an environmental solid and an aqueous phase. The soil organic carbon to water distribution coefficients of 57 neutral organic compounds consist of aliphatic and aromatic hydrocarbons, heterocyclic compounds and benzene derivatives were taken from Ref. [1]. The structures of molecules were drawn with HyperChem 4.5 package [2] and exported in a file format suitable for MOPAC 6.0 program [3] for geometry optimization semi-mpirical quantum method AM1. Then, theoretical molecular descriptors were derived for prediction of soil organic carbon to water distribution coefficients of compounds calculated by Dragon 6.0 package. For QSPR modeling, an ANN program was written in FORTRAN 77 in our laboratory. This network was feed-forward fully connected that has three layers with sigmoid transfer function. Descriptors appearing in multiple linear regression (MLR) model were used as inputs of network and signal of output node represent distribution coefficient of interested compound. Thus, this network has five nodes in input layer and one node in output layer. The value of each input was divided into its maximum value to bring them into dynamic range of sigmoid transfer function of network. The initial values of weights were randomly selected from a uniform distribution that ranged between -0.3 and +0.3. The initial values of biases were set to be one. These values were optimized during network training. Before training, network’s parameters would be optimized. These parameters are; number of nodes in hidden layer, weights and biases learning rates, and momentum values. Then optimized network was trained using 37 compounds introduced as training set for adjustment of weights and biases values. For evaluate prediction power of network during its training, after each 1000 training cycle, network was used to calculate distribution coefficient of molecules included in test set contains 10 compounds. The standard errors of obtained ANN model are 0.16, 0.19 and 0.27 for training, test and calibration sets respectively. Comparison between statistical results calculated from MLR and ANN models reveals that all statistics have been improved considerably in case of ANN model.