Neural network modeling for estimation of partition coefficient based on atom-type electrotopological state indices

Abstract

A method for predicting log P values for a diverse set of 1870 organic molecules has been developed based on atom-type electrotopological-state (E-state) indices and neural network modeling. An extended set of E-state indices, which included specific indices with a more detailed description of amino, carbonyl, and hydroxy groups, was used in the current study. For the training set of 1754 molecules the squared correlation coefficient and root-mean-squared error were r2 = 0.90 and RMS(LOO) = 0.46, respectively. Structural parameters which included molecular weight and 38 atom-type E-state indices were used as the inputs in 39-5-1 artificial neural networks. The results from multilinear regression analysis were r2 = 0.87 and RMS(LOO) = 0.55, respectively. For a test set of 35 nucleosides, 12 nucleoside bases, 19 drug compounds, and 50 general organic compounds (n = 116) not included in the training set, a predictive r2 = 0.94 and RMS = 0.41 were calculated by artificial neural networks. The results for the same set by multilinear regression were r2 = 0.86 and RMS = 0.72. The improved prediction ability of artificial neural networks can be attributed to the nonlinear properties of this method that allowed the detection of high-order relationships between E-state indices and the n-octanol/water partition coefficient. The present approach was found to be an accurate and fast method that can be used for the reliable estimation of log P values for even the most complex structures.