NEURAL NETWORKS PREDICTION OF PARTITION-COEFFICIENTS

A comparison has been made of the relative ability of neural networks and regression analysis to estimate the octanol-water partition coefficients (log P) of organic compounds, with a recently proposed model based on geometric and semi-empirical AMI parameters. The predictive power of the model is estimated by cross-validation. The regression analysis requires the use of a 17-parameter model including higher powers of a 8-parameter model, giving a standard error of prediction (SEP) of 0.371. The neural network approach gives comparable prediction (SEP = 0.379), even with the reduced 8-parameter model. With a 13-parameter model obtained by the addition of five new independent parameters, neural networks are found to give still better results (SEP = 0.300). The variability of the prediction made by neural networks has been related to the leverage of the compounds in the descriptor space of the multilinear model.