Development of a Model to Predict Partition Coefficient of Organic Pollutants in Cloud Point Extraction Process

Abstract

A quantitative structure property relationship (QSPR) study has been performed to establish a model to relate structural descriptors of 45 organic compounds to their partition coefficients in water-hexadecylpyridinium chloride (CPC) micelles at 298K using partial least squares (PLS). 510 of six different categories of structural descriptors were calculated by Dragon software. The descriptors with 0.9 mutually pair correlations and with less than 0.1 with dependent variables were excluded at the early stage of the preprocessing of the structural data matrix. The data set was randomly divided into two groups: training set (40 molecules) and test set (5 molecules). In the final model 50 of the most effective of the structural descriptors on the partition coefficient were remained to model building by PLS calibration method. The optimum number of latent variables 5, which spanned 80% of the original variations of data matrix, was selected using leave one out cross validation method. Prediction ability of the model was tested by prediction of the partition coefficients of five unknown compounds and the mean relative error of prediction was 3.6%. The outliers were treated using leverage and score plots of the first third principal components. The efficiency of the new model was compared with Abraham model and it was found that the proposed model has more prediction ability.