Pergamon 0045 -6535(94)00265 -7 Chemosphere, Vol. 29, No. 8, pp. 1611-1626, 1994 Copyright © 1994 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0045-6535/94 $7.00+0.00 ESTIMATION OF SOIL PARTITION COEFFICIENTS OF PESTICIDES FROM THEIR CHEMICAL STRUCTURE H. Lohninger Institute of General Chemistry Technical University Vienna Lehargasse 4/152, A-1060 Vienna, Austria email: hlohning@email.tuwien.ac.at (Received in Germany 16 May 1994; accepted 26 July 1994) The aim of this work was to investigate the estimation of soil partition coefficients (normalized to the organic carbon content) of pesticides solely from their chemical structure. Both multiple linear regression and neural networks have been applied to build a model from topological indices and structural fragments. The model has been derived from a data set of 120 pesticides and is based on eleven structural parameters. The application of the model to a test set of B1 other pesticides showed a performance which is comparable to conventional models based on water solubility. Keywords: soil sorption coefficients, pesticides, multiple linear regression, neural networks, RBF-networks, feature selection, QSPR INTRODUCTION Environmental protection has become a major issue in agriculture during the past decade. Especially the potential pollution of ground water by agrochemicals has drawn much attention. The pollution potential is mainly based on the tendency of a chemical to undergo leaching or run-off when applied to the soil. The leaching in turn can be described by some fundamental physical and chemical properties, such as water solubility, acidity, soil sorption coefficient, lipophilicity, or polarity of a chemical substance. It should be noted that the sorption behavior (expressed as sorption coefficient) is quite different for nonpolar, polar, and ionic compounds. The sorption mechanism is based on a wide variety of inter- molecular interactions, which are described in more detail by von Oepen et al. (1991). Thus, the distribution coefficient K d in the soil and its normalized counterpart, the soil sorption coefficient I~,c (Hamaker et al. 1972) play a major role in describing the pollution potential of pesticides. In general, there are two different approaches in estimating soil sorption coefficients: First, several authors used chemical and physical parameters in a Hansch-related approach to model soil sorption coefficients (Hansch et al. 1979, Zhou 1993, Briggs 1981, Pussemier et al. 1989, Bintein 1994). The main parameters for these models are hydrophobicity parameters, the water solubility, electronic parameters of the molecule, and stereoeffect parameters. One should note, that hydrophobicity parameters, as obtained by octanol/water partition data, are only valid for non-polar compounds. As polarity increases the sorption estimate becomes less reliable. 1611