QSPR methods are often used to estimate the physicochemical properties of organic compounds and to predict their behavior in the environment. QSPR models were developed for the prediction of octanol/organic carbone partition coefficient (Koc) of an heterogeneous set of pesticides. The approaches based on multilinear regression (MLR), artificial neural networks (ANN), every time associated with genetic algorithm (GA) selection of the most important variables, lead to models of very different qualities. The modeling of octanol/organic carbone partition coefficient of a heterogeneous mixture of pesticides show that the various statistics for the sets of training and validation (multiple coefficients of determination and prediction; roots of squared errors averages) attest to the superiority of non-linear models (ANN) and their relevance.