Comparison of different classification methods applied to a mode of toxic action data set

Successful discrimination of compounds by mode of toxic action (MOA) is a prerequisite for process-based quantitative structure- activity relationship (QSAR) approaches. A data set of 115 compounds comprising nine MOA classes and 24 descriptors has been studied with several classification methods: multinomial logistic regression (multinom), linear discriminant analysis (LDA), partial least squares (PLS), and counter-propagation neural networks (CPG NN). Variables were selected with stepwise methods and with a genetic algorithm (GA) for the CPG NN. Five-fold cross-validation was used for validating the models and the advantages and disadvantages of this validation method are critically discussed. Without variable selection the predictive power of the models ranges between 51% and 53% cross-validated overall correct classification. With appropriate parameter selection the predictive power slightly increased to 52-59%. The experimental data showed that a number of compounds were active in more than one MOA. Multinom and CPG NN models for multiple MOAs were derived for both single and multiple MOA data. The consideration of multiple MOAs resulted in a slight increase in predictive power even when all MOAs were modeled at the same time.