A Soft Computing Approach for Toxicity Prediction

This paper describes a hybrid method for supervised training of multivariate regression systems that can be an alternative to other methods. The proposed methodology relies on supervised clustering with genetic algorithms and local learning. G enetic A lgorithm d riven C lustering (GAdC) offers certain advantages related to robustness, generalization performance, feature selection, explanative behavior and the additional flexibility of defining the error function and the regularization constraints. In this contribution we present the use of GAdC for toxicity prediction of pesticides. Different molecular descriptors are computed and the correlation behavior of the different descriptors in the descriptor space is studied. Decreasing the number of descriptors leads to a faster and more accurate model.