Genetic algorithm driven clustering for toxicity prediction

The pace of technological advancement in today's society has generated an enormous demand for methods facilitating the intelligent testing of the toxicity of new chemicals. Until now it was common use to make predictions based on 'real' tests. Recent investigations support the general assumption that macroscopic properties like toxicity and ecotoxicity strongly depend on microscopic features and the structure of the molecule. The authors have developed a computationally intelligent method for supervised training of regression systems. Their method selects those features needed to predict and calculate the toxicity. The proposed methodology relies on supervised clustering with genetic algorithms and local learning. Different molecular descriptors are computed and the correlation behaviour of the different descriptors in the descriptor space is studied.