Prediction of the fish acute toxicity from heterogeneous data coming from notification files

Four descriptors (Molecular weight, log(Pow), hardness and free energy of solvation) were selected to predict, on a training set of heterogeneous chemical compounds, the fish acute toxicity. The data were extracted from 523 notification files of new chemicals stored at the French Department of the Environment. The selection of the descriptors was carried out by using a statistical technique coupling OLS regression and genetic algorithm. The limits of validity for the final equation are discussed by comparing the actual and predicted activities on several compounds.