Predicting toxicity against the fathead Minnow by Adaptive Fuzzy Partition

Recent progress in the development of powerful tools suitable to design and to classify large chemical libraries can be fruitfully extended also to the ecotoxicity domain. Amidst these methods, Fuzzy Logic concepts, based on the possibility to handle the concept of partial truth, constitute interesting approaches to developing general predictive models. In this work, a global strategy of Database Mining was applied on a data set of 568 chemicals, extracted from a toxicity database concerning the fathead minnow and divided into four classes, according to the toxicity ranges defined by the European Community legislation. Two large sets of molecular descriptors were tested on the 2D and 3D structures, and the best ones were selected with help of a procedure combining Genetic Algorithm concepts and stepwise method. After selecting the training set with a rational selection based on the Self Organizing Maps (SOM), structural-activity models were built by Adaptive Fuzzy Partition (AFP). This method consists in modeling relations between molecular descriptors and biological activities, by dynamically dividing the molecular descriptor hyperspace into a set of fuzzy subspaces. The best model was selected by a validation set, and its robustness was confirmed by predicting a test set of 80 chemicals never used to define the AFP models. An encouraging validation ratio of about 72% was obtained in the prediction of the experimental toxicity class. Furthermore, very similar results were obtained by using molecular descriptors computed on 2D or 3D structures.