|
1F - QSAR (TU13/5) Counter Propagation Neural Network as a Tool in QSAR Modelling. Vracko, Marjan1, Mazzatorta, Paolo1, 2, Novic, Marjana1, Roncaglioni, Alessandra 1, 2, Basak, Subhash3, Mills, Denise 3, 1 National Institute of Chemistry, Ljubljana, Slovenia2 Institute Mario Negri, Milan, Italy3 Natural Resources Research Institute, Duluth, USA ABSTRACT- We present the architecture and learning strategy of Counter Propagation Artificial Neural Network (CP ANN), which is a generalization of Kohonen ANN (or Self Organizing Map). CP ANN a suitable tool for clustering, classification and modeling, particularly when the data set is non-homogeneous. Further applications of CP ANN include training-test set division, outliers determination and descriptor selection. We discuss the models built with different sets: 42 benzenes considering the toxicity toward Ciliata Tetrahymena pyriformis; 225 phenols considering the same toxicity; 95 aromatic amines considering mutagenicity in Salmonella typhimurium TA98 and TA100; 562 compound considering the aquatic toxicity toward a fish fathead minnow; considering the classification of compounds, which are maybe active as endocrine disrupters. Descriptors for treated compounds were calculated with different software, considering constitutional, 2D and 3D description of molecular structures. The models were tested with leave-one-out cross validation method or with the test set, which was selected from the data set. In addition, we applied two randomization tests. First, we randomize the property values, and second, we add two random variables to the descriptor sets. [This work was supported by the European Union IMAGETOX Research Training Network (HPRN-CT-1999-00015), by Marie Curie Host Fellowship (HPMT-CT-2001-00240), and by Ministry of Education, Sport and Science of R Slovenia (P1-034507)]. Key words: Leave-one-out and randomisation tests, Counter propagation neural network, aquatic toxicity, mutagenicity, classification and endocrine disruption |
