|
(466) Computer-Assisted Toxicity Assessment Using Self Organized Maps. Espinosa, Gabriela1, Arenas, Alex1, Giralt, Francesc1, 1 ABSTRACT- The design of new drugs based on the forecast of activity from molecular structure information only is a major challenge in pharmacology and chemistry. The traditional methods that incorporate synthesis in the design process are usually laborious and expensive. In the other hand, computed assisted methods are time consuming and very sensitive to the selection of molecular input information. The main problem arises when trying to decide the molecular descriptors that should be used to build the computational model, i.e. the Quantitative Structure Activity Relationship (QSAR). Statistical pre-screening techniques are commonly used to select these descriptors. In the current study a neural network method is proposed to self-organize (cluster) compounds by activity with respect to carcinogenicity and toxicity, and to extract useful knowledge to better establish the relationships between molecular structure and activity. In particular, two neural network based QSARs are presented: One for the carcinogenicity of an homogeneous set TD50 of 104 nitrogenated aromatic compounds and a second for the toxicity of an heterogeneous set LD50 of 150 diverse organic compounds. A self-organized neural network has been used to classify the compounds by structural features and activity parameters. Each of the clusters obtained, labeled according to the prototypical activity of the represented compounds, has been used to extract knowledge about the relationship and significance of each molecular descriptor. This information is important to select the minimal set of descriptors needed to build a robust QSAR model. Key words: Neural Networks, , Self organized maps, , QSAR |
