Quantitative Characterization of Molecular Similarity Spaces: Tools for Computational Toxicology (1998-1999)

Abstract

During the third year of the project, our work on the first three tasks of the project; viz., a) characterization of molecular similarity spaces, b) selection of analogs, and c) similarity-based estimation of properties; has continued. However, the focus of our work has shifted to the fourth and final task of the project, viz., the application of neural networks in property estimation. In the area of Task 1, the effectiveness of theoretical molecular descriptors vis-a- vis experimental physicochemical properties in quantifying intermolecular similarity has been explored for several sets of compounds with varying physicochemical and biological properties. In Task 2, the various structure spaces developed in Task 1 have been used in the selection of analogs for specific probe compounds. In Task 3, we have used the /(-nearest neighbor (KNN) method to estimate properties of chemicals from various databases. For these experiments, k has been varied from 1-40. The results showed that, for different physicochemical, toxicological and biochemical properties, optimal property estimation is generally obtained in the range of k= 5-10. Finally, in Task 4, we have used neural networks for the prediction of toxicological endpoints. In addition, we examined several methods for feature (independent variable) selection using a machine learning techniques, GEFS (genetic ensemble feature selection), based on genetic algorithms. The results show that neural networks, in general, show slight improvement in modeling power over statistical methods, but the use of GEFS to select relevant features for modeling greatly improves the performance of the neural networks.