You are here:
QUANTITATIVE STRUCTURE ACTIVITY RELATIONSHIP (QSAR) MODELS TO PREDICT CHEMICAL TOXICITY FOR VARIOUS HEALTH ENDPOINTS
Citation:
HARTEN, P. F., D. M. YOUNG, T. M. MARTIN, R. VENKATAPATHY, AND S. T. DAS. QUANTITATIVE STRUCTURE ACTIVITY RELATIONSHIP (QSAR) MODELS TO PREDICT CHEMICAL TOXICITY FOR VARIOUS HEALTH ENDPOINTS. Presented at Midwest Chapter of SETC & Chicago REgional Chapter of SRA, Argonne, IL, March 14 - 16, 2007.
Impact/Purpose:
To inform the public.
Description:
Although ranking schemes based on exposure and toxicity have been developed to aid in the prioritization of research funds for identifying chemicals of regulatory concern, there are significant gaps in the availability of experimental toxicity data for most health endpoints. Predictive toxicological approaches such as Quantitative Structure-Activity Relationships (QSARs) provide a means to estimate the toxicities for chemicals that lack experimental data. QSARs are equations which relate the toxicity of a chemical to its physicochemical properties such as bond lengths, E-state indices, surface area and partial charges. The objective of this study is to construct an efficient Java-based software tool based on unsupervised learning methods such as cluster analysis and genetic algorithm to evaluate the QSAR method for predicting the toxicity of chemicals. We have used the acute aquatic toxicity data for fathead minnow to validate our software tool. Upon validation, the software tool will be expanded to include other health endpoint such as the oral rat lethal dose and tumor dose of chemicals. The training set used for model development contained approximately 350 chemicals. A hierarchical cluster analysis (based on Ward’s method) is used to divide the dataset into clusters with similar fragment-, 2D- and 3D-based physicochemical descriptors. An optimal model developed using a genetic algorithm for each cluster that meets the domain applicability criteria is used to predict the toxicity of new chemicals. The average predicted toxicity from clusters in extrapolation domain for chemicals in a validation dataset was used for validation against their respective experimental values. The validation results indicate that present software may be used to predict toxicities of chemicals for other cancer and non-cancer endpoints.