|
|
|
|
|
|
|
|
Alexander Tropsha was born in Moscow in 1960. He received his MS in Chemistry from Moscow State University in 1982 and PhD in Biochemistry and Pharmacology in 1986 from the same university. He immigrated to the US in 1989. In 1991, after two years of postdoctoral research at the University of North Carolina at Chapel Hill, he joined the UNC School of Pharmacy as an Assistant Professor and Director of the Laboratory for Molecular Modeling. Dr. Tropsha has since been promoted to the position of full Professor; he also holds position of the Associate Director of the Carolina Center for Genome Sciences.
The major area of Tropsha’s research is Biomolecular Informatics, which implies understanding relationships between structures (organic or macromolecular) and their properties (activity or function). In recent years, his group has developed several important methodologies and software tools for Computer Assisted Drug Design. Concurrently, they have developed a new approach to protein 3D structure analysis and prediction based on the principles of statistical geometry (Delaunay tessellation). This approach affords determination of key structural and sequence motifs responsible for protein function.
|
|
The statistical significance vs. mechanistic interpretation of ADME/tox models
Alex Tropsha, UNC
Alexander Tropsha (1), Ann Richard (2), Kun Wang(1), Maritja Wolf (2), Clarlynda Williams (2), Jamie Burch (2)
(1) Laboratory for Molecular Modeling, School of Pharmacy, UNC-Chapel Hill, Chapel Hill, NC 27599
(2) Mail Drop D343-03, National Center for Computational Toxicology (NCCT), Office of Research & Development, US Environmental Protection Agency, Research Triangle Park, NC 27711
Several major trends affecting public toxicity information resources have the potential to significantly alter the future of predictive toxicology. Standardized chemical structure annotation of toxicity databases and integration of diverse biological activity data afford a mine-able chemical semantic Internet. Formalized toxicity data models and public toxicity data schemas allow for flexible data mining and relational data searching across layers of chemical and biological information. Curated, systematically organized, and web-accessible toxicity and biological activity data in association with chemical structures is clearly the next frontier of advancement for QSPR and data mining technologies. The examples of such systems are provided by the DSSTox database and affiliated projects such as the Carcinogenic Potency Database (CPDB), PubChem, Leadscope ToxML, and the National Toxicology Program.
The importance of the combined toxico-cheminformatics and QSPR modeling is illustrated by the analysis of the CPDB. It contains TD50 data resulting from animal cancer tests of over 1000 chemicals as well as the information on species, strain, sex, shape of the dose-response, etc. Such extensively structured and mine-able database provides unique opportunities for developing QSPR models for subsets of compounds selected on the basis of biologically meaningful parameters. Rigorous QSAR analysis was applied to a subset of 693 compounds tested for their mutagenicity. 70 compounds were selected randomly for external validation. The remaining 623 compounds were split into diverse training set and test sets. Classification kNN QSPR approach afforded modes with the prediction accuracy for training, test and external validation sets as high as 0.917, 0.847, and 0.893, respectively. The analysis of chemical descriptors that afforded statistically significant and predictive QSAR models allowed for the mechanistic interpretation of such models in terms of chemical properties important for mutagenicity. We stress that only those QSPR models that have been rigorously validated both internally and externally should be considered for the mechanistic interpretation.
|
|
|
|
|
|
|
|
|
|
|
