A QSAR for baseline toxicity: Validation, domain of application, and prediction
Öberg, T.
Chemical Research in Toxicology 17, 1630-1637 (2004)

Abstract
The interest in modeling and application of structure-activity relationships has steadily increased in recent decades. It is generally acknowledged that these empirical relationships are valid only within the same domain for which they were developed. However, model validation is sometimes neglected, and the application domain is not always well-defined. The purpose of this paper is to outline how validation and domain definition can facilitate the modeling and prediction of baseline toxicity for a large database. A large number of theoretical descriptors (867) were generated from two-dimensional molecular structures for compounds present in the U.S. EPA's Fathead Minnow Database (611) and the Syracuse Research Corporation's PhysProp Database (25 000+). A quantitative structure-activity relationship model was developed for baseline toxicity (narcosis) toward the fathead minnow (Pimephales promelas) using a projection-based regression technique, PLSR (partial least squares regression). The PLSR model was subsequently validated with an external test set. The main factors of variation were related to size/shape and polar interactions. The prediction error was comparable to, or slightly better than, the ECOSAR procedures. A set of 16 805 compounds, drawn from the PhysProp Database, was projected onto the PLSR model. More than 90% (15 597) of the compounds fall within the valid model domain, defined by the residual standard deviation and the leverage. The predicted baseline toxicity indicates an acute hazard for two-thirds of these compounds, classes I-III in the OECD Globally Harmonized Classification System (LC₅₀ < 100 mg L^-1). Finally, the mode of action assigned in the U.S. EPA Fathead Minnow Database was investigated. Reclassification to narcosis as the mode of action is suggested for 92 compounds, mostly from the groups "unsure" and "mixed". The present classification into specific modes of action seems to be further strengthened by the findings in this investigation.

DOI: 10.1021/tx0498253

Full text as a PDF-file from ACS