|
Internet Electronic Journal of Molecular Design - IEJMD, ISSN 1538-6414, CODEN IEJMAT
| ABSTRACT - Internet Electron. J. Mol. Des. December 2004, Volume 3, Number 12, 802-821 |
Support Vector Machines Prediction of the Mechanism of
Toxic Action from Hydrophobicity and Experimental Toxicity
Against Pimephales promelas and Tetrahymena pyriformis
Ovidiu Ivanciuc
Internet Electron. J. Mol. Des. 2004, 3, 802-821
|
Abstract:
The prediction of the mechanism of action (MOA) using
structural descriptors has major applications in selecting the
appropriate quantitative structure-activity relationships
(QSAR) model, to identify chemicals with similar toxicity
mechanism, and in extrapolating toxic effects between different
species and exposure regimes. The SVM (support vector
machines) algorithm was recently proposed as an efficient and
flexible classification method for various bioinformatics and
cheminformatics applications. In this study we have
investigated the application of SVM for the classification of
337 organic compounds from eight MOA classes (nonpolar
narcosis, polar narcosis, ester narcosis, amine narcosis, weak
acid respiratory uncoupling, electrophilicity,
proelectrophilicity, and nucleophilicity). The MOA
classification was based on three indices, namely: log
Kow, the octanol-water partition
coefficient; log 1/IGC50, the 50% inhibitory
growth concentration against Tetrahymena pyriformis;
log 1/LC50, the 50% lethal concentration
against Pimephales promelas. The prediction power of
each SVM model was evaluated with a leave-5%-out cross-validation
procedure. In order to find classification models
with good predictive power, we have investigated a large
number of SVM models obtained with the dot, polynomial,
radial basis function, neural, and anova kernels. The MOA
classification performances of SVM models depend strongly
on the kernel type and various parameters that control the
kernel shape. The discrimination between nonpolar narcotic
compounds and the other chemicals can be obtained with radial
and anova SVM models, with a prediction accuracy of 0.80.
The separation of less reactive compounds (polar, ester, and
amine narcotics) from more reactive compounds (electrophiles,
proelectrophiles, and nucleophiles) is obtained with a slightly
higher error (prediction accuracy 0.71, obtained with radial
SVM models). SVM models that use as input parameters
hydrophobicity and experimental toxicity against
Pimephales promelas and Tetrahymena pyriformis
represent an effective MOA classification
method for a large diversity of organic compounds. This
approach can be used to predict the aquatic toxicity mechanism
and to select the appropriate QSAR model for new chemical
compounds.
|