Use of Molecular Similarity Indices for QSAR Training Set Selection
Ian T. Cousins, M Cronin, John C. Dearden, C. D. Watts · SAR and QSAR in environmental research · 1995
The objective of the work was to assess the utility of an objective measure of “similarity” between chemicals for assisting in the choice of appropriate chemicals for training sets used to construct QSARs for aquatic toxicity. Measures of molecular similarity were assessed using regression analysis and hierarchical cluster analysis. A group of 18 alcohols, all non-polar narcotics, were selected for the assessment and derivation of the basic QSAR. A good correlation was shown to exist for the selected chemicals between log P and toxicity. However, this correlation was significantly reduced when two chemicals which exhibit toxicity by a different mode of action were introduced into the training set. Measurement of molecular similarity was investigated as an objective means of identifying the two outlier chemicals. Similarity indices were calculated and autocorrelograms plotted for the selected alcohols and outlier chemicals using commercially available software. Statistical analysis of the similarity data was carried out to determine if the outlier chemicals could be identified. One of the two outlier chemicals proved to be clearly “dissimilar” to all of the alcohols but, the other outlier chemical proved to be “similar” to several alcohols.