A Simple Inter-Class Distance Parameter for Predictive SAR/QSAR Models

Abstract

The objective of many SAR or QSAR experiments is to develop statistical classi®cation models that can separate chemicals into classes of``active'' or``inactive'' under a toxicological endpoint. Several existing statistical methods are often used to provide either quantitative or qualitative measures of the distance or separation between the active and inactive classes, including sensitivity, accuracy and the receiver-operator characteristic (ROC) curve. The present study proposes a simple``distance parameter'' as a measure of the relative separation between classes of``active'' and`i nactive'' chemicals. The distance parameter can be used alone or with existing statistical measures for both the optimization of a single model with respect to maximizing sensitivity and speci®city, and also in the selection of the best model from among alternative models.