Assessing the reliability of methods for predicting drug metabolites

An ability to predict the metabolic fate of a drug is important to drug design. Programs for predicting drug metabolites are becoming available, as are databases that will facilitate the development of such programs. Objective analysis of the performance of these programs will require statistical methods. The development of appropriate statistical methods must address a fundamental data representation problem that arises from the fact that the basic metabolic information uses chemical graph representations rather than the usual vector representations that typify statistical methodology. This study addresses the representation problem by using concepts arising from molecular similarity analysis. The statistical methods that are developed are illustrated by evaluating the performance of MetabolExpert in predicting the metabolic fate of benzodiazepines.