ABSTRACT
Here we consider a narrowly defined but still frequently occurring problem in drug design that one might refer to as the unknown receptor 3D QSAR problem. Given are the experimentally determined biological activity measurements for a set of compounds and their chemical structures. These measured activities for all the compounds, which are presumed to be the result of reversible binding to some common site on some macromolecule, such as a protein, are expressed in terms of AGbind or some roughly equivalent scale, such as log Kj, or log ICso. While the three-dimensional structures of these small ligand molecules may be estimated with good molecular mechanics potentials, the three-dimensional structure of the receptor site is unknown. (Otherwise, we are dealing with structure-based binding modeling, quite a different problem, addressed by quite different methods: Goodsell and Olson, 1990; Jones et al., 1997; Mankino and Kuntz, 1997; Sandak et al., 1998). The objective is to use this given information to construct some sort of correlation between ligand structure, including three-dimensional features, and the observed activities. This relationship should map the three-dimensional structure of a ligand into a quantitative estimate of activity that agrees well in some sense
132 Crippen and Wildman
with the observations for the training set of compounds, and it should produce useful predictions for test compounds. If the only ligand features employed in the relationship come from the covalent bonding of the ligand, the problem belongs to traditional QSAR (Martin, 1978; Hansch and Leo, 1995).
