VRI: 3D QSAR at variable resolution

Abstract

VRI (Variable Resolution Invariants) is a new approach to quantitative structure–activity relations that makes use of three-dimensional features of molecules at different levels of spatial resolution as well as levels of resolution in atomic properties. These descriptors are independent of any numbering of the atoms of a molecule. They are also independent of rigid translation and rotation of a given conformer, which avoids problems with aligning different molecules or docking them with a receptor site model. Steric effects, stereospecificity, substituent effects, lipophilicity, and conformational flexibility are all dealt with in a single, natural formalism. Simple datasets can be fitted using a small number of descriptors corresponding to low-resolution descriptions of the molecules. More complicated data can require additional descriptors that recognize finer details of three-dimensional structure and physico-chemical properties. Overfitting due to the large number of descriptors is handled by partial least-squares analysis with crossvalidation. Performance in fitting and predicting is demonstrated on some simple illustrative cases, and on three standard sets of real data: steroids binding to human corticosteroid binding globulin and testosterone binding globulin, and inhibitors of dihydrofolate reductase. ©1999 John Wiley & Sons, Inc. J Comput Chem 20: 1577–1585, 1999