New QSAR methods applied to structure-activity mapping and combinatorial chemistry

A comparison is made of a number of computationally efficient molecular ind ices with a view to the screening of very large virtual data sets of molecu les. The use of Bayesian regularized neural networks is discussed, and thei r virtue in eliminating the need for validation sets, and potentially even test sets, is emphasized. The concept: of a virtual receptor is introduced, and this is illustrated by the results of screening a database of 40 000 m olecules.