Causality in structure—activity studies
Gilles Klopman, Alexander N. Kalos · Journal of Computational Chemistry · 1985
Abstract The problem of chance correlations in studies of structure—activity relationships using multivariate linear regression techniques is addressed. Random correlations are simulated by scrambling real potency values in data bases consisting of molecules possessing various biological properties. The statistical parameters of these correlations are compared to those obtained using the actual potency/compound data pairs and it is noted that the most noticeable differences are in the magnitude of the F values for the overall regression. Cutoff points for this parameter are suggested, which, when implemented in the recently introduced CASE method that correlates biological activity with substructural descriptors, can serve to weed out effectively the majority of meaningless correlations.