Dynamic QSAR: least squares fits with multiple predictors

Accounting for the multiplicity of conformers taking part in interactions carried out in complex reaction environments, the recently proposed dynamic QSAR method [O.G. Mekenyan, J.M. Ivanov, G.D. Veith, S.P. Bradbury, Quant. Structureactivity Relation. 13 (1994) 302–307] requires the least squares fit to be applied on a multiple predictor data sets. A parametric and model assessment of the least squares approach is proposed in case of such different data structure. The correlation between the experimental and calculated values is determined by three terms: the experimental error if multiple observations are taken into account, within group deviations if multiple predictors are taken into account and lack of fit between experimental and calculated means. To evaluate what a current regression model does accomplish with respect to those three terms, relative correlation coefficients are introduced. The approach and new statistical estimates are tested on simulated and real data sets.