Y - Randomization – A Useful Tool in QSAR Validation , or Folklore ?

Several variants of randomization procedures were compared as a tool in validation of multilinear regression (MLR) QSAR equations that are obtained by descriptor selection. Y-randomization, a method formerly said to be probably the most powerful validation procedure, was found to be overoptimistic. The statistical significance of a new MLR QSAR model should be checked by comparing its measure of fit to the average measure of fit of best random pseudomodels that are obtained using random pseudodescriptors instead of the original descriptors and applying descriptor selection as in building the original model. Application of this criterion to several recently published MLR QSAR equations identified dubious ones. Some progress also is reported towards the goal of obtaining the mean best r of random pseudomodels by calculation rather than by tedious multiple simulations on random number variables.