This paper emphasizes the importance of rigorous validation as a crucial, integral component of Quantitative Structure Property Relationship (QSPR) model development. We consider some examples of published QSPR models, which in spite of their high fitted accuracy for the training sets and apparent mechanistic appeal, fail rigorous validation tests, and, thus, may lack practical utility as reliable screening tools. We present a set of simple guidelines for developing validated and predictive QSPR models. To this end, we discuss several validation strategies including (1) randomization of the modelled property, also called Y-scrambling, (2) multiple leave-many-out crossvalidations, and (3) external validation using rational division of a dataset into training and test sets. We also highlight the need to establish the domain of model applicability in the chemical space to flag molecules for which predictions may be unreliable, and discuss some algorithms that can be used for this purpose. We advocate the broad use of these guidelines in the development of predictive QSPR models.

No abstract

Computational ADME (absorption, distribution, metabolism, and excretion) models may be used early in the drug discovery process in order to flag drug candidates with potentially problematic ADME profiles. We report the development, validation, and application of quantitative structure-property relationship (QSPR) models of metabolic turnover rate for compounds in human S9 homogenate. Biological data were obtained from uniform bioassays of 631 diverse chemicals proprietary to GlaxoSmithKline (GSK). The models were built with topological molecular descriptors such as molecular connectivity indices or atom pairs using the k-nearest neighbor variable selection optimization method developed at the University of North Carolina (Zheng, W.; Tropsha, A. A novel variable selection QSAR approach based on the k-nearest neighbor principle. J. Chem. Inf. Comput. Sci., 2000, 40, 185-194.). For the purpose of validation, the whole data set was divided into training and test sets. The training set QSPR models were characterized by high internal accuracy with leave-one-out cross-validated R(2) (q(2)) values ranging between 0.5 and 0.6. The test set compounds were correctly classified as stable or unstable in S9 assay with an accuracy above 85%. These models were additionally validated by in silico metabolic stability screening of 107 new chemicals under development in several drug discovery programs at GSK. One representative model generated with MolConnZ descriptors predicted 40 compounds to be metabolically stable (turnover rate less than 25%), and 33 of them were indeed found to be stable experimentally. This success (83% concordance) in correctly picking chemicals that are metabolically stable in the human S9 homogenate spells a rapid, computational screen for generating components of the ADME profile in a drug discovery process.

A model, VLOGP, has been developed for assessment of n-octanol/water partition coefficient, log P, of chemicals from their structures. Unlike group contribution methods, VLOGP is based on linear free energy relationship (LFER) approach and employs information-rich electrotopological structure quantifiers derived solely from molecular topology. VLOGP, a robust and cross-validated model derived from accurately measured experimental log P values of 6675 diverse chemicals, has a coefficient of determination, R2, of 0.986 and a standard error of estimate of 0.20. When applied to the training set, the largest deviation observed between experimental and calculated log P was 0.42. VLOGP is different from other log P predictors in that its application domain, called Optimum Prediction Space (OPS), has been quantitatively defined, i.e., structures to which the model should not be applied for predicting log P can be identified. A computer-assisted implementation of this model within HDi's toxicity assessment software package, TOPKAT 3.0, automatically checks whether the submitted structure is inside the OPS or not. VLOGP was applied to a set of 113 chemicals not included in the training set. It was observed that for the structures inside the OPS the average deviation between experimental and model-calculated log P values is 0.27, whereas the corresponding deviation for structures outside the OPS is 1.35. This demonstrates the necessity of identifying the structures to which a model is not applicable before accepting a model-based predicted log P value. For a set of 47 nucleosides, the performance of VLOGP was compared with that of four published log P predictors; a standard deviation of 0.33 was obtained with VLOGP, whereas the standard deviation from other log P predictors ranged between 0.46 and 1.20.