Quantitative structure-activity relationship model validation

In quantitative structure-activity relationship (QSAR) modeling, it is of utmost importance to validate the resulting model. Thus one has to consider simultaneously the model's fit and predictive ability and assess the optimal balance between these. Several tools are available for model validation. These are reviewed here in the context of QSAR. The most demanding model validation principle is (1) external validation, which consists of making predictions for an independent set of data not used in the model calibration. Ideally, such a prediction set should be selected according to a statistical experimental design. External validation, however, is usually not tractable with QSAR, notably because of resources needed for making a test set of new compounds. Hence, alternatives for QSAR validation are of interest. The additional model validation principles discussed here are (2) cross-validation, (3) cross-validation together with permutation testing, and (4) plots of the model, the underlying data, and the residuals. The validation principles (2) to (4) are illustrated by means of 4 QSAR datasets from environmental toxicology and chemistry. For the data analysis, the multivariate partial least squares (PLS) projection to latent structure technique is used.