QSARs for chemical mutagens from structure: ridge regression fitting and diagnostics.

QSAR models have been developed for a diverse set of mutagens using computed molecular descriptors. Such models can be used in predicting mutagenicity from structure. All common methods-regression, neural nets, k-nearest neighbors-are 'linear smoothers'-weighted averages of the activities in the calibration data with weights dependent on the descriptors. While they have been studied extensively, a vital but overlooked area is 'case diagnostics', pointers to compounds that are poorly fitted, or are unusually influential in fitting the model. This is particularly true where the measured activity is binary-present or absent. We illustrate the use of numeric and graphic diagnostics, particularly that of the FF plot, with a data set with 508 compounds and 307 structural descriptors used to predict mutagenicity.