QSAR applicabilty domain estimation by projection of the training set descriptor space: a review.

As the use of Quantitative Structure Activity Relationship (QSAR) models for chemical management increases, the reliability of the predictions from such models is a matter of growing concern. The OECD QSAR Validation Principles recommend that a model should be used within its applicability domain (AD). The Setubal Workshop report provided conceptual guidance on defining a (Q)SAR AD, but it is difficult to use directly. The practical application of the AD concept requires an operational definition that permits the design of an automatic (computerised), quantitative procedure to determine a models AD. An attempt is made to address this need, and methods and criteria for estimating AD through training set interpolation in descriptor space are reviewed. It is proposed that response space should be included in the training set representation. Thus, training set chemicals are points in n-dimensional descriptor space and m-dimensional model response space. Four major approaches for estimating interpolation regions in a multivariate space are reviewed and compared: range, distance, geometrical, and probability density distribution.

An approach to determining applicability domains for QSAR group contribution models: an analysis of SRC KOWWIN.

QSAR model predictions are most reliable if they come from the models applicability domain. The Setubal Workshop report provides a conceptual guidance for defining a (Q)SAR applicability domain. However, an operational definition is necessary for applying this guidance in practice. It should also permit the design of an automatic (computerised) procedure for determining a models applicability domain. This paper attempts to address this need for models that use a large number of descriptors (for example, group contribution-based models). The high dimensionality of these models imposes specific computational restrictions on estimating the interpolation region. The Syracuse Research Corporation KOWWIN model for prediction of the n-octanol/water partition coefficient is analysed as a case study. This is a linear regression model that uses 508 fragment counts and correction factors as descriptors, and is based on the group contribution approach. We conclude that the applicability domain estimation by descriptor ranges, combined with Principal Component rotation as a data pre-processing step, is an acceptable compromise between estimation accuracy and the amount of data in the training set.