Toward better QSAR/QSPR modeling: simultaneous outlier detection and variable selection using distribution of model features

Building a robust and reliable QSAR/QSPR model should greatly consider two aspects: selecting the optimal variable subset from a large pool of molecular descriptors and detecting outliers from a pool of samples. The two problems have the specific similarity and complementarity to some extent. Given a particular learning algorithm on a particular data set, one should consider how the interaction could happen between variable selection and outlier detection. In this paper, we describe a consistent methodology for simultaneously performing variable subset selection and outlier detection using the idea of statistical distribution which can be simulated by the establishment of many cross-predictive linear models. The approach exploits the fact that the distribution of linear model coefficients provides a mechanism for ranking and interpreting the effects of variable, while the distribution of prediction errors provides a mechanism for differentiating the outliers from normal samples. The use of statistic of these distributions, namely mean value and standard deviation, inherently provides a feasible way to effectively describe the information contained by the original samples. Several examples are used to demonstrate the prediction ability of our proposed approach through the comparison of different approaches as well as their combinations.