Maximum-likelihood estimation of predictive uncertainty in probabilistic QSAR modeling
Abstract
Quantitative Structure-Activity Relationships (QSAR) models routinely predict biological activities of interest and the resulting predictions are meaningful only if their uncertainty is properly characterized. Consequently, the availability of a methodology allowing the estimation of predictive uncertainty for QSAR predictions is an issue of practical relevance both for scientific research and for regulatory purposes. In this paper, we present a novel statistical methodology based on maximum likelihood estimation that enables the QSAR prediction of biological activities and the probabilistic assessment of the uncertainty attached to individual predictions. In the formulation of the methodology, chemicals are positioned in a descriptor hyperspace whose coordinates are defined by a set of orthogonal axis such as the principal components determined by Partial Least Squares (PLS) regressions. Each training set chemical contributes in predicting the probability for the query chemical to be active and such a contribution is weighted with respect to the Euclidean distance separating the two chemicals. This study explains how uncertainty can be probabilistically assessed during QSAR modeling of molecular databases and compares the performance of the predictive methodology to published data.