Combining statistical and physical considerations in deriving targeted QSPRs using very large molecular descriptor databases

The use of 3-D molecular descriptors in Quantitative Structure Property Relationships (QSPR) is considered. Such descriptors are known to have high level of uncertainty, as different minimization algorithms tend to yield different 3-D structures, which in turn yield different descriptor values. An algorithm has been developed in which the uncertainty in a 3-D descriptor is determined by the difference between the values obtained when using minimized structures coming from different sources for the same compound. This uncertainty can be used as an estimate in the descriptor "noise"level for determining its "signal to noise"ratio. A descriptor of a low signal-to-noise ratio should not be included in QSPRs even if it is highly correlated with the property values of the training set.