Analysis and refinement of the training set in predicting a variety of constant pure compound properties by the targeted QSPR method

In the present study new statistical indicators are introduced, which enable a reliable estimation of the prediction uncertainty for the (unknown) property of the target compound based on the training set data. It is shown that while increasing the number of descriptors in the QSPR enables better representation of the training set data, it may significantly deteriorate the prediction of the target compound property value. If necessary, improved prediction is achievable by using the statistical information to refine the training set, rather than by increasing the number of the descriptors used. It is demonstrated that by proper adjustment of the training set, the great majority of the constant properties can be predicted within the experimental error level.