Search for Optimal Molecular Descriptors

We illustrate, on a set of heptane isomers and van der Waals molecular areas as a property, a search for optimal descriptors to be used in multivariate analyses. We employed molecular connectivities as the basis descriptors and examined their various combinations as alternative descriptors that could lead to a better regression. The critical step in the search appears to be the selection of the first descriptor to which, subsequently, other descriptors are made orthogonal. A well constructed descriptor can yield a regression equation superior to regressions using several standard descriptors. The approach is illustrated by considering, among others, the molecular ID numbers, the Wiener numbers and Balaban’s J index, as well as novel graphical bond orders based on the above indices. A by-product of the systematic search for better structure-property relationship is a novel property-property correlation (between octane number and molecular cavity areas)