Two QSPR Methodologies, the Random, and the Super-Descriptors

(Received April 21, 2010) ABSTRACT The full combinatorial search methodology has been used with a set of molecular connectivity indices plus five experimental parameters and the molar mass to extract the best descriptors for twelve properties of a set of organic solvents. The performance of the full combinatorial methodology is compared with the performance of the greedy search methodology obtained in a previous paper. The molecular connectivity indices used with both methodologies belong to different configurations as they can encode different hydrogen and core electron contributions and can give rise to configuration-dependent descriptors. The indices of the best configuration-dependent descriptors for the twelve properties, either molecular connectivity indices and/or experimental indices have then been pooled together and used to derive super-descriptors. These super-descriptors achieve, a better description for four properties, among which, an impressive description for the melting points. A thorough investigation has also been performed on the model quality of random indices, which have been used to derive either ‘zero-level’ descriptors, or semi-random descriptors. This has not only allowed to have a concrete idea of the model quality of random indices but also do draw interesting considerations about the quality of semi-random descriptors that is, descriptors based both on random numbers and on molecular connectivity indices and/or experimental parameters. In fact, a few properties can advantageously be described with this type of descriptors. This last investigation has allowed to better focus the validity of the q leave-oneout statistics and of the Topliss-Costello rule. On the other side, the full combinatorial technique, either with normal descriptors or with super-descriptors has shown the real limits of the greedy search algorithm, has confirmed previous conclusions about the contributions of the hydrogen atoms, and has underlined the importance of pseudoconnectivity indices, experimental indices, and of some ad hoc parameters.