Original scientific paper

Comparative Regression Analysis. Regressions Based on a Single Descriptor

Milan Randić ; Department of Mathematics and Computer Science, Drake University, Des Moines, Iowa 50311, USA

Abstract

In this article we report the results on over 750 linear regressions based on a single descriptor, using some twenty molecular properties of alkanes and some forty distinct molecular descriptors. It is shown that, among the numerous descriptors, less than a dozen descriptors outperform all the others. Hence, while the search for graph invariants has been fruitful, few of the constructed descriptors show sufficiently novel characteristics and have displayed sufficiently different behavior in correlations with physico-chemical properties. At the same time, we find that some properties are more susceptible to a successful single-descriptor regression analysis, while few properties remain difficult to characterize by a single descriptor. At the end of this report, we have listed a few challenges to be considered by those involved in structure-property-activity studies. We have drawn attention to some properties of alkanes (octanes in particular) for which better regression results are warranted. We also recommend that the performance of novel indices be compared with the known results, like those reported here. If a novel index shows a better behavior with respect to any of the properties already reported in the literature, they certainly deserve publicity. If an index shows a performance comparable to some existing descriptors, they ought to have other advantages in order to replace the existing descriptors. The burden of »proof« ought to be on the »inventor« of a novel topological index. Vague statements that an index may show promise in structure-property studies ought to be replaced by a comparative study, such as shown here for octanes. While the present paper answers the questions which are the best single simple descriptors for correlations of octane physico-chemical properties, our restriction to the use of a single descriptor in structure-property correlations neither signifies that we expect all properties to be well represented by simple regressions, nor that the best single descriptors will necessarily remain the dominant descriptor in multiple regression analysis.

Keywords

Hrčak ID:

137063

URI

https://hrcak.srce.hr/137063

Publication date:

1.7.1993.

Visits: 1.628 *