Use of artificial neural networks to predict the gas chromatographic retention index data of alkylbenzenes on carbowax-20M

Comput Chem. 2000 Mar;24(2):171-9. doi: 10.1016/s0097-8485(99)00058-3.

Abstract

Quantitative structure-activity relationships (QSARs) quantify the connection between the structure and properties of molecules and allow the prediction of properties from structural parameters. Models of relationships between structure and retention index of alkylbenzenes were constructed by means of a multilayer neural network using extended delta-bar-delta (EDBD) algorithms. The 165 group data belong to 129 alkylbenzenes at different temperatures on carbowax-20M. We proposed a new method to describe the structure of the alkylbenzene with a simple set of six numeric code depending on its molecular formula. A set of six numbers and the temperature were used as input parameters to predict the retention indices. The performance of different order of structural coding was investigated. The networks' architecture and the learning times were optimized. The optimum ANNs could give excellent prediction results. In addition, the multiple linear regression (MLR) and nonlinear multivariate regression were applied. We have shown in our studies that, based on the structural numeric codes, ANNs give more accurate predictions of retention index data of alkykbenzenes than regression analysis.