Modelling of gas chromatographic retention indices using counterpropagation neural networks

Unspecific fragmentation of organic substances in the ion source of MS detector hinders identification of organic substances in gas chromatographic separation. In such instances theoretical prediction of the retention indices could be a useful tool. A new method for theoretical prediction of gas chromatographic retention indices is described. Artificial neural networks were trained in counterpropagation mode to predict retention data. Extensive data sets of simple organic compounds with known retention indices taken from the literature were serving for training and test sets. The structure of molecules was described with a 12-dimensional vector the components of which were topological and chemical parameters. Various geometries of artificial neural networks were tested and different divisions into training and testing sets tried. The ANN with the configuration of 15 × 15 neurons has been chosen for routine work. The average RMS value was 36.6 retention time units.