Simulation of mass spectral intensities by regression analysis of calculated structural characteristics

Abstract

Intensities of low-resolution, electron-impact mass spectral peaks were modelled with equations found by multiple linear regression analysis of experimental spectra and molecular descriptors derived from connection tables. Spectra of 246 hydrocarbons, including acyclic alkanes and alkenes, and cyclic alkanes and alkenes, were used. A randomly selected sample of twenty compounds was reserved for later evaluation of predictive capability. Molecular structure descriptors including molecular connectivity indices, substructure counts and molecular connectivity environments, numbers of double bonds, numbers of rings plus double bonds (unsaturations), path counts, and molecular weights were used as potential independent variables in a stepwise multiple linear regression analysis program to simulate logarithmically transformed peak intensities. Simulated spectra were compared to observed spectra by using correlation coefficient and Euclidean distance criteria. Simulated spectra were treated as unknowns in a file-searching experiment. The simulated spectra produced by the model equations were very similar to the actual observed spectra for both the training set of compounds and the set of unknown compounds.