Empirical pattern recognition/expert system for molecular weight estimation of low resolution mass spectra

Abstract A fast, personal-computer based method of estimating molecular weights of organic compounds from low resolution mass spectra has been redesigned and implemented with a rule-based expert system. It has a sequential design with a pattern recognition classifier followed by filter and molecular weight estimator modules for each of six classes. The classes are nonhalobenzenes, chlorobenzenes, bromo- and bromochloroalkanes/alkenes, mono- and di-chloroalkanes/alkenes, tri-, tetra- and pentachloroalkanes/alkenes and unknowns. The classifier was derived from 106 NIST/EPA/MSDC reference spectra. The filters employ computed series of allowed molecular weights and selected base peaks for each class, except unknown, to reduce misclassification. Empirical linear corrections from the training spectra are applied to two mass spectral features, MAXMASS and HIMAX1, to yield estimates and lower limits to the molecular weights. Extensive testing of the system was conducted with 32 test, 99 randomly chosen and 37 field gas chromatographic-mass spectrometric (GC-MS) spectra and results were compared to those from STIRS. The median absolute deviations from the true molecular weights of the test, random and field GC-MS spectra with the expert system were all 1 dalton (average 5.6, 7.3, 5.9 daltons, respectively). This approach also was evaluated with 400 spectra of volatile and nonvolatile compounds of pharmaceutical interest. The median and average absolute deviations from the true molecular weights of the 400 spectra were 2 and 10 daltons. Classification of the evaluation spectra, including many incomplete spectra, was very good with accuracies of 97 (test, random and pharmaceutical) and 95% (field GC-MS).