COMPUTER PREDICTION OF MOLECULAR WEIGHTS FROM UNKNOWN MASS SPECTRA

Gas chromatography/mass spectrometry (GC/MS) has become a tool of choice in analyzing complex mixture samples from chemical, forensic, environmental, clinical, and food science areas. The analysis of such mixtures can produce hundreds of mass spectra daily. Some of these spectra cannot be identified by computer matching against reference mass spectra due to a limited number of mass spectra in the existing reference libraries. The high volume of these mass spectra which must be interpreted has caused a shortage of experienced mass spectrometrists. Several computer programs have been developed to assist in mass spectral interpretation. None of these, however, have a useful capability for determining molecular weight, which is probably the single most important piece of information necessary for the interpretation of unknown mass spectra.;The reliability of the molecular weight predictions has been tested with 268 spectra of compounds with molecular ions and 71 of compounds without molecular ions. The method has achieved 97.0% reliability in predicting the correct molecular weight for a mass spectrum containing the molecular ion and 59.2% reliability for a mass spectrum without a molecular ion, as the first choice. For an unknown mass spectrum at random, of which 15% have no molecular ion, there is a 91.3% chance that the highest confidence value assignment is the correct molecular weight while the first and second highest confidence value assignments have a 94.7% chance of being correct.;This thesis uses the "Self-Training Interpretive and Retrieval System" (STIRS), a computer program for aiding mass spectral interpretation, to provide quantitative information about the molecular weight. In each of several different mass spectral data classes, STIRS retrieves the 15 reference mass spectra which best match the unknown in mass spectral behavior. Also combination match classes, made up of combinations of these data classes, retrieve compounds whose mass spectra most closely match the overall spectral behavior of the unknown. A new data class for the reference spectra without molecular ions has been added to STIRS as well as a routine which calculates the number of bromine and chlorine atoms in the unknown based on isotopic abundances. The statistical treatment of the primary neutral loss patterns of the best matching compounds retrieved in the combination match classes, based on a random drawing model, allows a calculation of the probability that each primary neutral loss occurs in the unknown. These probabilities and the abundance values of corresponding peaks in the unknown spectrum are used to compute the relative confidence value of each candidate molecular weight. The bromine and chlorine atom content of the unknown from the halogen routine is used to reinforce the confidence values.